v2.5.0-task-model: lightweight task model services + tasks table

Task model infrastructure for cheap LLM calls (auto-naming, search
rewrite, tags, summaries) via a dedicated llama-server instance at
TASK_MODEL_URL, falling back to LLAMA_SWAP_URL with FAST_MODEL when
unset. Replaces the inline fetch in auto_name.ts with taskModelCompletion.

Adds search query rewriting: on step 0 when web tools are enabled, the
user's message is summarized into a search intent hint appended to the
system prompt, improving web_search relevance.

Schema: tasks table for provider dispatch and arena, sessions.tags column.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

.env.example

@@ -11,6 +11,11 @@ POSTGRES_PASSWORD=CHANGE_ME
 # point BooCode at a different SearXNG instance.
 SEARXNG_URL=http://100.114.205.53:8888
 
+# Task model: lightweight model for auto-naming, search rewrite, etc.
+# Direct llama-server instance (NOT llama-swap). Falls back to LLAMA_SWAP_URL
+# with FAST_MODEL when unset.
+# TASK_MODEL_URL=http://100.90.172.55:7995
+
 # v1.13.15-tools: BOOCODE_TOOLS narrows the tool whitelist sent to the LLM.
 # Unset (default) → all tools (~21k schema). Useful primarily for single-purpose
 # sessions where the model only needs read-only filesystem access.

LICENSE

@@ -0,0 +1,661 @@
+                    GNU AFFERO GENERAL PUBLIC LICENSE
+                       Version 3, 19 November 2007
+
+ Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+                            Preamble
+
+  The GNU Affero General Public License is a free, copyleft license for
+software and other kinds of works, specifically designed to ensure
+cooperation with the community in the case of network server software.
+
+  The licenses for most software and other practical works are designed
+to take away your freedom to share and change the works.  By contrast,
+our General Public Licenses are intended to guarantee your freedom to
+share and change all versions of a program--to make sure it remains free
+software for all its users.
+
+  When we speak of free software, we are referring to freedom, not
+price.  Our General Public Licenses are designed to make sure that you
+have the freedom to distribute copies of free software (and charge for
+them if you wish), that you receive source code or can get it if you
+want it, that you can change the software or use pieces of it in new
+free programs, and that you know you can do these things.
+
+  Developers that use our General Public Licenses protect your rights
+with two steps: (1) assert copyright on the software, and (2) offer
+you this License which gives you legal permission to copy, distribute
+and/or modify the software.
+
+  A secondary benefit of defending all users' freedom is that
+improvements made in alternate versions of the program, if they
+receive widespread use, become available for other developers to
+incorporate.  Many developers of free software are heartened and
+encouraged by the resulting cooperation.  However, in the case of
+software used on network servers, this result may fail to come about.
+The GNU General Public License permits making a modified version and
+letting the public access it on a server without ever releasing its
+source code to the public.
+
+  The GNU Affero General Public License is designed specifically to
+ensure that, in such cases, the modified source code becomes available
+to the community.  It requires the operator of a network server to
+provide the source code of the modified version running there to the
+users of that server.  Therefore, public use of a modified version, on
+a publicly accessible server, gives the public access to the source
+code of the modified version.
+
+  An older license, called the Affero General Public License and
+published by Affero, was designed to accomplish similar goals.  This is
+a different license, not a version of the Affero GPL, but Affero has
+released a new version of the Affero GPL which permits relicensing under
+this license.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.
+
+                       TERMS AND CONDITIONS
+
+  0. Definitions.
+
+  "This License" refers to version 3 of the GNU Affero General Public License.
+
+  "Copyright" also means copyright-like laws that apply to other kinds of
+works, such as semiconductor masks.
+
+  "The Program" refers to any copyrightable work licensed under this
+License.  Each licensee is addressed as "you".  "Licensees" and
+"recipients" may be individuals or organizations.
+
+  To "modify" a work means to copy from or adapt all or part of the work
+in a fashion requiring copyright permission, other than the making of an
+exact copy.  The resulting work is called a "modified version" of the
+earlier work or a work "based on" the earlier work.
+
+  A "covered work" means either the unmodified Program or a work based
+on the Program.
+
+  To "propagate" a work means to do anything with it that, without
+permission, would make you directly or secondarily liable for
+infringement under applicable copyright law, except executing it on a
+computer or modifying a private copy.  Propagation includes copying,
+distribution (with or without modification), making available to the
+public, and in some countries other activities as well.
+
+  To "convey" a work means any kind of propagation that enables other
+parties to make or receive copies.  Mere interaction with a user through
+a computer network, with no transfer of a copy, is not conveying.
+
+  An interactive user interface displays "Appropriate Legal Notices"
+to the extent that it includes a convenient and prominently visible
+feature that (1) displays an appropriate copyright notice, and (2)
+tells the user that there is no warranty for the work (except to the
+extent that warranties are provided), that licensees may convey the
+work under this License, and how to view a copy of this License.  If
+the interface presents a list of user commands or options, such as a
+menu, a prominent item in the list meets this criterion.
+
+  1. Source Code.
+
+  The "source code" for a work means the preferred form of the work
+for making modifications to it.  "Object code" means any non-source
+form of a work.
+
+  A "Standard Interface" means an interface that either is an official
+standard defined by a recognized standards body, or, in the case of
+interfaces specified for a particular programming language, one that
+is widely used among developers working in that language.
+
+  The "System Libraries" of an executable work include anything, other
+than the work as a whole, that (a) is included in the normal form of
+packaging a Major Component, but which is not part of that Major
+Component, and (b) serves only to enable use of the work with that
+Major Component, or to implement a Standard Interface for which an
+implementation is available to the public in source code form.  A
+"Major Component", in this context, means a major essential component
+(kernel, window system, and so on) of the specific operating system
+(if any) on which the executable work runs, or a compiler used to
+produce the work, or an object code interpreter used to run it.
+
+  The "Corresponding Source" for a work in object code form means all
+the source code needed to generate, install, and (for an executable
+work) run the object code and to modify the work, including scripts to
+control those activities.  However, it does not include the work's
+System Libraries, or general-purpose tools or generally available free
+programs which are used unmodified in performing those activities but
+which are not part of the work.  For example, Corresponding Source
+includes interface definition files associated with source files for
+the work, and the source code for shared libraries and dynamically
+linked subprograms that the work is specifically designed to require,
+such as by intimate data communication or control flow between those
+subprograms and other parts of the work.
+
+  The Corresponding Source need not include anything that users
+can regenerate automatically from other parts of the Corresponding
+Source.
+
+  The Corresponding Source for a work in source code form is that
+same work.
+
+  2. Basic Permissions.
+
+  All rights granted under this License are granted for the term of
+copyright on the Program, and are irrevocable provided the stated
+conditions are met.  This License explicitly affirms your unlimited
+permission to run the unmodified Program.  The output from running a
+covered work is covered by this License only if the output, given its
+content, constitutes a covered work.  This License acknowledges your
+rights of fair use or other equivalent, as provided by copyright law.
+
+  You may make, run and propagate covered works that you do not
+convey, without conditions so long as your license otherwise remains
+in force.  You may convey covered works to others for the sole purpose
+of having them make modifications exclusively for you, or provide you
+with facilities for running those works, provided that you comply with
+the terms of this License in conveying all material for which you do
+not control copyright.  Those thus making or running the covered works
+for you must do so exclusively on your behalf, under your direction
+and control, on terms that prohibit them from making any copies of
+your copyrighted material outside their relationship with you.
+
+  Conveying under any other circumstances is permitted solely under
+the conditions stated below.  Sublicensing is not allowed; section 10
+makes it unnecessary.
+
+  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
+
+  No covered work shall be deemed part of an effective technological
+measure under any applicable law fulfilling obligations under article
+11 of the WIPO copyright treaty adopted on 20 December 1996, or
+similar laws prohibiting or restricting circumvention of such
+measures.
+
+  When you convey a covered work, you waive any legal power to forbid
+circumvention of technological measures to the extent such circumvention
+is effected by exercising rights under this License with respect to
+the covered work, and you disclaim any intention to limit operation or
+modification of the work as a means of enforcing, against the work's
+users, your or third parties' legal rights to forbid circumvention of
+technological measures.
+
+  4. Conveying Verbatim Copies.
+
+  You may convey verbatim copies of the Program's source code as you
+receive it, in any medium, provided that you conspicuously and
+appropriately publish on each copy an appropriate copyright notice;
+keep intact all notices stating that this License and any
+non-permissive terms added in accord with section 7 apply to the code;
+keep intact all notices of the absence of any warranty; and give all
+recipients a copy of this License along with the Program.
+
+  You may charge any price or no price for each copy that you convey,
+and you may offer support or warranty protection for a fee.
+
+  5. Conveying Modified Source Versions.
+
+  You may convey a work based on the Program, or the modifications to
+produce it from the Program, in the form of source code under the
+terms of section 4, provided that you also meet all of these conditions:
+
+    a) The work must carry prominent notices stating that you modified
+    it, and giving a relevant date.
+
+    b) The work must carry prominent notices stating that it is
+    released under this License and any conditions added under section
+    7.  This requirement modifies the requirement in section 4 to
+    "keep intact all notices".
+
+    c) You must license the entire work, as a whole, under this
+    License to anyone who comes into possession of a copy.  This
+    License will therefore apply, along with any applicable section 7
+    additional terms, to the whole of the work, and all its parts,
+    regardless of how they are packaged.  This License gives no
+    permission to license the work in any other way, but it does not
+    invalidate such permission if you have separately received it.
+
+    d) If the work has interactive user interfaces, each must display
+    Appropriate Legal Notices; however, if the Program has interactive
+    interfaces that do not display Appropriate Legal Notices, your
+    work need not make them do so.
+
+  A compilation of a covered work with other separate and independent
+works, which are not by their nature extensions of the covered work,
+and which are not combined with it such as to form a larger program,
+in or on a volume of a storage or distribution medium, is called an
+"aggregate" if the compilation and its resulting copyright are not
+used to limit the access or legal rights of the compilation's users
+beyond what the individual works permit.  Inclusion of a covered work
+in an aggregate does not cause this License to apply to the other
+parts of the aggregate.
+
+  6. Conveying Non-Source Forms.
+
+  You may convey a covered work in object code form under the terms
+of sections 4 and 5, provided that you also convey the
+machine-readable Corresponding Source under the terms of this License,
+in one of these ways:
+
+    a) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by the
+    Corresponding Source fixed on a durable physical medium
+    customarily used for software interchange.
+
+    b) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by a
+    written offer, valid for at least three years and valid for as
+    long as you offer spare parts or customer support for that product
+    model, to give anyone who possesses the object code either (1) a
+    copy of the Corresponding Source for all the software in the
+    product that is covered by this License, on a durable physical
+    medium customarily used for software interchange, for a price no
+    more than your reasonable cost of physically performing this
+    conveying of source, or (2) access to copy the
+    Corresponding Source from a network server at no charge.
+
+    c) Convey individual copies of the object code with a copy of the
+    written offer to provide the Corresponding Source.  This
+    alternative is allowed only occasionally and noncommercially, and
+    only if you received the object code with such an offer, in accord
+    with subsection 6b.
+
+    d) Convey the object code by offering access from a designated
+    place (gratis or for a charge), and offer equivalent access to the
+    Corresponding Source in the same way through the same place at no
+    further charge.  You need not require recipients to copy the
+    Corresponding Source along with the object code.  If the place to
+    copy the object code is a network server, the Corresponding Source
+    may be on a different server (operated by you or a third party)
+    that supports equivalent copying facilities, provided you maintain
+    clear directions next to the object code saying where to find the
+    Corresponding Source.  Regardless of what server hosts the
+    Corresponding Source, you remain obligated to ensure that it is
+    available for as long as needed to satisfy these requirements.
+
+    e) Convey the object code using peer-to-peer transmission, provided
+    you inform other peers where the object code and Corresponding
+    Source of the work are being offered to the general public at no
+    charge under subsection 6d.
+
+  A separable portion of the object code, whose source code is excluded
+from the Corresponding Source as a System Library, need not be
+included in conveying the object code work.
+
+  A "User Product" is either (1) a "consumer product", which means any
+tangible personal property which is normally used for personal, family,
+or household purposes, or (2) anything designed or sold for incorporation
+into a dwelling.  In determining whether a product is a consumer product,
+doubtful cases shall be resolved in favor of coverage.  For a particular
+product received by a particular user, "normally used" refers to a
+typical or common use of that class of product, regardless of the status
+of the particular user or of the way in which the particular user
+actually uses, or expects or is expected to use, the product.  A product
+is a consumer product regardless of whether the product has substantial
+commercial, industrial or non-consumer uses, unless such uses represent
+the only significant mode of use of the product.
+
+  "Installation Information" for a User Product means any methods,
+procedures, authorization keys, or other information required to install
+and execute modified versions of a covered work in that User Product from
+a modified version of its Corresponding Source.  The information must
+suffice to ensure that the continued functioning of the modified object
+code is in no case prevented or interfered with solely because
+modification has been made.
+
+  If you convey an object code work under this section in, or with, or
+specifically for use in, a User Product, and the conveying occurs as
+part of a transaction in which the right of possession and use of the
+User Product is transferred to the recipient in perpetuity or for a
+fixed term (regardless of how the transaction is characterized), the
+Corresponding Source conveyed under this section must be accompanied
+by the Installation Information.  But this requirement does not apply
+if neither you nor any third party retains the ability to install
+modified object code on the User Product (for example, the work has
+been installed in ROM).
+
+  The requirement to provide Installation Information does not include a
+requirement to continue to provide support service, warranty, or updates
+for a work that has been modified or installed by the recipient, or for
+the User Product in which it has been modified or installed.  Access to a
+network may be denied when the modification itself materially and
+adversely affects the operation of the network or violates the rules and
+protocols for communication across the network.
+
+  Corresponding Source conveyed, and Installation Information provided,
+in accord with this section must be in a format that is publicly
+documented (and with an implementation available to the public in
+source code form), and must require no special password or key for
+unpacking, reading or copying.
+
+  7. Additional Terms.
+
+  "Additional permissions" are terms that supplement the terms of this
+License by making exceptions from one or more of its conditions.
+Additional permissions that are applicable to the entire Program shall
+be treated as though they were included in this License, to the extent
+that they are valid under applicable law.  If additional permissions
+apply only to part of the Program, that part may be used separately
+under those permissions, but the entire Program remains governed by
+this License without regard to the additional permissions.
+
+  When you convey a copy of a covered work, you may at your option
+remove any additional permissions from that copy, or from any part of
+it.  (Additional permissions may be written to require their own
+removal in certain cases when you modify the work.)  You may place
+additional permissions on material, added by you to a covered work,
+for which you have or can give appropriate copyright permission.
+
+  Notwithstanding any other provision of this License, for material you
+add to a covered work, you may (if authorized by the copyright holders of
+that material) supplement the terms of this License with terms:
+
+    a) Disclaiming warranty or limiting liability differently from the
+    terms of sections 15 and 16 of this License; or
+
+    b) Requiring preservation of specified reasonable legal notices or
+    author attributions in that material or in the Appropriate Legal
+    Notices displayed by works containing it; or
+
+    c) Prohibiting misrepresentation of the origin of that material, or
+    requiring that modified versions of such material be marked in
+    reasonable ways as different from the original version; or
+
+    d) Limiting the use for publicity purposes of names of licensors or
+    authors of the material; or
+
+    e) Declining to grant rights under trademark law for use of some
+    trade names, trademarks, or service marks; or
+
+    f) Requiring indemnification of licensors and authors of that
+    material by anyone who conveys the material (or modified versions of
+    it) with contractual assumptions of liability to the recipient, for
+    any liability that these contractual assumptions directly impose on
+    those licensors and authors.
+
+  All other non-permissive additional terms are considered "further
+restrictions" within the meaning of section 10.  If the Program as you
+received it, or any part of it, contains a notice stating that it is
+governed by this License along with a term that is a further
+restriction, you may remove that term.  If a license document contains
+a further restriction but permits relicensing or conveying under this
+License, you may add to a covered work material governed by the terms
+of that license document, provided that the further restriction does
+not survive such relicensing or conveying.
+
+  If you add terms to a covered work in accord with this section, you
+must place, in the relevant source files, a statement of the
+additional terms that apply to those files, or a notice indicating
+where to find the applicable terms.
+
+  Additional terms, permissive or non-permissive, may be stated in the
+form of a separately written license, or stated as exceptions;
+the above requirements apply either way.
+
+  8. Termination.
+
+  You may not propagate or modify a covered work except as expressly
+provided under this License.  Any attempt otherwise to propagate or
+modify it is void, and will automatically terminate your rights under
+this License (including any patent licenses granted under the third
+paragraph of section 11).
+
+  However, if you cease all violation of this License, then your
+license from a particular copyright holder is reinstated (a)
+provisionally, unless and until the copyright holder explicitly and
+finally terminates your license, and (b) permanently, if the copyright
+holder fails to notify you of the violation by some reasonable means
+prior to 60 days after the cessation.
+
+  Moreover, your license from a particular copyright holder is
+reinstated permanently if the copyright holder notifies you of the
+violation by some reasonable means, this is the first time you have
+received notice of violation of this License (for any work) from that
+copyright holder, and you cure the violation prior to 30 days after
+your receipt of the notice.
+
+  Termination of your rights under this section does not terminate the
+licenses of parties who have received copies or rights from you under
+this License.  If your rights have been terminated and not permanently
+reinstated, you do not qualify to receive new licenses for the same
+material under section 10.
+
+  9. Acceptance Not Required for Having Copies.
+
+  You are not required to accept this License in order to receive or
+run a copy of the Program.  Ancillary propagation of a covered work
+occurring solely as a consequence of using peer-to-peer transmission
+to receive a copy likewise does not require acceptance.  However,
+nothing other than this License grants you permission to propagate or
+modify any covered work.  These actions infringe copyright if you do
+not accept this License.  Therefore, by modifying or propagating a
+covered work, you indicate your acceptance of this License to do so.
+
+  10. Automatic Licensing of Downstream Recipients.
+
+  Each time you convey a covered work, the recipient automatically
+receives a license from the original licensors, to run, modify and
+propagate that work, subject to this License.  You are not responsible
+for enforcing compliance by third parties with this License.
+
+  An "entity transaction" is a transaction transferring control of an
+organization, or substantially all assets of one, or subdividing an
+organization, or merging organizations.  If propagation of a covered
+work results from an entity transaction, each party to that
+transaction who receives a copy of the work also receives whatever
+licenses to the work the party's predecessor in interest had or could
+give under the previous paragraph, plus a right to possession of the
+Corresponding Source of the work from the predecessor in interest, if
+the predecessor has it or can get it with reasonable efforts.
+
+  You may not impose any further restrictions on the exercise of the
+rights granted or affirmed under this License.  For example, you may
+not impose a license fee, royalty, or other charge for exercise of
+rights granted under this License, and you may not initiate litigation
+(including a cross-claim or counterclaim in a lawsuit) alleging that
+any patent claim is infringed by making, using, selling, offering for
+sale, or importing the Program or any portion of it.
+
+  11. Patents.
+
+  A "contributor" is a copyright holder who authorizes use under this
+License of the Program or a work on which the Program is based.  The
+work thus licensed is called the contributor's "contributor version".
+
+  A contributor's "essential patent claims" are all patent claims
+owned or controlled by the contributor, whether already acquired or
+hereafter acquired, that would be infringed by some manner, permitted
+by this License, of making, using, or selling its contributor version,
+but do not include claims that would be infringed only as a
+consequence of further modification of the contributor version.  For
+purposes of this definition, "control" includes the right to grant
+patent sublicenses in a manner consistent with the requirements of
+this License.
+
+  Each contributor grants you a non-exclusive, worldwide, royalty-free
+patent license under the contributor's essential patent claims, to
+make, use, sell, offer for sale, import and otherwise run, modify and
+propagate the contents of its contributor version.
+
+  In the following three paragraphs, a "patent license" is any express
+agreement or commitment, however denominated, not to enforce a patent
+(such as an express permission to practice a patent or covenant not to
+sue for patent infringement).  To "grant" such a patent license to a
+party means to make such an agreement or commitment not to enforce a
+patent against the party.
+
+  If you convey a covered work, knowingly relying on a patent license,
+and the Corresponding Source of the work is not available for anyone
+to copy, free of charge and under the terms of this License, through a
+publicly available network server or other readily accessible means,
+then you must either (1) cause the Corresponding Source to be so
+available, or (2) arrange to deprive yourself of the benefit of the
+patent license for this particular work, or (3) arrange, in a manner
+consistent with the requirements of this License, to extend the patent
+license to downstream recipients.  "Knowingly relying" means you have
+actual knowledge that, but for the patent license, your conveying the
+covered work in a country, or your recipient's use of the covered work
+in a country, would infringe one or more identifiable patents in that
+country that you have reason to believe are valid.
+
+  If, pursuant to or in connection with a single transaction or
+arrangement, you convey, or propagate by procuring conveyance of, a
+covered work, and grant a patent license to some of the parties
+receiving the covered work authorizing them to use, propagate, modify
+or convey a specific copy of the covered work, then the patent license
+you grant is automatically extended to all recipients of the covered
+work and works based on it.
+
+  A patent license is "discriminatory" if it does not include within
+the scope of its coverage, prohibits the exercise of, or is
+conditioned on the non-exercise of one or more of the rights that are
+specifically granted under this License.  You may not convey a covered
+work if you are a party to an arrangement with a third party that is
+in the business of distributing software, under which you make payment
+to the third party based on the extent of your activity of conveying
+the work, and under which the third party grants, to any of the
+parties who would receive the covered work from you, a discriminatory
+patent license (a) in connection with copies of the covered work
+conveyed by you (or copies made from those copies), or (b) primarily
+for and in connection with specific products or compilations that
+contain the covered work, unless you entered into that arrangement,
+or that patent license was granted, prior to 28 March 2007.
+
+  Nothing in this License shall be construed as excluding or limiting
+any implied license or other defenses to infringement that may
+otherwise be available to you under applicable patent law.
+
+  12. No Surrender of Others' Freedom.
+
+  If conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot convey a
+covered work so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you may
+not convey it at all.  For example, if you agree to terms that obligate you
+to collect a royalty for further conveying from those to whom you convey
+the Program, the only way you could satisfy both those terms and this
+License would be to refrain entirely from conveying the Program.
+
+  13. Remote Network Interaction; Use with the GNU General Public License.
+
+  Notwithstanding any other provision of this License, if you modify the
+Program, your modified version must prominently offer all users
+interacting with it remotely through a computer network (if your version
+supports such interaction) an opportunity to receive the Corresponding
+Source of your version by providing access to the Corresponding Source
+from a network server at no charge, through some standard or customary
+means of facilitating copying of software.  This Corresponding Source
+shall include the Corresponding Source for any work covered by version 3
+of the GNU General Public License that is incorporated pursuant to the
+following paragraph.
+
+  Notwithstanding any other provision of this License, you have
+permission to link or combine any covered work with a work licensed
+under version 3 of the GNU General Public License into a single
+combined work, and to convey the resulting work.  The terms of this
+License will continue to apply to the part which is the covered work,
+but the work with which it is combined will remain governed by version
+3 of the GNU General Public License.
+
+  14. Revised Versions of this License.
+
+  The Free Software Foundation may publish revised and/or new versions of
+the GNU Affero General Public License from time to time.  Such new versions
+will be similar in spirit to the present version, but may differ in detail to
+address new problems or concerns.
+
+  Each version is given a distinguishing version number.  If the
+Program specifies that a certain numbered version of the GNU Affero General
+Public License "or any later version" applies to it, you have the
+option of following the terms and conditions either of that numbered
+version or of any later version published by the Free Software
+Foundation.  If the Program does not specify a version number of the
+GNU Affero General Public License, you may choose any version ever published
+by the Free Software Foundation.
+
+  If the Program specifies that a proxy can decide which future
+versions of the GNU Affero General Public License can be used, that proxy's
+public statement of acceptance of a version permanently authorizes you
+to choose that version for the Program.
+
+  Later license versions may give you additional or different
+permissions.  However, no additional obligations are imposed on any
+author or copyright holder as a result of your choosing to follow a
+later version.
+
+  15. Disclaimer of Warranty.
+
+  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
+APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
+HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
+OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
+THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
+IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
+ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
+
+  16. Limitation of Liability.
+
+  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
+WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
+THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
+GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
+USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
+DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
+PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
+EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
+SUCH DAMAGES.
+
+  17. Interpretation of Sections 15 and 16.
+
+  If the disclaimer of warranty and limitation of liability provided
+above cannot be given local legal effect according to their terms,
+reviewing courts shall apply local law that most closely approximates
+an absolute waiver of all civil liability in connection with the
+Program, unless a warranty or assumption of liability accompanies a
+copy of the Program in return for a fee.
+
+                     END OF TERMS AND CONDITIONS
+
+            How to Apply These Terms to Your New Programs
+
+  If you develop a new program, and you want it to be of the greatest
+possible use to the public, the best way to achieve this is to make it
+free software which everyone can redistribute and change under these terms.
+
+  To do so, attach the following notices to the program.  It is safest
+to attach them to the start of each source file to most effectively
+state the exclusion of warranty; and each file should have at least
+the "copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the program's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU Affero General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU Affero General Public License for more details.
+
+    You should have received a copy of the GNU Affero General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+Also add information on how to contact you by electronic and paper mail.
+
+  If your software can interact with users remotely through a computer
+network, you should also make sure that it provides a way for users to
+get its source.  For example, if your program is a web application, its
+interface could display a "Source" link that leads users to an archive
+of the code.  There are many ways you could offer source, and different
+solutions will be better for different programs; see section 13 for the
+specific requirements.
+
+  You should also get your employer (if you work as a programmer) or school,
+if any, to sign a "copyright disclaimer" for the program, if necessary.
+For more information on this, and how to apply and follow the GNU AGPL, see
+<https://www.gnu.org/licenses/>.

apps/booterm/package.json

@@ -23,5 +23,6 @@
     "@types/pg": "^8.11.10",
     "tsx": "^4.16.2",
     "typescript": "^5.5.0"
-  }
+  },
+  "license": "AGPL-3.0-only"
 }

apps/coder/package.json

@@ -29,5 +29,6 @@
     "tsx": "^4.16.2",
     "typescript": "^5.5.0",
     "vitest": "^3.0.0"
-  }
+  },
+  "license": "AGPL-3.0-only"
 }

apps/coder/src/index.ts

@@ -85,7 +85,9 @@ async function main() {
         type: 'permission_requested',
         task_id: prompt.taskId,
         session_id: prompt.sessionId,
+        kind: prompt.kind,
         tool_title: prompt.toolTitle,
+        ...(prompt.input ? { input: prompt.input } : {}),
         options: prompt.options.map((o) => ({ option_id: o.optionId, label: o.label })),
       } as WsFrame);
     },

apps/coder/src/routes/messages.ts

@@ -5,6 +5,33 @@ import type { Broker } from '@boocode/server/broker';
 import type { WsFrame } from '@boocode/server/ws-frames';
 import { resolveChatId } from './chat-resolve.js';
 
+const AnswerUserInputBody = z.object({
+  tool_call_id: z.string().min(1),
+  answers: z
+    .array(
+      z.object({
+        question: z.string(),
+        selected_options: z.array(z.string()),
+        free_text: z.string().nullable(),
+      }),
+    )
+    .min(1)
+    .max(3),
+});
+
+const AskUserInputArgs = z.object({
+  questions: z
+    .array(
+      z.object({
+        question: z.string(),
+        type: z.enum(['single_select', 'multi_select']),
+        options: z.array(z.string()).min(1),
+      }),
+    )
+    .min(1)
+    .max(3),
+});
+
 const SendBody = z.object({
   content: z.string().min(1).max(64_000),
   pane_id: z.string().min(1).max(200),
@@ -219,6 +246,138 @@ export function registerMessageRoutes(
     },
   );
 
+  // POST /api/chats/:id/answer_user_input — answer a pending ask_user_input
+  app.post<{ Params: { id: string } }>(
+    '/api/chats/:id/answer_user_input',
+    async (req, reply) => {
+      const parsed = AnswerUserInputBody.safeParse(req.body);
+      if (!parsed.success) {
+        reply.code(400);
+        return { error: 'invalid_body', details: parsed.error.flatten() };
+      }
+      const { tool_call_id, answers } = parsed.data;
+
+      const chatRows = await sql<{ id: string; session_id: string }[]>`
+        SELECT id, session_id FROM chats WHERE id = ${req.params.id} AND status = 'open'
+      `;
+      if (chatRows.length === 0) {
+        reply.code(404);
+        return { error: 'chat_not_found' };
+      }
+      const chat = chatRows[0]!;
+      const sessionId = chat.session_id;
+
+      const callerRows = await sql<{
+        message_id: string;
+        payload: { id: string; name: string; args: Record<string, unknown> };
+      }[]>`
+        SELECT p.message_id, p.payload
+        FROM message_parts p
+        JOIN messages m ON m.id = p.message_id
+        WHERE m.chat_id = ${chat.id}
+          AND m.role = 'assistant'
+          AND p.kind = 'tool_call'
+          AND p.payload->>'id' = ${tool_call_id}
+        ORDER BY m.created_at DESC
+        LIMIT 1
+      `;
+      if (!callerRows[0]) {
+        reply.code(404);
+        return { error: 'unknown_tool_call_id' };
+      }
+      const foundCall = callerRows[0].payload;
+      if (foundCall.name !== 'ask_user_input') {
+        reply.code(400);
+        return { error: 'tool_call_not_ask_user_input' };
+      }
+
+      const argsParsed = AskUserInputArgs.safeParse(foundCall.args);
+      if (!argsParsed.success) {
+        reply.code(400);
+        return { error: 'mismatched_answer_shape', detail: 'tool_call args invalid' };
+      }
+      const questions = argsParsed.data.questions;
+      if (answers.length !== questions.length) {
+        reply.code(400);
+        return { error: 'mismatched_answer_shape', detail: `expected ${questions.length} answer(s), got ${answers.length}` };
+      }
+      for (let i = 0; i < questions.length; i++) {
+        const q = questions[i]!;
+        const a = answers[i]!;
+        for (const sel of a.selected_options) {
+          if (!q.options.includes(sel)) {
+            reply.code(400);
+            return { error: 'mismatched_answer_shape', detail: `answer ${i + 1} option not in question: ${sel}` };
+          }
+        }
+        if (q.type === 'single_select' && a.selected_options.length > 1) {
+          reply.code(400);
+          return { error: 'mismatched_answer_shape', detail: `answer ${i + 1} multi on single_select` };
+        }
+        if (a.selected_options.length === 0 && (!a.free_text || !a.free_text.trim())) {
+          reply.code(400);
+          return { error: 'mismatched_answer_shape', detail: `answer ${i + 1} is empty` };
+        }
+      }
+
+      const toolRows = await sql<{
+        message_id: string;
+        payload: { tool_call_id: string; output: unknown };
+      }[]>`
+        SELECT p.message_id, p.payload
+        FROM message_parts p
+        JOIN messages m ON m.id = p.message_id
+        WHERE m.chat_id = ${chat.id}
+          AND m.role = 'tool'
+          AND p.kind = 'tool_result'
+          AND p.payload->>'tool_call_id' = ${tool_call_id}
+        ORDER BY m.created_at DESC
+        LIMIT 1
+      `;
+      if (!toolRows[0]) {
+        reply.code(404);
+        return { error: 'unknown_tool_call_id', detail: 'tool message not found' };
+      }
+      if (toolRows[0].payload?.output !== null) {
+        reply.code(409);
+        return { error: 'tool_call_already_answered' };
+      }
+
+      const answerSet = { answers };
+      const newToolResults = { tool_call_id, output: answerSet, truncated: false };
+      const toolMessageId = toolRows[0].message_id;
+
+      const result = await sql.begin(async (tx) => {
+        await tx`DELETE FROM message_parts WHERE message_id = ${toolMessageId} AND kind = 'tool_result'`;
+        await tx`
+          INSERT INTO message_parts (message_id, sequence, kind, payload)
+          VALUES (${toolMessageId}, 0, 'tool_result', ${tx.json(newToolResults as never)})
+        `;
+        const [assistantMsg] = await tx<{ id: string }[]>`
+          INSERT INTO messages (session_id, chat_id, role, content, status, created_at)
+          VALUES (${sessionId}, ${chat.id}, 'assistant', '', 'streaming', clock_timestamp())
+          RETURNING id
+        `;
+        await tx`UPDATE sessions SET updated_at = clock_timestamp() WHERE id = ${sessionId}`;
+        await tx`UPDATE chats SET updated_at = clock_timestamp() WHERE id = ${chat.id}`;
+        return { tool_message_id: toolMessageId, assistant_message_id: assistantMsg!.id };
+      });
+
+      broker.publishFrame(sessionId, {
+        type: 'tool_result',
+        tool_message_id: result.tool_message_id,
+        tool_call_id,
+        chat_id: chat.id,
+        output: answerSet,
+        truncated: false,
+      } as unknown as WsFrame);
+      inference.enqueue(sessionId, chat.id, result.assistant_message_id, 'default');
+
+      reply.code(202);
+      return result;
+    },
+  );
+
   // POST /api/sessions/:sessionId/stop — cancel active inference
   app.post<{ Params: { sessionId: string } }>(
     '/api/sessions/:sessionId/stop',

apps/coder/src/routes/tasks.ts

@@ -19,6 +19,7 @@ const CreateBody = z.object({
 
 const PermissionBody = z.object({
   option_id: z.string().max(200).nullable(),
+  updated_input: z.record(z.unknown()).optional(),
 });
 
 const ListQuery = z.object({
@@ -164,7 +165,7 @@ export function registerTaskRoutes(app: FastifyInstance, sql: Sql, inference: In
       return { error: 'invalid body', details: parsed.error.flatten() };
     }
 
-    const ok = respondToPermission(req.params.id, parsed.data.option_id);
+    const ok = respondToPermission(req.params.id, parsed.data.option_id, parsed.data.updated_input as Record<string, unknown> | undefined);
     if (!ok) {
       reply.code(404);
       return { error: 'no pending permission' };

apps/coder/src/services/acp-dispatch.ts

@@ -17,6 +17,8 @@ import {
   type WriteTextFileResponse,
   type CreateTerminalRequest,
   type CreateTerminalResponse,
+  type CreateElicitationRequest,
+  type CreateElicitationResponse,
   type SessionConfigOption,
   type ClientSideConnection as ConnectionType,
 } from '@agentclientprotocol/sdk';
@@ -26,7 +28,7 @@ import { spawn } from 'node:child_process';
 import { findThoughtLevelConfigId } from './acp-derive.js';
 import { resolveAcpSpawnArgs } from './acp-spawn.js';
 import { createAcpNdJsonStream } from './acp-stream.js';
-import { waitForPermissionResponse, cancelPendingPermission } from './permission-waiter.js';
+import { waitForPermissionResponse, waitForElicitationResponse, cancelPendingPermission } from './permission-waiter.js';
 import { mergeTaskCommands, getTaskCommands } from './agent-commands-cache.js';
 import { readWorktreeTextFile, writeWorktreeTextFile } from './acp-client-fs.js';
 import {
@@ -254,6 +256,12 @@ class AcpStreamContext {
       createTerminal: async (_params: CreateTerminalRequest): Promise<CreateTerminalResponse> => {
         return { terminalId: 'noop' };
       },
+      unstable_createElicitation: async (params: CreateElicitationRequest): Promise<CreateElicitationResponse> => {
+        if (taskId && sessionId) {
+          return waitForElicitationResponse(taskId, sessionId, agent, modeId, params);
+        }
+        return { action: 'decline' };
+      },
     };
   }
 }

apps/coder/src/services/permission-waiter.ts

@@ -1,12 +1,13 @@
 /**
- * Blocks ACP dispatch on permission prompts until the user responds via API.
+ * Blocks ACP dispatch on permission/elicitation prompts until the user responds via API.
  */
-import type { RequestPermissionRequest, RequestPermissionResponse } from '@agentclientprotocol/sdk';
+import type { RequestPermissionRequest, RequestPermissionResponse, CreateElicitationRequest, CreateElicitationResponse } from '@agentclientprotocol/sdk';
 import { isUnattendedMode } from './provider-manifest.js';
 
 const DEFAULT_TIMEOUT_MS = 120_000;
 
 interface PendingPermission {
+  type: 'permission';
   request: RequestPermissionRequest;
   sessionId: string;
   resolve: (response: RequestPermissionResponse) => void;
@@ -14,11 +15,27 @@ interface PendingPermission {
   timer: ReturnType<typeof setTimeout>;
 }
 
-const pendingByTask = new Map<string, PendingPermission>();
+interface PendingElicitation {
+  type: 'elicitation';
+  request: CreateElicitationRequest;
+  sessionId: string;
+  resolve: (response: CreateElicitationResponse) => void;
+  reject: (err: Error) => void;
+  timer: ReturnType<typeof setTimeout>;
+}
+
+type PendingEntry = PendingPermission | PendingElicitation;
+
+const pendingByTask = new Map<string, PendingEntry>();
+
+export type PermissionKind = 'tool' | 'question' | 'plan' | 'elicitation';
 
 export interface PermissionPrompt {
   taskId: string;
+  kind: PermissionKind;
   toolTitle?: string;
+  description?: string;
+  input?: Record<string, unknown>;
   options: Array<{ optionId: string; label: string }>;
 }
 
@@ -33,10 +50,25 @@ export function setPermissionHooks(next: PermissionHooks): void {
   hooks = next;
 }
 
+function resolveKind(params: RequestPermissionRequest): PermissionKind {
+  const input = params.toolCall?.rawInput;
+  if (input && typeof input === 'object' && !Array.isArray(input) && 'questions' in input && Array.isArray((input as Record<string, unknown>).questions)) {
+    return 'question';
+  }
+  return 'tool';
+}
+
 function toPrompt(taskId: string, params: RequestPermissionRequest): PermissionPrompt {
+  const kind = resolveKind(params);
+  const rawInput = params.toolCall?.rawInput;
+  const input = rawInput && typeof rawInput === 'object' && !Array.isArray(rawInput)
+    ? rawInput as Record<string, unknown>
+    : undefined;
   return {
     taskId,
+    kind,
     toolTitle: params.toolCall?.title ?? undefined,
+    ...(input ? { input } : {}),
     options: params.options.map((o) => ({
       optionId: o.optionId,
       label: o.name,
@@ -73,24 +105,33 @@ export function waitForPermissionResponse(
       resolve({ outcome: { outcome: 'cancelled' } });
     }, timeoutMs);
 
-    pendingByTask.set(taskId, { request: params, sessionId, resolve, reject, timer });
+    pendingByTask.set(taskId, { type: 'permission', request: params, sessionId, resolve, reject, timer });
 
     const prompt = toPrompt(taskId, params);
     void hooks.onPrompt?.({ ...prompt, sessionId });
   });
 }
 
-export function respondToPermission(taskId: string, optionId: string | null): boolean {
+export function respondToPermission(taskId: string, optionId: string | null, updatedInput?: Record<string, unknown>): boolean {
   const pending = pendingByTask.get(taskId);
   if (!pending) return false;
 
   clearTimeout(pending.timer);
   pendingByTask.delete(taskId);
 
-  if (optionId) {
+  if (pending.type === 'elicitation') {
-    pending.resolve({ outcome: { outcome: 'selected', optionId } });
+    if (updatedInput) {
+      const content = updatedInput as { [key: string]: string | number | boolean | string[] };
+      pending.resolve({ action: 'accept', content });
+    } else {
+      pending.resolve({ action: 'decline' });
+    }
   } else {
-    pending.resolve({ outcome: { outcome: 'cancelled' } });
+    if (optionId) {
+      pending.resolve({ outcome: { outcome: 'selected', optionId } });
+    } else {
+      pending.resolve({ outcome: { outcome: 'cancelled' } });
+    }
   }
 
   void hooks.onResolved?.(taskId, pending.sessionId);
@@ -100,14 +141,67 @@ export function respondToPermission(taskId: string, optionId: string | null): bo
 export function getPendingPermission(taskId: string): PermissionPrompt | null {
   const pending = pendingByTask.get(taskId);
   if (!pending) return null;
+  if (pending.type === 'elicitation') {
+    return elicitationToPrompt(taskId, pending.request);
+  }
   return toPrompt(taskId, pending.request);
 }
 
+function elicitationToPrompt(taskId: string, params: CreateElicitationRequest): PermissionPrompt {
+  const input: Record<string, unknown> = { message: params.message };
+  if ('requestedSchema' in params) {
+    input.requestedSchema = params.requestedSchema;
+  }
+  return {
+    taskId,
+    kind: 'elicitation',
+    toolTitle: params.message,
+    input,
+    options: [],
+  };
+}
+
+export function waitForElicitationResponse(
+  taskId: string,
+  sessionId: string,
+  provider: string,
+  modeId: string | undefined,
+  params: CreateElicitationRequest,
+  timeoutMs = DEFAULT_TIMEOUT_MS,
+): Promise<CreateElicitationResponse> {
+  if (isUnattendedMode(provider, modeId)) {
+    return Promise.resolve({ action: 'decline' });
+  }
+
+  return new Promise((resolve, reject) => {
+    const existing = pendingByTask.get(taskId);
+    if (existing) {
+      clearTimeout(existing.timer);
+      existing.reject(new Error('superseded by newer elicitation request'));
+    }
+
+    const timer = setTimeout(() => {
+      pendingByTask.delete(taskId);
+      void hooks.onResolved?.(taskId, sessionId);
+      resolve({ action: 'cancel' });
+    }, timeoutMs);
+
+    pendingByTask.set(taskId, { type: 'elicitation', request: params, sessionId, resolve, reject, timer });
+
+    const prompt = elicitationToPrompt(taskId, params);
+    void hooks.onPrompt?.({ ...prompt, sessionId });
+  });
+}
+
 export function cancelPendingPermission(taskId: string): void {
   const pending = pendingByTask.get(taskId);
   if (!pending) return;
   clearTimeout(pending.timer);
   pendingByTask.delete(taskId);
-  pending.resolve({ outcome: { outcome: 'cancelled' } });
+  if (pending.type === 'elicitation') {
+    pending.resolve({ action: 'cancel' });
+  } else {
+    pending.resolve({ outcome: { outcome: 'cancelled' } });
+  }
   void hooks.onResolved?.(taskId, pending.sessionId);
 }

apps/server/package.json

@@ -5,23 +5,74 @@
   "type": "module",
   "main": "dist/index.js",
   "exports": {
-    ".": { "types": "./dist/index.d.ts", "default": "./dist/index.js" },
+    ".": {
-    "./inference": { "types": "./dist/services/inference/index.d.ts", "default": "./dist/services/inference/index.js" },
+      "types": "./dist/index.d.ts",
-    "./tools": { "types": "./dist/services/tools.d.ts", "default": "./dist/services/tools.js" },
+      "default": "./dist/index.js"
-    "./broker": { "types": "./dist/services/broker.d.ts", "default": "./dist/services/broker.js" },
+    },
-    "./compaction": { "types": "./dist/services/compaction.d.ts", "default": "./dist/services/compaction.js" },
+    "./inference": {
-    "./model-context": { "types": "./dist/services/model-context.d.ts", "default": "./dist/services/model-context.js" },
+      "types": "./dist/services/inference/index.d.ts",
-    "./system-prompt": { "types": "./dist/services/system-prompt.d.ts", "default": "./dist/services/system-prompt.js" },
+      "default": "./dist/services/inference/index.js"
-    "./agents": { "types": "./dist/services/agents.d.ts", "default": "./dist/services/agents.js" },
+    },
-    "./truncate": { "types": "./dist/services/truncate.d.ts", "default": "./dist/services/truncate.js" },
+    "./tools": {
-    "./path-guard": { "types": "./dist/services/path_guard.d.ts", "default": "./dist/services/path_guard.js" },
+      "types": "./dist/services/tools.d.ts",
-    "./file-ops": { "types": "./dist/services/file_ops.d.ts", "default": "./dist/services/file_ops.js" },
+      "default": "./dist/services/tools.js"
-    "./types": { "types": "./dist/types/api.d.ts", "default": "./dist/types/api.js" },
+    },
-    "./ws-frames": { "types": "./dist/types/ws-frames.d.ts", "default": "./dist/types/ws-frames.js" },
+    "./broker": {
-    "./db": { "types": "./dist/db.d.ts", "default": "./dist/db.js" },
+      "types": "./dist/services/broker.d.ts",
-    "./config": { "types": "./dist/config.d.ts", "default": "./dist/config.js" },
+      "default": "./dist/services/broker.js"
-    "./skills": { "types": "./dist/services/skills.d.ts", "default": "./dist/services/skills.js" },
+    },
-    "./skill-invoke": { "types": "./dist/services/skill-invoke.d.ts", "default": "./dist/services/skill-invoke.js" }
+    "./compaction": {
+      "types": "./dist/services/compaction.d.ts",
+      "default": "./dist/services/compaction.js"
+    },
+    "./model-context": {
+      "types": "./dist/services/model-context.d.ts",
+      "default": "./dist/services/model-context.js"
+    },
+    "./system-prompt": {
+      "types": "./dist/services/system-prompt.d.ts",
+      "default": "./dist/services/system-prompt.js"
+    },
+    "./agents": {
+      "types": "./dist/services/agents.d.ts",
+      "default": "./dist/services/agents.js"
+    },
+    "./truncate": {
+      "types": "./dist/services/truncate.d.ts",
+      "default": "./dist/services/truncate.js"
+    },
+    "./path-guard": {
+      "types": "./dist/services/path_guard.d.ts",
+      "default": "./dist/services/path_guard.js"
+    },
+    "./file-ops": {
+      "types": "./dist/services/file_ops.d.ts",
+      "default": "./dist/services/file_ops.js"
+    },
+    "./types": {
+      "types": "./dist/types/api.d.ts",
+      "default": "./dist/types/api.js"
+    },
+    "./ws-frames": {
+      "types": "./dist/types/ws-frames.d.ts",
+      "default": "./dist/types/ws-frames.js"
+    },
+    "./db": {
+      "types": "./dist/db.d.ts",
+      "default": "./dist/db.js"
+    },
+    "./config": {
+      "types": "./dist/config.d.ts",
+      "default": "./dist/config.js"
+    },
+    "./skills": {
+      "types": "./dist/services/skills.d.ts",
+      "default": "./dist/services/skills.js"
+    },
+    "./skill-invoke": {
+      "types": "./dist/services/skill-invoke.d.ts",
+      "default": "./dist/services/skill-invoke.js"
+    }
   },
   "scripts": {
     "dev": "tsx watch src/index.ts",
@@ -36,6 +87,7 @@
     "@modelcontextprotocol/sdk": "^1.29.0",
     "ai": "^6.0.190",
     "fastify": "^4.28.1",
+    "parse5": "^8.0.1",
     "postgres": "^3.4.4",
     "ws": "^8.18.0",
     "zod": "^3.23.8"
@@ -46,5 +98,6 @@
     "tsx": "^4.16.2",
     "typescript": "^5.5.0",
     "vitest": "^3.2.4"
-  }
+  },
+  "license": "AGPL-3.0-only"
 }

apps/server/src/config.ts

@@ -25,6 +25,8 @@ const ConfigSchema = z.object({
   // v2.0.5: cheaper model for titles, summaries, labeling. Falls back to
   // session model (auto_name) or DEFAULT_MODEL when unset.
   FAST_MODEL: z.string().optional(),
+  TASK_MODEL_URL: z.string().url().optional(),
+  LLAMA_SIDECAR_URL: z.string().url().optional(),
 });
 
 export type Config = z.infer<typeof ConfigSchema>;

apps/server/src/index.ts

@@ -28,7 +28,7 @@ import { cleanupTruncations } from './services/truncate.js';
 import { loadMcpConfig } from './services/mcp-config.js';
 import { initialize as initMcp, getTools as getMcpTools, shutdown as shutdownMcp } from './services/mcp-client.js';
 import { appendMcpTools } from './services/tools.js';
-import { refreshToolNames } from './services/agents.js';
+import { refreshToolNames, getAgentsForProject } from './services/agents.js';
 
 async function main() {
   const config = loadConfig();
@@ -91,6 +91,20 @@ async function main() {
   }
   app.addHook('onClose', async () => { await shutdownMcp(); });
 
+  // Boot-time guard: if any agent has llama_extra_args but LLAMA_SIDECAR_URL
+  // is unset, fail fast. Silent fallback would defeat per-agent flags.
+  if (!config.LLAMA_SIDECAR_URL) {
+    const { agents } = await getAgentsForProject('');
+    const offending = agents.find(a => a.llama_extra_args && a.llama_extra_args.length > 0);
+    if (offending) {
+      app.log.fatal(
+        { agent: offending.name },
+        `Agent "${offending.name}" has llama_extra_args but LLAMA_SIDECAR_URL is not set`,
+      );
+      process.exit(1);
+    }
+  }
+
   await app.register(fastifyWebsocket);
 
   app.get('/api/health', async () => {

apps/server/src/schema.sql

@@ -344,6 +344,7 @@ INSERT INTO settings (key, value) VALUES ('theme_mode', '"dark"') ON CONFLICT (k
 ALTER TABLE projects ADD COLUMN IF NOT EXISTS default_system_prompt TEXT NOT NULL DEFAULT '';
 ALTER TABLE projects ADD COLUMN IF NOT EXISTS default_web_search_enabled BOOLEAN NOT NULL DEFAULT false;
 ALTER TABLE sessions ADD COLUMN IF NOT EXISTS web_search_enabled BOOLEAN;
+ALTER TABLE sessions ADD COLUMN IF NOT EXISTS tags TEXT[] DEFAULT '{}';
 
 -- v1.11: anchored rolling compaction.
 --   compacted_at  — marks rows that are "behind the curtain" of the latest
@@ -366,3 +367,39 @@ ALTER TABLE messages ADD COLUMN IF NOT EXISTS summary BOOLEAN NOT NULL DEFAULT F
 ALTER TABLE messages ADD COLUMN IF NOT EXISTS tail_start_id UUID REFERENCES messages(id) ON DELETE SET NULL;
 ALTER TABLE chats ADD COLUMN IF NOT EXISTS needs_compaction BOOLEAN NOT NULL DEFAULT FALSE;
 CREATE INDEX IF NOT EXISTS idx_messages_chat_compacted ON messages (chat_id, compacted_at);
+
+-- tasks table (provider dispatch, arena)
+CREATE TABLE IF NOT EXISTS tasks (
+  id                UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+  project_id        UUID NOT NULL REFERENCES projects(id) ON DELETE CASCADE,
+  session_id        UUID REFERENCES sessions(id) ON DELETE CASCADE,
+  parent_task_id    UUID REFERENCES tasks(id),
+  arena_id          UUID,
+  state             TEXT NOT NULL DEFAULT 'pending'
+                    CHECK (state IN ('pending','running','completed','failed','blocked','cancelled')),
+  input             TEXT NOT NULL,
+  output_summary    TEXT,
+  agent             TEXT,
+  model             TEXT,
+  mode_id           TEXT,
+  thinking_option_id TEXT,
+  feature_values    JSONB,
+  execution_path    TEXT CHECK (execution_path IS NULL OR execution_path IN ('native','acp','pty','qwen')),
+  worktree_path     TEXT,
+  cost_tokens       INTEGER,
+  started_at        TIMESTAMPTZ,
+  ended_at          TIMESTAMPTZ,
+  created_at        TIMESTAMPTZ NOT NULL DEFAULT clock_timestamp()
+);
+
+-- Fix tasks FK to cascade on session delete (existing tables without CASCADE)
+DO $$ BEGIN
+  IF EXISTS (
+    SELECT 1 FROM pg_constraint WHERE conname = 'tasks_session_id_fkey'
+    AND confdeltype != 'c'
+  ) THEN
+    ALTER TABLE tasks DROP CONSTRAINT tasks_session_id_fkey;
+    ALTER TABLE tasks ADD CONSTRAINT tasks_session_id_fkey
+      FOREIGN KEY (session_id) REFERENCES sessions(id) ON DELETE CASCADE;
+  END IF;
+END $$;

apps/server/src/services/__tests__/html-to-md.test.ts

@@ -0,0 +1,223 @@
+import { describe, expect, it } from 'vitest';
+import { htmlToMarkdown } from '../web/html-to-md.js';
+
+describe('htmlToMarkdown', () => {
+  it('converts h1 heading', () => {
+    expect(htmlToMarkdown('<h1>Title</h1>')).toBe('# Title');
+  });
+
+  it('converts h1 through h6', () => {
+    const html = '<h1>One</h1><h2>Two</h2><h3>Three</h3><h4>Four</h4><h5>Five</h5><h6>Six</h6>';
+    const md = htmlToMarkdown(html);
+    expect(md).toContain('# One');
+    expect(md).toContain('## Two');
+    expect(md).toContain('### Three');
+    expect(md).toContain('#### Four');
+    expect(md).toContain('##### Five');
+    expect(md).toContain('###### Six');
+  });
+
+  it('converts anchor with href', () => {
+    expect(htmlToMarkdown('<a href="https://example.com">click here</a>'))
+      .toBe('[click here](https://example.com)');
+  });
+
+  it('converts anchor without href to plain text', () => {
+    expect(htmlToMarkdown('<a>just text</a>')).toBe('just text');
+  });
+
+  it('converts bold and italic', () => {
+    expect(htmlToMarkdown('<strong>bold</strong>')).toBe('**bold**');
+    expect(htmlToMarkdown('<b>bold</b>')).toBe('**bold**');
+    expect(htmlToMarkdown('<em>italic</em>')).toBe('*italic*');
+    expect(htmlToMarkdown('<i>italic</i>')).toBe('*italic*');
+  });
+
+  it('handles combined bold+italic', () => {
+    const md = htmlToMarkdown('<strong><em>bold italic</em></strong>');
+    expect(md).toBe('***bold italic***');
+  });
+
+  it('converts unordered list', () => {
+    const html = '<ul><li>one</li><li>two</li><li>three</li></ul>';
+    const md = htmlToMarkdown(html);
+    expect(md).toContain('* one');
+    expect(md).toContain('* two');
+    expect(md).toContain('* three');
+  });
+
+  it('converts ordered list', () => {
+    const html = '<ol><li>first</li><li>second</li></ol>';
+    const md = htmlToMarkdown(html);
+    expect(md).toContain('1. first');
+    expect(md).toContain('2. second');
+  });
+
+  it('handles nested lists', () => {
+    const html = '<ul><li>outer<ul><li>inner</li></ul></li></ul>';
+    const md = htmlToMarkdown(html);
+    expect(md).toContain('* outer');
+    expect(md).toContain('  * inner');
+  });
+
+  it('converts 3-column GFM table with header', () => {
+    const html = `
+      <table>
+        <thead><tr><th>Name</th><th>Age</th><th>City</th></tr></thead>
+        <tbody>
+          <tr><td>Alice</td><td>30</td><td>NYC</td></tr>
+          <tr><td>Bob</td><td>25</td><td>LA</td></tr>
+        </tbody>
+      </table>`;
+    const md = htmlToMarkdown(html);
+    expect(md).toContain('| Name | Age | City |');
+    expect(md).toContain('| --- | --- | --- |');
+    expect(md).toContain('| Alice | 30 | NYC |');
+    expect(md).toContain('| Bob | 25 | LA |');
+  });
+
+  it('escapes pipe characters in table cells', () => {
+    const html = '<table><tr><th>A</th></tr><tr><td>x | y</td></tr></table>';
+    const md = htmlToMarkdown(html);
+    expect(md).toContain('x \\| y');
+  });
+
+  it('converts blockquote', () => {
+    const html = '<blockquote><p>quoted text</p></blockquote>';
+    const md = htmlToMarkdown(html);
+    expect(md).toContain('> quoted text');
+  });
+
+  it('converts multi-line blockquote', () => {
+    const html = '<blockquote><p>line one</p><p>line two</p></blockquote>';
+    const md = htmlToMarkdown(html);
+    expect(md).toContain('> line one');
+    expect(md).toContain('> line two');
+  });
+
+  it('converts fenced code block', () => {
+    const html = '<pre><code>const x = 1;</code></pre>';
+    const md = htmlToMarkdown(html);
+    expect(md).toContain('```\nconst x = 1;\n```');
+  });
+
+  it('preserves language hint from code class', () => {
+    const html = '<pre><code class="language-py">print("hello")</code></pre>';
+    const md = htmlToMarkdown(html);
+    expect(md).toContain('```py\nprint("hello")\n```');
+  });
+
+  it('converts inline code', () => {
+    expect(htmlToMarkdown('use <code>npm install</code> to install'))
+      .toContain('`npm install`');
+  });
+
+  it('decodes HTML entities', () => {
+    expect(htmlToMarkdown('&amp; &lt; &gt; &quot;')).toBe('& < > "');
+  });
+
+  it('decodes numeric character references', () => {
+    expect(htmlToMarkdown('&#39;')).toBe("'");
+  });
+
+  it('decodes &nbsp; as space', () => {
+    const md = htmlToMarkdown('hello&nbsp;world');
+    expect(md).toMatch(/hello\s+world/);
+  });
+
+  it('skips script content', () => {
+    const html = '<p>before</p><script>alert("xss")</script><p>after</p>';
+    const md = htmlToMarkdown(html);
+    expect(md).not.toContain('alert');
+    expect(md).toContain('before');
+    expect(md).toContain('after');
+  });
+
+  it('skips style content', () => {
+    const html = '<p>text</p><style>body { color: red }</style>';
+    const md = htmlToMarkdown(html);
+    expect(md).not.toContain('color');
+    expect(md).toContain('text');
+  });
+
+  it('does not throw on malformed HTML', () => {
+    expect(() => htmlToMarkdown('<p>unclosed <b>bold <i>italic')).not.toThrow();
+    const md = htmlToMarkdown('<p>unclosed <b>bold <i>italic');
+    expect(md).toContain('bold');
+    expect(md).toContain('italic');
+  });
+
+  it('returns empty string for empty input', () => {
+    expect(htmlToMarkdown('')).toBe('');
+  });
+
+  it('returns empty string for whitespace-only input', () => {
+    expect(htmlToMarkdown('   \n\n  ')).toBe('');
+  });
+
+  it('converts hr to horizontal rule', () => {
+    const md = htmlToMarkdown('<p>above</p><hr><p>below</p>');
+    expect(md).toContain('---');
+  });
+
+  it('converts br to newline', () => {
+    const md = htmlToMarkdown('line one<br>line two');
+    expect(md).toContain('line one\nline two');
+  });
+
+  it('handles ol with start attribute', () => {
+    const html = '<ol start="5"><li>five</li><li>six</li></ol>';
+    const md = htmlToMarkdown(html);
+    expect(md).toContain('5. five');
+    expect(md).toContain('6. six');
+  });
+
+  it('collapses excessive blank lines', () => {
+    const html = '<p>one</p><p></p><p></p><p></p><p>two</p>';
+    const md = htmlToMarkdown(html);
+    const blankRuns = md.match(/\n{3,}/g);
+    expect(blankRuns).toBeNull();
+  });
+
+  // Golden test: small Hacker News-style snippet
+  it('golden: HN-style snippet produces structured markdown', () => {
+    const html = `
+      <html>
+      <head><title>Test Page</title></head>
+      <body>
+        <h1>Welcome</h1>
+        <p>This is a <strong>test</strong> page with <a href="https://example.com">a link</a>.</p>
+        <h2>Features</h2>
+        <ul>
+          <li>Fast</li>
+          <li>Reliable</li>
+          <li>Secure</li>
+        </ul>
+        <h2>Data</h2>
+        <table>
+          <thead><tr><th>Metric</th><th>Value</th></tr></thead>
+          <tbody>
+            <tr><td>Uptime</td><td>99.9%</td></tr>
+            <tr><td>Latency</td><td>42ms</td></tr>
+          </tbody>
+        </table>
+        <blockquote><p>This tool is amazing.</p></blockquote>
+        <pre><code class="language-js">console.log("hello");</code></pre>
+        <script>evil();</script>
+      </body>
+      </html>`;
+    const md = htmlToMarkdown(html);
+    expect(md).toContain('# Welcome');
+    expect(md).toContain('**test**');
+    expect(md).toContain('[a link](https://example.com)');
+    expect(md).toContain('## Features');
+    expect(md).toContain('* Fast');
+    expect(md).toContain('| Metric | Value |');
+    expect(md).toContain('| --- | --- |');
+    expect(md).toContain('| Uptime | 99.9% |');
+    expect(md).toContain('> This tool is amazing.');
+    expect(md).toContain('```js\nconsole.log("hello");\n```');
+    expect(md).not.toContain('evil');
+    expect(md).not.toContain('<title>');
+  });
+});

apps/server/src/services/__tests__/llama-args-validator.test.ts

@@ -0,0 +1,160 @@
+import { describe, expect, it } from 'vitest';
+import {
+  validateExtraArgs,
+  isManagedFlag,
+  stripShadowingFlags,
+} from '../inference/llama-args-validator.js';
+import { parseAgentsMd } from '../agents.js';
+
+describe('validateExtraArgs', () => {
+  describe('deny list — each alias rejected', () => {
+    const denied = [
+      '-m', '--model',
+      '-mu', '--model-url',
+      '-dr', '--docker-repo',
+      '-hf', '-hfr', '--hf-repo',
+      '-hff', '--hf-file',
+      '-hfv', '-hfrv', '--hf-repo-v',
+      '-hffv', '--hf-file-v',
+      '-hft', '--hf-token',
+      '-mm', '--mmproj',
+      '-mmu', '--mmproj-url',
+      '--host', '--port', '--path', '--api-prefix', '--reuse-port',
+      '--api-key', '--api-key-file',
+      '--ssl-key-file', '--ssl-cert-file',
+      '--webui', '--no-webui', '--ui', '--no-ui',
+      '--ui-config', '--ui-config-file',
+      '--ui-mcp-proxy', '--no-ui-mcp-proxy',
+      '--models-dir', '--models-preset', '--models-max',
+      '--models-autoload', '--no-models-autoload',
+    ];
+    for (const flag of denied) {
+      it(`rejects ${flag}`, () => {
+        expect(() => validateExtraArgs([flag])).toThrow(/managed/);
+      });
+    }
+  });
+
+  describe('safe flags accepted', () => {
+    const safe = [
+      '-c', '--ctx-size', '-ngl', '--gpu-layers',
+      '--top-k', '--cache-type-k', '--jinja', '--no-jinja',
+      '--spec-draft-n-max', '-fa', '--flash-attn',
+      '-t', '--threads', '-np', '--parallel',
+    ];
+    for (const flag of safe) {
+      it(`accepts ${flag}`, () => {
+        expect(() => validateExtraArgs([flag])).not.toThrow();
+        expect(validateExtraArgs([flag])).toEqual([flag]);
+      });
+    }
+  });
+
+  it('handles --flag=value shape (denies the flag part)', () => {
+    expect(() => validateExtraArgs(['--model=evil.gguf'])).toThrow(/managed/);
+  });
+
+  it('handles --flag=value shape (accepts safe flag)', () => {
+    expect(validateExtraArgs(['--ctx-size=4096'])).toEqual(['--ctx-size=4096']);
+  });
+
+  it('returns empty array for undefined input', () => {
+    expect(validateExtraArgs(undefined)).toEqual([]);
+  });
+
+  it('returns empty array for empty input', () => {
+    expect(validateExtraArgs([])).toEqual([]);
+  });
+
+  it('treats negative numbers as values, not flags', () => {
+    expect(validateExtraArgs(['--seed', '-1'])).toEqual(['--seed', '-1']);
+  });
+});
+
+describe('isManagedFlag', () => {
+  it('returns true for denied flags', () => {
+    expect(isManagedFlag('--model')).toBe(true);
+    expect(isManagedFlag('-m')).toBe(true);
+    expect(isManagedFlag('--api-key')).toBe(true);
+    expect(isManagedFlag('--port')).toBe(true);
+  });
+
+  it('returns false for safe flags', () => {
+    expect(isManagedFlag('-c')).toBe(false);
+    expect(isManagedFlag('--ctx-size')).toBe(false);
+    expect(isManagedFlag('--top-k')).toBe(false);
+  });
+});
+
+describe('stripShadowingFlags', () => {
+  it('strips auto -c when user supplies -c', () => {
+    const result = stripShadowingFlags(['-c', '4096', '--top-k', '40']);
+    expect(result).toEqual(['--top-k', '40']);
+  });
+
+  it('retains both when no overlap', () => {
+    const result = stripShadowingFlags(['--top-k', '40', '--top-p', '0.95']);
+    expect(result).toEqual(['--top-k', '40', '--top-p', '0.95']);
+  });
+
+  it('strips --ctx-size=value form', () => {
+    const result = stripShadowingFlags(['--ctx-size=4096']);
+    expect(result).toEqual([]);
+  });
+
+  it('strips boolean --jinja flag (no value consumed)', () => {
+    const result = stripShadowingFlags(['--jinja', '--top-k', '40']);
+    expect(result).toEqual(['--top-k', '40']);
+  });
+
+  it('respects stripContext=false to keep context flags', () => {
+    const result = stripShadowingFlags(['-c', '4096'], { stripContext: false });
+    expect(result).toEqual(['-c', '4096']);
+  });
+
+  it('strips cache flags by default', () => {
+    const result = stripShadowingFlags(['--cache-type-k', 'q8_0']);
+    expect(result).toEqual([]);
+  });
+
+  it('strips spec flags by default', () => {
+    const result = stripShadowingFlags(['--spec-draft-n-max', '16']);
+    expect(result).toEqual([]);
+  });
+});
+
+describe('AGENTS.md frontmatter validation', () => {
+  it('rejects agent with managed flag in llama_extra_args', () => {
+    const md = `## Evil Agent
+---
+llama_extra_args: ["--model", "evil.gguf"]
+---
+You are evil.`;
+    const { agents, errors } = parseAgentsMd(md);
+    expect(agents).toHaveLength(0);
+    expect(errors).toHaveLength(1);
+    expect(errors[0]!.reason).toContain('managed');
+  });
+
+  it('accepts agent with safe llama_extra_args', () => {
+    const md = `## Good Agent
+---
+llama_extra_args: ["--top-k", "20"]
+---
+You are good.`;
+    const { agents, errors } = parseAgentsMd(md);
+    expect(errors).toHaveLength(0);
+    expect(agents).toHaveLength(1);
+    expect(agents[0]!.llama_extra_args).toEqual(['--top-k', '20']);
+  });
+
+  it('agent without llama_extra_args has null field', () => {
+    const md = `## Simple Agent
+---
+temperature: 0.5
+---
+You are simple.`;
+    const { agents } = parseAgentsMd(md);
+    expect(agents[0]!.llama_extra_args).toBeNull();
+  });
+});

apps/server/src/services/__tests__/provider.test.ts

@@ -0,0 +1,58 @@
+import { describe, expect, it } from 'vitest';
+import { resolveRoute, upstreamModel } from '../inference/provider.js';
+
+describe('resolveRoute', () => {
+  it('routes to swap when agent is null', () => {
+    expect(resolveRoute(null)).toEqual({ route: 'swap', flags: null });
+  });
+
+  it('routes to swap when agent has no llama_extra_args', () => {
+    expect(resolveRoute({ llama_extra_args: null })).toEqual({ route: 'swap', flags: null });
+  });
+
+  it('routes to swap when agent has empty llama_extra_args', () => {
+    expect(resolveRoute({ llama_extra_args: [] })).toEqual({ route: 'swap', flags: null });
+  });
+
+  it('routes to sidecar when agent has llama_extra_args', () => {
+    const result = resolveRoute({ llama_extra_args: ['--top-k', '20'] });
+    expect(result.route).toBe('sidecar');
+    expect(result.flags).toEqual(['--top-k', '20']);
+  });
+});
+
+describe('upstreamModel', () => {
+  const swapConfig = { LLAMA_SWAP_URL: 'http://localhost:8401' };
+  const fullConfig = {
+    LLAMA_SWAP_URL: 'http://localhost:8401',
+    LLAMA_SIDECAR_URL: 'http://localhost:8402',
+  };
+
+  it('returns a model for swap route (no agent)', () => {
+    const model = upstreamModel(swapConfig, 'test-model');
+    expect(model).toBeDefined();
+    expect((model as any).modelId).toBe('test-model');
+  });
+
+  it('returns a model for swap route (agent without extra args)', () => {
+    const model = upstreamModel(swapConfig, 'test-model', { llama_extra_args: null });
+    expect(model).toBeDefined();
+  });
+
+  it('returns a model for sidecar route', () => {
+    const model = upstreamModel(fullConfig, 'test-model', { llama_extra_args: ['--top-k', '20'] });
+    expect(model).toBeDefined();
+    expect((model as any).modelId).toBe('test-model');
+  });
+
+  it('throws when sidecar route requested but URL missing', () => {
+    expect(() =>
+      upstreamModel(swapConfig, 'test-model', { llama_extra_args: ['--top-k', '20'] }),
+    ).toThrow(/LLAMA_SIDECAR_URL/);
+  });
+
+  it('routes to swap for empty llama_extra_args array', () => {
+    const model = upstreamModel(swapConfig, 'test-model', { llama_extra_args: [] });
+    expect(model).toBeDefined();
+  });
+});

apps/server/src/services/__tests__/tool-call-parser.test.ts

@@ -1,25 +1,24 @@
-// v1.13.16: covers the Qwen/Hermes <tool_call> parser, the new Anthropic
-// <invoke> parser, the partial-opener detector for both flavors, the unified
-// extraction helper, and the unknown-tool error formatter that downstream
-// dispatch uses to give the model a recovery hint when it drifts to a
-// Claude Code tool name like read_file instead of BooCode's view_file.
-
 import { describe, expect, it } from 'vitest';
 import {
   parseXmlToolCall,
   parseInvokeToolCall,
   partialXmlOpenerStart,
   extractToolCallBlocks,
+  parseToolCallsFromText,
+  stripToolMarkup,
+  hasToolSignal,
   XML_TOOL_OPEN,
   XML_TOOL_CLOSE,
   INVOKE_TOOL_OPEN,
   INVOKE_TOOL_CLOSE,
-} from '../inference/xml-parser.js';
+  TOOL_XML_SIGNALS,
-import {
+  BUDGET_EXHAUSTED_NUDGE,
-  levenshtein,
+  DUPLICATE_CALL_NUDGE,
-  suggestToolName,
+  TOOL_ERROR_NUDGE,
-  formatUnknownToolError,
+  TOOL_ERROR_PREFIXES,
-} from '../inference/tool-suggestions.js';
+} from '../inference/tool-call-parser.js';
+
+// ── Ported from xml-parser.test.ts ───────────────────────────────────────
 
 describe('parseXmlToolCall (Qwen/Hermes <tool_call>)', () => {
   it('parses a well-formed single-parameter call', () => {
@@ -66,7 +65,6 @@ describe('parseXmlToolCall (Qwen/Hermes <tool_call>)', () => {
 });
 
 describe('parseInvokeToolCall (Anthropic <invoke>) — v1.13.16', () => {
-  // Spec case 1
   it('parses a well-formed single-parameter call (spec case 1)', () => {
     const block = '<invoke name="view_file"><parameter name="path">/tmp/foo</parameter></invoke>';
     expect(parseInvokeToolCall(block)).toEqual({
@@ -75,7 +73,6 @@ describe('parseInvokeToolCall (Anthropic <invoke>) — v1.13.16', () => {
     });
   });
 
-  // Spec case 2
   it('parses a multi-parameter call (spec case 2)', () => {
     const block = '<invoke name="grep"><parameter name="pattern">foo</parameter><parameter name="path">src/</parameter></invoke>';
     expect(parseInvokeToolCall(block)).toEqual({
@@ -84,7 +81,6 @@ describe('parseInvokeToolCall (Anthropic <invoke>) — v1.13.16', () => {
     });
   });
 
-  // Spec case 3
   it('tolerates newlines and spaces in attributes (spec case 3)', () => {
     const block = `<invoke
       name="view_file"
@@ -99,7 +95,6 @@ describe('parseInvokeToolCall (Anthropic <invoke>) — v1.13.16', () => {
     });
   });
 
-  // Spec case 4 (parser portion — the not-found enrichment is tested below)
   it('parses a call whose name is not a registered BooCode tool (spec case 4)', () => {
     const block = '<invoke name="read_file"><parameter name="path">/tmp/foo</parameter></invoke>';
     expect(parseInvokeToolCall(block)).toEqual({
@@ -187,7 +182,6 @@ describe('partialXmlOpenerStart (v1.13.16 — both flavors)', () => {
 });
 
 describe('extractToolCallBlocks (v1.13.16 — unified extraction)', () => {
-  // Spec case 1 (extraction-level)
   it('extracts a single <invoke> block (spec case 1)', () => {
     const input = '<invoke name="view_file"><parameter name="path">/tmp/foo</parameter></invoke>';
     const result = extractToolCallBlocks(input);
@@ -196,7 +190,6 @@ describe('extractToolCallBlocks (v1.13.16 — unified extraction)', () => {
     expect(result.remaining).toBe('');
   });
 
-  // Spec case 5: opener arrives in one chunk, closer in the next.
   it('holds the partial <invoke> chunk when the closer has not arrived (spec case 5, first chunk)', () => {
     const firstChunk = '<invoke name="view_file"><parameter name="path">/tmp/foo</parameter>';
     const result = extractToolCallBlocks(firstChunk);
@@ -215,7 +208,6 @@ describe('extractToolCallBlocks (v1.13.16 — unified extraction)', () => {
     expect(r2.remaining).toBe('');
   });
 
-  // Spec case 6: prose interleaving
   it('flushes prose around a recognized block but not the markup itself (spec case 6)', () => {
     const input = 'I will read the file.\n<invoke name="view_file"><parameter name="path">/tmp/foo</parameter></invoke>\nThanks.';
     const result = extractToolCallBlocks(input);
@@ -224,7 +216,6 @@ describe('extractToolCallBlocks (v1.13.16 — unified extraction)', () => {
     expect(result.remaining).toBe('');
   });
 
-  // Spec case 7 regression
   it('extracts a <tool_call> Qwen block alongside the new code path (spec case 7 regression)', () => {
     const input = '<tool_call><function=view_file><parameter=path>/tmp/foo</parameter></function></tool_call>';
     const result = extractToolCallBlocks(input);
@@ -310,86 +301,245 @@ describe('extractToolCallBlocks (v1.13.16 — unified extraction)', () => {
   });
 });
 
-describe('levenshtein', () => {
+// ── New tests: Unsloth-ported functions ──────────────────────────────────
-  it('returns 0 for identical strings', () => {
+
-    expect(levenshtein('view_file', 'view_file')).toBe(0);
+describe('hasToolSignal', () => {
+  it('returns true for <tool_call>', () => {
+    expect(hasToolSignal('prefix <tool_call> suffix')).toBe(true);
   });
 
-  it('returns the length when one string is empty', () => {
+  it('returns true for <function=', () => {
-    expect(levenshtein('', 'view_file')).toBe(9);
+    expect(hasToolSignal('prefix <function=view_file> suffix')).toBe(true);
-    expect(levenshtein('view_file', '')).toBe(9);
   });
 
-  it('computes a small distance for a single-character substitution', () => {
+  it('returns true for <invoke', () => {
-    expect(levenshtein('cat', 'bat')).toBe(1);
+    expect(hasToolSignal('prefix <invoke name="x"> suffix')).toBe(true);
   });
 
-  it('computes a known case: read_file → view_file is 4', () => {
+  it('returns false for near-miss <tool>', () => {
-    // r→v, e→i, a→e, d→w → 4 substitutions, same length
+    expect(hasToolSignal('prefix <tool> suffix')).toBe(false);
-    expect(levenshtein('read_file', 'view_file')).toBe(4);
+  });
+
+  it('returns false for near-miss <function>', () => {
+    expect(hasToolSignal('prefix <function> suffix')).toBe(false);
+  });
+
+  it('returns false for near-miss <tool_call_thing>', () => {
+    expect(hasToolSignal('<tool_call_thing>')).toBe(false);
+  });
+
+  it('returns false for plain text', () => {
+    expect(hasToolSignal('just some text')).toBe(false);
   });
 });
 
-describe('suggestToolName (v1.13.16)', () => {
+describe('stripToolMarkup', () => {
-  const tools = [
+  it('strips closed <tool_call> blocks', () => {
-    'view_file',
+    const input = 'before <tool_call>{"name":"x"}</tool_call> after';
-    'list_dir',
+    expect(stripToolMarkup(input)).toBe('before  after');
-    'grep',
-    'find_files',
-    'view_truncated_output',
-    'ask_user_input',
-    'web_search',
-  ];
-
-  it('suggests the closest match when distance is small', () => {
-    expect(suggestToolName('view_files', tools)).toBe('view_file');
   });
 
-  it('suggests via substring match when distance alone would miss', () => {
+  it('strips closed <function=...> blocks', () => {
-    // 'file' is a substring of multiple tools; closest by distance wins.
+    const input = 'before <function=x><parameter=y>z</parameter></function> after';
-    expect(suggestToolName('file', tools)).toBe('view_file');
+    expect(stripToolMarkup(input)).toBe('before  after');
   });
 
-  it('returns null when nothing is close', () => {
+  it('strips closed <invoke> blocks', () => {
-    expect(suggestToolName('xxxx_yyyy_zzzz', tools)).toBeNull();
+    const input = 'before <invoke name="x"><parameter name="y">z</parameter></invoke> after';
+    expect(stripToolMarkup(input)).toBe('before  after');
   });
 
-  it('is case-insensitive in the distance check', () => {
+  it('leaves trailing unclosed block when final=false', () => {
-    expect(suggestToolName('VIEW_FILE', tools)).toBe('view_file');
+    const input = 'text <tool_call>{"name":"x"';
+    expect(stripToolMarkup(input)).toBe('text <tool_call>{"name":"x"');
+  });
+
+  it('strips trailing unclosed <tool_call> when final=true', () => {
+    const input = 'text <tool_call>{"name":"x"';
+    expect(stripToolMarkup(input, { final: true })).toBe('text');
+  });
+
+  it('strips trailing unclosed <function= when final=true', () => {
+    const input = 'text <function=run_bash><parameter=command>ls';
+    expect(stripToolMarkup(input, { final: true })).toBe('text');
+  });
+
+  it('strips trailing unclosed <invoke when final=true', () => {
+    const input = 'text <invoke name="x"><parameter name="y">val';
+    expect(stripToolMarkup(input, { final: true })).toBe('text');
+  });
+
+  it('trims whitespace when final=true', () => {
+    const input = '  text  <tool_call>partial';
+    expect(stripToolMarkup(input, { final: true })).toBe('text');
+  });
+
+  it('strips multiple closed blocks', () => {
+    const input = '<tool_call>a</tool_call> mid <tool_call>b</tool_call>';
+    expect(stripToolMarkup(input)).toBe(' mid ');
   });
 });
 
-describe('formatUnknownToolError (v1.13.16)', () => {
+describe('parseToolCallsFromText', () => {
-  const tools = ['view_file', 'list_dir', 'grep', 'find_files'];
+  describe('pattern 1: <tool_call>{json}</tool_call>', () => {
+    it('parses a well-formed JSON tool call', () => {
+      const input = '<tool_call>{"name":"web_search","arguments":{"query":"hello"}}</tool_call>';
+      const calls = parseToolCallsFromText(input);
+      expect(calls).toHaveLength(1);
+      expect(calls[0]!.id).toBe('call_0');
+      expect(calls[0]!.type).toBe('function');
+      expect(calls[0]!.function.name).toBe('web_search');
+      expect(JSON.parse(calls[0]!.function.arguments)).toEqual({ query: 'hello' });
+    });
 
-  it('includes the wrong name and the available tools list', () => {
+    it('handles string arguments field', () => {
-    const msg = formatUnknownToolError('read_file', tools);
+      const input = '<tool_call>{"name":"x","arguments":"already a string"}</tool_call>';
-    expect(msg).toContain("Tool 'read_file' not found");
+      const calls = parseToolCallsFromText(input);
-    expect(msg).toContain('Available tools:');
+      expect(calls[0]!.function.arguments).toBe('already a string');
-    expect(msg).toContain('view_file');
+    });
-    expect(msg).toContain('find_files');
+
+    it('handles balanced braces inside JSON strings', () => {
+      const input = '<tool_call>{"name":"x","arguments":{"q":"} { extra "}}</tool_call>';
+      const calls = parseToolCallsFromText(input);
+      expect(calls).toHaveLength(1);
+      const parsed = JSON.parse(calls[0]!.function.arguments);
+      expect(parsed.q).toBe('} { extra ');
+    });
+
+    it('respects idOffset', () => {
+      const input = '<tool_call>{"name":"a","arguments":{}}</tool_call>';
+      const calls = parseToolCallsFromText(input, { idOffset: 5 });
+      expect(calls[0]!.id).toBe('call_5');
+    });
+
+    it('parses multiple JSON tool calls', () => {
+      const input =
+        '<tool_call>{"name":"a","arguments":{}}</tool_call>' +
+        '<tool_call>{"name":"b","arguments":{}}</tool_call>';
+      const calls = parseToolCallsFromText(input);
+      expect(calls).toHaveLength(2);
+      expect(calls[0]!.id).toBe('call_0');
+      expect(calls[1]!.id).toBe('call_1');
+    });
+
+    it('skips malformed JSON', () => {
+      const input = '<tool_call>{not json}</tool_call>';
+      const calls = parseToolCallsFromText(input);
+      expect(calls).toHaveLength(0);
+    });
+
+    it('handles missing closing tag', () => {
+      const input = '<tool_call>{"name":"x","arguments":{"q":"hello"}}';
+      const calls = parseToolCallsFromText(input);
+      expect(calls).toHaveLength(1);
+      expect(calls[0]!.function.name).toBe('x');
+    });
   });
 
-  it('includes a suggestion when the drifted name is within threshold', () => {
+  describe('pattern 2: <function=name><parameter=key>value', () => {
-    // distance(view_files, view_file) = 1 (one extra char)
+    it('parses a single-parameter function call', () => {
-    const msg = formatUnknownToolError('view_files', tools);
+      const input = '<function=view_file><parameter=path>/tmp/foo</parameter></function>';
-    expect(msg).toContain('Did you mean: view_file?');
+      const calls = parseToolCallsFromText(input);
+      expect(calls).toHaveLength(1);
+      expect(calls[0]!.function.name).toBe('view_file');
+      expect(JSON.parse(calls[0]!.function.arguments)).toEqual({ path: '/tmp/foo' });
+    });
+
+    it('single-param fast path preserves embedded </parameter>', () => {
+      const input = '<function=run_bash><parameter=command>echo "</parameter>"</parameter></function>';
+      const calls = parseToolCallsFromText(input);
+      expect(calls).toHaveLength(1);
+      expect(JSON.parse(calls[0]!.function.arguments).command).toBe('echo "</parameter>"');
+    });
+
+    it('multi-param: value of first stops at start of second', () => {
+      const input = '<function=grep><parameter=pattern>foo</parameter><parameter=path>src/</parameter></function>';
+      const calls = parseToolCallsFromText(input);
+      expect(calls).toHaveLength(1);
+      const args = JSON.parse(calls[0]!.function.arguments);
+      expect(args.pattern).toBe('foo');
+      expect(args.path).toBe('src/');
+    });
+
+    it('tolerates missing closing tags', () => {
+      const input = '<function=view_file><parameter=path>/tmp/foo';
+      const calls = parseToolCallsFromText(input);
+      expect(calls).toHaveLength(1);
+      expect(calls[0]!.function.name).toBe('view_file');
+      expect(JSON.parse(calls[0]!.function.arguments)).toEqual({ path: '/tmp/foo' });
+    });
+
+    it('does not fire when pattern 1 found results', () => {
+      const input = '<tool_call>{"name":"a","arguments":{}}</tool_call><function=b><parameter=x>y</parameter></function>';
+      const calls = parseToolCallsFromText(input);
+      expect(calls).toHaveLength(1);
+      expect(calls[0]!.function.name).toBe('a');
+    });
   });
 
-  it('omits the suggestion clause when no tool is close enough', () => {
+  describe('pattern 3: <invoke name="..."><parameter name="...">value (Anthropic)', () => {
-    const msg = formatUnknownToolError('zzzzzzz', tools);
+    it('parses a single-parameter invoke call', () => {
-    expect(msg).toContain("Tool 'zzzzzzz' not found");
+      const input = '<invoke name="view_file"><parameter name="path">/tmp/foo</parameter></invoke>';
-    expect(msg).toContain('Available tools:');
+      const calls = parseToolCallsFromText(input);
-    expect(msg).not.toContain('Did you mean');
+      expect(calls).toHaveLength(1);
-  });
+      expect(calls[0]!.function.name).toBe('view_file');
+      expect(JSON.parse(calls[0]!.function.arguments)).toEqual({ path: '/tmp/foo' });
+    });
 
-  // The drift incident in the recon (chat 30d8…1be7167, msg 7ff558f4) had the
+    it('parses multi-parameter invoke call', () => {
-  // model emit <invoke name="read_file">. lev(read_file, view_file) = 4, so
+      const input = '<invoke name="grep"><parameter name="pattern">foo</parameter><parameter name="path">src/</parameter></invoke>';
-  // the spec's threshold (<=3) doesn't suggest view_file — the model still
+      const calls = parseToolCallsFromText(input);
-  // gets the available-tools list to pick from. This pins that behavior so a
+      expect(calls).toHaveLength(1);
-  // future loosening of the threshold is a deliberate choice.
+      const args = JSON.parse(calls[0]!.function.arguments);
-  it('does not suggest view_file for the read_file drift case (distance is 4, over threshold)', () => {
+      expect(args.pattern).toBe('foo');
-    const msg = formatUnknownToolError('read_file', tools);
+      expect(args.path).toBe('src/');
-    expect(msg).not.toContain('Did you mean');
+    });
+
+    it('does not fire when pattern 1 found results', () => {
+      const input = '<tool_call>{"name":"a","arguments":{}}</tool_call><invoke name="b"><parameter name="x">y</parameter></invoke>';
+      const calls = parseToolCallsFromText(input);
+      expect(calls).toHaveLength(1);
+      expect(calls[0]!.function.name).toBe('a');
+    });
+
+    it('does not fire when pattern 2 found results', () => {
+      const input = '<function=a><parameter=x>y</parameter></function><invoke name="b"><parameter name="x">y</parameter></invoke>';
+      const calls = parseToolCallsFromText(input);
+      expect(calls).toHaveLength(1);
+      expect(calls[0]!.function.name).toBe('a');
+    });
+
+    it('tolerates missing closing tags', () => {
+      const input = '<invoke name="view_file"><parameter name="path">/tmp/foo';
+      const calls = parseToolCallsFromText(input);
+      expect(calls).toHaveLength(1);
+      expect(JSON.parse(calls[0]!.function.arguments)).toEqual({ path: '/tmp/foo' });
+    });
+
+    it('supports single-quoted attributes', () => {
+      const input = "<invoke name='view_file'><parameter name='path'>/tmp/foo</parameter></invoke>";
+      const calls = parseToolCallsFromText(input);
+      expect(calls).toHaveLength(1);
+      expect(calls[0]!.function.name).toBe('view_file');
+    });
+  });
+});
+
+describe('constants', () => {
+  it('TOOL_XML_SIGNALS includes all three signal prefixes', () => {
+    expect(TOOL_XML_SIGNALS).toContain('<tool_call>');
+    expect(TOOL_XML_SIGNALS).toContain('<function=');
+    expect(TOOL_XML_SIGNALS).toContain('<invoke');
+  });
+
+  it('nudge constants are non-empty strings', () => {
+    expect(BUDGET_EXHAUSTED_NUDGE.length).toBeGreaterThan(0);
+    expect(DUPLICATE_CALL_NUDGE.length).toBeGreaterThan(0);
+    expect(TOOL_ERROR_NUDGE.length).toBeGreaterThan(0);
+  });
+
+  it('TOOL_ERROR_PREFIXES is a non-empty tuple', () => {
+    expect(TOOL_ERROR_PREFIXES.length).toBeGreaterThan(0);
+    expect(TOOL_ERROR_PREFIXES).toContain('Error');
   });
 });

apps/server/src/services/__tests__/tool-suggestions.test.ts

@@ -0,0 +1,82 @@
+import { describe, expect, it } from 'vitest';
+import {
+  levenshtein,
+  suggestToolName,
+  formatUnknownToolError,
+} from '../inference/tool-suggestions.js';
+
+describe('levenshtein', () => {
+  it('returns 0 for identical strings', () => {
+    expect(levenshtein('view_file', 'view_file')).toBe(0);
+  });
+
+  it('returns the length when one string is empty', () => {
+    expect(levenshtein('', 'view_file')).toBe(9);
+    expect(levenshtein('view_file', '')).toBe(9);
+  });
+
+  it('computes a small distance for a single-character substitution', () => {
+    expect(levenshtein('cat', 'bat')).toBe(1);
+  });
+
+  it('computes a known case: read_file → view_file is 4', () => {
+    expect(levenshtein('read_file', 'view_file')).toBe(4);
+  });
+});
+
+describe('suggestToolName (v1.13.16)', () => {
+  const tools = [
+    'view_file',
+    'list_dir',
+    'grep',
+    'find_files',
+    'view_truncated_output',
+    'ask_user_input',
+    'web_search',
+  ];
+
+  it('suggests the closest match when distance is small', () => {
+    expect(suggestToolName('view_files', tools)).toBe('view_file');
+  });
+
+  it('suggests via substring match when distance alone would miss', () => {
+    expect(suggestToolName('file', tools)).toBe('view_file');
+  });
+
+  it('returns null when nothing is close', () => {
+    expect(suggestToolName('xxxx_yyyy_zzzz', tools)).toBeNull();
+  });
+
+  it('is case-insensitive in the distance check', () => {
+    expect(suggestToolName('VIEW_FILE', tools)).toBe('view_file');
+  });
+});
+
+describe('formatUnknownToolError (v1.13.16)', () => {
+  const tools = ['view_file', 'list_dir', 'grep', 'find_files'];
+
+  it('includes the wrong name and the available tools list', () => {
+    const msg = formatUnknownToolError('read_file', tools);
+    expect(msg).toContain("Tool 'read_file' not found");
+    expect(msg).toContain('Available tools:');
+    expect(msg).toContain('view_file');
+    expect(msg).toContain('find_files');
+  });
+
+  it('includes a suggestion when the drifted name is within threshold', () => {
+    const msg = formatUnknownToolError('view_files', tools);
+    expect(msg).toContain('Did you mean: view_file?');
+  });
+
+  it('omits the suggestion clause when no tool is close enough', () => {
+    const msg = formatUnknownToolError('zzzzzzz', tools);
+    expect(msg).toContain("Tool 'zzzzzzz' not found");
+    expect(msg).toContain('Available tools:');
+    expect(msg).not.toContain('Did you mean');
+  });
+
+  it('does not suggest view_file for the read_file drift case (distance is 4, over threshold)', () => {
+    const msg = formatUnknownToolError('read_file', tools);
+    expect(msg).not.toContain('Did you mean');
+  });
+});

apps/server/src/services/agents.ts

@@ -2,6 +2,7 @@ import { promises as fs } from 'node:fs';
 import { join } from 'node:path';
 import type { Agent, AgentsResponse, AgentParseError } from '../types/api.js';
 import { ALL_TOOLS, resolveToolTier } from './tools.js';
+import { validateExtraArgs } from './inference/llama-args-validator.js';
 
 // v1.8.1: global agents live at /data/AGENTS.md inside the container
 // (./data:/data:ro mount on the host). Per-project AGENTS.md at the project
@@ -97,6 +98,7 @@ interface ParsedFrontmatter {
   // (200) in the outer loop. Integer ≥ 0; steps: 0 means "no tool calls
   // allowed" — the model responds text-only.
   steps?: number;
+  llama_extra_args?: string[];
 }
 
 function stripQuotes(s: string): string {
@@ -227,6 +229,34 @@ function parseFrontmatter(yaml: string): { data: ParsedFrontmatter; errors: stri
       } else {
         errors.push(`steps must be a non-negative integer (got "${valueRaw}")`);
       }
+    } else if (key === 'llama_extra_args') {
+      if (valueRaw === '') {
+        data.llama_extra_args = [];
+        // No arrayKey support — llama_extra_args uses inline list only.
+      } else if (valueRaw.startsWith('[') && valueRaw.endsWith(']')) {
+        const inner = valueRaw.slice(1, -1);
+        const parsed = inner
+          .split(',')
+          .map((s) => stripQuotes(s.trim()))
+          .filter((s) => s.length > 0);
+        try {
+          validateExtraArgs(parsed);
+          data.llama_extra_args = parsed;
+        } catch (err) {
+          errors.push(err instanceof Error ? err.message : String(err));
+        }
+      } else {
+        const parsed = valueRaw
+          .split(',')
+          .map((s) => stripQuotes(s.trim()))
+          .filter((s) => s.length > 0);
+        try {
+          validateExtraArgs(parsed);
+          data.llama_extra_args = parsed;
+        } catch (err) {
+          errors.push(err instanceof Error ? err.message : String(err));
+        }
+      }
     }
     // Unknown keys silently ignored — forward-compat.
   }
@@ -328,6 +358,7 @@ function parseAgentSection(section: RawSection): Omit<Agent, 'source'> {
     model: typeof fm.model === 'string' && fm.model.length > 0 ? fm.model : null,
     max_tool_calls: typeof fm.max_tool_calls === 'number' ? fm.max_tool_calls : null,
     steps: typeof fm.steps === 'number' ? fm.steps : null,
+    llama_extra_args: Array.isArray(fm.llama_extra_args) ? fm.llama_extra_args : null,
   };
 }
 

apps/server/src/services/auto_name.ts

@@ -1,9 +1,10 @@
 import type { InferenceContext } from './inference/index.js';
+import { taskModelCompletion } from './task-model.js';
 
 const NAMING_SYSTEM_PROMPT =
-  'You name chat sessions based on what the assistant did. Summarize the topic or outcome — do NOT copy the first few words verbatim. Reply directly with no thinking, reasoning, or explanation. Output ONLY the title, 4 words max, no quotes, no punctuation, no prefix like "Title:".';
+  'You name chat sessions. Reply with ONLY the title. 4 to 6 words. No quotes, no punctuation, no prefix.';
 
-const MAX_TITLE_CHARS = 60;
+const MAX_TITLE_CHARS = 80;
 
 function cleanTitle(raw: string): string {
   let name = raw.trim();
@@ -18,27 +19,7 @@ function cleanTitle(raw: string): string {
   return name;
 }
 
-interface NamingResponse {
+// TODO: wire suggestTags after task model validation
-  choices?: Array<{
-    message?: {
-      content?: string;
-      reasoning_content?: string;
-    };
-  }>;
-}
-
-function pickTitleSource(data: NamingResponse): string {
-  const choice = data.choices?.[0]?.message;
-  if (!choice) return '';
-  if (choice.content && choice.content.trim().length > 0) return choice.content;
-  const reasoning = choice.reasoning_content ?? '';
-  if (reasoning.length === 0) return '';
-  const lines = reasoning
-    .split('\n')
-    .map((l) => l.trim())
-    .filter((l) => l.length > 0);
-  return lines[lines.length - 1] ?? '';
-}
 
 export async function maybeAutoNameChat(
   ctx: InferenceContext,
@@ -64,13 +45,6 @@ export async function maybeAutoNameChat(
   if (!chat) return;
   if (chat.name !== null && chat.name !== '') return;
 
-  const sessionRows = await ctx.sql<{ model: string }[]>`
-    SELECT model FROM sessions WHERE id = ${sessionId}
-  `;
-  // v2.0.5: prefer FAST_MODEL for cheap LLM calls (titles, summaries).
-  const model = ctx.config.FAST_MODEL ?? sessionRows[0]?.model;
-  if (!model) return;
-
   const assistantMsg = await ctx.sql<{ content: string }[]>`
     SELECT content FROM messages
     WHERE chat_id = ${chatId}
@@ -84,32 +58,12 @@ export async function maybeAutoNameChat(
 
   const assistantText = assistantMsg[0].content.slice(0, 2000);
 
-  const body = {
+  const raw = await taskModelCompletion({
-    model,
+    system: NAMING_SYSTEM_PROMPT,
-    messages: [
+    user: assistantText,
-      { role: 'system', content: NAMING_SYSTEM_PROMPT },
+    maxTokens: 30,
-      {
-        role: 'user',
-        content: assistantText,
-      },
-    ],
-    max_tokens: 30,
     temperature: 0.3,
-    stream: false,
-    chat_template_kwargs: { enable_thinking: false },
-  };
-
-  const res = await fetch(`${ctx.config.LLAMA_SWAP_URL}/v1/chat/completions`, {
-    method: 'POST',
-    headers: { 'Content-Type': 'application/json' },
-    body: JSON.stringify(body),
   });
-  if (!res.ok) {
-    const text = await res.text().catch(() => '');
-    throw new Error(`naming request failed: ${res.status} ${text.slice(0, 200)}`);
-  }
-  const data = (await res.json()) as NamingResponse;
-  const raw = pickTitleSource(data);
   const name = cleanTitle(raw);
   if (!name) {
     ctx.log.warn({ chatId, raw }, 'auto-name: empty title from model');

apps/server/src/services/inference/error-handler.ts

@@ -9,6 +9,7 @@ import * as modelContext from '../model-context.js';
 import { maybeFlagForCompaction } from './payload.js';
 import { insertParts, partsFromAssistantMessage } from './parts.js';
 import type { PartInsert } from './parts.js';
+import { stripToolMarkup } from './tool-call-parser.js';
 import type { InferenceContext, StreamResult, TurnArgs } from './turn.js';
 
 export async function handleAbortOrError(
@@ -21,6 +22,7 @@ export async function handleAbortOrError(
   const isAbort = err instanceof Error && err.name === 'AbortError';
   const finalStatus = isAbort ? 'cancelled' : 'failed';
   const errMsg = err instanceof Error ? err.message : String(err);
+  accumulated = stripToolMarkup(accumulated, { final: true });
   // v1.8.2: persist a structured error metadata blob on genuine failures so
   // the bubble can render the reason on reload without re-deriving from the
   // (one-shot) WS error frame. User-initiated abort skips this — there's no
@@ -101,7 +103,8 @@ export async function finalizeCompletion(
   session: Session
 ): Promise<void> {
   const { sessionId, chatId, assistantMessageId } = args;
-  const { content, finishReason, promptTokens, completionTokens } = result;
+  const content = stripToolMarkup(result.content, { final: true });
+  const { finishReason, promptTokens, completionTokens } = result;
 
   // v1.11.3: see executeToolPhase for the rationale.
   const mctx = await modelContext.getModelContext(session.model);

apps/server/src/services/inference/llama-args-validator.ts

@@ -0,0 +1,142 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+// Ported from studio/backend/core/inference/llama_server_args.py.
+// Original: https://github.com/unslothai/unsloth/blob/main/studio/backend/core/inference/llama_server_args.py
+
+// Each group is the full set of aliases (short + long) for one hard-denied
+// flag, taken from the llama-server README. Flags NOT in this list pass
+// through and override auto-set values via llama.cpp's last-wins CLI parsing.
+const DENYLIST_GROUPS: ReadonlyArray<ReadonlySet<string>> = [
+  // Model identity
+  new Set(['-m', '--model']),
+  new Set(['-mu', '--model-url']),
+  new Set(['-dr', '--docker-repo']),
+  new Set(['-hf', '-hfr', '--hf-repo']),
+  new Set(['-hff', '--hf-file']),
+  new Set(['-hfv', '-hfrv', '--hf-repo-v']),
+  new Set(['-hffv', '--hf-file-v']),
+  new Set(['-hft', '--hf-token']),
+  new Set(['-mm', '--mmproj']),
+  new Set(['-mmu', '--mmproj-url']),
+  // Networking
+  new Set(['--host']),
+  new Set(['--port']),
+  new Set(['--path']),
+  new Set(['--api-prefix']),
+  new Set(['--reuse-port']),
+  // Auth / TLS
+  new Set(['--api-key']),
+  new Set(['--api-key-file']),
+  new Set(['--ssl-key-file']),
+  new Set(['--ssl-cert-file']),
+  // Single-model server / UI
+  new Set(['--webui', '--no-webui']),
+  new Set(['--ui', '--no-ui']),
+  new Set(['--ui-config']),
+  new Set(['--ui-config-file']),
+  new Set(['--ui-mcp-proxy', '--no-ui-mcp-proxy']),
+  new Set(['--models-dir']),
+  new Set(['--models-preset']),
+  new Set(['--models-max']),
+  new Set(['--models-autoload', '--no-models-autoload']),
+];
+
+const DENYLIST: ReadonlySet<string> = new Set(
+  DENYLIST_GROUPS.flatMap((g) => [...g]),
+);
+
+function flagName(token: string): string | null {
+  if (!token.startsWith('-') || token === '-' || token === '--') return null;
+  if (token.length >= 2 && (token[1]!.match(/\d/) || token[1] === '.')) return null;
+  return token.split('=', 1)[0]!;
+}
+
+export function validateExtraArgs(args?: Iterable<string>): string[] {
+  if (!args) return [];
+  const out: string[] = [];
+  for (const raw of args) {
+    const token = String(raw);
+    const flag = flagName(token);
+    if (flag !== null && DENYLIST.has(flag)) {
+      throw new Error(
+        `llama-server flag '${flag}' is managed and cannot be passed as an extra arg`,
+      );
+    }
+    out.push(token);
+  }
+  return out;
+}
+
+export function isManagedFlag(flag: string): boolean {
+  return DENYLIST.has(flag);
+}
+
+// Shadowing flag groups: pass-through flags that shadow first-class settings.
+const CONTEXT_FLAGS = new Set(['-c', '--ctx-size']);
+const CACHE_FLAGS = new Set(['-ctk', '--cache-type-k', '-ctv', '--cache-type-v']);
+const SPEC_FLAGS = new Set([
+  '--spec-default',
+  '--spec-type',
+  '--spec-ngram-size-n',
+  '--spec-ngram-size',
+  '--draft-min',
+  '--draft-max',
+  '--spec-draft-n-max',
+  '--spec-draft-n-min',
+  '--spec-draft-p-min',
+  '--spec-draft-p-split',
+  '--spec-ngram-mod-n-match',
+  '--spec-ngram-mod-n-min',
+  '--spec-ngram-mod-n-max',
+]);
+const TEMPLATE_FLAGS = new Set([
+  '--chat-template',
+  '--chat-template-file',
+  '--chat-template-kwargs',
+  '--jinja',
+  '--no-jinja',
+]);
+
+const BOOLEAN_SHADOWING_FLAGS = new Set([
+  '--spec-default', '--jinja', '--no-jinja',
+]);
+
+export interface StripOptions {
+  stripContext?: boolean;
+  stripCache?: boolean;
+  stripSpec?: boolean;
+  stripTemplate?: boolean;
+}
+
+export function stripShadowingFlags(
+  args: Iterable<string>,
+  opts?: StripOptions,
+): string[] {
+  const shadowing = new Set<string>();
+  if (opts?.stripContext !== false) for (const f of CONTEXT_FLAGS) shadowing.add(f);
+  if (opts?.stripCache !== false) for (const f of CACHE_FLAGS) shadowing.add(f);
+  if (opts?.stripSpec !== false) for (const f of SPEC_FLAGS) shadowing.add(f);
+  if (opts?.stripTemplate !== false) for (const f of TEMPLATE_FLAGS) shadowing.add(f);
+
+  const tokens = [...args].map(String);
+  const out: string[] = [];
+  let i = 0;
+  const n = tokens.length;
+  while (i < n) {
+    const tok = tokens[i]!;
+    const flag = flagName(tok);
+    if (flag === null || !shadowing.has(flag)) {
+      out.push(tok);
+      i++;
+      continue;
+    }
+    if (BOOLEAN_SHADOWING_FLAGS.has(flag) || tok.includes('=')) {
+      i++;
+    } else if (i + 1 < n && flagName(tokens[i + 1]!) === null) {
+      i += 2;
+    } else {
+      i++;
+    }
+  }
+  return out;
+}

apps/server/src/services/inference/provider.ts

@@ -6,29 +6,79 @@ import type { LanguageModel } from 'ai';
 // upstream without touching env vars. No apiKey — llama-swap is unauth in our
 // Tailscale topology and exposing it over the public internet is gated by
 // Authelia at the Caddy layer, not by API keys.
+//
+// v2.4.1-sidecar: when the agent has llama_extra_args, route through
+// llama-sidecar instead. A fresh provider is created per call (not cached)
+// because the X-Agent-Flags header varies per agent. The llama-swap path
+// stays cached since it has no per-request headers.
 
-const cache = new Map<string, ReturnType<typeof createOpenAICompatible>>();
+const swapCache = new Map<string, ReturnType<typeof createOpenAICompatible>>();
 
-function getProvider(baseURL: string): ReturnType<typeof createOpenAICompatible> {
+function getSwapProvider(baseURL: string): ReturnType<typeof createOpenAICompatible> {
-  let provider = cache.get(baseURL);
+  let provider = swapCache.get(baseURL);
   if (!provider) {
     provider = createOpenAICompatible({
       name: 'llama-swap',
       baseURL: baseURL.endsWith('/v1') ? baseURL : `${baseURL}/v1`,
-      // v1.13.7: @ai-sdk/openai-compatible defaults includeUsage=false, which
-      // omits `stream_options.include_usage` from the request body. Without
-      // it, llama.cpp / llama-swap never emits the trailing usage block, so
-      // `result.usage` resolves with inputTokens=outputTokens=undefined and
-      // tokens_used / ctx_used land as NULL in every messages row. Setting
-      // true here re-enables the per-stream usage payload across all models
-      // served via the llama-swap provider.
       includeUsage: true,
     });
-    cache.set(baseURL, provider);
+    swapCache.set(baseURL, provider);
   }
   return provider;
 }
 
-export function upstreamModel(baseURL: string, modelId: string): LanguageModel {
+function sidecarProvider(
-  return getProvider(baseURL).chatModel(modelId);
+  baseURL: string,
+  flags: string[],
+): ReturnType<typeof createOpenAICompatible> {
+  return createOpenAICompatible({
+    name: 'llama-sidecar',
+    baseURL: baseURL.endsWith('/v1') ? baseURL : `${baseURL}/v1`,
+    includeUsage: true,
+    headers: {
+      'X-Agent-Flags': flags.join(' '),
+    },
+  });
+}
+
+export type InferenceRoute = 'swap' | 'sidecar';
+
+export interface RoutingInfo {
+  route: InferenceRoute;
+  flags: string[] | null;
+}
+
+interface AgentLike {
+  llama_extra_args: string[] | null;
+}
+
+interface ConfigLike {
+  LLAMA_SWAP_URL: string;
+  LLAMA_SIDECAR_URL?: string;
+}
+
+export function resolveRoute(agent: AgentLike | null): RoutingInfo {
+  const flags = agent?.llama_extra_args;
+  if (flags && flags.length > 0) {
+    return { route: 'sidecar', flags };
+  }
+  return { route: 'swap', flags: null };
+}
+
+export function upstreamModel(
+  config: ConfigLike,
+  modelId: string,
+  agent?: AgentLike | null,
+): LanguageModel {
+  const { route, flags } = resolveRoute(agent ?? null);
+  if (route === 'sidecar') {
+    const url = config.LLAMA_SIDECAR_URL;
+    if (!url) {
+      throw new Error(
+        `Agent has llama_extra_args but LLAMA_SIDECAR_URL is not set`,
+      );
+    }
+    return sidecarProvider(url, flags!).chatModel(modelId);
+  }
+  return getSwapProvider(config.LLAMA_SWAP_URL).chatModel(modelId);
 }

apps/server/src/services/inference/stream-phase.ts

@@ -7,9 +7,7 @@ import * as modelContext from '../model-context.js';
 import { toolJsonSchemas, type ToolJsonSchema } from '../tools.js';
 import { matchToolGlob } from '../agents.js';
 import type { OpenAiMessage } from './payload.js';
-// v1.13.16: extractToolCallBlocks replaces the inline opener-search loop and
+import { extractToolCallBlocks } from './tool-call-parser.js';
-// recognizes both Qwen <tool_call> and Anthropic <invoke> markup in one pass.
-import { extractToolCallBlocks } from './xml-parser.js';
 import { DB_FLUSH_INTERVAL_MS, type StreamPhaseState } from './types.js';
 import type {
   InferenceContext,
@@ -159,7 +157,8 @@ export async function streamCompletion(
   opts: StreamOptions,
   onDelta: (content: string) => void,
   onUsage: ((prompt: number | null, completion: number | null) => void) | undefined,
-  signal?: AbortSignal
+  signal?: AbortSignal,
+  agent?: Agent | null,
 ): Promise<StreamResult> {
   const aiMessages = toModelMessages(messages);
   const hasTools = opts.tools !== null && opts.tools.length > 0;
@@ -197,7 +196,7 @@ export async function streamCompletion(
   };
 
   const result = streamText({
-    model: upstreamModel(ctx.config.LLAMA_SWAP_URL, model),
+    model: upstreamModel(ctx.config, model, agent ?? null),
     messages: aiMessages,
     ...(aiTools
       ? { tools: aiTools, toolChoice: 'auto' as const, experimental_repairToolCall: repairToolCall }
@@ -460,7 +459,8 @@ export async function executeStreamPhase(
           }, USAGE_THROTTLE_MS - elapsed);
         }
       },
-      signal
+      signal,
+      agent,
     );
   } finally {
     if (pendingFlushTimer) {

apps/server/src/services/inference/tool-call-parser.ts

@@ -0,0 +1,426 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+// Ported from studio/backend/core/inference/tool_call_parser.py.
+// Original: https://github.com/unslothai/unsloth/blob/main/studio/backend/core/inference/tool_call_parser.py
+
+// ── Constants ────────────────────────────────────────────────────────────
+
+export const XML_TOOL_OPEN = '<tool_call>';
+export const XML_TOOL_CLOSE = '</tool_call>';
+export const INVOKE_TOOL_OPEN = '<invoke';
+export const INVOKE_TOOL_CLOSE = '</invoke>';
+
+export const TOOL_XML_SIGNALS = [XML_TOOL_OPEN, '<function=', INVOKE_TOOL_OPEN] as const;
+
+export const TOOL_ERROR_PREFIXES = [
+  'Error',
+  'Search failed',
+  'Execution error',
+  'Blocked:',
+  'Exit code',
+  'Failed to fetch',
+  'Failed to resolve',
+  'No query provided',
+] as const;
+
+export const DUPLICATE_CALL_NUDGE =
+  'You already made this exact call. Do not repeat the same tool ' +
+  'call. Try a different approach: fetch a URL from previous ' +
+  'results, use Python to process data you already have, or ' +
+  'provide your final answer now.';
+
+export const TOOL_ERROR_NUDGE =
+  '\n\nThe tool call encountered an issue. Please try a different ' +
+  'approach or rephrase your request.';
+
+export const BUDGET_EXHAUSTED_NUDGE =
+  'You have used all available tool calls. Based on everything you ' +
+  'have found so far, provide your final answer now. Do not call ' +
+  'any more tools.';
+
+// ── Strip patterns ───────────────────────────────────────────────────────
+
+const TOOL_CLOSED_PATS = [
+  /<tool_call>.*?<\/tool_call>/gs,
+  /<function=\w+>.*?<\/function>/gs,
+  /<invoke\s[^>]*>.*?<\/invoke>/gs,
+];
+
+const TOOL_ALL_PATS = [
+  ...TOOL_CLOSED_PATS,
+  /<tool_call>.*$/gs,
+  /<function=\w+>.*$/gs,
+  /<invoke\s[^>]*>.*$/gs,
+];
+
+// ── Strip / signal ───────────────────────────────────────────────────────
+
+export function stripToolMarkup(text: string, opts?: { final?: boolean }): string {
+  const pats = opts?.final ? TOOL_ALL_PATS : TOOL_CLOSED_PATS;
+  for (const pat of pats) {
+    text = text.replace(pat, '');
+  }
+  return opts?.final ? text.trim() : text;
+}
+
+export function hasToolSignal(text: string): boolean {
+  return TOOL_XML_SIGNALS.some((s) => text.includes(s));
+}
+
+// ── parseToolCallsFromText (Unsloth port + Anthropic extension) ──────────
+
+export interface OpenAiToolCall {
+  id: string;
+  type: 'function';
+  function: { name: string; arguments: string };
+}
+
+const TC_JSON_START_RE = /<tool_call>\s*\{/g;
+const TC_FUNC_START_RE = /<function=(\w+)>\s*/g;
+const TC_END_TAG_RE = /<\/tool_call>/;
+const TC_FUNC_CLOSE_RE = /\s*<\/function>\s*$/;
+const TC_PARAM_START_RE = /<parameter=(\w+)>\s*/g;
+const TC_PARAM_CLOSE_RE = /\s*<\/parameter>\s*$/;
+
+const TC_INVOKE_START_RE = /<invoke\s+name\s*=\s*(?:"([^"]*)"|'([^']*)')\s*>/g;
+const TC_INVOKE_CLOSE_RE = /\s*<\/invoke>\s*$/;
+const TC_INVOKE_PARAM_RE = /<parameter\s+name\s*=\s*(?:"([^"]*)"|'([^']*)')\s*>/g;
+const TC_INVOKE_PARAM_CLOSE_RE = /\s*<\/parameter>\s*$/;
+
+function scanBalancedBraces(content: string, start: number): number {
+  let depth = 0;
+  let i = start;
+  let inString = false;
+  while (i < content.length) {
+    const ch = content[i]!;
+    if (inString) {
+      if (ch === '\\' && i + 1 < content.length) {
+        i += 2;
+        continue;
+      }
+      if (ch === '"') inString = false;
+    } else if (ch === '"') {
+      inString = true;
+    } else if (ch === '{') {
+      depth++;
+    } else if (ch === '}') {
+      depth--;
+      if (depth === 0) return i;
+    }
+    i++;
+  }
+  return -1;
+}
+
+export function parseToolCallsFromText(
+  content: string,
+  opts?: { idOffset?: number },
+): OpenAiToolCall[] {
+  const toolCalls: OpenAiToolCall[] = [];
+  const idOffset = opts?.idOffset ?? 0;
+
+  // Pattern 1: <tool_call>{json}</tool_call> -- balanced-brace JSON scanner.
+  // Skips braces inside JSON strings so nested objects parse correctly.
+  TC_JSON_START_RE.lastIndex = 0;
+  let m: RegExpExecArray | null;
+  while ((m = TC_JSON_START_RE.exec(content)) !== null) {
+    const braceStart = m.index + m[0].length - 1;
+    const braceEnd = scanBalancedBraces(content, braceStart);
+    if (braceEnd === -1) continue;
+    const jsonStr = content.slice(braceStart, braceEnd + 1);
+    try {
+      const obj = JSON.parse(jsonStr) as Record<string, unknown>;
+      const name = typeof obj.name === 'string' ? obj.name : '';
+      let args: string;
+      const rawArgs = obj.arguments ?? {};
+      if (typeof rawArgs === 'string') {
+        args = rawArgs;
+      } else {
+        args = JSON.stringify(rawArgs);
+      }
+      toolCalls.push({
+        id: `call_${idOffset + toolCalls.length}`,
+        type: 'function',
+        function: { name, arguments: args },
+      });
+    } catch {
+      // malformed JSON -- skip
+    }
+  }
+
+  // Pattern 2: <function=name><parameter=key>value -- closing tags optional.
+  // Body boundary uses </tool_call> or next <function= (not </function>,
+  // because code parameter values can contain that literal).
+  if (toolCalls.length === 0) {
+    TC_FUNC_START_RE.lastIndex = 0;
+    const funcStarts: Array<{ match: RegExpExecArray; name: string }> = [];
+    while ((m = TC_FUNC_START_RE.exec(content)) !== null) {
+      funcStarts.push({ match: m, name: m[1]! });
+    }
+    for (let idx = 0; idx < funcStarts.length; idx++) {
+      const { match: fm, name: funcName } = funcStarts[idx]!;
+      const bodyStart = fm.index + fm[0].length;
+      const nextFunc = idx + 1 < funcStarts.length
+        ? funcStarts[idx + 1]!.match.index
+        : content.length;
+      const endTag = TC_END_TAG_RE.exec(content.slice(bodyStart));
+      let bodyEnd = endTag ? bodyStart + endTag.index : content.length;
+      bodyEnd = Math.min(bodyEnd, nextFunc);
+      let body = content.slice(bodyStart, bodyEnd);
+      body = body.replace(TC_FUNC_CLOSE_RE, '');
+
+      const args: Record<string, string> = {};
+      TC_PARAM_START_RE.lastIndex = 0;
+      const paramStarts: Array<{ match: RegExpExecArray; name: string }> = [];
+      let pm: RegExpExecArray | null;
+      while ((pm = TC_PARAM_START_RE.exec(body)) !== null) {
+        paramStarts.push({ match: pm, name: pm[1]! });
+      }
+      if (paramStarts.length === 1) {
+        // Single param: take everything to body end so embedded
+        // </parameter> in code strings is preserved.
+        const p = paramStarts[0]!;
+        let val = body.slice(p.match.index + p.match[0].length);
+        val = val.replace(TC_PARAM_CLOSE_RE, '');
+        args[p.name] = val.trim();
+      } else {
+        for (let pidx = 0; pidx < paramStarts.length; pidx++) {
+          const p = paramStarts[pidx]!;
+          const valStart = p.match.index + p.match[0].length;
+          const nextParam = pidx + 1 < paramStarts.length
+            ? paramStarts[pidx + 1]!.match.index
+            : body.length;
+          let val = body.slice(valStart, nextParam);
+          val = val.replace(TC_PARAM_CLOSE_RE, '');
+          args[p.name] = val.trim();
+        }
+      }
+
+      toolCalls.push({
+        id: `call_${idOffset + toolCalls.length}`,
+        type: 'function',
+        function: { name: funcName, arguments: JSON.stringify(args) },
+      });
+    }
+  }
+
+  // Pattern 3: <invoke name="..."><parameter name="...">value -- Anthropic
+  // shape that qwen3.6 drifts to from Claude Code documentation residue.
+  // Closing tags optional; same single-param fast path as pattern 2.
+  if (toolCalls.length === 0) {
+    TC_INVOKE_START_RE.lastIndex = 0;
+    const invokeStarts: Array<{ match: RegExpExecArray; name: string }> = [];
+    while ((m = TC_INVOKE_START_RE.exec(content)) !== null) {
+      const name = (m[1] ?? m[2] ?? '').trim();
+      if (name) invokeStarts.push({ match: m, name });
+    }
+    for (let idx = 0; idx < invokeStarts.length; idx++) {
+      const { match: im, name: invokeName } = invokeStarts[idx]!;
+      const bodyStart = im.index + im[0].length;
+      const nextInvoke = idx + 1 < invokeStarts.length
+        ? invokeStarts[idx + 1]!.match.index
+        : content.length;
+      const closeTag = content.slice(bodyStart).match(/<\/invoke>/);
+      let bodyEnd = closeTag ? bodyStart + (closeTag.index ?? 0) : content.length;
+      bodyEnd = Math.min(bodyEnd, nextInvoke);
+      let body = content.slice(bodyStart, bodyEnd);
+      body = body.replace(TC_INVOKE_CLOSE_RE, '');
+
+      const args: Record<string, string> = {};
+      TC_INVOKE_PARAM_RE.lastIndex = 0;
+      const paramStarts: Array<{ match: RegExpExecArray; name: string }> = [];
+      let pm: RegExpExecArray | null;
+      while ((pm = TC_INVOKE_PARAM_RE.exec(body)) !== null) {
+        const pname = (pm[1] ?? pm[2] ?? '').trim();
+        if (pname) paramStarts.push({ match: pm, name: pname });
+      }
+      if (paramStarts.length === 1) {
+        const p = paramStarts[0]!;
+        let val = body.slice(p.match.index + p.match[0].length);
+        val = val.replace(TC_INVOKE_PARAM_CLOSE_RE, '');
+        args[p.name] = val.trim();
+      } else {
+        for (let pidx = 0; pidx < paramStarts.length; pidx++) {
+          const p = paramStarts[pidx]!;
+          const valStart = p.match.index + p.match[0].length;
+          const nextParam = pidx + 1 < paramStarts.length
+            ? paramStarts[pidx + 1]!.match.index
+            : body.length;
+          let val = body.slice(valStart, nextParam);
+          val = val.replace(TC_INVOKE_PARAM_CLOSE_RE, '');
+          args[p.name] = val.trim();
+        }
+      }
+
+      toolCalls.push({
+        id: `call_${idOffset + toolCalls.length}`,
+        type: 'function',
+        function: { name: invokeName, arguments: JSON.stringify(args) },
+      });
+    }
+  }
+
+  return toolCalls;
+}
+
+// ── BooCode streaming helpers ────────────────────────────────────────────
+
+export interface ParsedCall {
+  name: string;
+  args: Record<string, unknown>;
+}
+
+const PLACEHOLDER_LITERALS = new Set(['...', 'placeholder', '<path>', '<file>']);
+const ANGLE_BRACKET_SENTINEL_RE = /^<[^>]+>$/;
+
+export function isPlaceholderArgValue(value: unknown): boolean {
+  if (typeof value !== 'string') return false;
+  const trimmed = value.trim();
+  if (trimmed === '') return true;
+  if (PLACEHOLDER_LITERALS.has(trimmed)) return true;
+  if (ANGLE_BRACKET_SENTINEL_RE.test(trimmed)) return true;
+  return false;
+}
+
+function hasPlaceholderArgs(args: Record<string, unknown>): boolean {
+  for (const value of Object.values(args)) {
+    if (isPlaceholderArgValue(value)) return true;
+  }
+  return false;
+}
+
+function logRejectedPlaceholder(parsed: ParsedCall): void {
+  console.debug(
+    { toolName: parsed.name, args: parsed.args },
+    'rejected placeholder tool call at parse time',
+  );
+}
+
+const QWEN_FUNCTION_RE = /<function\s*=\s*([^>\s]+)\s*>/;
+const QWEN_PARAM_RE = /<parameter\s*=\s*([^>\s]+)\s*>([\s\S]*?)<\/parameter>/g;
+
+export function parseXmlToolCall(block: string): ParsedCall | null {
+  const nameMatch = block.match(QWEN_FUNCTION_RE);
+  if (!nameMatch || !nameMatch[1]) return null;
+  const name = nameMatch[1].trim();
+  if (!name) return null;
+  const args: Record<string, unknown> = {};
+  for (const m of block.matchAll(QWEN_PARAM_RE)) {
+    const key = (m[1] ?? '').trim();
+    if (!key) continue;
+    const raw = (m[2] ?? '').trim();
+    try {
+      args[key] = JSON.parse(raw);
+    } catch {
+      args[key] = raw;
+    }
+  }
+  return { name, args };
+}
+
+const INVOKE_NAME_RE =
+  /<invoke\s+name\s*=\s*("([^"]*)"|'([^']*)')\s*>/;
+const INVOKE_PARAM_RE =
+  /<parameter\s+name\s*=\s*("([^"]*)"|'([^']*)')\s*>([\s\S]*?)<\/parameter>/g;
+
+export function parseInvokeToolCall(block: string): ParsedCall | null {
+  const nameMatch = block.match(INVOKE_NAME_RE);
+  if (!nameMatch) return null;
+  const name = (nameMatch[2] ?? nameMatch[3] ?? '').trim();
+  if (!name) return null;
+  const args: Record<string, unknown> = {};
+  for (const m of block.matchAll(INVOKE_PARAM_RE)) {
+    const key = ((m[2] ?? m[3] ?? '') as string).trim();
+    if (!key) continue;
+    const raw = (m[4] ?? '').trim();
+    try {
+      args[key] = JSON.parse(raw);
+    } catch {
+      args[key] = raw;
+    }
+  }
+  return { name, args };
+}
+
+const ALL_OPENERS = [XML_TOOL_OPEN, INVOKE_TOOL_OPEN] as const;
+
+export function partialXmlOpenerStart(s: string): number {
+  let earliest = -1;
+  for (const op of ALL_OPENERS) {
+    const idx = s.indexOf(op);
+    if (idx === -1) continue;
+    if (earliest === -1 || idx < earliest) earliest = idx;
+  }
+  if (earliest !== -1) return earliest;
+  const lastLt = s.lastIndexOf('<');
+  if (lastLt === -1) return -1;
+  const suffix = s.slice(lastLt);
+  for (const op of ALL_OPENERS) {
+    if (op.startsWith(suffix) && suffix.length < op.length) return lastLt;
+  }
+  return -1;
+}
+
+export interface ToolCallExtraction {
+  flushed: string;
+  calls: ParsedCall[];
+  remaining: string;
+}
+
+interface OpenerSpec {
+  open: string;
+  close: string;
+  parse: (block: string) => ParsedCall | null;
+}
+
+const OPENER_SPECS: ReadonlyArray<OpenerSpec> = [
+  { open: XML_TOOL_OPEN, close: XML_TOOL_CLOSE, parse: parseXmlToolCall },
+  { open: INVOKE_TOOL_OPEN, close: INVOKE_TOOL_CLOSE, parse: parseInvokeToolCall },
+];
+
+export function extractToolCallBlocks(buffer: string): ToolCallExtraction {
+  let flushed = '';
+  const calls: ParsedCall[] = [];
+  let pos = 0;
+
+  while (pos < buffer.length) {
+    let next: { spec: OpenerSpec; openIdx: number; closeIdx: number } | null = null;
+    for (const spec of OPENER_SPECS) {
+      const openIdx = buffer.indexOf(spec.open, pos);
+      if (openIdx === -1) continue;
+      const closeIdx = buffer.indexOf(spec.close, openIdx);
+      if (closeIdx === -1) continue;
+      if (next === null || openIdx < next.openIdx) {
+        next = { spec, openIdx, closeIdx };
+      }
+    }
+    if (next === null) break;
+
+    if (next.openIdx > pos) {
+      flushed += buffer.slice(pos, next.openIdx);
+    }
+    const blockEnd = next.closeIdx + next.spec.close.length;
+    const block = buffer.slice(next.openIdx, blockEnd);
+    const parsed = next.spec.parse(block);
+    if (parsed) {
+      if (hasPlaceholderArgs(parsed.args)) {
+        logRejectedPlaceholder(parsed);
+        flushed += block;
+      } else {
+        calls.push(parsed);
+      }
+    }
+    pos = blockEnd;
+  }
+
+  const tail = buffer.slice(pos);
+  const partialIdx = partialXmlOpenerStart(tail);
+  if (partialIdx === -1) {
+    flushed += tail;
+    return { flushed, calls, remaining: '' };
+  }
+  if (partialIdx > 0) {
+    flushed += tail.slice(0, partialIdx);
+  }
+  return { flushed, calls, remaining: tail.slice(partialIdx) };
+}

apps/server/src/services/inference/tool-phase.ts

@@ -14,6 +14,7 @@ import { formatUnknownToolError } from './tool-suggestions.js';
 // Resolves the grant root before pausing the loop so the user is never
 // prompted about paths we couldn't grant anyway (e.g. /etc/passwd).
 import { resolveGrantRoot } from '../grant_resolver.js';
+import { stripToolMarkup } from './tool-call-parser.js';
 import type {
   InferenceContext,
   StreamResult,
@@ -100,7 +101,8 @@ export async function executeToolPhase(
   projectRoot: string
 ): Promise<ToolPhaseResult> {
   const { sessionId, chatId, assistantMessageId } = args;
-  const { content, toolCalls, promptTokens, completionTokens } = result;
+  const content = stripToolMarkup(result.content, { final: true });
+  const { toolCalls, promptTokens, completionTokens } = result;
 
   // v1.11.3: ctx_max comes from llama-swap /upstream/<model>/props, not the
   // streaming completion (which doesn't emit n_ctx). getModelContext caches

apps/server/src/services/inference/turn.ts

@@ -14,6 +14,7 @@ import type {
 import { ALL_TOOLS } from '../tools.js';
 import { resolveProjectRoot } from '../path_guard.js';
 import { maybeAutoNameChat } from '../auto_name.js';
+import { rewriteSearchQuery } from '../task-search-rewrite.js';
 import { getAgentById } from '../agents.js';
 import * as compaction from '../compaction.js';
 import type { Broker } from '../broker.js';
@@ -254,6 +255,16 @@ export async function runAssistantTurn(
     const webToolsEnabled =
       iterSession.web_search_enabled ?? iterProject.default_web_search_enabled ?? false;
 
+    if (stepNumber === 0 && webToolsEnabled && messages.length >= 2) {
+      const lastUserMsg = [...messages].reverse().find((m) => m.role === 'user');
+      if (lastUserMsg?.content) {
+        const hint = await rewriteSearchQuery(lastUserMsg.content);
+        if (hint && messages[0]?.role === 'system' && messages[0].content) {
+          messages[0].content += `\n\nThe user's search intent can be summarized as: "${hint}"`;
+        }
+      }
+    }
+
     const iterArgs: TurnArgs = { sessionId, chatId, assistantMessageId, toolsUsed, recentToolCalls, signal };
     const state: StreamPhaseState = { accumulated: '', startedAt: null };
     let result: StreamResult;

apps/server/src/services/inference/xml-parser.ts

@@ -1,204 +0,0 @@
-// v1.10.5: XML-tag tool-call fallback. Some models emit
-// <tool_call><function=foo><parameter=key>value</parameter></function></tool_call>
-// in plain content instead of using the OpenAI tool_calls JSON channel.
-// The streaming loop in stream-phase.ts extracts these blocks via these helpers.
-//
-// v1.13.16: also recognize Anthropic <invoke name="..."><parameter name="...">
-// markup. qwen3.6-35b-a3b-mxfp4 drifts to this format when prompted as an
-// "Architect"-style agent because Claude Code documentation in its
-// pre-training data uses this shape. Both formats route through the same
-// synthetic ToolCall path with shared xml_call_${idx} IDs; downstream
-// dispatch handles unknown tool names with a richer error (see
-// tool-suggestions.ts + tool-phase.ts).
-
-export const XML_TOOL_OPEN = '<tool_call>';
-export const XML_TOOL_CLOSE = '</tool_call>';
-
-// v1.13.16: Anthropic <invoke> opener is matched by prefix (not the full
-// `<invoke ...>` tag) because attributes follow. Closer is the literal tag.
-export const INVOKE_TOOL_OPEN = '<invoke';
-export const INVOKE_TOOL_CLOSE = '</invoke>';
-
-export interface ParsedCall {
-  name: string;
-  args: Record<string, unknown>;
-}
-
-const PLACEHOLDER_LITERALS = new Set(['...', 'placeholder', '<path>', '<file>']);
-const ANGLE_BRACKET_SENTINEL_RE = /^<[^>]+>$/;
-
-/** True when a string arg looks like a model placeholder, not a real path/value. */
-export function isPlaceholderArgValue(value: unknown): boolean {
-  if (typeof value !== 'string') return false;
-  const trimmed = value.trim();
-  if (trimmed === '') return true;
-  if (PLACEHOLDER_LITERALS.has(trimmed)) return true;
-  if (ANGLE_BRACKET_SENTINEL_RE.test(trimmed)) return true;
-  return false;
-}
-
-function hasPlaceholderArgs(args: Record<string, unknown>): boolean {
-  for (const value of Object.values(args)) {
-    if (isPlaceholderArgValue(value)) return true;
-  }
-  return false;
-}
-
-function logRejectedPlaceholder(parsed: ParsedCall): void {
-  // Pure helper — no Fastify logger here (stream-phase.ts stays unchanged).
-  console.debug(
-    { toolName: parsed.name, args: parsed.args },
-    'rejected placeholder tool call at parse time',
-  );
-}
-
-// v1.10.5: Qwen-flavor parser. Tightened in v1.13.16 to tolerate whitespace
-// around `=` (e.g. `<function = view_file>`). Name capture is non-whitespace,
-// non-`>` so a stray space doesn't get absorbed into the function name.
-const QWEN_FUNCTION_RE = /<function\s*=\s*([^>\s]+)\s*>/;
-const QWEN_PARAM_RE = /<parameter\s*=\s*([^>\s]+)\s*>([\s\S]*?)<\/parameter>/g;
-
-export function parseXmlToolCall(block: string): ParsedCall | null {
-  const nameMatch = block.match(QWEN_FUNCTION_RE);
-  if (!nameMatch || !nameMatch[1]) return null;
-  const name = nameMatch[1].trim();
-  if (!name) return null;
-  const args: Record<string, unknown> = {};
-  for (const m of block.matchAll(QWEN_PARAM_RE)) {
-    const key = (m[1] ?? '').trim();
-    if (!key) continue;
-    const raw = (m[2] ?? '').trim();
-    try {
-      args[key] = JSON.parse(raw);
-    } catch {
-      args[key] = raw;
-    }
-  }
-  return { name, args };
-}
-
-// v1.13.16: Anthropic-flavor parser. Same JSON-parse-with-string-fallback
-// shape as parseXmlToolCall so the dispatch layer doesn't need to care which
-// flavor produced the call.
-const INVOKE_NAME_RE =
-  /<invoke\s+name\s*=\s*("([^"]*)"|'([^']*)')\s*>/;
-const INVOKE_PARAM_RE =
-  /<parameter\s+name\s*=\s*("([^"]*)"|'([^']*)')\s*>([\s\S]*?)<\/parameter>/g;
-
-export function parseInvokeToolCall(block: string): ParsedCall | null {
-  const nameMatch = block.match(INVOKE_NAME_RE);
-  if (!nameMatch) return null;
-  const name = (nameMatch[2] ?? nameMatch[3] ?? '').trim();
-  if (!name) return null;
-  const args: Record<string, unknown> = {};
-  for (const m of block.matchAll(INVOKE_PARAM_RE)) {
-    const key = ((m[2] ?? m[3] ?? '') as string).trim();
-    if (!key) continue;
-    const raw = (m[4] ?? '').trim();
-    try {
-      args[key] = JSON.parse(raw);
-    } catch {
-      args[key] = raw;
-    }
-  }
-  return { name, args };
-}
-
-// Locate the first character that begins (or completely contains) an
-// unfinished opener (either flavor) in `s`. Returns -1 when `s` can be
-// flushed to the client in full without risking a partial tag leak.
-//   Case 1: a full opener (`<tool_call>` or `<invoke`) with no matching
-//           closer — caller must keep everything from that index forward
-//           until the next chunk arrives with the closer.
-//   Case 2: `s` ends with a strict prefix of either opener (e.g. `<tool_c`
-//           or `<invo`). Caller must keep just that suffix in the buffer.
-// Note: case 1 assumes the calling loop already extracted every complete
-// block before reaching this check.
-const ALL_OPENERS = [XML_TOOL_OPEN, INVOKE_TOOL_OPEN] as const;
-
-export function partialXmlOpenerStart(s: string): number {
-  let earliest = -1;
-  for (const op of ALL_OPENERS) {
-    const idx = s.indexOf(op);
-    if (idx === -1) continue;
-    if (earliest === -1 || idx < earliest) earliest = idx;
-  }
-  if (earliest !== -1) return earliest;
-  const lastLt = s.lastIndexOf('<');
-  if (lastLt === -1) return -1;
-  const suffix = s.slice(lastLt);
-  for (const op of ALL_OPENERS) {
-    if (op.startsWith(suffix) && suffix.length < op.length) return lastLt;
-  }
-  return -1;
-}
-
-// v1.13.16: unified extraction. Replaces the inline loop that used to live
-// in stream-phase.ts. Pure function — returns the visible text to flush,
-// the parsed tool-call payloads in source order, and the buffer remainder
-// to retain for the next streaming chunk. Parse failures are silently
-// dropped (matches the pre-v1.13.16 behavior — leaking partial XML to the
-// chat looks worse than swallowing a bad block).
-export interface ToolCallExtraction {
-  flushed: string;
-  calls: ParsedCall[];
-  remaining: string;
-}
-
-interface OpenerSpec {
-  open: string;
-  close: string;
-  parse: (block: string) => ParsedCall | null;
-}
-
-const OPENER_SPECS: ReadonlyArray<OpenerSpec> = [
-  { open: XML_TOOL_OPEN, close: XML_TOOL_CLOSE, parse: parseXmlToolCall },
-  { open: INVOKE_TOOL_OPEN, close: INVOKE_TOOL_CLOSE, parse: parseInvokeToolCall },
-];
-
-export function extractToolCallBlocks(buffer: string): ToolCallExtraction {
-  let flushed = '';
-  const calls: ParsedCall[] = [];
-  let pos = 0;
-
-  while (pos < buffer.length) {
-    let next: { spec: OpenerSpec; openIdx: number; closeIdx: number } | null = null;
-    for (const spec of OPENER_SPECS) {
-      const openIdx = buffer.indexOf(spec.open, pos);
-      if (openIdx === -1) continue;
-      const closeIdx = buffer.indexOf(spec.close, openIdx);
-      if (closeIdx === -1) continue;
-      if (next === null || openIdx < next.openIdx) {
-        next = { spec, openIdx, closeIdx };
-      }
-    }
-    if (next === null) break;
-
-    if (next.openIdx > pos) {
-      flushed += buffer.slice(pos, next.openIdx);
-    }
-    const blockEnd = next.closeIdx + next.spec.close.length;
-    const block = buffer.slice(next.openIdx, blockEnd);
-    const parsed = next.spec.parse(block);
-    if (parsed) {
-      if (hasPlaceholderArgs(parsed.args)) {
-        logRejectedPlaceholder(parsed);
-        flushed += block;
-      } else {
-        calls.push(parsed);
-      }
-    }
-    pos = blockEnd;
-  }
-
-  const tail = buffer.slice(pos);
-  const partialIdx = partialXmlOpenerStart(tail);
-  if (partialIdx === -1) {
-    flushed += tail;
-    return { flushed, calls, remaining: '' };
-  }
-  if (partialIdx > 0) {
-    flushed += tail.slice(0, partialIdx);
-  }
-  return { flushed, calls, remaining: tail.slice(partialIdx) };
-}

apps/server/src/services/system-prompt.ts

@@ -21,6 +21,7 @@ import { createHash } from 'node:crypto';
 import { readFile, stat } from 'node:fs/promises';
 import type { Agent, Project, Session } from '../types/api.js';
 import { getAgentsMtimes } from './agents.js';
+import { resolveRoute } from './inference/provider.js';
 
 const BASE_SYSTEM_PROMPT = (projectPath: string) =>
   `You are BooCode Chat, a code investigation assistant. The user is working on a project located at ${projectPath}. Use the file-read tools (view_file, list_dir, grep, find_files) to investigate code when needed. Be concise. Cite file paths and line numbers when discussing code. Do not hallucinate file contents — read the file first. Tool results may be truncated; if so, narrow your query rather than guessing.`;
@@ -98,6 +99,7 @@ export interface PrefixFingerprint {
   has_agent_system_prompt: boolean;
   has_session_override: boolean;
   has_project_override: boolean;
+  route: 'swap' | 'sidecar';
 }
 
 export interface PrefixDrift {
@@ -125,6 +127,7 @@ interface ObservedInputs {
   has_agent_system_prompt: boolean;
   has_session_override: boolean;
   has_project_override: boolean;
+  route: 'swap' | 'sidecar';
 }
 
 interface ObserverEntry {
@@ -183,6 +186,7 @@ export async function buildSystemPromptWithFingerprint(
     has_agent_system_prompt: !!(agent && agent.system_prompt.trim().length > 0),
     has_session_override: sessionPrompt.length > 0,
     has_project_override: projectPrompt.length > 0,
+    route: resolveRoute(agent).route,
   };
 
   const fingerprint: PrefixFingerprint = {
@@ -199,6 +203,7 @@ export async function buildSystemPromptWithFingerprint(
     has_agent_system_prompt: inputs.has_agent_system_prompt,
     has_session_override: inputs.has_session_override,
     has_project_override: inputs.has_project_override,
+    route: inputs.route,
   };
 
   let drift: PrefixDrift | null = null;

apps/server/src/services/task-model.ts

@@ -0,0 +1,68 @@
+import { loadConfig, type Config } from '../config.js';
+
+const TIMEOUT_MS = 10_000;
+
+export async function taskModelCompletion(opts: {
+  system: string;
+  user: string;
+  maxTokens?: number;
+  temperature?: number;
+  fallbackModel?: string;
+}): Promise<string> {
+  const config = loadConfig();
+  const maxTokens = opts.maxTokens ?? 30;
+  const temperature = opts.temperature ?? 0.3;
+
+  const { url, model } = resolveEndpoint(config, opts.fallbackModel);
+
+  try {
+    const res = await fetch(`${url}/v1/chat/completions`, {
+      method: 'POST',
+      headers: { 'Content-Type': 'application/json' },
+      body: JSON.stringify({
+        model,
+        messages: [
+          { role: 'system', content: opts.system },
+          { role: 'user', content: opts.user },
+        ],
+        max_tokens: maxTokens,
+        temperature,
+        stream: false,
+        chat_template_kwargs: { enable_thinking: false },
+      }),
+      signal: AbortSignal.timeout(TIMEOUT_MS),
+    });
+    if (!res.ok) {
+      const text = await res.text().catch(() => '');
+      console.warn(`task-model: ${res.status} ${text.slice(0, 200)}`);
+      return '';
+    }
+    const data = (await res.json()) as {
+      choices?: Array<{
+        message?: { content?: string; reasoning_content?: string };
+      }>;
+    };
+    const choice = data.choices?.[0]?.message;
+    if (!choice) return '';
+    const content = (choice.content ?? '').trim();
+    if (content.length > 0) return content;
+    const reasoning = choice.reasoning_content ?? '';
+    if (reasoning.length === 0) return '';
+    const lines = reasoning.split('\n').map((l) => l.trim()).filter((l) => l.length > 0);
+    return lines[lines.length - 1] ?? '';
+  } catch (err) {
+    console.warn('task-model: request failed', err);
+    return '';
+  }
+}
+
+function resolveEndpoint(
+  config: Config,
+  fallbackModel?: string,
+): { url: string; model: string } {
+  if (config.TASK_MODEL_URL) {
+    return { url: config.TASK_MODEL_URL, model: 'gemma-3-270m-it' };
+  }
+  const model = config.FAST_MODEL ?? fallbackModel ?? config.DEFAULT_MODEL;
+  return { url: config.LLAMA_SWAP_URL, model };
+}

apps/server/src/services/task-search-rewrite.ts

@@ -0,0 +1,19 @@
+import { taskModelCompletion } from './task-model.js';
+
+const SYSTEM_PROMPT =
+  'You rewrite user messages into concise web search queries. Reply with ONLY the search query. 3 to 6 words. No quotes, no explanation.';
+
+const MAX_INPUT_CHARS = 500;
+const FALLBACK_CHARS = 60;
+
+export async function rewriteSearchQuery(userMessage: string): Promise<string> {
+  const input = userMessage.slice(0, MAX_INPUT_CHARS);
+  const result = await taskModelCompletion({
+    system: SYSTEM_PROMPT,
+    user: input,
+    maxTokens: 20,
+    temperature: 0.2,
+  });
+  if (result.length > 0) return result;
+  return userMessage.slice(0, FALLBACK_CHARS).trim();
+}

apps/server/src/services/task-summary.ts

@@ -0,0 +1,24 @@
+import { taskModelCompletion } from './task-model.js';
+
+const SYSTEM_PROMPT =
+  'Summarize this conversation in one sentence, 15 words max. No quotes, no prefix.';
+
+const MAX_INPUT_CHARS = 1000;
+
+export async function oneLineSummary(
+  messages: Array<{ role: string; content: string }>,
+): Promise<string> {
+  const lastPairs = messages.slice(-6);
+  let input = lastPairs
+    .map((m) => `${m.role}: ${m.content}`)
+    .join('\n');
+  if (input.length > MAX_INPUT_CHARS) {
+    input = input.slice(0, MAX_INPUT_CHARS);
+  }
+  return taskModelCompletion({
+    system: SYSTEM_PROMPT,
+    user: input,
+    maxTokens: 30,
+    temperature: 0.3,
+  });
+}

apps/server/src/services/task-tags.ts

@@ -0,0 +1,22 @@
+import { taskModelCompletion } from './task-model.js';
+
+const SYSTEM_PROMPT =
+  'You tag chat sessions. Reply with 1 to 3 lowercase tags separated by commas. Tags should describe the topic. No explanation. Examples: "docker, deployment", "python, debugging", "react, styling".';
+
+export async function suggestTags(
+  userMessage: string,
+  assistantReply: string,
+): Promise<string[]> {
+  const input = `User: ${userMessage.slice(0, 300)}\nAssistant: ${assistantReply.slice(0, 300)}`;
+  const result = await taskModelCompletion({
+    system: SYSTEM_PROMPT,
+    user: input,
+    maxTokens: 30,
+    temperature: 0.3,
+  });
+  if (result.length === 0) return [];
+  return result
+    .split(',')
+    .map((t) => t.trim().toLowerCase())
+    .filter((t) => t.length > 0 && t.length <= 30);
+}

apps/server/src/services/web/html-to-md.ts

@@ -0,0 +1,347 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+// Ported from studio/backend/core/inference/_html_to_md.py.
+// Original: https://github.com/unslothai/unsloth/blob/main/studio/backend/core/inference/_html_to_md.py
+
+import { parse, type DefaultTreeAdapterTypes } from 'parse5';
+
+type Document = DefaultTreeAdapterTypes.Document;
+type ChildNode = DefaultTreeAdapterTypes.ChildNode;
+type Element = DefaultTreeAdapterTypes.Element;
+type TextNode = DefaultTreeAdapterTypes.TextNode;
+
+const SKIP_TAGS = new Set([
+  'script', 'style', 'head', 'noscript', 'svg', 'math', 'nav', 'footer',
+]);
+
+const BLOCK_TAGS = new Set([
+  'p', 'div', 'section', 'article', 'main', 'aside', 'figure',
+  'figcaption', 'details', 'summary', 'dl', 'dt', 'dd',
+]);
+
+const HEADING_TAGS = new Set(['h1', 'h2', 'h3', 'h4', 'h5', 'h6']);
+
+const INLINE_EMPHASIS: Record<string, string> = {
+  strong: '**', b: '**', em: '*', i: '*',
+};
+
+function isElement(node: ChildNode): node is Element {
+  return 'tagName' in node;
+}
+
+function isText(node: ChildNode): node is TextNode {
+  return node.nodeName === '#text';
+}
+
+class MarkdownRenderer {
+  private out: string[] = [];
+
+  private inLink = false;
+  private linkHref: string | null = null;
+  private linkTextParts: string[] = [];
+
+  private listStack: string[] = [];
+  private olCounter: number[] = [];
+
+  private inTable = false;
+  private currentRow: string[] = [];
+  private cellParts: string[] = [];
+  private inCell = false;
+  private headerRowDone = false;
+  private rowHasTh = false;
+  private isFirstRow = false;
+
+  private inPre = false;
+  private preParts: string[] = [];
+  private preLanguage: string | null = null;
+  private inInlineCode = false;
+
+  private bqStack: string[][] = [];
+
+  private emit(text: string): void {
+    if (this.inLink) {
+      this.linkTextParts.push(text);
+    } else if (this.inCell) {
+      this.cellParts.push(text);
+    } else if (this.inPre) {
+      this.preParts.push(text);
+    } else if (this.bqStack.length > 0) {
+      this.bqStack[this.bqStack.length - 1]!.push(text);
+    } else {
+      this.out.push(text);
+    }
+  }
+
+  private prefixBlockquote(content: string): string {
+    content = content.replace(/[ \t]+$/gm, '');
+    content = content.replace(/\n{3,}/g, '\n\n').trim();
+    if (!content) return '';
+    return content.split('\n').map(line =>
+      line.trim() ? '> ' + line : '>'
+    ).join('\n');
+  }
+
+  private finishCell(): void {
+    if (!this.inCell) return;
+    this.inCell = false;
+    let cellText = this.cellParts.join('').trim().replace(/\n/g, ' ');
+    cellText = cellText.replace(/\|/g, '\\|');
+    this.currentRow.push(cellText);
+    this.cellParts = [];
+  }
+
+  private finishRow(): void {
+    if (this.currentRow.length === 0) return;
+    const line = '| ' + this.currentRow.join(' | ') + ' |';
+    this.emit(line + '\n');
+    if (!this.headerRowDone && (this.rowHasTh || this.isFirstRow)) {
+      const sep = '| ' + this.currentRow.map(() => '---').join(' | ') + ' |';
+      this.emit(sep + '\n');
+      this.headerRowDone = true;
+    }
+    this.isFirstRow = false;
+    this.currentRow = [];
+    this.rowHasTh = false;
+  }
+
+  private finishLink(): void {
+    const text = this.linkTextParts.join('').replace(/\s+/g, ' ').trim();
+    const href = this.linkHref ?? '';
+    this.inLink = false;
+    if (href && text) {
+      this.emit(`[${text}](${href})`);
+    } else if (text) {
+      this.emit(text);
+    }
+  }
+
+  private getAttr(el: Element, name: string): string | undefined {
+    return el.attrs.find(a => a.name === name)?.value;
+  }
+
+  private handleOpen(el: Element): void {
+    const tag = el.tagName.toLowerCase();
+
+    if (HEADING_TAGS.has(tag)) {
+      const level = parseInt(tag[1]!, 10);
+      this.emit('\n\n' + '#'.repeat(level) + ' ');
+    } else if (tag === 'a') {
+      this.linkHref = this.getAttr(el, 'href') ?? null;
+      this.linkTextParts = [];
+      this.inLink = true;
+    } else if (tag in INLINE_EMPHASIS) {
+      this.emit(INLINE_EMPHASIS[tag]!);
+    } else if (tag === 'br') {
+      this.emit('\n');
+    } else if (BLOCK_TAGS.has(tag)) {
+      this.emit('\n\n');
+    } else if (tag === 'hr') {
+      this.emit('\n\n---\n\n');
+    } else if (tag === 'blockquote') {
+      this.emit('\n\n');
+      this.bqStack.push([]);
+    } else if (tag === 'ul') {
+      this.listStack.push('ul');
+      this.emit('\n');
+    } else if (tag === 'ol') {
+      this.listStack.push('ol');
+      const startAttr = this.getAttr(el, 'start');
+      let start = 1;
+      if (startAttr != null) {
+        const parsed = parseInt(startAttr, 10);
+        if (!isNaN(parsed)) start = parsed;
+      }
+      this.olCounter.push(start - 1);
+      this.emit('\n');
+    } else if (tag === 'li') {
+      const indent = '  '.repeat(Math.max(0, this.listStack.length - 1));
+      if (this.listStack.length > 0 && this.listStack[this.listStack.length - 1] === 'ol') {
+        if (this.olCounter.length > 0) {
+          this.olCounter[this.olCounter.length - 1]!++;
+          this.emit(`\n${indent}${this.olCounter[this.olCounter.length - 1]}. `);
+        } else {
+          this.emit(`\n${indent}1. `);
+        }
+      } else {
+        this.emit(`\n${indent}* `);
+      }
+    } else if (tag === 'pre') {
+      this.preParts = [];
+      this.inPre = true;
+      this.preLanguage = null;
+      const codeChild = el.childNodes.find(
+        (c): c is Element => isElement(c) && c.tagName === 'code'
+      );
+      if (codeChild) {
+        const cls = this.getAttr(codeChild, 'class') ?? '';
+        const langMatch = cls.match(/(?:^|\s)language-(\S+)/);
+        if (langMatch) this.preLanguage = langMatch[1]!;
+      }
+    } else if (tag === 'code' && !this.inPre) {
+      this.inInlineCode = true;
+      this.emit('`');
+    } else if (tag === 'table') {
+      this.inTable = true;
+      this.headerRowDone = false;
+      this.isFirstRow = true;
+      this.emit('\n\n');
+    } else if (tag === 'tr') {
+      this.finishCell();
+      this.finishRow();
+    } else if (tag === 'th' || tag === 'td') {
+      this.finishCell();
+      this.cellParts = [];
+      this.inCell = true;
+      if (tag === 'th') this.rowHasTh = true;
+    }
+  }
+
+  private handleClose(tag: string): void {
+    tag = tag.toLowerCase();
+
+    if (HEADING_TAGS.has(tag)) {
+      this.emit('\n\n');
+    } else if (tag === 'a') {
+      this.finishLink();
+    } else if (tag in INLINE_EMPHASIS) {
+      this.emit(INLINE_EMPHASIS[tag]!);
+    } else if (BLOCK_TAGS.has(tag)) {
+      this.emit('\n\n');
+    } else if (tag === 'blockquote') {
+      if (this.bqStack.length > 0) {
+        const content = this.bqStack.pop()!.join('');
+        const prefixed = this.prefixBlockquote(content);
+        if (prefixed) this.emit('\n\n' + prefixed + '\n\n');
+      }
+    } else if (tag === 'ul') {
+      if (this.listStack.length > 0 && this.listStack[this.listStack.length - 1] === 'ul') {
+        this.listStack.pop();
+      }
+      this.emit('\n');
+    } else if (tag === 'ol') {
+      if (this.listStack.length > 0 && this.listStack[this.listStack.length - 1] === 'ol') {
+        this.listStack.pop();
+        if (this.olCounter.length > 0) this.olCounter.pop();
+      }
+      this.emit('\n');
+    } else if (tag === 'pre') {
+      const raw = this.preParts.join('');
+      this.inPre = false;
+      const lang = this.preLanguage ?? '';
+      const block = '```' + lang + '\n' + raw + '\n```';
+      this.emit('\n\n' + block + '\n\n');
+      this.preLanguage = null;
+    } else if (tag === 'code' && !this.inPre) {
+      this.inInlineCode = false;
+      this.emit('`');
+    } else if (tag === 'th' || tag === 'td') {
+      this.finishCell();
+    } else if (tag === 'tr') {
+      this.finishCell();
+      this.finishRow();
+    } else if (tag === 'table') {
+      this.finishCell();
+      this.finishRow();
+      this.inTable = false;
+      this.emit('\n');
+    }
+  }
+
+  private handleText(data: string): void {
+    if (this.inPre) {
+      this.preParts.push(data);
+      return;
+    }
+    if (this.inInlineCode) {
+      this.emit(data);
+      return;
+    }
+    const text = data.replace(/\s+/g, ' ');
+    if (this.inTable && !this.inCell && !text.trim()) return;
+    this.emit(text);
+  }
+
+  walk(node: ChildNode | Document): void {
+    if (isText(node as ChildNode)) {
+      this.handleText((node as TextNode).value);
+      return;
+    }
+    if (node.nodeName === '#comment') return;
+
+    if (isElement(node as ChildNode)) {
+      const el = node as Element;
+      const tag = el.tagName.toLowerCase();
+      if (SKIP_TAGS.has(tag)) return;
+      if (tag === 'img') return;
+
+      this.handleOpen(el);
+
+      if (tag === 'pre') {
+        for (const child of el.childNodes) {
+          if (isElement(child) && child.tagName === 'code') {
+            for (const grandchild of child.childNodes) {
+              this.walk(grandchild);
+            }
+          } else {
+            this.walk(child);
+          }
+        }
+      } else {
+        for (const child of el.childNodes) {
+          this.walk(child);
+        }
+      }
+
+      this.handleClose(tag);
+      return;
+    }
+
+    if ('childNodes' in node) {
+      for (const child of (node as Document).childNodes) {
+        this.walk(child);
+      }
+    }
+  }
+
+  getOutput(): string {
+    return this.out.join('');
+  }
+}
+
+function cleanup(text: string): string {
+  const lines = text.split('\n');
+  const out: string[] = [];
+  let inFence = false;
+  let blankRun = 0;
+
+  for (const line of lines) {
+    const stripped = line.replace(/[ \t]+$/, '');
+    if (stripped.startsWith('```')) {
+      inFence = !inFence;
+      blankRun = 0;
+      out.push(stripped);
+      continue;
+    }
+    if (inFence) {
+      out.push(line);
+      continue;
+    }
+    if (!stripped) {
+      blankRun++;
+      if (blankRun <= 1) out.push('');
+      continue;
+    }
+    blankRun = 0;
+    out.push(stripped);
+  }
+
+  return out.join('\n').trim();
+}
+
+export function htmlToMarkdown(sourceHtml: string): string {
+  sourceHtml = sourceHtml.replace(/\r\n/g, '\n').replace(/\r/g, '\n');
+  const doc = parse(sourceHtml);
+  const renderer = new MarkdownRenderer();
+  renderer.walk(doc);
+  return cleanup(renderer.getOutput());
+}

apps/server/src/services/web/index.ts

@@ -0,0 +1 @@
+export { htmlToMarkdown } from './html-to-md.js';

apps/server/src/services/web_fetch.ts

@@ -12,6 +12,7 @@ import { z } from 'zod';
 import { isPublicUrl } from './url_guard.js';
 import type { ToolDef } from './tools.js';
 import { truncateIfNeeded } from './truncate.js';
+import { htmlToMarkdown } from './web/index.js';
 
 const WebFetchInput = z.object({
   url: z.string().min(1).max(2048),
@@ -38,29 +39,9 @@ export type WebFetchOutput =
     }
   | { error: string; reason: string; content_type?: string };
 
-function stripHtml(html: string): { text: string; title: string | undefined } {
+function extractTitle(html: string): string | undefined {
-  // Title first, before we destroy the markup. Trim collapsed whitespace.
   const titleMatch = html.match(/<title[^>]*>([\s\S]*?)<\/title>/i);
-  const title = titleMatch?.[1]?.replace(/\s+/g, ' ').trim() || undefined;
+  return titleMatch?.[1]?.replace(/\s+/g, ' ').trim() || undefined;
-  // Drop script + style + comments entirely (their CONTENT must not leak —
-  // a regex tag stripper alone would expose inline JS as plain text).
-  const text = html
-    .replace(/<script\b[^>]*>[\s\S]*?<\/script>/gi, ' ')
-    .replace(/<style\b[^>]*>[\s\S]*?<\/style>/gi, ' ')
-    .replace(/<noscript\b[^>]*>[\s\S]*?<\/noscript>/gi, ' ')
-    .replace(/<!--[\s\S]*?-->/g, ' ')
-    .replace(/<[^>]+>/g, ' ')
-    // Minimal entity decode — full coverage would need a table; covering
-    // the five common ones plus &nbsp; is enough for snippet readability.
-    .replace(/&nbsp;/g, ' ')
-    .replace(/&amp;/g, '&')
-    .replace(/&lt;/g, '<')
-    .replace(/&gt;/g, '>')
-    .replace(/&quot;/g, '"')
-    .replace(/&#39;/g, "'")
-    .replace(/\s+/g, ' ')
-    .trim();
-  return { text, title };
 }
 
 // v1.11.10: streaming body reader. Aborts the response stream the instant
@@ -211,9 +192,8 @@ export async function executeWebFetch(
   let textRaw: string;
   let title: string | undefined;
   if (contentType.includes('text/html') || contentType.includes('application/xhtml')) {
-    const stripped = stripHtml(body);
+    title = extractTitle(body);
-    textRaw = stripped.text;
+    textRaw = htmlToMarkdown(body);
-    title = stripped.title;
   } else if (
     contentType.includes('text/plain') ||
     contentType.includes('text/markdown') ||

apps/server/src/types/api.ts

@@ -113,6 +113,7 @@ export interface Agent {
   // v1.14.0: per-agent step cap for the outer inference loop. null means
   // bounded only by MAX_STEPS (200). 0 means "no tool calls allowed."
   steps: number | null;
+  llama_extra_args: string[] | null;
 }
 
 // One entry per malformed `## Name` block. Per-block errors don't fail the

apps/server/src/types/ws-frames.ts

@@ -272,7 +272,9 @@ export const PermissionRequestedFrame = z.object({
   type: z.literal('permission_requested'),
   task_id: Uuid,
   session_id: Uuid,
+  kind: z.enum(['tool', 'question', 'plan', 'elicitation']).optional(),
   tool_title: z.string().optional(),
+  input: z.record(z.unknown()).optional(),
   options: z.array(PermissionOptionShape),
 });
 

apps/web/package.json

@@ -43,5 +43,6 @@
     "tailwindcss": "^4.3.0",
     "typescript": "^5.5.0",
     "vite": "^5.3.4"
-  }
+  },
+  "license": "AGPL-3.0-only"
 }

apps/web/src/api/client.ts

@@ -319,10 +319,10 @@ export const api = {
       }),
     getTaskPermission: (taskId: string) =>
       request<PermissionPrompt>(`/api/coder/tasks/${taskId}/permission`),
-    respondTaskPermission: (taskId: string, optionId: string | null) =>
+    respondTaskPermission: (taskId: string, optionId: string | null, updatedInput?: Record<string, unknown>) =>
       request<{ ok: boolean }>(`/api/coder/tasks/${taskId}/permission`, {
         method: 'POST',
-        body: JSON.stringify({ option_id: optionId }),
+        body: JSON.stringify({ option_id: optionId, ...(updatedInput ? { updated_input: updatedInput } : {}) }),
       }),
     getTaskCommands: (taskId: string) =>
       request<{ taskId: string; commands: AgentCommand[] }>(`/api/coder/tasks/${taskId}/commands`),

apps/web/src/api/types.ts

@@ -250,9 +250,13 @@ export interface AgentSessionConfig {
   thinkingOptionId: string | null;
 }
 
+export type PermissionKind = 'tool' | 'question' | 'plan' | 'elicitation';
+
 export interface PermissionPrompt {
   taskId: string;
+  kind?: PermissionKind;
   toolTitle?: string;
+  input?: Record<string, unknown>;
   options: Array<{ optionId: string; label: string }>;
 }
 

apps/web/src/api/ws-frames.ts

@@ -272,7 +272,9 @@ export const PermissionRequestedFrame = z.object({
   type: z.literal('permission_requested'),
   task_id: Uuid,
   session_id: Uuid,
+  kind: z.enum(['tool', 'question', 'plan', 'elicitation']).optional(),
   tool_title: z.string().optional(),
+  input: z.record(z.unknown()).optional(),
   options: z.array(PermissionOptionShape),
 });
 

apps/web/src/components/AskUserInputCard.tsx

@@ -1,7 +1,6 @@
 import { useMemo, useState } from 'react';
 import { Check } from 'lucide-react';
 import { toast } from 'sonner';
-import { api } from '@/api/client';
 import { RadioGroup, RadioGroupItem } from '@/components/ui/radio-group';
 import { Button } from '@/components/ui/button';
 import type {
@@ -22,6 +21,7 @@ interface Props {
   toolCall: ToolCall;
   toolResult: ToolResult | null;
   chatId: string;
+  apiPrefix?: string;
 }
 
 function parseQuestions(raw: unknown): AskUserQuestion[] {
@@ -63,7 +63,7 @@ function parseAnswerSet(raw: unknown): AskUserAnswerSet | null {
   return { answers };
 }
 
-export function AskUserInputCard({ toolCall, toolResult, chatId }: Props) {
+export function AskUserInputCard({ toolCall, toolResult, chatId, apiPrefix = '' }: Props) {
   const questions = useMemo(() => parseQuestions(toolCall.args), [toolCall.args]);
 
   if (questions.length === 0) {
@@ -74,9 +74,6 @@ export function AskUserInputCard({ toolCall, toolResult, chatId }: Props) {
     );
   }
 
-  // Tool result with a non-null output means the answer is already submitted.
-  // The pending sentinel uses output=null, so this branch only triggers after
-  // the real WS tool_result frame lands.
   const answered = toolResult && toolResult.output !== null;
   if (answered) {
     const answerSet = parseAnswerSet(toolResult!.output);
@@ -84,7 +81,7 @@ export function AskUserInputCard({ toolCall, toolResult, chatId }: Props) {
   }
 
   return (
-    <PendingView questions={questions} toolCallId={toolCall.id} chatId={chatId} />
+    <PendingView questions={questions} toolCallId={toolCall.id} chatId={chatId} apiPrefix={apiPrefix} />
   );
 }
 
@@ -92,10 +89,12 @@ function PendingView({
   questions,
   toolCallId,
   chatId,
+  apiPrefix = '',
 }: {
   questions: AskUserQuestion[];
   toolCallId: string;
   chatId: string;
+  apiPrefix?: string;
 }) {
   // Per-question selections + free text. Selections are option arrays so the
   // multi_select case is uniform; single_select just constrains to length 1.
@@ -133,9 +132,16 @@ function PendingView({
     if (submitting) return;
     setSubmitting(true);
     try {
-      await api.chats.answerUserInput(chatId, toolCallId, answers);
+      const url = `${apiPrefix}/api/chats/${chatId}/answer_user_input`;
-      // Card stays mounted; the incoming WS tool_result frame will flip it
+      const res = await fetch(url, {
-      // into AnsweredView via the parent prop change.
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify({ tool_call_id: toolCallId, answers }),
+      });
+      if (!res.ok) {
+        const body = await res.json().catch(() => ({})) as { error?: string; detail?: string };
+        throw new Error(body.detail ?? body.error ?? `HTTP ${res.status}`);
+      }
     } catch (err) {
       toast.error(err instanceof Error ? err.message : 'submit failed');
       setSubmitting(false);

apps/web/src/components/PermissionCard.tsx

@@ -1,14 +1,105 @@
-import { ShieldAlert } from 'lucide-react';
+import { useState } from 'react';
+import { ShieldAlert, MessageCircleQuestion } from 'lucide-react';
 import type { PermissionPrompt } from '@/api/types';
+import { RadioGroup, RadioGroupItem } from '@/components/ui/radio-group';
+import { Button } from '@/components/ui/button';
 import { cn } from '@/lib/utils';
 
 interface Props {
   prompt: PermissionPrompt;
-  onRespond: (optionId: string | null) => void;
+  onRespond: (optionId: string | null, updatedInput?: Record<string, unknown>) => void;
   busy?: boolean;
 }
 
+// ---------------------------------------------------------------------------
+// Question detection — ACP's RequestPermissionRequest carries the tool input
+// in `input`. Claude Code's AskUserQuestion puts { questions: [...] } there.
+// ---------------------------------------------------------------------------
+
+interface Question {
+  question: string;
+  header?: string;
+  options: string[];
+  multiSelect: boolean;
+}
+
+function parseQuestions(input: Record<string, unknown> | undefined): Question[] | null {
+  if (!input) return null;
+  const raw = input.questions;
+  if (!Array.isArray(raw)) return null;
+  const out: Question[] = [];
+  for (const item of raw) {
+    if (!item || typeof item !== 'object') continue;
+    const q = item as { question?: unknown; header?: unknown; options?: unknown; multiSelect?: unknown };
+    if (typeof q.question !== 'string') continue;
+    const opts = Array.isArray(q.options)
+      ? q.options.filter((o): o is string => typeof o === 'string')
+      : [];
+    out.push({
+      question: q.question,
+      header: typeof q.header === 'string' ? q.header : undefined,
+      options: opts,
+      multiSelect: q.multiSelect === true,
+    });
+  }
+  return out.length > 0 ? out : null;
+}
+
+// ---------------------------------------------------------------------------
+// Elicitation detection — ACP's createElicitation carries a JSON Schema in
+// `input.requestedSchema`. For now, render each property as a text input.
+// ---------------------------------------------------------------------------
+
+interface ElicitationField {
+  key: string;
+  title: string;
+  description?: string;
+  type: string;
+  enumValues?: string[];
+}
+
+function parseElicitation(input: Record<string, unknown> | undefined): { message: string; fields: ElicitationField[] } | null {
+  if (!input) return null;
+  const schema = input.requestedSchema;
+  if (!schema || typeof schema !== 'object') return null;
+  const s = schema as Record<string, unknown>;
+  const props = s.properties;
+  if (!props || typeof props !== 'object') return null;
+  const fields: ElicitationField[] = [];
+  for (const [key, val] of Object.entries(props as Record<string, unknown>)) {
+    if (!val || typeof val !== 'object') continue;
+    const p = val as Record<string, unknown>;
+    fields.push({
+      key,
+      title: typeof p.title === 'string' ? p.title : key,
+      description: typeof p.description === 'string' ? p.description : undefined,
+      type: typeof p.type === 'string' ? p.type : 'string',
+      enumValues: Array.isArray(p.enum) ? p.enum.filter((e): e is string => typeof e === 'string') : undefined,
+    });
+  }
+  if (fields.length === 0) return null;
+  return { message: typeof input.message === 'string' ? input.message : '', fields };
+}
+
 export function PermissionCard({ prompt, onRespond, busy }: Props) {
+  const isQuestion = prompt.kind === 'question';
+  const isElicitation = prompt.kind === 'elicitation';
+
+  if (isQuestion) {
+    const questions = parseQuestions(prompt.input);
+    if (questions) {
+      return <QuestionView questions={questions} prompt={prompt} onRespond={onRespond} busy={busy} />;
+    }
+  }
+
+  if (isElicitation) {
+    const elicitation = parseElicitation(prompt.input);
+    if (elicitation) {
+      return <ElicitationView elicitation={elicitation} prompt={prompt} onRespond={onRespond} busy={busy} />;
+    }
+  }
+
+  // Standard tool permission — approve/deny buttons
   return (
     <div className="mx-3 my-2 rounded-md border border-amber-500/40 bg-amber-500/10 px-3 py-2 text-sm">
       <div className="flex items-start gap-2">
@@ -47,3 +138,286 @@ export function PermissionCard({ prompt, onRespond, busy }: Props) {
     </div>
   );
 }
+
+// ---------------------------------------------------------------------------
+// QuestionView — renders Claude's AskUserQuestion as interactive radio/checkbox
+// ---------------------------------------------------------------------------
+
+function QuestionView({
+  questions,
+  prompt,
+  onRespond,
+  busy,
+}: {
+  questions: Question[];
+  prompt: PermissionPrompt;
+  onRespond: Props['onRespond'];
+  busy?: boolean;
+}) {
+  const [selections, setSelections] = useState<string[][]>(() => questions.map(() => []));
+  const [freeTexts, setFreeTexts] = useState<string[]>(() => questions.map(() => ''));
+  const [submitting, setSubmitting] = useState(false);
+  const disabled = busy || submitting;
+
+  const allComplete = questions.every((_, i) =>
+    selections[i]!.length > 0 || freeTexts[i]!.trim().length > 0,
+  );
+
+  function buildAnswers(): Record<string, string> {
+    const answers: Record<string, string> = {};
+    for (let i = 0; i < questions.length; i++) {
+      const q = questions[i]!;
+      const key = q.question;
+      const selected = selections[i]!;
+      const free = freeTexts[i]!.trim();
+      if (free) {
+        answers[key] = free;
+      } else if (selected.length > 0) {
+        answers[key] = selected.join(', ');
+      }
+    }
+    return answers;
+  }
+
+  function handleSubmit() {
+    if (!allComplete || submitting) return;
+    setSubmitting(true);
+    const answers = buildAnswers();
+    const firstAllow = prompt.options.find((o) =>
+      o.label.toLowerCase().includes('allow') || o.label.toLowerCase().includes('yes'),
+    );
+    onRespond(firstAllow?.optionId ?? prompt.options[0]?.optionId ?? null, {
+      ...prompt.input,
+      answers,
+    });
+  }
+
+  function pickSingle(qIdx: number, option: string) {
+    setSelections((prev) => prev.map((arr, i) => (i === qIdx ? [option] : arr)));
+    if (questions.length === 1 && !freeTexts[0]!.trim()) {
+      setSubmitting(true);
+      const firstAllow = prompt.options.find((o) =>
+        o.label.toLowerCase().includes('allow') || o.label.toLowerCase().includes('yes'),
+      );
+      onRespond(firstAllow?.optionId ?? prompt.options[0]?.optionId ?? null, {
+        ...prompt.input,
+        answers: { [questions[0]!.question]: option },
+      });
+    }
+  }
+
+  function toggleMulti(qIdx: number, option: string) {
+    setSelections((prev) =>
+      prev.map((arr, i) => {
+        if (i !== qIdx) return arr;
+        return arr.includes(option) ? arr.filter((o) => o !== option) : [...arr, option];
+      }),
+    );
+  }
+
+  return (
+    <div className="mx-3 my-2 rounded-lg border bg-muted/20 text-sm">
+      <div className="px-4 py-3 space-y-4">
+        {questions.map((q, i) => (
+          <div key={i} className="space-y-2">
+            {questions.length > 1 && (
+              <div className="text-[10px] uppercase tracking-wide text-muted-foreground/70">
+                {q.header ?? `Question ${i + 1}`}
+              </div>
+            )}
+            <div className="font-medium leading-snug">{q.question}</div>
+            {q.options.length > 0 && !q.multiSelect && (
+              <RadioGroup
+                value={selections[i]![0] ?? ''}
+                onValueChange={(v) => pickSingle(i, v)}
+                disabled={disabled}
+                className="gap-1.5"
+              >
+                {q.options.map((opt, j) => {
+                  const id = `q${i}-opt${j}`;
+                  return (
+                    <label
+                      key={j}
+                      htmlFor={id}
+                      className="flex items-start gap-2 text-sm leading-snug cursor-pointer rounded px-1 py-0.5 hover:bg-muted/40"
+                    >
+                      <RadioGroupItem id={id} value={opt} className="mt-0.5" />
+                      <span>{opt}</span>
+                    </label>
+                  );
+                })}
+              </RadioGroup>
+            )}
+            {q.options.length > 0 && q.multiSelect && (
+              <div className="grid gap-1.5">
+                {q.options.map((opt, j) => {
+                  const id = `q${i}-opt${j}`;
+                  const checked = selections[i]!.includes(opt);
+                  return (
+                    <label
+                      key={j}
+                      htmlFor={id}
+                      className="flex items-start gap-2 text-sm leading-snug cursor-pointer rounded px-1 py-0.5 hover:bg-muted/40"
+                    >
+                      <input
+                        id={id}
+                        type="checkbox"
+                        checked={checked}
+                        disabled={disabled}
+                        onChange={() => toggleMulti(i, opt)}
+                        className="mt-1 size-3.5 rounded border-input accent-primary"
+                      />
+                      <span>{opt}</span>
+                    </label>
+                  );
+                })}
+              </div>
+            )}
+            <div className="pt-1 space-y-1">
+              <div className="text-[10px] uppercase tracking-wide text-muted-foreground/70">
+                Or type a custom answer
+              </div>
+              <input
+                type="text"
+                value={freeTexts[i]}
+                disabled={disabled}
+                placeholder="Free text…"
+                onChange={(e) =>
+                  setFreeTexts((prev) => prev.map((t, idx) => (idx === i ? e.target.value : t)))
+                }
+                className="w-full rounded border border-input bg-background px-2 py-1 text-sm outline-none focus-visible:ring-2 focus-visible:ring-ring/40 disabled:opacity-60"
+              />
+            </div>
+          </div>
+        ))}
+      </div>
+      {(questions.length > 1 || freeTexts.some((t) => t.trim())) && (
+        <div className="flex justify-between items-center border-t px-4 py-2">
+          <button
+            type="button"
+            disabled={disabled}
+            onClick={() => onRespond(null)}
+            className="text-xs text-destructive hover:underline disabled:opacity-40"
+          >
+            Dismiss
+          </button>
+          <Button
+            type="button"
+            size="sm"
+            disabled={!allComplete || disabled}
+            onClick={handleSubmit}
+          >
+            {submitting ? 'Submitting…' : 'Submit'}
+          </Button>
+        </div>
+      )}
+    </div>
+  );
+}
+
+// ---------------------------------------------------------------------------
+// ElicitationView — renders ACP elicitation forms (JSON Schema-driven)
+// ---------------------------------------------------------------------------
+
+function ElicitationView({
+  elicitation,
+  prompt,
+  onRespond,
+  busy,
+}: {
+  elicitation: { message: string; fields: ElicitationField[] };
+  prompt: PermissionPrompt;
+  onRespond: Props['onRespond'];
+  busy?: boolean;
+}) {
+  const [values, setValues] = useState<Record<string, string>>(() => {
+    const init: Record<string, string> = {};
+    for (const f of elicitation.fields) init[f.key] = '';
+    return init;
+  });
+  const [submitting, setSubmitting] = useState(false);
+  const disabled = busy || submitting;
+
+  const allFilled = elicitation.fields.every((f) => (values[f.key] ?? '').trim().length > 0);
+
+  function handleSubmit() {
+    if (!allFilled || submitting) return;
+    setSubmitting(true);
+    const content: Record<string, unknown> = {};
+    for (const f of elicitation.fields) {
+      const raw = values[f.key]!.trim();
+      if (f.type === 'number' || f.type === 'integer') {
+        content[f.key] = Number(raw);
+      } else if (f.type === 'boolean') {
+        content[f.key] = raw === 'true' || raw === 'yes' || raw === '1';
+      } else {
+        content[f.key] = raw;
+      }
+    }
+    const firstAllow = prompt.options[0];
+    onRespond(firstAllow?.optionId ?? null, content);
+  }
+
+  return (
+    <div className="mx-3 my-2 rounded-lg border bg-muted/20 text-sm">
+      <div className="px-4 py-3 space-y-3">
+        <div className="flex items-start gap-2">
+          <MessageCircleQuestion className="size-4 text-blue-500 shrink-0 mt-0.5" />
+          <p className="font-medium leading-snug">{elicitation.message}</p>
+        </div>
+        {elicitation.fields.map((f) => (
+          <div key={f.key} className="space-y-1">
+            <label className="text-xs font-medium text-muted-foreground">{f.title}</label>
+            {f.description && (
+              <p className="text-[11px] text-muted-foreground/70">{f.description}</p>
+            )}
+            {f.enumValues ? (
+              <RadioGroup
+                value={values[f.key] ?? ''}
+                onValueChange={(v) => setValues((prev) => ({ ...prev, [f.key]: v }))}
+                disabled={disabled}
+                className="gap-1.5"
+              >
+                {f.enumValues.map((opt, j) => {
+                  const id = `e-${f.key}-${j}`;
+                  return (
+                    <label key={j} htmlFor={id} className="flex items-start gap-2 text-sm cursor-pointer rounded px-1 py-0.5 hover:bg-muted/40">
+                      <RadioGroupItem id={id} value={opt} className="mt-0.5" />
+                      <span>{opt}</span>
+                    </label>
+                  );
+                })}
+              </RadioGroup>
+            ) : (
+              <input
+                type={f.type === 'number' || f.type === 'integer' ? 'number' : 'text'}
+                value={values[f.key] ?? ''}
+                disabled={disabled}
+                onChange={(e) => setValues((prev) => ({ ...prev, [f.key]: e.target.value }))}
+                className="w-full rounded border border-input bg-background px-2 py-1 text-sm outline-none focus-visible:ring-2 focus-visible:ring-ring/40 disabled:opacity-60"
+              />
+            )}
+          </div>
+        ))}
+      </div>
+      <div className="flex justify-between items-center border-t px-4 py-2">
+        <button
+          type="button"
+          disabled={disabled}
+          onClick={() => onRespond(null)}
+          className="text-xs text-destructive hover:underline disabled:opacity-40"
+        >
+          Dismiss
+        </button>
+        <Button
+          type="button"
+          size="sm"
+          disabled={!allFilled || disabled}
+          onClick={handleSubmit}
+        >
+          {submitting ? 'Submitting…' : 'Submit'}
+        </Button>
+      </div>
+    </div>
+  );
+}

apps/web/src/components/panes/CoderMessageList.tsx

@@ -230,6 +230,7 @@ export function CoderMessageList({ messages, chatId, footer }: Props) {
                   toolCall={item.run.call}
                   toolResult={item.run.result}
                   chatId={chatId}
+                  apiPrefix="/api/coder"
                 />
               );
             }

apps/web/src/components/panes/CoderPane.tsx

@@ -290,7 +290,9 @@ function useCoderMessages(sessionId: string, chatId: string | undefined, handler
         } else if (frame.type === 'permission_requested') {
           handlersRef.current.onPermissionRequested?.({
             taskId: frame.task_id,
+            kind: frame.kind,
             toolTitle: frame.tool_title,
+            ...(frame.input ? { input: frame.input as Record<string, unknown> } : {}),
             options: (frame.options ?? []).map((o: { option_id: string; label: string }) => ({
               optionId: o.option_id,
               label: o.label,
@@ -565,11 +567,11 @@ export function CoderPane({
     setProviderCommands(commands);
   }, []);
 
-  const handlePermissionRespond = useCallback(async (optionId: string | null) => {
+  const handlePermissionRespond = useCallback(async (optionId: string | null, updatedInput?: Record<string, unknown>) => {
     if (!permissionPrompt) return;
     setPermissionBusy(true);
     try {
-      await api.coder.respondTaskPermission(permissionPrompt.taskId, optionId);
+      await api.coder.respondTaskPermission(permissionPrompt.taskId, optionId, updatedInput);
       setPermissionPrompt(null);
     } finally {
       setPermissionBusy(false);
@@ -716,7 +718,7 @@ export function CoderPane({
       {permissionPrompt && (
         <PermissionCard
           prompt={permissionPrompt}
-          onRespond={(id) => void handlePermissionRespond(id)}
+          onRespond={(id, input) => void handlePermissionRespond(id, input)}
           busy={permissionBusy}
         />
       )}

data/AGENTS.md

@@ -7,7 +7,7 @@ top_p: 0.95
 top_k: 20
 min_p: 0.0
 presence_penalty: 0.0
-tools: [find_files, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_semantic_neighborhoods, get_symbol_info, grep, list_dir, search_symbols, view_file, watch_changes]
+tools: [find_files, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_semantic_neighborhoods, get_symbol_info, grep, list_dir, search_symbols, view_file, watch_changes, request_read_access, view_truncated_output, ask_user_input, git_status, get_blast_radius, get_hot_files, get_middleware, get_routes]
 description: Reviews code for bugs, security issues, and maintainability. Read-only.
 ---
 You review code. Find real problems, not style nits.
@@ -46,7 +46,7 @@ top_p: 0.95
 top_k: 20
 min_p: 0.0
 presence_penalty: 0.0
-tools: [find_files, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_semantic_neighborhoods, get_symbol_info, grep, list_dir, search_symbols, view_file, watch_changes]
+tools: [find_files, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_semantic_neighborhoods, get_symbol_info, grep, list_dir, search_symbols, view_file, watch_changes, request_read_access, view_truncated_output, ask_user_input, git_status, get_blast_radius, get_hot_files, get_middleware, get_routes]
 description: Diagnoses bugs from error messages, logs, or described symptoms.
 ---
 You diagnose bugs. Form a hypothesis, prove it with evidence from the code.
@@ -72,7 +72,7 @@ top_k: 20
 min_p: 0.0
 presence_penalty: 0.0
 steps: 5
-tools: [find_files, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_semantic_neighborhoods, get_symbol_info, grep, list_dir, search_symbols, view_file, watch_changes]
+tools: [find_files, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_semantic_neighborhoods, get_symbol_info, grep, list_dir, search_symbols, view_file, watch_changes, request_read_access, view_truncated_output, ask_user_input, git_status, get_blast_radius, get_hot_files, get_middleware, get_routes]
 description: Proposes refactors for clarity, deduplication, or decoupling. Read-only — outputs plans, not edits.
 ---
 You propose refactors. You do not apply them. The user applies via OpenCode or Claude Code.
@@ -115,7 +115,7 @@ top_k: 20
 min_p: 0.0
 presence_penalty: 1.5
 steps: 20
-tools: [find_files, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_semantic_neighborhoods, get_symbol_info, grep, list_dir, search_symbols, view_file, watch_changes]
+tools: [find_files, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_semantic_neighborhoods, get_symbol_info, grep, list_dir, search_symbols, view_file, watch_changes, request_read_access, view_truncated_output, ask_user_input, git_status, get_blast_radius, get_hot_files, get_middleware, get_routes]
 description: Designs new features, modules, or architectural changes. Outputs a build plan.
 ---
 You design. You produce build plans, not code.
@@ -157,7 +157,7 @@ top_p: 0.95
 top_k: 20
 min_p: 0.0
 presence_penalty: 0.0
-tools: [find_files, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_semantic_neighborhoods, get_symbol_info, grep, list_dir, search_symbols, view_file, watch_changes]
+tools: [find_files, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_semantic_neighborhoods, get_symbol_info, grep, list_dir, search_symbols, view_file, watch_changes, request_read_access, view_truncated_output, ask_user_input, git_status, get_blast_radius, get_hot_files, get_middleware, get_routes]
 description: Audits code for security vulnerabilities. Read-only.
 ---
 You audit for security issues. Concrete findings only, no generic warnings.
@@ -240,7 +240,7 @@ top_p: 0.95
 top_k: 20
 min_p: 0.0
 presence_penalty: 0.0
-tools: [find_files, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_semantic_neighborhoods, get_symbol_info, grep, list_dir, search_symbols, view_file, watch_changes]
+tools: [find_files, get_codebase_overview, get_dependencies, get_file_analysis, get_framework_analysis, get_semantic_neighborhoods, get_symbol_info, grep, list_dir, search_symbols, view_file, watch_changes, request_read_access, view_truncated_output, ask_user_input, git_status, get_blast_radius, get_hot_files, get_middleware, get_routes]
 description: Discovers and maps unfamiliar codebases. Reads architecture, traces data flow, identifies key symbols.
 ---
 You map codebases. Start broad, then drill into specifics.

docker-compose.yml

@@ -11,6 +11,7 @@ services:
       CONTAINER_GUIDANCE_FILE: /app/BOOCHAT.md
       DATABASE_URL: postgres://boocode:${POSTGRES_PASSWORD}@boocode_db:5432/boochat
       BOOCODER_URL: http://100.114.205.53:9502
+      LLAMA_SIDECAR_URL: http://100.101.41.16:8402
     volumes:
       - /opt:/opt
       - /opt/projects:/opt/projects:rw

package.json

@@ -10,5 +10,6 @@
   },
   "devDependencies": {
     "typescript": "^5.5.0"
-  }
+  },
+  "license": "AGPL-3.0-only"
 }

pnpm-lock.yaml

@@ -155,6 +155,9 @@ importers:
       fastify:
         specifier: ^4.28.1
         version: 4.29.1
+      parse5:
+        specifier: ^8.0.1
+        version: 8.0.1
       postgres:
         specifier: ^3.4.4
         version: 3.4.9
@@ -2382,6 +2385,10 @@ packages:
     resolution: {integrity: sha512-QyL119InA+XXEkNLNTPCXPugSvOfhwv0JOlGNzvxs0hZaiHLNvXSpudUWsOlsXGWJh8G6ckCScEkVHfX3kw/2Q==}
     engines: {node: '>=10.13.0'}
 
+  entities@8.0.0:
+    resolution: {integrity: sha512-zwfzJecQ/Uej6tusMqwAqU/6KL2XaB2VZ2Jg54Je6ahNBGNH6Ek6g3jjNCF0fG9EWQKGZNddNjU5F1ZQn/sBnA==}
+    engines: {node: '>=20.19.0'}
+
   env-paths@2.2.1:
     resolution: {integrity: sha512-+h1lkLKhZMTYjog1VEpJNG7NZJWcuc2DDk/qsqSTRRCOXiLjeQ1d1/udrUGhqMxUgAlwKNZ0cf2uqan5GLuS2A==}
     engines: {node: '>=6'}
@@ -3274,6 +3281,9 @@ packages:
     resolution: {integrity: sha512-TXfryirbmq34y8QBwgqCVLi+8oA3oWx2eAnSn62ITyEhEYaWRlVZ2DvMM9eZbMs/RfxPu/PK/aBLyGj4IrqMHw==}
     engines: {node: '>=18'}
 
+  parse5@8.0.1:
+    resolution: {integrity: sha512-z1e/HMG90obSGeidlli3hj7cbocou0/wa5HacvI3ASx34PecNjNQeaHNo5WIZpWofN9kgkqV1q5YvXe3F0FoPw==}
+
   parseurl@1.3.3:
     resolution: {integrity: sha512-CiyeOxFT/JZyN5m0z9PfXw4SCBJ6Sygz1Dpl0wqjlhDEGGBP1GnsUVEL0p63hoG1fcj3fHynXi9NYO4nWOL+qQ==}
     engines: {node: '>= 0.8'}
@@ -6110,6 +6120,8 @@ snapshots:
       graceful-fs: 4.2.11
       tapable: 2.3.3
 
+  entities@8.0.0: {}
+
   env-paths@2.2.1: {}
 
   error-ex@1.3.4:
@@ -7267,6 +7279,10 @@ snapshots:
 
   parse-ms@4.0.0: {}
 
+  parse5@8.0.1:
+    dependencies:
+      entities: 8.0.0
+
   parseurl@1.3.3: {}
 
   path-browserify@1.0.1: {}