Checkpoint of in-progress backend work present in the tree, not authored this session: auto_name, inference tool-phase/turn, secret_guard, provider-registry, plus a new agent-allowlist test (7 tests, passing).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
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>
Converts the ad-hoc executeToolPhase → runAssistantTurn recursion into an
explicit while (stepNumber < effectiveCap) loop. A step is one stream-and-
tool-execute iteration; the loop terminates on non-tool finish, step-cap hit,
doom-loop, budget exhaustion, abort, or synthesis success.
MAX_STEPS = 200 hard ceiling (4x old effective limit from budget). Per-agent
steps: field in AGENTS.md frontmatter sets tighter caps (Refactorer: 5,
Architect: 20, others: unset = bounded only by MAX_STEPS). Resolution:
effectiveCap = Math.min(agent.steps ?? Infinity, MAX_STEPS).
executeToolPhase no longer recurses — returns ToolPhaseResult struct
(action: 'continue' | 'paused' | 'synthesis_done') so the caller decides
whether to continue or break. steps: 0 handled as "no tool calls allowed"
via runTextOnlyTurn (one text-only stream phase, tool calls ignored with
warn log).
Step-cap hits produce a sentinel summary (reuses cap_hit kind so
CapHitSentinel.tsx renders without frontend changes; text distinguishes
"Step limit reached" from "Tool budget exhausted"). Doom-loop check migrated
to top of loop body — same predicate, same threshold (3), break instead of
return.
step_start parts are in the schema CHECK but not emitted as message_parts —
writing before the stream phase creates a sequence-0 collision with
partsFromAssistantMessage. Structured log line emitted instead. Adversarial
review caught the collision pre-deploy.
332/332 server tests passing. No frontend changes. No schema changes.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Recon during planning disproved the original v1.13.7 (DB-cache) premise:
buildSystemPrompt already runs over inputs mtime-cached at the file layer
(BOOCHAT.md in system-prompt.ts:25, AGENTS.md global+per-project in
agents.ts:245), and DB scalars are byte-stable until edited. The output
is microsecond pure-string concat with no I/O. Skills aren't in the
prefix; tools live in a separate request body field alpha-sorted by
v1.13.3.
This batch closes the verification gap with instrumentation, not
implementation:
- system-prompt.ts: buildSystemPromptWithFingerprint canonical impl
computes SHA-256 over the assembled prefix, runs a per-session
Map<sessionId, lastHash> observer, emits PrefixFingerprint per call
and PrefixDrift (with field-level changed_inputs) on hash change.
buildSystemPrompt is now a thin shim returning .prompt.
- agents.ts: getAgentsMtimes accessor — cache-read only, no I/O.
- payload.ts: buildMessagesPayload takes optional log argument; when
passed, emits prefix-fingerprint (info) + prefix-drift (warn).
- turn.ts + sentinel-summaries.ts: pass ctx.log at 3 production call
sites; sentinel summaries log too so any drift across cap-hit /
doom-loop paths surfaces.
- system-prompt.test.ts: 4 new tests (byte-identical, no-drift-on-
stable, drift-fires-with-changed-inputs, cross-session-no-drift).
194/194 tests pass (was 190).
Smoke: 5 messages in a fresh session produced 7 prefix-fingerprint
logs (extras from buildMessagesPayload being called from sentinel
summary paths), all with identical prefix_hash and prefix_length=2907,
zero prefix-drift. Prefix is byte-stable in steady-state.
Decision: original system_prompt_cache DB table from the roadmap is
permanently dropped. The v1.12.0 mtime caches at the input layer plus
alpha tool ordering at the request body (v1.13.3) already address the
load-bearing cache-stability surfaces. Instrumentation stays so the
claim can be re-verified at any time.
Pass 1 — ask_user_input correlation port (messages.ts:478, :549):
- The two correlation queries that backed the elicitation flow used to scan
messages.tool_calls and messages.tool_results JSON columns directly. They
now JOIN message_parts on payload->>'id' (for the caller assistant) and
payload->>'tool_call_id' (for the pending tool row). Semantics preserved:
ORDER BY m.created_at DESC LIMIT 1 still picks the latest issuance, the
already-answered 409 guard now reads payload.output, and the UPDATE +
parts replace inside sql.begin is unchanged from v1.13.0.
- Pre-v1.13.0 history has no parts rows and is unreachable to this lookup
path (404). Acceptable per dispatch decision — no pending elicitation
from before v1.13.0 will still be open. JSON-column fallback can land as
a hotfix if it ever surfaces.
Pass 2 — reasoning_parts wired end-to-end:
- types.ts/StreamResult gains `reasoning: string`. stream-phase.ts accumulates
reasoning-delta text per stream (replacing the v1.13.1-A counter-only
diagnostic) and returns it on the result.
- parts.ts/partsFromAssistantMessage gains an optional `reasoning` param.
When present it emits a kind='reasoning' part at sequence 0, ahead of
the text and tool_call parts.
- error-handler.ts/finalizeCompletion and tool-phase.ts/executeToolPhase
both thread result.reasoning into the dual-write call so reasoning-channel
models (qwen3.6) get persistent reasoning rows.
- payload.ts: loadContext SELECT pulls reasoning_parts from the v1.13.1-B
view; OpenAiMessage gains an optional `reasoning` field; buildMessagesPayload
collapses reasoning_parts into a single string per assistant message.
- stream-phase.ts/toModelMessages converts assistant messages with reasoning
into an AI SDK ModelMessage content array starting with a ReasoningPart,
matching the @ai-sdk/provider-utils AssistantContent union. Reasoning
models can now replay prior reasoning context across tool-call boundaries.
- types/api.ts and apps/web/src/api/types.ts Message interface gain
reasoning_parts (optional, nullable). Frontend doesn't render this yet —
field reserved for a v1.14 UI surface.
Tests: 2 new in parts.test.ts cover reasoning-at-sequence-0 with and
without text content. 172 tests pass (170 prior + 2 new).
Smoke verified against the live container:
- A reasoning-prompt ("walk through 17 × 23 step by step") produced one
message with kind='reasoning' (361 chars) at sequence 0 and kind='text'
(429 chars) at sequence 1. Adapter log confirmed reasoning capture.
- The new correlation SQL was validated against existing tool_call /
tool_result parts: returns the expected message_id + payload shape with
pending state correctly identified via payload.output IS NULL.
- ask_user_input end-to-end through the UI is Sam's smoke — the Prompt
Builder agent does not always trigger ask_user_input for these prompts,
so synthetic verification via SQL substituted for traffic-driven cover.
Annotation: the v1.13.1-A abort-throw site in stream-phase.ts got a
one-liner comment ("AI SDK v6 fullStream returns normally on abort; check
signal explicitly.") to prevent a future refactor removing it.
v1.13.2 drops the dual-write + the JSON columns + collapses the view.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- sentinel-summaries.ts: runCapHitSummary, insertCapHitSentinel,
runDoomLoopSummary, insertDoomLoopSentinel
- inference.ts → inference/turn.ts: residue is runAssistantTurn,
runInference, createInferenceRunner orchestration only
- inference/index.ts: re-export shim preserves the public surface
(createInferenceRunner, runInference, runAssistantTurn,
detectDoomLoop, DOOM_LOOP_THRESHOLD, buildMessagesPayload, plus
type-side InferenceContext/InferenceFrame/StreamResult/TurnArgs/
FramePublisher)
- src/index.ts + auto_name.ts + the two vitest test files updated to
import from ./services/inference/index.js explicitly (NodeNext ESM
doesn't honor directory-index resolution)
Final tally: 11 files under services/inference/, the largest being
sentinel-summaries.ts at 523 LoC (two near-clone summary paths kept
side-by-side until a third sentinel justifies factoring out a shared
runWrapUpSummary). turn.ts is now 326 LoC, the next-largest is
stream-phase.ts at 380. Public import surface unchanged.
tool-phase.ts → turn.ts back-edge for runAssistantTurn remains
(cycle is safe; resolved at call time).
Prepares the file structure for v1.13 AI SDK migration — streamText
swap targets stream-phase.ts only.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
