docs(changelog): v2.6.4-agent-sessions-fk

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

CHANGELOG.md

@@ -2,6 +2,18 @@
 
 All notable changes per release tag. Most recent on top, ordered by tag creation date (which matches the git history). Tag names follow `vMAJOR.MINOR.PATCH-slug` — the slug describes what shipped, so the tag name alone is enough to recall the batch.
 
+## v2.6.4-agent-sessions-fk — 2026-05-31
+
+Follow-up to `v2.6.3-chatkey-and-skills` (P1.5-b): the live `agent_sessions.session_id` foreign key is converged from `ON DELETE CASCADE` to `ON DELETE SET NULL`, matching the schema's stated intent. The P1.5-b re-key block re-adds `session_id_fkey` as `SET NULL`, but the whole block is guarded on `chat_id_fkey`'s absence — so a database already re-keyed to `(chat_id, agent)` while `session_id_fkey` was still `CASCADE` never re-enters it, leaving the live FK at `CASCADE` and diverging from both `worktree_id` (already `SET NULL`) and the `v2.6.3` changelog's own claim that `session_id` is informational `SET NULL`. The fix adds a standalone `confdeltype`-guarded `DO` block (mirroring the `session_worktrees` defang) that flips `session_id_fkey` `CASCADE → SET NULL` independently of the re-key gate; it is idempotent — fires only while the FK is still `'c'`, a no-op on a fresh deploy (already `'n'`) and on every re-run. The live DB was converged by hand with the identical statements, so `applySchema` and the hand-applied state match (`\d agent_sessions` now shows `session_id ... ON DELETE SET NULL`). Also bundles a CLAUDE.md doc-sync (committed separately): per-session SSE (P1.5-a) and the `(chat_id, agent)` re-key reflected in the engineering notes, the stale root `AGENTS.md` navigation pointer dropped, and new conventions for `data/AGENTS.md` parsing and the `data/skills/<vendor>/` layout.
+
+## v2.6.3-chatkey-and-skills — 2026-05-31
+
+Three threads. **agent_sessions re-keyed to `(chat_id, agent)` (P1.5-b):** the tab (a chat) is now the agent-context unit, so two opencode tabs in one BooCode session are two independent contexts that share one worktree. `chat_id` is threaded end-to-end — `tasks.chat_id` added, stamped by the coder message + skills routes from the frontend tab, read by `runOpenCodeServerTask` which falls back to resolve-or-create a chat for session-less creators (arena/MCP/new_task/generic `/api/tasks`) so `ensureSession` never receives a degenerate `(null, agent)` key. A new first-class `worktrees` table (one-per-session, survives session delete via `session_id ON DELETE SET NULL`) supersedes `session_worktrees`, which is defanged (CASCADE dropped, not yet removed); `agent_sessions.chat_id` CASCADEs from `chats` (closing a tab ends its context) while `worktree_id`/`session_id` are informational `SET NULL`. The migration is idempotent with a backfill-verify gate; the live re-key was applied against an empty table after the 35-chat test session `20d28876` was deleted (backed up first). This corrects and supersedes an earlier draft that wrongly keyed on `(worktree_id, agent)`; the delete-guard from `v2.6.2-delete-guard-and-sse` is repointed here from `session_worktrees` to `worktrees` (`worktree_path`→`path`). **dcp-strip cross-chunk fix:** the `<dcp-message-id>` tag streams split across SSE deltas, which the per-chunk strip from `v2.6.1-phase1-opencode` missed — a stateful `makeDcpStreamStripper` at the dispatcher boundary holds back partial-tag tails so neither live frames nor persisted content carry the tag (11 unit tests). **Agent-judgment skills:** `committing-changes` (segment by concern, stage explicitly, present-and-stop, never push) and `using-worktrees` (the when-to-isolate heuristic, autonomous-when-clear vs committing's command-gate) land in `data/skills/boocode/` with eval.yamls, plus a parser-safe `data/AGENTS.md` preamble pointing at both.
+
+## v2.6.2-delete-guard-and-sse — 2026-05-30
+
+Two coder-side batches under one tag. **Session-delete work-loss guard:** deleting a BooChat session CASCADE-wipes its `session_worktrees` row, which would silently orphan uncommitted/unpushed/unmerged work — so the server's `DELETE /api/sessions/:id` now gates before the delete. It reads `session_worktrees` from the shared DB first (no row → chat-only session → delete immediately, zero round-trip), and for worktree-backed sessions calls a new BooCoder endpoint (`/worktree-risk`) that runs git on the host, since the container can't see `/tmp/booworktrees` — only the host systemd service can. `checkWorktreeWorkAtRisk` reports dirty/unpushed/unmerged via the audited `hostExec`+`shellEscape` path, default branch detected from `refs/remotes/origin/HEAD` (never the worktree's own branch, never hardcoded); any at-risk worktree returns 409 with per-worktree `RiskReport[]`, `force=true` bypasses, and the check is fail-closed (BooCoder unreachable also blocks — force still escapes). The sidebar renders a block dialog distinguishing work-at-risk (Commit/Stash/Force; stash uses `-u` and re-blocks on remaining commits) from couldn't-verify (Cancel/Force), and Commit never auto-commits. A follow-up fix gates the `unpushed` arm behind an actual upstream (`atRisk = dirty || unmerged > 0 || (hasUpstream && unpushed > 0)`) so the no-upstream `session-<id>` branches stop flagging every pristine worktree-backed session — no protection lost, since real local work always also surfaces as `unmerged > 0`. **Per-session SSE (P1.5-a):** replaces the single global SSE loop scoped to the most-recent worktree directory — the known limit flagged in `v2.6.1-phase1-opencode` — with one `event.subscribe({directory})` per live opencode session, so sessions in different worktrees stream concurrently instead of the second silently dropping the first's events. Each session owns an `AbortController` wired into `subscribe(…, {signal})`, which also fixes a latent Phase-1 bug where switching directories left the old loop parked forever in its `for await` (zombie loops); a `sessionID` demux guard drops cross-session events so two sessions sharing a worktree (possible after P1.5-b) don't double-process deltas. The opencode SDK was confirmed to open an independent SSE connection per `subscribe()` call, so N concurrent dir-scoped streams are supported.
+
 ## v2.6.1-phase1-opencode — 2026-05-30
 
 v2.6 Phase 1: opencode runs as a warm HTTP server (`apps/coder/src/services/backends/opencode-server.ts`) — one `opencode serve` per BooCoder process, one opencode session per BooCode session resumed across turns via the new `agent_sessions` table, with a single SSE read loop, reasoning dedup ported from Paseo, an inactivity watchdog, and a stale-session guard (crashed-not-resumed + a `config_hash` fingerprint over `opencode_server|<model>`, deliberately excluding the ephemeral server port so cross-restart resume survives). Builds on the `v2.6.0-phase0-foundations` schema/interface scaffold. The batch's hard-won fixes: opencode streams `session.next.*` events (not `message.part.*`), and `event.subscribe()` must pass the session's worktree `directory` or events route to the server CWD and turns come back empty; model strings must be `llama-swap/`-prefixed and present in opencode's own config, with `agent-probe` now populating `available_agents.models` via `mergeLlamaSwap` so the frontend stops sending an empty model; `session_worktrees`/`agent_sessions` FKs are `ON DELETE CASCADE` so session deletion no longer 500s. Also bundled: dcp-message-id tag stripping from opencode text output, a reopen-closed-pane control, the `[+]`/split-pane button separation, auto-name using the session's loaded model, and a `systematic-debugging` slash command. Smoke 1 verified end-to-end (two turns, session reuse, turn 2 ~9x faster). Known Phase 1 limit: one SSE stream scoped to the most-recent session's directory — concurrent opencode sessions in different worktrees collide (warns; per-session SSE is Phase 2).

CLAUDE.md

@@ -2,7 +2,7 @@
 
 This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
 
-**Cursor agents:** start with `AGENTS.md` (navigation) and `docs/ARCHITECTURE.md` (diagram). This file is the deep engineering reference.
+**Cursor agents:** start with `docs/ARCHITECTURE.md` (diagram). This file is the deep engineering reference. (Note: the root navigation `AGENTS.md` was removed in v1.12; `data/AGENTS.md` is the agent *registry*, not navigation.)
 
 ## What is BooCode
 
@@ -90,9 +90,9 @@ Route registration: all routes registered in `index.ts` via `register*Routes(app
 - **Provider snapshot lifecycle** (`apps/coder/src/services/`): `provider-config.ts` (Zod config, never-throws on bad input) → `provider-config-registry.ts` (`buildResolvedRegistry`, singleton) → `provider-snapshot.ts` (two-tier probe: tier-1 fast presence, tier-2 cold ACP probe skipped unless force / stale `PROVIDER_PROBE_TTL_MS` 24h / dbEmpty; cached). Verify live: `curl http://100.114.205.53:9502/api/providers/snapshot` — returns providers + models + commands, the exact shape `AgentComposerBar` renders.
 - `PATCH /api/providers/config` replaces a provider id's override object **wholesale** (per-id shallow merge) — to flip one field send `{...existing, enabled}`, or a custom ACP entry's `command`/`label` is wiped and it drops out of the resolved registry. `data/coder-providers.json` is **gitignored** (it's live runtime config — the coder reads AND writes it on UI toggles); the tracked reference is `data/coder-providers.example.json`. The loader falls back to `{providers:{}}` (built-ins only) when the live file is absent, so a fresh checkout needs no copy.
 - **opencode** runs as a warm HTTP server (v2.6 Phase 1, `services/backends/opencode-server.ts` — `opencode serve` per BooCoder process, one opencode session per BooCode session, resumed via `agent_sessions`). goose/qwen/claude still dispatch **one-shot** ACP/PTY with no ctx/token usage; only native `boocode` (llama-swap engine) tracks ctx. Paseo's per-provider native clients (design §12) deliberately not ported.
-- **opencode SSE** (`opencode-server.ts`): live streaming arrives as `session.next.text.delta` / `session.next.reasoning.delta` / `session.next.tool.{called,success,failed}` — NOT `message.part.*` (those are terminal/post-hoc). `client.event.subscribe({ directory })` MUST pass the session's worktree directory; omit it and opencode scopes events to the server's `process.cwd()` → zero session events (empty turns, 180s watchdog timeout). One SSE stream at a time scoped to the last session's dir — concurrent opencode sessions in different worktrees collide (known Phase 1 limit, warns). Turn completes on `session.idle`; `promptAsync` is fire-and-forget (204).
+- **opencode SSE** (`opencode-server.ts`): live streaming arrives as `session.next.text.delta` / `session.next.reasoning.delta` / `session.next.tool.{called,success,failed}` — NOT `message.part.*` (those are terminal/post-hoc). `client.event.subscribe({ directory })` MUST pass the session's worktree directory; omit it and opencode scopes events to the server's `process.cwd()` → zero session events (empty turns, 180s watchdog timeout). Per-session SSE (P1.5-a): each live session owns its own `event.subscribe({directory})` loop + AbortController, so concurrent sessions in different worktrees stream independently; a `sessionID` demux guard drops cross-session events when two share a dir. Turn completes on `session.idle`; `promptAsync` is fire-and-forget (204).
 - **opencode model strings** must be provider-prefixed (`llama-swap/<model>`) AND exist in `~/.config/opencode/opencode.json` `provider.llama-swap.models` — not merely loadable by llama-swap. `parseModel` infers `llama-swap/` for a bare id; the dispatcher coalesces empty→DEFAULT_MODEL then prefixes. `agent-probe` populates opencode's `available_agents.models` via `mergeLlamaSwap` (fetches `/v1/models`); empty model list → frontend sends `''` → no inference (`input:0`, empty turn).
-- **agent_sessions resume**: `config_hash = sha256('opencode_server|<model>')` — must NOT include the server port (random per boot; including it breaks cross-restart resume). `session_worktrees` + `agent_sessions` FKs to `sessions(id)` are `ON DELETE CASCADE` (else DELETE /api/sessions/:id 500s on FK violation). The `@opencode-ai/sdk` v2 client takes flattened params (`{sessionID, directory, parts, model:{providerID,modelID}}`), imports `createOpencodeClient` from `@opencode-ai/sdk/v2/client`.
+- **agent_sessions resume**: `config_hash = sha256('opencode_server|<model>')` — must NOT include the server port (random per boot; including it breaks cross-restart resume). P1.5-b: `agent_sessions` is keyed `(chat_id, agent)` — the tab/chat is the context unit (two opencode tabs in one session = two contexts sharing one worktree). `chat_id` CASCADEs from `chats`; `session_id`/`worktree_id` are informational `SET NULL`. The `worktrees` table (one-per-session, `session_id` SET NULL so it survives session delete) supersedes the defanged `session_worktrees`. `tasks.chat_id` threads the tab id to the dispatcher; `runOpenCodeServerTask` falls back to resolve-or-create a chat when it's null (arena/MCP/new_task). The `@opencode-ai/sdk` v2 client takes flattened params (`{sessionID, directory, parts, model:{providerID,modelID}}`), imports `createOpencodeClient` from `@opencode-ai/sdk/v2/client`.
 
 ### Frontend (`apps/web/src/`)
 
@@ -192,6 +192,8 @@ BooCoder at port 9502: `curl http://100.114.205.53:9502/api/health`. Runs as `bo
 - **ReadableStream test stubs** use `pull()` (not `start()`) so chunks are produced lazily — `start()` enqueues everything and calls `controller.close()` before the consumer reads, so a subsequent `reader.cancel()` finds the stream already closed and the `cancel()` callback never fires. Also provide MORE chunks than the test will consume so the source stays in 'readable' state when cancel runs (e.g. cap test reads ~6 chunks, stub provides 10).
 - Tool-name whitelists must derive from `ALL_TOOLS` in `services/tools.ts`, never hardcoded. `services/agents.ts` `ALL_TOOL_NAMES` had this drift class until v1.12 — same pattern applies to any future tool-aware code.
 - Agent registry lives at `data/AGENTS.md` (global, bind-mounted at `/data/AGENTS.md`). No per-project `AGENTS.md` in this repo — removed in v1.12 to eliminate the two-files-must-stay-in-sync drift. The `getAgentsForProject` per-project override mechanism remains for *other* projects.
+- `data/AGENTS.md` is PARSED (`agents.ts` `splitSections`/`parseAgentSection`): each `## <Name>` is one agent and must be followed by a `---` frontmatter fence or the block throws; content before the first `## ` is discarded. Do NOT add free-form `## ` rule sections — they break the registry. Cross-cutting agent rules go in CLAUDE.md or a parser-ignored preamble.
+- Skills live in `data/skills/<vendor>/`; Sam's own namespace is `boocode/` (`committing-changes`, `using-worktrees`, `improving-boocode-guidance`) — `SKILL.md` + optional `eval.yaml` (gerund names; eval = `skill:` + `tasks:` of `prompt`+`grader`, incl. a negative-trigger task). `data/skills/` is canonical; a divergent mirror at `/opt/skills/` exists.
 - MCP stdio transport uses newline-delimited JSON (NDJSON), NOT LSP-style `Content-Length` headers. The `codecontext/shim.go` framing implementation is the reference; per the MCP spec (modelcontextprotocol.io/specification/server/transports).
 - **Workspace dependency pattern** (`apps/coder` → `@boocode/server`): the consuming package adds `"@boocode/server": "workspace:*"` in `package.json`. The provider's `package.json` needs `exports` with `types` + `default` conditions per subpath: `"./inference": { "types": "./dist/.../index.d.ts", "default": "./dist/.../index.js" }`. Without the `types` condition, NodeNext resolution can't find `.d.ts` files and tsc fails with "Cannot find module" in the consumer.
 - **JSONB columns**: use `sql.json(value as never)` — NOT `${JSON.stringify(value)}::jsonb` which double-serializes (stores a JSON string instead of a JSON object/array). Pattern established in `parts.ts`, `settings.ts`.

apps/coder/src/index.ts

@@ -30,6 +30,7 @@ import { registerInboxRoutes } from './routes/inbox.js';
 import { registerStatsRoutes } from './routes/stats.js';
 import { registerArenaRoutes } from './routes/arena.js';
 import { registerProviderRoutes } from './routes/providers.js';
+import { registerWorktreeSafetyRoutes } from './routes/worktree-safety.js';
 import { registerWebSocket } from './routes/ws.js';
 // Phase 4: dispatcher + agent probe
 import { createDispatcher } from './services/dispatcher.js';
@@ -195,6 +196,7 @@ async function main() {
   registerStatsRoutes(app, sql);
   registerArenaRoutes(app, sql);
   registerProviderRoutes(app, sql, config);
+  registerWorktreeSafetyRoutes(app, sql);
   registerWebSocket(app, sql, broker);
 
   // Serve static frontend (built web app). In production, the dist/ is

apps/coder/src/routes/messages.ts

@@ -224,8 +224,8 @@ export function registerMessageRoutes(
         // External provider: create a task for the dispatcher
         const projectId = sessionRows[0]!.project_id;
         const [task] = await sql<{ id: string; state: string }[]>`
-          INSERT INTO tasks (project_id, input, agent, model, mode_id, thinking_option_id, session_id)
+          INSERT INTO tasks (project_id, input, agent, model, mode_id, thinking_option_id, session_id, chat_id)
-          VALUES (${projectId}, ${content}, ${provider}, ${model ?? null}, ${mode_id ?? null}, ${thinking_option_id ?? null}, ${sessionId})
+          VALUES (${projectId}, ${content}, ${provider}, ${model ?? null}, ${mode_id ?? null}, ${thinking_option_id ?? null}, ${sessionId}, ${chatId})
           RETURNING id, state
         `;
         reply.code(202);

apps/coder/src/routes/skills.ts

@@ -91,8 +91,8 @@ export function registerSkillRoutes(
 
         const taskInput = `${body}\n\n---\n\n${userText}`;
         const [task] = await sql<{ id: string; state: string }[]>`
-          INSERT INTO tasks (project_id, input, agent, model, mode_id, thinking_option_id, session_id)
+          INSERT INTO tasks (project_id, input, agent, model, mode_id, thinking_option_id, session_id, chat_id)
-          VALUES (${sessionRows[0]!.project_id}, ${taskInput}, ${provider}, ${model ?? null}, ${mode_id ?? null}, ${thinking_option_id ?? null}, ${sessionId})
+          VALUES (${sessionRows[0]!.project_id}, ${taskInput}, ${provider}, ${model ?? null}, ${mode_id ?? null}, ${thinking_option_id ?? null}, ${sessionId}, ${chatId})
           RETURNING id, state
         `;
         await sql`UPDATE chats SET updated_at = clock_timestamp() WHERE id = ${chatId}`;

apps/coder/src/routes/worktree-safety.ts

@@ -0,0 +1,45 @@
+/**
+ * Session-delete work-loss guard (coder side).
+ *
+ * Session delete itself lives in apps/server (Docker), which CANNOT see the
+ * host worktree dirs (/tmp/booworktrees) or run git on them. Only BooCoder
+ * (host systemd) can. So the server's DELETE route calls these endpoints
+ * pre-delete to learn whether a session's worktree holds work at risk, and to
+ * stash it. The server owns the gate; coder owns the git truth.
+ */
+import type { FastifyInstance } from 'fastify';
+import type { Sql } from '../db.js';
+import { checkWorktreeWorkAtRisk, stashWorktree } from '../services/worktrees.js';
+
+export function registerWorktreeSafetyRoutes(app: FastifyInstance, sql: Sql): void {
+  // GET risk for a session's worktree(s). One row per session today (PK on
+  // session_id); the loop already handles the Phase-1.5 multi-worktree case.
+  app.get<{ Params: { sessionId: string } }>(
+    '/api/sessions/:sessionId/worktree-risk',
+    async (req) => {
+      const rows = await sql<{ worktree_path: string }[]>`
+        SELECT path AS worktree_path FROM worktrees WHERE session_id = ${req.params.sessionId}
+      `;
+      const reports = [];
+      for (const row of rows) {
+        reports.push(await checkWorktreeWorkAtRisk(row.worktree_path));
+      }
+      return { reports };
+    },
+  );
+
+  // Stash a session's worktree(s) — clears the dirty risk; recoverable.
+  app.post<{ Params: { sessionId: string } }>(
+    '/api/sessions/:sessionId/worktree-stash',
+    async (req) => {
+      const rows = await sql<{ worktree_path: string }[]>`
+        SELECT path AS worktree_path FROM worktrees WHERE session_id = ${req.params.sessionId}
+      `;
+      const results = [];
+      for (const row of rows) {
+        results.push({ worktreePath: row.worktree_path, ...(await stashWorktree(row.worktree_path)) });
+      }
+      return { results };
+    },
+  );
+}

apps/coder/src/schema.sql

@@ -83,16 +83,20 @@ CREATE TABLE IF NOT EXISTS session_worktrees (
   base_commit TEXT,
   created_at TIMESTAMPTZ NOT NULL DEFAULT clock_timestamp()
 );
--- Migrate existing FK to CASCADE (idempotent: drops the old constraint if present).
+-- P1.5-b: DEFANG the CASCADE — a session delete must no longer wipe its worktree
+-- row. This table is SUPERSEDED by `worktrees` below; all readers are repointed
+-- this phase, so the row just persists (dead) on session delete until a later
+-- cleanup drops the table. session_id is this table's PRIMARY KEY, so it cannot be
+-- nullable → SET NULL is invalid and NO ACTION/RESTRICT would block deletes; the
+-- only valid defang is to drop the FK with no replacement. Idempotent: only fires
+-- while the FK is still ON DELETE CASCADE ('c').
 DO $$ BEGIN
   IF EXISTS (
     SELECT 1 FROM pg_constraint
     WHERE conname = 'session_worktrees_session_id_fkey'
-      AND confdeltype <> 'c'
+      AND confdeltype = 'c'
   ) THEN
     ALTER TABLE session_worktrees DROP CONSTRAINT session_worktrees_session_id_fkey;
-    ALTER TABLE session_worktrees ADD CONSTRAINT session_worktrees_session_id_fkey
-      FOREIGN KEY (session_id) REFERENCES sessions(id) ON DELETE CASCADE;
   END IF;
 END $$;
 
@@ -127,6 +131,101 @@ END $$;
 -- v2.6: config fingerprint for stale-session detection (auto-recover on model change).
 ALTER TABLE agent_sessions ADD COLUMN IF NOT EXISTS config_hash TEXT;
 
+-- ─── P1.5-b (corrected): worktrees entity + re-key agent_sessions to (chat_id, agent) ───
+-- The TAB (a chat) is the context unit: two opencode tabs in one session = two
+-- independent contexts sharing one worktree. So agent_sessions keys on
+-- (chat_id, agent), NOT (worktree_id, agent) or (session_id, agent). The
+-- `worktrees` table is one-per-session (selectable later) and only referenced
+-- informationally by agent_sessions.worktree_id (SET NULL); chat_id is the key.
+--
+-- PREREQUISITE: the unmigratable test session (35 chats, 1 agent_sessions row that
+-- maps to no single chat) is DELETED before this runs, so agent_sessions is empty
+-- and the chat_id backfill is N/A. If a row with NULL chat_id remains, the verify
+-- gate below RAISEs and aborts — delete the offending session first.
+
+-- worktree as a first-class entity; survives session delete (session_id SET NULL).
+CREATE TABLE IF NOT EXISTS worktrees (
+  id          UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+  session_id  UUID REFERENCES sessions(id) ON DELETE SET NULL,
+  project_id  UUID,
+  path        TEXT NOT NULL,
+  branch      TEXT,
+  base_commit TEXT,
+  slug        TEXT,
+  status      TEXT NOT NULL DEFAULT 'active' CHECK (status IN ('active','archived')),
+  created_at  TIMESTAMPTZ NOT NULL DEFAULT clock_timestamp()
+);
+CREATE UNIQUE INDEX IF NOT EXISTS worktrees_active_path_uidx ON worktrees(path) WHERE status='active';
+
+-- Migrate any surviving session_worktrees rows → worktrees (idempotent; 0 rows
+-- after the test-session delete, kept for generality / fresh-DB safety).
+INSERT INTO worktrees (session_id, path, branch, base_commit, status)
+SELECT sw.session_id, sw.worktree_path, 'session-' || sw.session_id, sw.base_commit, 'active'
+FROM session_worktrees sw
+WHERE NOT EXISTS (SELECT 1 FROM worktrees w WHERE w.session_id = sw.session_id AND w.status='active');
+
+-- Dispatch hint: which chat (tab) a task belongs to. The coder message route and
+-- skills route set it from the frontend tab; session-less creators (arena, MCP,
+-- new_task, generic /api/tasks) leave it NULL and the dispatcher creates a chat.
+ALTER TABLE tasks ADD COLUMN IF NOT EXISTS chat_id UUID REFERENCES chats(id) ON DELETE SET NULL;
+
+-- Re-key columns on agent_sessions.
+ALTER TABLE agent_sessions ADD COLUMN IF NOT EXISTS chat_id UUID;
+ALTER TABLE agent_sessions ADD COLUMN IF NOT EXISTS worktree_id UUID;
+
+-- BACKFILL-VERIFY GATE: the new PK is (chat_id, agent), so chat_id must be
+-- non-null on every row before the swap. With the test session deleted this is a
+-- 0-row assertion; if any row has NULL chat_id (an unmigratable pre-existing row),
+-- abort loudly rather than create a degenerate (NULL, agent) key.
+DO $$
+DECLARE n int;
+BEGIN
+  SELECT count(*) INTO n FROM agent_sessions WHERE chat_id IS NULL;
+  IF n > 0 THEN
+    RAISE EXCEPTION 'P1.5-b: % agent_sessions row(s) have NULL chat_id — delete the unmigratable session(s) before applying', n;
+  END IF;
+END $$;
+
+-- Swap PK (session_id,agent) → (chat_id,agent) + FKs (run-once, guarded on the new
+-- FK's absence). chat_id CASCADEs from chats (closing a tab ends its context);
+-- worktree_id is informational SET NULL; session_id defanged to nullable SET NULL.
+DO $$ BEGIN
+  IF NOT EXISTS (SELECT 1 FROM pg_constraint WHERE conname = 'agent_sessions_chat_id_fkey') THEN
+    ALTER TABLE agent_sessions DROP CONSTRAINT IF EXISTS agent_sessions_pkey;
+    ALTER TABLE agent_sessions DROP CONSTRAINT IF EXISTS agent_sessions_session_id_fkey;
+    ALTER TABLE agent_sessions ALTER COLUMN session_id DROP NOT NULL;
+    ALTER TABLE agent_sessions ALTER COLUMN chat_id SET NOT NULL;
+    ALTER TABLE agent_sessions ADD CONSTRAINT agent_sessions_pkey PRIMARY KEY (chat_id, agent);
+    ALTER TABLE agent_sessions ADD CONSTRAINT agent_sessions_chat_id_fkey
+      FOREIGN KEY (chat_id) REFERENCES chats(id) ON DELETE CASCADE;
+    ALTER TABLE agent_sessions ADD CONSTRAINT agent_sessions_session_id_fkey
+      FOREIGN KEY (session_id) REFERENCES sessions(id) ON DELETE SET NULL;
+    ALTER TABLE agent_sessions ADD CONSTRAINT agent_sessions_worktree_id_fkey
+      FOREIGN KEY (worktree_id) REFERENCES worktrees(id) ON DELETE SET NULL;
+  END IF;
+END $$;
+
+-- P1.5-b follow-up: converge agent_sessions.session_id FK CASCADE → SET NULL.
+-- The re-key block above re-adds session_id_fkey as SET NULL, but it is guarded on
+-- chat_id_fkey's ABSENCE — so a DB already re-keyed to (chat_id, agent) while
+-- session_id_fkey was still ON DELETE CASCADE never re-enters that block and stays
+-- 'c'. This standalone guard flips it to SET NULL ('n'), matching worktree_id.
+-- Idempotent (mirrors the session_worktrees defang's confdeltype check): only fires
+-- while the FK is still CASCADE — a no-op on a fresh deploy (already 'n' from the
+-- re-key block) and on every re-run thereafter.
+DO $$ BEGIN
+  IF EXISTS (
+    SELECT 1 FROM pg_constraint
+    WHERE conname = 'agent_sessions_session_id_fkey'
+      AND confdeltype = 'c'
+  ) THEN
+    ALTER TABLE agent_sessions ALTER COLUMN session_id DROP NOT NULL;
+    ALTER TABLE agent_sessions DROP CONSTRAINT agent_sessions_session_id_fkey;
+    ALTER TABLE agent_sessions ADD CONSTRAINT agent_sessions_session_id_fkey
+      FOREIGN KEY (session_id) REFERENCES sessions(id) ON DELETE SET NULL;
+  END IF;
+END $$;
+
 -- v2.6: attribution for DiffPanel badges (Phase 1 UX reads this).
 ALTER TABLE pending_changes ADD COLUMN IF NOT EXISTS agent TEXT;
 

apps/coder/src/services/__tests__/dcp-strip.test.ts

@@ -0,0 +1,73 @@
+import { describe, it, expect } from 'vitest';
+import { stripDcpTags, makeDcpStreamStripper } from '../dcp-strip.js';
+
+// Feed chunks through a fresh stripper and return the fully reassembled output
+// (everything emitted during streaming + the final flush) — i.e. what the
+// dispatcher would accumulate into the persisted message content.
+function run(chunks: string[]): string {
+  const s = makeDcpStreamStripper();
+  let out = '';
+  for (const c of chunks) out += s.push(c);
+  out += s.flush();
+  return out;
+}
+
+describe('stripDcpTags (one-shot)', () => {
+  it('removes a complete tag', () => {
+    expect(stripDcpTags('Yes — "Test".\n\n<dcp-message-id>m0019</dcp-message-id>')).toBe(
+      'Yes — "Test".\n\n',
+    );
+  });
+  it('leaves text without a tag untouched', () => {
+    expect(stripDcpTags('no tag here')).toBe('no tag here');
+  });
+});
+
+describe('per-chunk strip is INSUFFICIENT (documents the bug)', () => {
+  it('a tag split across chunks survives a naive per-chunk .replace()', () => {
+    const chunks = ['Yes.\n\n<dcp', '-message', '-id>m0019</dcp', '-message-id>'];
+    const naive = chunks.map(stripDcpTags).join('');
+    // The reassembled content still contains the tag — this is the screenshot bug.
+    expect(naive).toContain('<dcp-message-id>m0019</dcp-message-id>');
+  });
+});
+
+describe('makeDcpStreamStripper (cross-chunk fix)', () => {
+  it('strips a tag split across chunks (the real opencode case)', () => {
+    expect(run(['Yes.\n\n<dcp', '-message', '-id>m0019</dcp', '-message-id>'])).toBe('Yes.\n\n');
+  });
+
+  it('strips a tag split at EVERY character boundary', () => {
+    const full = 'Answer.<dcp-message-id>m0019</dcp-message-id>';
+    expect(run([...full])).toBe('Answer.');
+  });
+
+  it('strips a tag delivered whole in one chunk', () => {
+    expect(run(['Answer.<dcp-message-id>m0019</dcp-message-id>'])).toBe('Answer.');
+  });
+
+  it('passes through text with no tag', () => {
+    expect(run(['hello ', 'world'])).toBe('hello world');
+  });
+
+  it('does NOT swallow legitimate < content (code/HTML/generics)', () => {
+    expect(run(['use ', '<div>', ' and ', 'Array<', 'string>'])).toBe('use <div> and Array<string>');
+  });
+
+  it('handles a lone < that is not a dcp tag, split across chunks', () => {
+    expect(run(['a <', 'b c'])).toBe('a <b c');
+  });
+
+  it('emits surrounding text and strips a mid-text tag', () => {
+    expect(run(['before ', '<dcp-message-id>', 'm1', '</dcp-message-id>', ' after'])).toBe(
+      'before  after',
+    );
+  });
+
+  it('flushes a truncated/never-closed partial tag without leaking it as a complete tag', () => {
+    // If the stream ends mid-tag, flush strips complete tags; an incomplete
+    // remnant is returned as-is (no complete tag ever existed to render).
+    const out = run(['done.<dcp-message-id>m00']);
+    expect(out).not.toContain('</dcp-message-id>');
+  });
+});

apps/coder/src/services/agent-backend.ts

@@ -37,8 +37,15 @@ export interface EnsureSessionOpts {
   agent: string;
   /** Resolved model id. */
   model: string;
+  /** P1.5-b: the chat (tab) this turn belongs to. agent_sessions is keyed
+   *  (chat_id, agent) — the tab/chat is the context unit. Always non-null:
+   *  the dispatcher creates a chat for session-less tasks before calling. */
+  chatId: string;
   /** Shared per-session worktree (one per `sessions.id`, not per pane). */
   worktreePath: string;
+  /** P1.5-b: the `worktrees.id` for this session's worktree — stored on the
+   *  agent_sessions row informationally (NOT the key). */
+  worktreeId: string;
   projectId: string;
 }
 
@@ -47,6 +54,10 @@ export interface AgentSessionHandle {
   sessionId: string;
   agent: string;
   backend: AgentBackendKind;
+  /** P1.5-b: the chat (tab) this session is keyed on (with agent). */
+  chatId: string;
+  /** P1.5-b: the worktree this session's chat runs in (informational link). */
+  worktreeId: string;
   /** Provider's own session id (resume token); null until the backend assigns one. */
   agentSessionId: string | null;
   /** opencode HTTP server port; null for ACP backends. */

apps/coder/src/services/backends/opencode-server.ts

@@ -3,7 +3,9 @@
  *
  * Warm, multi-turn backend for the `opencode` agent. One `opencode serve` HTTP
  * server per BooCoder process; one opencode session per BooCode session (resumed
- * on switch-back); a single SSE read loop demuxes all sessions' events.
+ * on switch-back); one SSE read loop PER session, each scoped to that session's
+ * worktree directory so sessions in different directories stream concurrently
+ * (P1.5-a — replaced the Phase-1 single-stream-last-directory model).
  *
  * Implements the Phase 0 `AgentBackend` interface. Emits transport-agnostic
  * `AgentEvent`s — the dispatcher (Phase 1.7, NOT wired in this batch) maps them
@@ -73,6 +75,9 @@ interface SessionState {
   activeTurn: TurnState | null;
   /** Inactivity backstop timer for the active turn; null when no turn in flight. */
   watchdog: ReturnType<typeof setTimeout> | null;
+  /** Per-session SSE subscription handle. Non-null while the loop is running;
+   *  aborting it tears down the underlying fetch and exits the loop. */
+  sseAbort: AbortController | null;
 }
 
 export interface OpenCodeServerBackendDeps {
@@ -94,7 +99,6 @@ export class OpenCodeServerBackend implements AgentBackend {
   private port: number | null = null;
   private up = false;
   private serverStarting: Promise<void> | null = null;
-  private sseRunning = false;
 
   /** opencode session id → demux state. Maintained by ensureSession; read by the SSE loop. */
   private readonly byOpencodeId = new Map<string, SessionState>();
@@ -150,37 +154,58 @@ export class OpenCodeServerBackend implements AgentBackend {
 
   // ─── SSE read loop + demux + translate (1.3) + dedup (1.4) ───────────────────
 
-  /** Per-directory SSE subscription. opencode scopes events by directory (defaults
+  /** Per-session SSE subscription, scoped to the session's worktree directory.
-   *  to process.cwd if omitted) — so we must subscribe with the same directory used
+   *  opencode scopes events by the `directory` query param (defaults to the
-   *  to create the session. Called from ensureSession; reconnects while up. */
+   *  server's cwd if omitted), so two sessions in different worktrees each get
-  private startEventLoop(directory: string): void {
+   *  their own dir-scoped stream and never drop each other's events. Idempotent:
-    if (this.sseRunning) return;
+   *  a no-op if this session's loop is already running. Started from ensureSession
-    this.sseRunning = true;
+   *  (and defensively from prompt) once worktreePath is known. */
-    this.sseDirectory = directory;
+  private startSessionEventLoop(state: SessionState): void {
-    void this.runEventLoop(directory);
+    if (state.sseAbort) return; // already running
+    const abort = new AbortController();
+    state.sseAbort = abort;
+    void this.runSessionEventLoop(state, abort).finally(() => {
+      // Only clear if this controller is still the live one (a later restart may
+      // have already installed a new one).
+      if (state.sseAbort === abort) state.sseAbort = null;
+    });
   }
 
-  private sseDirectory: string | null = null;
+  private async runSessionEventLoop(state: SessionState, abort: AbortController): Promise<void> {
-
+    const signal = abort.signal;
-  private async runEventLoop(directory: string): Promise<void> {
+    while (this.up && this.client && !signal.aborted) {
-    while (this.up && this.client) {
       try {
-        const sub = await this.client.event.subscribe({ directory });
+        // Re-read worktreePath each (re)subscribe so a directory refresh is picked
+        // up on reconnect. Passing `signal` lets close/dispose tear down a stream
+        // that's parked in `for await` between events.
+        const sub = await this.client.event.subscribe(
+          { directory: state.worktreePath },
+          { signal },
+        );
         for await (const ev of sub.stream) {
+          if (signal.aborted) break;
+          // Dir-scoped streams should only carry this session's events, but two
+          // sessions sharing a worktree (possible post-P1.5-b) each receive BOTH
+          // sessions' events — so drop anything that isn't ours, else the other
+          // session's deltas get processed twice (once per loop).
+          const sid = eventSessionId(ev);
+          if (sid != null && sid !== state.agentSessionId) continue;
           this.dispatchEvent(ev);
         }
-        if (this.up) {
+        if (this.up && !signal.aborted) {
-          await this.reconcileInFlight();
+          await this.reconcile(state); // recover an idle/error lost during the gap
           await sleep(SSE_RECONNECT_DELAY_MS);
         }
       } catch (err) {
-        if (!this.up) break;
+        if (!this.up || signal.aborted) break;
-        this.log.warn({ err: errMsg(err) }, 'opencode-server: event loop error; reconnecting');
+        this.log.warn(
-        await this.reconcileInFlight();
+          { err: errMsg(err), agentSessionId: state.agentSessionId },
+          'opencode-server: session event loop error; reconnecting',
+        );
+        await this.reconcile(state);
         await sleep(SSE_RECONNECT_DELAY_MS);
       }
     }
-    this.sseRunning = false;
   }
 
   /** Demux one event to the owning session's active turn. Unknown/between-turns → drop. */
@@ -354,13 +379,6 @@ export class OpenCodeServerBackend implements AgentBackend {
     }
   }
 
-  /** Reconcile every in-flight turn against the server (called after an SSE drop). */
-  private async reconcileInFlight(): Promise<void> {
-    const states = [...this.byOpencodeId.values()].filter((s) => s.activeTurn);
-    if (states.length === 0) return;
-    await Promise.allSettled(states.map((s) => this.reconcile(s)));
-  }
-
   /**
    * Ask the server whether this session's turn already finished — recovers a
    * session.idle/error lost during an SSE gap. Returns true if it settled the turn.
@@ -405,9 +423,12 @@ export class OpenCodeServerBackend implements AgentBackend {
     if (!this.client) throw new Error('opencode-server: client not ready after ensureServer');
 
     const configHash = sessionConfigHash(opts.model);
+    // P1.5-b: agent_sessions is keyed (chat_id, agent) — the tab/chat is the
+    // context unit (two tabs in one session = two contexts sharing one worktree).
+    // session_id + worktree_id are retained as informational (SET NULL) columns.
     const [row] = await this.sql<{ agent_session_id: string | null; status: string; config_hash: string | null }[]>`
       SELECT agent_session_id, status, config_hash FROM agent_sessions
-      WHERE session_id = ${sessionId} AND agent = ${opts.agent}
+      WHERE chat_id = ${opts.chatId} AND agent = ${opts.agent}
     `;
     let agentSessionId = row?.agent_session_id ?? null;
 
@@ -429,10 +450,12 @@ export class OpenCodeServerBackend implements AgentBackend {
       agentSessionId = created.data.id;
       await this.sql`
         INSERT INTO agent_sessions
-          (session_id, agent, backend, agent_session_id, server_port, status, last_active_at, config_hash)
+          (chat_id, session_id, worktree_id, agent, backend, agent_session_id, server_port, status, last_active_at, config_hash)
         VALUES
-          (${sessionId}, ${opts.agent}, 'opencode_server', ${agentSessionId}, ${this.port}, 'active', clock_timestamp(), ${configHash})
+          (${opts.chatId}, ${sessionId}, ${opts.worktreeId}, ${opts.agent}, 'opencode_server', ${agentSessionId}, ${this.port}, 'active', clock_timestamp(), ${configHash})
-        ON CONFLICT (session_id, agent) DO UPDATE SET
+        ON CONFLICT (chat_id, agent) DO UPDATE SET
+          session_id = EXCLUDED.session_id,
+          worktree_id = EXCLUDED.worktree_id,
           backend = 'opencode_server',
           agent_session_id = EXCLUDED.agent_session_id,
           server_port = EXCLUDED.server_port,
@@ -444,42 +467,21 @@ export class OpenCodeServerBackend implements AgentBackend {
       await this.sql`
         UPDATE agent_sessions
         SET status = 'active', last_active_at = clock_timestamp(), server_port = ${this.port}, config_hash = ${configHash}
-        WHERE session_id = ${sessionId} AND agent = ${opts.agent}
+        WHERE chat_id = ${opts.chatId} AND agent = ${opts.agent}
       `;
     }
 
     // Both branches above guarantee agentSessionId is non-null.
     const ocSessionId = agentSessionId!;
 
-    // Start (or re-start) the SSE event loop scoped to this session's directory.
-    // opencode scopes events by the `directory` query param; without it events
-    // default to the server's CWD which doesn't match our worktree paths.
-    //
-    // KNOWN Phase 1 LIMITATION: one SSE stream at a time, scoped to a single
-    // directory. Under 1.9 concurrency, if two opencode sessions use different
-    // worktree directories simultaneously, re-subscribing for the second drops
-    // the first session's events (the watchdog backstop prevents a full hang,
-    // but streamed content is lost). Phase 2 should move to per-session SSE
-    // subscriptions or a directory-agnostic event path.
-    if (!this.sseRunning || this.sseDirectory !== opts.worktreePath) {
-      if (this.sseRunning && this.sseDirectory && this.sseDirectory !== opts.worktreePath) {
-        this.log.warn(
-          { prev: this.sseDirectory, next: opts.worktreePath },
-          'opencode-server: SSE directory changed — concurrent sessions will lose events from the previous directory',
-        );
-      }
-      this.sseRunning = false;
-      this.startEventLoop(opts.worktreePath);
-    }
-
     // Register / refresh the demux entry the SSE loop keys on. Preserve an existing
     // entry (and any in-flight turn) — just refresh the routing fields.
-    const existing = this.byOpencodeId.get(ocSessionId);
+    let state = this.byOpencodeId.get(ocSessionId);
-    if (existing) {
+    if (state) {
-      existing.boocodeSessionId = sessionId;
+      state.boocodeSessionId = sessionId;
-      existing.worktreePath = opts.worktreePath;
+      state.worktreePath = opts.worktreePath;
     } else {
-      this.byOpencodeId.set(ocSessionId, {
+      state = {
         boocodeSessionId: sessionId,
         agentSessionId: ocSessionId,
         worktreePath: opts.worktreePath,
@@ -487,13 +489,22 @@ export class OpenCodeServerBackend implements AgentBackend {
         partTypeById: new Map(),
         activeTurn: null,
         watchdog: null,
-      });
+        sseAbort: null,
+      };
+      this.byOpencodeId.set(ocSessionId, state);
     }
 
+    // Start this session's own SSE loop, scoped to its worktree directory. Both
+    // fresh-create and resume reach here; idempotent, so a re-ensure (e.g. a
+    // second turn) won't spawn a duplicate loop.
+    this.startSessionEventLoop(state);
+
     return {
       sessionId,
       agent: opts.agent,
       backend: 'opencode_server',
+      chatId: opts.chatId,
+      worktreeId: opts.worktreeId,
       agentSessionId: ocSessionId,
       serverPort: this.port,
     };
@@ -516,12 +527,17 @@ export class OpenCodeServerBackend implements AgentBackend {
         partTypeById: new Map(),
         activeTurn: null,
         watchdog: null,
+        sseAbort: null,
       };
       this.byOpencodeId.set(oc, state);
     }
     const session = state;
     // Authoritative per-turn directory for SDK routing + reconcile.
     session.worktreePath = ctx.worktreePath;
+    // Defensive: ensureSession normally starts the loop, but if prompt is reached
+    // with a freshly-created state (no loop yet), start it so the turn streams.
+    // Idempotent when ensureSession already started one.
+    this.startSessionEventLoop(session);
     const client = this.client;
 
     return await new Promise<TurnResult>((resolve) => {
@@ -577,15 +593,21 @@ export class OpenCodeServerBackend implements AgentBackend {
   // ─── teardown ────────────────────────────────────────────────────────────────
 
   async closeSession(handle: AgentSessionHandle): Promise<void> {
-    if (handle.agentSessionId) this.byOpencodeId.delete(handle.agentSessionId);
+    if (handle.agentSessionId) {
+      // Stop this session's SSE loop before dropping its demux entry.
+      this.byOpencodeId.get(handle.agentSessionId)?.sseAbort?.abort();
+      this.byOpencodeId.delete(handle.agentSessionId);
+    }
     await this.sql`
       UPDATE agent_sessions SET status = 'closed'
-      WHERE session_id = ${handle.sessionId} AND agent = ${handle.agent}
+      WHERE chat_id = ${handle.chatId} AND agent = ${handle.agent}
     `.catch(() => {});
   }
 
   async dispose(): Promise<void> {
     this.up = false;
+    // Abort every per-session SSE loop so none survive the teardown.
+    for (const st of this.byOpencodeId.values()) st.sseAbort?.abort();
     const child = this.child;
     this.child = null;
     this.client = null;
@@ -602,6 +624,20 @@ export class OpenCodeServerBackend implements AgentBackend {
 
 // ─── helpers ──────────────────────────────────────────────────────────────────
 
+/** Extract the opencode sessionID an event belongs to, across event shapes.
+ *  Most carry `properties.sessionID`; `message.part.updated` nests it under
+ *  `properties.part.sessionID`. Returns null when the event has no session
+ *  (the per-session loop then leaves it to dispatchEvent, which drops it). */
+function eventSessionId(ev: Event): string | null {
+  const props = (ev as { properties?: unknown }).properties;
+  if (!props || typeof props !== 'object') return null;
+  if (ev.type === 'message.part.updated') {
+    const part = (props as { part?: { sessionID?: string } }).part;
+    return part?.sessionID ?? null;
+  }
+  return (props as { sessionID?: string }).sessionID ?? null;
+}
+
 /** BooCoder model string "provider/model" → opencode's structured {providerID, modelID}. */
 function parseModel(model: string | undefined): { providerID: string; modelID: string } | undefined {
   if (!model || !model.trim()) return undefined;

apps/coder/src/services/dcp-strip.ts

@@ -0,0 +1,77 @@
+/**
+ * Strip opencode-dcp plugin tags (`<dcp-message-id>mNNNN</dcp-message-id>`) that
+ * the @tarquinen/opencode-dcp plugin appends to assistant text and which
+ * otherwise render as literal text in the UI.
+ *
+ * Why a streaming stripper and not a per-chunk `.replace()`: opencode streams
+ * assistant text token-by-token, so the tag arrives SPLIT across many SSE deltas
+ * (`<dcp`, `-message`, `-id>`, `m0019`, `</dcp`, …). A per-chunk regex never sees
+ * a complete tag in any single fragment, so the fragments pass through and the
+ * dispatcher reassembles the full tag in the persisted/displayed content. The
+ * stripper below buffers across chunks: it emits everything that cannot be part
+ * of a forming tag and holds back only a trailing partial-tag prefix until the
+ * next chunk resolves it — without holding back legitimate `<…>` content.
+ */
+
+const DCP_TAG_RE = /<dcp-message-id>[^<]*<\/dcp-message-id>/g;
+const OPEN = '<dcp-message-id>';
+const CLOSE = '</dcp-message-id>';
+
+/** One-shot strip of COMPLETE tags. Safe for non-streaming / final content. */
+export function stripDcpTags(s: string): string {
+  return s.replace(DCP_TAG_RE, '');
+}
+
+/**
+ * Could `tail` (a substring starting at a `<`) still grow into a complete dcp
+ * tag on a future chunk? If so the caller must hold it back rather than emit it.
+ * Returns false for unrelated `<` content (`<div>`, `<T>`, …) so those stream
+ * normally.
+ */
+function isPartialDcp(tail: string): boolean {
+  // A prefix of the opening marker: '<', '<d', …, '<dcp-message-id'.
+  if (OPEN.startsWith(tail)) return true;
+  // Opening marker fully seen — content (and maybe a forming close) still streaming.
+  if (tail.startsWith(OPEN)) {
+    const rest = tail.slice(OPEN.length);
+    const lt = rest.indexOf('<');
+    if (lt === -1) return true; // still inside the [^<]* content run
+    return CLOSE.startsWith(rest.slice(lt)); // a partial close marker forming
+  }
+  return false;
+}
+
+export interface DcpStreamStripper {
+  /** Feed one text chunk; returns the portion safe to emit now (may be ''). */
+  push(chunk: string): string;
+  /** Stream end: returns whatever was held back, with complete tags stripped. */
+  flush(): string;
+}
+
+/** Stateful, cross-chunk-safe dcp stripper. One instance per turn. */
+export function makeDcpStreamStripper(): DcpStreamStripper {
+  let buf = '';
+  return {
+    push(chunk: string): string {
+      buf += chunk;
+      buf = buf.replace(DCP_TAG_RE, ''); // drop any now-complete tags
+      // Find the earliest `<` whose suffix is a forming dcp tag; hold from there,
+      // emit everything before it (real text, including unrelated `<…>`).
+      for (let i = buf.indexOf('<'); i !== -1; i = buf.indexOf('<', i + 1)) {
+        if (isPartialDcp(buf.slice(i))) {
+          const emit = buf.slice(0, i);
+          buf = buf.slice(i);
+          return emit;
+        }
+      }
+      const emit = buf;
+      buf = '';
+      return emit;
+    },
+    flush(): string {
+      const out = stripDcpTags(buf);
+      buf = '';
+      return out;
+    },
+  };
+}

apps/coder/src/services/dispatcher.ts

@@ -4,6 +4,7 @@ import type { Broker } from '@boocode/server/broker';
 import type { WsFrame } from '@boocode/server/ws-frames';
 import type { Config } from '../config.js';
 import { createWorktree, diffWorktree, cleanupWorktree, ensureSessionWorktree } from './worktrees.js';
+import { makeDcpStreamStripper } from './dcp-strip.js';
 import { dispatchViaAcp } from './acp-dispatch.js';
 import { getResolvedRegistry } from './provider-config-registry.js';
 import { dispatchViaPty } from './pty-dispatch.js';
@@ -77,8 +78,9 @@ export function createDispatcher(deps: Deps): { start(): void; stop(): Promise<v
         mode_id: string | null;
         thinking_option_id: string | null;
         session_id: string | null;
+        chat_id: string | null;
       }[]>`
-        SELECT id, project_id, input, agent, model, mode_id, thinking_option_id, session_id
+        SELECT id, project_id, input, agent, model, mode_id, thinking_option_id, session_id, chat_id
         FROM tasks
         WHERE state = 'pending'
         ORDER BY created_at
@@ -109,6 +111,7 @@ export function createDispatcher(deps: Deps): { start(): void; stop(): Promise<v
     mode_id: string | null;
     thinking_option_id: string | null;
     session_id: string | null;
+    chat_id: string | null;
   }): Promise<void> {
     const taskId = task.id;
 
@@ -510,6 +513,7 @@ export function createDispatcher(deps: Deps): { start(): void; stop(): Promise<v
       mode_id: string | null;
       thinking_option_id: string | null;
       session_id: string | null;
+      chat_id: string | null;
     },
     installPath: string | null,
   ): Promise<void> {
@@ -542,10 +546,18 @@ export function createDispatcher(deps: Deps): { start(): void; stop(): Promise<v
         WHERE id = ${taskId}
       `;
 
-      // Resolve session + chat (mirrors runExternalAgent).
+      // Resolve session + chat. P1.5-b: the chat (tab) is the context key, so the
+      // chat_id MUST be non-null and stable before ensureSession. The coder message
+      // route + skills route stamp task.chat_id with the frontend tab's chat — use
+      // it directly. Session-less creators (arena, MCP, new_task, generic
+      // /api/tasks) leave it null; fall back to resolving/creating a real chat so
+      // ensureSession never receives a degenerate (null, agent) key.
       let sessionId: string;
       let chatId: string;
-      if (task.session_id) {
+      if (task.chat_id && task.session_id) {
+        sessionId = task.session_id;
+        chatId = task.chat_id;
+      } else if (task.session_id) {
         sessionId = task.session_id;
         const chats = await sql<{ id: string }[]>`
           SELECT id FROM chats WHERE session_id = ${sessionId} AND status = 'open' ORDER BY created_at DESC LIMIT 1
@@ -586,7 +598,7 @@ export function createDispatcher(deps: Deps): { start(): void; stop(): Promise<v
 
       // Persistent, session-keyed worktree (shared across turns; NOT torn down
       // per turn — Phase 3 reaps it). Captures base_commit for a stable diff.
-      const { worktreePath, baseCommit } = await ensureSessionWorktree(sql, projectPath, sessionId, {
+      const { worktreeId, worktreePath, baseCommit } = await ensureSessionWorktree(sql, projectPath, sessionId, {
         signal: ac.signal,
       });
       log.info({ taskId, worktreePath }, 'dispatcher: session worktree ready');
@@ -620,21 +632,30 @@ export function createDispatcher(deps: Deps): { start(): void; stop(): Promise<v
       const textChunks: string[] = [];
       const reasoningChunks: string[] = [];
       const toolSnaps = new Map<string, AcpToolSnapshot>();
+      // opencode's dcp plugin appends <dcp-message-id>…</dcp-message-id> to the
+      // text, streamed split across deltas — a per-chunk regex misses it (see
+      // dcp-strip.ts). Buffer text through a cross-chunk stripper so neither the
+      // live `delta` frames nor the persisted content ever carry the tag.
+      const dcp = makeDcpStreamStripper();
 
       // Map transport-agnostic AgentEvents → the SAME WS frames the ACP path emits.
       // This boundary is where message_id/chat_id get attached (the backend never
       // owns them).
       const onEvent = (e: AgentEvent): void => {
         switch (e.type) {
-          case 'text':
+          case 'text': {
-            textChunks.push(e.text);
+            const safe = dcp.push(e.text);
+            if (safe) {
+              textChunks.push(safe);
               broker.publishFrame(sessionId, {
                 type: 'delta',
                 message_id: assistantId,
                 chat_id: chatId,
-              content: e.text,
+                content: safe,
               } as WsFrame);
+            }
             break;
+          }
           case 'reasoning':
             reasoningChunks.push(e.text);
             broker.publishFrame(sessionId, {
@@ -670,7 +691,9 @@ export function createDispatcher(deps: Deps): { start(): void; stop(): Promise<v
       const handle = await backend.ensureSession(sessionId, {
         agent,
         model,
+        chatId,
         worktreePath,
+        worktreeId,
         projectId: task.project_id,
       });
       const result = await backend.prompt(handle, task.input, {
@@ -680,6 +703,18 @@ export function createDispatcher(deps: Deps): { start(): void; stop(): Promise<v
         onEvent,
       });
 
+      // Flush any text held back mid-tag at stream end (complete tags stripped).
+      const dcpTail = dcp.flush();
+      if (dcpTail) {
+        textChunks.push(dcpTail);
+        broker.publishFrame(sessionId, {
+          type: 'delta',
+          message_id: assistantId,
+          chat_id: chatId,
+          content: dcpTail,
+        } as WsFrame);
+      }
+
       const assistantContent = textChunks.join('').slice(0, 50_000);
       const reasoningText = reasoningChunks.join('').slice(0, 200_000);
       const outputSummary = (result.ok ? textChunks.join('') : result.error ?? 'opencode turn failed').slice(0, 500);

apps/coder/src/services/worktrees.ts

@@ -119,16 +119,18 @@ export async function cleanupWorktree(
 // ─── v2.6: session-keyed persistent worktree ────────────────────────────────
 
 export interface SessionWorktree {
+  /** P1.5-b: the `worktrees.id` — stored on agent_sessions informationally. */
+  worktreeId: string;
   worktreePath: string;
   baseCommit: string | null;
 }
 
 /**
- * v2.6: create-or-reuse ONE worktree per BooCode session (shared across all
+ * v2.6 / P1.5-b: create-or-reuse ONE worktree per BooCode session (shared across
- * agents/turns in the session), recorded in `session_worktrees`. Unlike the
+ * all tabs/agents in the session), recorded in `worktrees` (was the superseded
- * per-task `createWorktree`, this persists — it is NOT torn down per turn
+ * `session_worktrees`). Persists — NOT torn down per turn (cleanup is Phase 3) —
- * (cleanup is Phase 3). Captures the project's current HEAD as `base_commit`
+ * and now survives session delete (`worktrees.session_id` is ON DELETE SET NULL).
- * so the accumulating diff has a stable baseline across turns.
+ * Captures the project's current HEAD as `base_commit` for a stable diff baseline.
  *
  * Distinct path namespace (`session-<id>` branch, `/sess-<id>` dir) so it never
  * collides with the per-task worktrees that arena/new_task/MCP still use.
@@ -139,11 +141,13 @@ export async function ensureSessionWorktree(
   sessionId: string,
   opts?: { signal?: AbortSignal },
 ): Promise<SessionWorktree> {
-  const [existing] = await sql<{ worktree_path: string; base_commit: string | null }[]>`
+  const [existing] = await sql<{ id: string; path: string; base_commit: string | null }[]>`
-    SELECT worktree_path, base_commit FROM session_worktrees WHERE session_id = ${sessionId}
+    SELECT id, path, base_commit FROM worktrees
+    WHERE session_id = ${sessionId} AND status = 'active'
+    LIMIT 1
   `;
   if (existing) {
-    return { worktreePath: existing.worktree_path, baseCommit: existing.base_commit };
+    return { worktreeId: existing.id, worktreePath: existing.path, baseCommit: existing.base_commit };
   }
 
   const worktreePath = `${WORKTREE_BASE}/sess-${sessionId}`;
@@ -167,21 +171,191 @@ export async function ensureSessionWorktree(
     throw new Error(`Failed to create session worktree: ${result.stderr.trim() || result.stdout.trim()}`);
   }
 
-  // Persist. ON CONFLICT keeps the first writer's row if two turns race the create.
+  // Insert-or-get: WHERE NOT EXISTS keeps the first writer's row if two turns race
-  await sql`
+  // the create (the partial unique on active path also backstops it).
-    INSERT INTO session_worktrees (session_id, worktree_path, base_commit)
+  const [inserted] = await sql<{ id: string; path: string; base_commit: string | null }[]>`
-    VALUES (${sessionId}, ${worktreePath}, ${baseCommit})
+    INSERT INTO worktrees (session_id, path, branch, base_commit, status)
-    ON CONFLICT (session_id) DO NOTHING
+    SELECT ${sessionId}, ${worktreePath}, ${branchName}, ${baseCommit}, 'active'
+    WHERE NOT EXISTS (
+      SELECT 1 FROM worktrees WHERE session_id = ${sessionId} AND status = 'active'
+    )
+    RETURNING id, path, base_commit
   `;
-  const [row] = await sql<{ worktree_path: string; base_commit: string | null }[]>`
+  if (inserted) {
-    SELECT worktree_path, base_commit FROM session_worktrees WHERE session_id = ${sessionId}
+    return { worktreeId: inserted.id, worktreePath: inserted.path, baseCommit: inserted.base_commit };
+  }
+  // Lost the race — another turn inserted first; read its row.
+  const [row] = await sql<{ id: string; path: string; base_commit: string | null }[]>`
+    SELECT id, path, base_commit FROM worktrees
+    WHERE session_id = ${sessionId} AND status = 'active'
+    LIMIT 1
   `;
   return {
-    worktreePath: row?.worktree_path ?? worktreePath,
+    worktreeId: row!.id,
+    worktreePath: row?.path ?? worktreePath,
     baseCommit: row?.base_commit ?? baseCommit,
   };
 }
 
+// ─── Session-delete work-loss guard ─────────────────────────────────────────
+
+/**
+ * Risk report for a single worktree, returned by checkWorktreeWorkAtRisk.
+ * `atRisk` is the gate the server reads before allowing a session delete.
+ * A git error never silently passes — it forces `atRisk` true and surfaces
+ * the message in `error` (fail-closed).
+ */
+export interface RiskReport {
+  worktreePath: string;
+  branch: string;
+  dirty: boolean;   // uncommitted working-tree changes (incl. untracked)
+  unpushed: number; // commits ahead of upstream, or -1 if no upstream is set
+  unmerged: number; // commits on this branch not in the project default branch
+  atRisk: boolean;  // dirty || unmerged > 0 || (upstream && unpushed > 0) || git error
+  error?: string;   // populated on a git failure; presence forces atRisk
+}
+
+/**
+ * Resolve the project's default branch as a git-usable ref (e.g. "origin/main").
+ *
+ * `refs/remotes/origin/HEAD` lives in the repo's COMMON git dir and is shared
+ * across every linked worktree, so reading it from the session worktree returns
+ * the REMOTE's default branch — never this worktree's own `session-<id>` branch
+ * (that would be `symbolic-ref HEAD`, a different ref). Falls back to probing
+ * common defaults by verified existence when origin/HEAD isn't set (e.g. a repo
+ * that never ran `git remote set-head`). Returns null if none resolve, in which
+ * case the unmerged check is skipped (dirty + unpushed still protect the work).
+ */
+async function detectDefaultBranchRef(
+  worktreePath: string,
+  opts?: { signal?: AbortSignal },
+): Promise<string | null> {
+  const head = await hostExec(
+    `git -C ${shellEscape(worktreePath)} symbolic-ref --short refs/remotes/origin/HEAD`,
+    { signal: opts?.signal, timeoutMs: 10_000 },
+  );
+  if (head.exitCode === 0) {
+    const ref = head.stdout.trim(); // e.g. "origin/main"
+    if (ref) {
+      const verify = await hostExec(
+        `git -C ${shellEscape(worktreePath)} rev-parse --verify --quiet ${shellEscape(ref + '^{commit}')}`,
+        { signal: opts?.signal, timeoutMs: 10_000 },
+      );
+      if (verify.exitCode === 0 && verify.stdout.trim()) return ref;
+    }
+  }
+  // origin/HEAD unset or unresolvable — probe common defaults. Prefer the
+  // remote-tracking ref (always resolvable in a fresh worktree) over the local
+  // head, which may not exist if the default branch lives only in the main tree.
+  for (const cand of ['origin/main', 'origin/master', 'main', 'master']) {
+    const verify = await hostExec(
+      `git -C ${shellEscape(worktreePath)} rev-parse --verify --quiet ${shellEscape(cand + '^{commit}')}`,
+      { signal: opts?.signal, timeoutMs: 10_000 },
+    );
+    if (verify.exitCode === 0 && verify.stdout.trim()) return cand;
+  }
+  return null;
+}
+
+/**
+ * Inspect a worktree for work that would be lost if its session were deleted.
+ * Three checks, all via the audited hostExec + shellEscape path (every
+ * interpolated value — paths, refs — is single-quote-escaped; no bare
+ * interpolation). Any unexpected git failure is treated as at-risk, never a
+ * silent pass.
+ */
+export async function checkWorktreeWorkAtRisk(
+  worktreePath: string,
+  opts?: { signal?: AbortSignal },
+): Promise<RiskReport> {
+  // Branch name — also doubles as the "is this still a git worktree?" probe.
+  const br = await hostExec(
+    `git -C ${shellEscape(worktreePath)} rev-parse --abbrev-ref HEAD`,
+    { signal: opts?.signal, timeoutMs: 10_000 },
+  );
+  if (br.exitCode !== 0) {
+    return {
+      worktreePath,
+      branch: '',
+      dirty: false,
+      unpushed: 0,
+      unmerged: 0,
+      atRisk: true,
+      error: `git rev-parse failed: ${br.stderr.trim() || 'not a git worktree'}`,
+    };
+  }
+  const branch = br.stdout.trim();
+
+  // (a) Uncommitted (dirty working tree, including untracked files).
+  const st = await hostExec(
+    `git -C ${shellEscape(worktreePath)} status --porcelain`,
+    { signal: opts?.signal, timeoutMs: 15_000 },
+  );
+  if (st.exitCode !== 0) {
+    return {
+      worktreePath,
+      branch,
+      dirty: false,
+      unpushed: 0,
+      unmerged: 0,
+      atRisk: true,
+      error: `git status failed: ${st.stderr.trim()}`,
+    };
+  }
+  const dirty = st.stdout.trim().length > 0;
+
+  // (b) Unpushed commits. No upstream configured => work exists only locally;
+  // treat as unpushed-by-definition (-1) rather than an error.
+  const up = await hostExec(
+    `git -C ${shellEscape(worktreePath)} rev-list --count ${shellEscape('@{u}..HEAD')}`,
+    { signal: opts?.signal, timeoutMs: 15_000 },
+  );
+  const unpushed = up.exitCode === 0 ? (parseInt(up.stdout.trim() || '0', 10) || 0) : -1;
+
+  // (c) Unmerged commits — on this branch but not in the project default branch.
+  const defaultRef = await detectDefaultBranchRef(worktreePath, opts);
+  let unmerged = 0;
+  if (defaultRef) {
+    const rl = await hostExec(
+      `git -C ${shellEscape(worktreePath)} rev-list --count ${shellEscape(defaultRef + '..HEAD')}`,
+      { signal: opts?.signal, timeoutMs: 15_000 },
+    );
+    if (rl.exitCode === 0) unmerged = parseInt(rl.stdout.trim() || '0', 10) || 0;
+  }
+
+  // unpushed only contributes when an upstream actually exists. Session branches
+  // (session-<id>) never have one (unpushed === -1), and any real local-only work
+  // there already surfaces as unmerged > 0 — so the no-upstream case adds no
+  // protection, only friction (it flagged every pristine worktree-backed session).
+  // The unpushed > 0 arm stays forward-compatible with P1.5 pushable branches.
+  const hasUpstream = unpushed !== -1;
+  const atRisk = dirty || unmerged > 0 || (hasUpstream && unpushed > 0);
+  return { worktreePath, branch, dirty, unpushed, unmerged, atRisk };
+}
+
+/**
+ * Stash a worktree's uncommitted changes (including untracked, via -u) so the
+ * working tree is clean. Stash entries live in the repo's common git dir, so
+ * they survive worktree-dir removal — this is the recoverable, safe-by-default
+ * escape. Note it only clears the *dirty* risk; unpushed/unmerged commits
+ * remain on the branch, so a re-attempted delete may still block on those.
+ */
+export async function stashWorktree(
+  worktreePath: string,
+  opts?: { signal?: AbortSignal },
+): Promise<{ stashed: boolean; error?: string }> {
+  const r = await hostExec(
+    `git -C ${shellEscape(worktreePath)} stash push -u -m ${shellEscape('boocode: pre-delete stash')}`,
+    { signal: opts?.signal, timeoutMs: 30_000 },
+  );
+  if (r.exitCode !== 0) {
+    return { stashed: false, error: r.stderr.trim() || r.stdout.trim() };
+  }
+  // "No local changes to save" => exit 0, nothing stashed — not an error.
+  const stashed = !/no local changes to save/i.test(r.stdout);
+  return { stashed };
+}
+
 /** Minimal shell escape for paths (single-quote wrapping). */
 function shellEscape(s: string): string {
   // Replace single quotes with escaped version, wrap in single quotes

apps/server/src/routes/sessions.ts

@@ -3,7 +3,7 @@ import { z } from 'zod';
 import type { Sql } from '../db.js';
 import type { Config } from '../config.js';
 import type { Broker } from '../services/broker.js';
-import type { Session } from '../types/api.js';
+import type { Session, WorktreeRiskReport } from '../types/api.js';
 import { getSetting } from './settings.js';
 
 const CreateBody = z.object({
@@ -426,10 +426,55 @@ export function registerSessionRoutes(
     }
   );
 
-  app.delete<{ Params: { id: string } }>(
+  app.delete<{ Params: { id: string }; Querystring: { force?: string } }>(
     '/api/sessions/:id',
     async (req, reply) => {
       const id = req.params.id;
+      const force = req.query.force === 'true' || req.query.force === '1';
+
+      // Session-delete work-loss guard. The check MUST run BEFORE the DELETE:
+      // worktrees.session_id is ON DELETE SET NULL (P1.5-b), so once the session
+      // is gone the worktree rows no longer point back to it — read them while
+      // the link still exists.
+      //
+      // Optimization: read worktrees (P1.5-b — was session_worktrees) from our
+      // own (shared) DB first. No row => chat-only session => nothing on disk =>
+      // delete immediately, zero round-trip. Only worktree-backed sessions pay
+      // the host git check.
+      if (!force) {
+        const worktrees = await sql<{ path: string }[]>`
+          SELECT path FROM worktrees WHERE session_id = ${id}
+        `;
+        if (worktrees.length > 0) {
+          // Worktree dirs live on the host; only BooCoder can run git on them.
+          const origin = process.env.BOOCODER_URL ?? 'http://boocoder:3000';
+          let reports: WorktreeRiskReport[];
+          try {
+            const res = await fetch(`${origin}/api/sessions/${id}/worktree-risk`);
+            if (!res.ok) {
+              // Fail-closed: can't verify => don't risk silent loss. Force escapes.
+              reply.code(409);
+              return {
+                error: 'could not verify worktree safety (BooCoder check failed). Use force to delete anyway.',
+                reports: [] as WorktreeRiskReport[],
+              };
+            }
+            reports = ((await res.json()) as { reports?: WorktreeRiskReport[] }).reports ?? [];
+          } catch {
+            // Fail-closed: BooCoder unreachable. Force bypasses this path entirely.
+            reply.code(409);
+            return {
+              error: 'BooCoder unreachable; cannot verify worktree safety. Use force to delete anyway.',
+              reports: [] as WorktreeRiskReport[],
+            };
+          }
+          if (reports.some((r) => r.atRisk)) {
+            reply.code(409);
+            return { error: 'This session has work at risk in its worktree.', reports };
+          }
+        }
+      }
+
       const deleted = await sql<{ project_id: string }[]>`
         DELETE FROM sessions WHERE id = ${id} RETURNING project_id
       `;

apps/server/src/types/api.ts

@@ -25,6 +25,20 @@ export interface AvailableProject {
 
 export type SessionStatus = 'open' | 'archived';
 
+// Session-delete work-loss guard. Returned (as `reports`) in the 409 body when
+// a delete is blocked because the session's worktree holds work at risk. The
+// shape is produced by BooCoder's checkWorktreeWorkAtRisk and passed through
+// verbatim; mirrored byte-for-byte in apps/web/src/api/types.ts for the dialog.
+export interface WorktreeRiskReport {
+  worktreePath: string;
+  branch: string;
+  dirty: boolean;
+  unpushed: number; // commits ahead of upstream, or -1 if no upstream
+  unmerged: number; // commits not in the project default branch
+  atRisk: boolean;
+  error?: string;
+}
+
 export interface Session {
   id: string;
   project_id: string;

apps/web/src/api/client.ts

@@ -151,8 +151,17 @@ export const api = {
         method: 'PATCH',
         body: JSON.stringify(body),
       }),
-    remove: (id: string) =>
+    // force=true bypasses the server-side worktree work-loss guard. A blocked
-      request<void>(`/api/sessions/${id}`, { method: 'DELETE' }),
+    // delete throws ApiError(409) whose body carries { error, reports }.
+    remove: (id: string, force = false) =>
+      request<void>(`/api/sessions/${id}${force ? '?force=true' : ''}`, { method: 'DELETE' }),
+    // Stash the session's worktree (uncommitted changes) on the host, via the
+    // BooCoder proxy. Recoverable escape from the work-at-risk dialog.
+    worktreeStash: (id: string) =>
+      request<{ results: { worktreePath: string; stashed: boolean; error?: string }[] }>(
+        `/api/coder/sessions/${id}/worktree-stash`,
+        { method: 'POST' },
+      ),
     archive: (id: string) =>
       request<void>(`/api/sessions/${id}/archive`, { method: 'POST' }),
     unarchive: (id: string) =>

apps/web/src/api/types.ts

@@ -34,6 +34,19 @@ export interface AvailableProject {
 
 export type SessionStatus = 'open' | 'archived';
 
+// Session-delete work-loss guard. Mirror of WorktreeRiskReport in
+// apps/server/src/types/api.ts — edit both copies together. Arrives as the
+// `reports` field of the 409 body when a delete is blocked.
+export interface WorktreeRiskReport {
+  worktreePath: string;
+  branch: string;
+  dirty: boolean;
+  unpushed: number; // commits ahead of upstream, or -1 if no upstream
+  unmerged: number; // commits not in the project default branch
+  atRisk: boolean;
+  error?: string;
+}
+
 export interface Session {
   id: string;
   project_id: string;

apps/web/src/components/ProjectSidebar.tsx

@@ -19,12 +19,12 @@ import {
   DialogDescription,
 } from '@/components/ui/dialog';
 import { AddProjectModal } from './AddProjectModal';
-import { api } from '@/api/client';
+import { api, ApiError } from '@/api/client';
 import { useSidebar } from '@/hooks/useSidebar';
 import { useSidebarDrawer } from '@/hooks/useSidebarDrawer';
 import { useViewport } from '@/hooks/useViewport';
 import { usePullToRefresh } from '@/hooks/usePullToRefresh';
-import type { SidebarProject } from '@/api/types';
+import type { SidebarProject, WorktreeRiskReport } from '@/api/types';
 import { giteaUrlFor } from '@/lib/projectUrls';
 import { isCoderSessionName } from '@/lib/coder-session';
 import { cn } from '@/lib/utils';
@@ -110,6 +110,16 @@ export function ProjectSidebar() {
   const [renamingProject, setRenamingProject] = useState<string | null>(null);
   const [renameProjectValue, setRenameProjectValue] = useState('');
   const [archiveProjectConfirm, setArchiveProjectConfirm] = useState<{ id: string; name: string } | null>(null);
+  // Work-at-risk dialog: shown when a delete is blocked (409) because the
+  // session's worktree holds uncommitted/unpushed/unmerged work.
+  const [riskState, setRiskState] = useState<{
+    sessionId: string;
+    projectId: string;
+    name: string;
+    message: string;
+    reports: WorktreeRiskReport[];
+  } | null>(null);
+  const [riskBusy, setRiskBusy] = useState(false);
   const navigate = useNavigate();
   const location = useLocation();
   const lastToastedError = useRef<string | null>(null);
@@ -174,16 +184,81 @@ export function ProjectSidebar() {
     }
   }
 
-  async function handleDeleteSession(sessionId: string, projectId: string) {
+  async function handleDeleteSession(
+    sessionId: string,
+    projectId: string,
+    name: string,
+    force = false,
+  ) {
     try {
-      await api.sessions.remove(sessionId);
+      await api.sessions.remove(sessionId, force);
       // Server publishes session_deleted via WS; useUserEvents delivers it.
+      setRiskState(null);
       if (activeSession === sessionId) navigate(`/project/${projectId}`);
     } catch (err) {
+      // 409 => the server's work-loss guard blocked the delete. Open the
+      // work-at-risk dialog with the per-worktree reports instead of toasting.
+      if (
+        err instanceof ApiError &&
+        err.status === 409 &&
+        err.body && typeof err.body === 'object' && 'reports' in err.body
+      ) {
+        const body = err.body as { error?: string; reports?: WorktreeRiskReport[] };
+        setRiskState({
+          sessionId,
+          projectId,
+          name,
+          message: body.error ?? 'This session has work at risk.',
+          reports: body.reports ?? [],
+        });
+        return;
+      }
       toast.error(err instanceof Error ? err.message : 'failed to delete session');
     }
   }
 
+  // Stash the worktree's uncommitted changes (recoverable), then re-attempt the
+  // delete. If unpushed/unmerged commits remain, the retry 409s again and the
+  // dialog re-renders with the narrowed risk.
+  async function handleStashAndRetry() {
+    if (!riskState || riskBusy) return;
+    setRiskBusy(true);
+    try {
+      const { results } = await api.sessions.worktreeStash(riskState.sessionId);
+      const failed = results.find((r) => r.error);
+      if (failed) {
+        toast.error(`stash failed: ${failed.error}`);
+        return;
+      }
+      await handleDeleteSession(riskState.sessionId, riskState.projectId, riskState.name, false);
+    } catch (err) {
+      toast.error(err instanceof Error ? err.message : 'stash failed');
+    } finally {
+      setRiskBusy(false);
+    }
+  }
+
+  // Explicit, destructive override — deletes despite work at risk.
+  async function handleForceDelete() {
+    if (!riskState || riskBusy) return;
+    setRiskBusy(true);
+    try {
+      await handleDeleteSession(riskState.sessionId, riskState.projectId, riskState.name, true);
+    } finally {
+      setRiskBusy(false);
+    }
+  }
+
+  // Route the user to commit it themselves — never auto-commit. Opens the
+  // session workspace where they can use a terminal or agent pane.
+  function handleGoCommit() {
+    if (!riskState) return;
+    const sessionId = riskState.sessionId;
+    setRiskState(null);
+    navigate(`/session/${sessionId}`);
+    toast.info('Open a terminal or agent in this session, commit and push your work, then delete again.');
+  }
+
   async function handleRenameSession(sessionId: string) {
     const trimmed = renameValue.trim();
     setRenamingSession(null);
@@ -216,6 +291,20 @@ export function ProjectSidebar() {
       )
     : 'w-60 shrink-0 border-r bg-sidebar text-sidebar-foreground flex flex-col h-screen';
 
+  // Work-at-risk dialog framing. The server returns 409 in two distinct
+  // situations: (1) work genuinely at risk (reports has ≥1 atRisk entry), or
+  // (2) it couldn't verify (BooCoder down/errored → reports is empty). These
+  // are different user stories — "your work is in danger" vs "the checker is
+  // offline" — so the dialog must not show one generic message for both.
+  const atRiskReports = riskState?.reports.filter((r) => r.atRisk) ?? [];
+  const verifyFailed = riskState !== null && atRiskReports.length === 0;
+  const anyDirty = atRiskReports.some((r) => r.dirty);
+  // Commit-based risk (unpushed/unmerged) that stash can NOT clear. When this is
+  // all that remains (e.g. after a stash cleared the dirty changes), the dialog
+  // explains why it re-blocked and hides the Stash button so it doesn't look
+  // like stash "didn't work".
+  const anyCommits = atRiskReports.some((r) => r.unpushed !== 0 || r.unmerged > 0);
+
   return (
     <aside className={asideCls}>
       <div className="px-4 py-3 border-b flex items-center justify-between">
@@ -499,7 +588,7 @@ export function ProjectSidebar() {
                   const projectId = projects.find((p) =>
                     p.recent_sessions.some((s) => s.id === deleteConfirm.id)
                   )?.id;
-                  if (projectId) void handleDeleteSession(deleteConfirm.id, projectId);
+                  if (projectId) void handleDeleteSession(deleteConfirm.id, projectId, deleteConfirm.name);
                 }
                 setDeleteConfirm(null);
               }}
@@ -509,6 +598,77 @@ export function ProjectSidebar() {
           </div>
         </DialogContent>
       </Dialog>
+
+      <Dialog open={riskState !== null} onOpenChange={(open) => { if (!open && !riskBusy) setRiskState(null); }}>
+        <DialogContent>
+          <DialogHeader>
+            <DialogTitle>
+              {verifyFailed ? 'Could not verify worktree safety' : 'This session has work at risk'}
+            </DialogTitle>
+            <DialogDescription>
+              {verifyFailed ? (
+                <>
+                  {riskState?.message ?? 'The worktree safety check is unavailable.'} Your work may be
+                  fine, but it couldn&apos;t be checked — only force-delete if you&apos;re sure.
+                </>
+              ) : anyDirty && anyCommits ? (
+                <>
+                  Deleting {riskState ? `"${riskState.name}"` : 'this session'} would orphan uncommitted
+                  changes <em>and</em> commits that aren&apos;t pushed or merged. Stash clears the
+                  changes (recoverable), but the commits will still block — push them or force-delete.
+                </>
+              ) : anyDirty ? (
+                <>
+                  Deleting {riskState ? `"${riskState.name}"` : 'this session'} would orphan uncommitted
+                  changes in its worktree. Stash them (recoverable), commit them, or force-delete.
+                </>
+              ) : (
+                <>
+                  Deleting {riskState ? `"${riskState.name}"` : 'this session'} would orphan commits that
+                  aren&apos;t pushed or merged. Stashing won&apos;t recover these — push them, or
+                  force-delete.
+                </>
+              )}
+            </DialogDescription>
+          </DialogHeader>
+          {!verifyFailed && (
+            <div className="flex flex-col gap-2 py-1 text-sm">
+              {atRiskReports.map((r) => (
+                <div key={r.worktreePath} className="rounded border border-border/60 px-3 py-2">
+                  <div className="font-mono text-xs text-muted-foreground truncate" title={r.worktreePath}>
+                    {r.branch || r.worktreePath}
+                  </div>
+                  <ul className="mt-1 list-disc pl-5 text-foreground/90">
+                    {r.error && <li className="text-destructive">git error: {r.error}</li>}
+                    {r.dirty && <li>uncommitted changes</li>}
+                    {r.unpushed === -1 && <li>local-only branch (no upstream)</li>}
+                    {r.unpushed > 0 && <li>{r.unpushed} unpushed commit{r.unpushed === 1 ? '' : 's'}</li>}
+                    {r.unmerged > 0 && <li>{r.unmerged} unmerged commit{r.unmerged === 1 ? '' : 's'}</li>}
+                  </ul>
+                </div>
+              ))}
+            </div>
+          )}
+          <div className="flex flex-wrap gap-2 justify-end pt-2">
+            <Button variant="outline" disabled={riskBusy} onClick={() => setRiskState(null)}>
+              Cancel
+            </Button>
+            {!verifyFailed && (
+              <Button variant="outline" disabled={riskBusy} onClick={handleGoCommit}>
+                Commit&hellip;
+              </Button>
+            )}
+            {!verifyFailed && anyDirty && (
+              <Button variant="outline" disabled={riskBusy} onClick={() => void handleStashAndRetry()}>
+                Stash &amp; delete
+              </Button>
+            )}
+            <Button variant="destructive" disabled={riskBusy} onClick={() => void handleForceDelete()}>
+              Force delete
+            </Button>
+          </div>
+        </DialogContent>
+      </Dialog>
     </aside>
   );
 }

data/AGENTS.md

@@ -1,5 +1,11 @@
 # Agents
 
+Operating rules for every agent in this registry. Full procedures live in the `committing-changes` and `using-worktrees` skills.
+
+**Committing** — Commit only on Sam's explicit command, never autonomously and never on apply; never `git push` (Sam pushes manually, Gitea + GitHub mirror). Stage by concern (named files or `git add -p`), never `git add -A`; never stage Sam's unrelated work. Identity `indifferentketchup` / `sam@indifferentketchup.com`, never a personal Gmail. Freeform scope-prefix messages, explain *why* for non-obvious changes, no emojis. Full workflow: invoke `committing-changes`.
+
+**Worktrees** — Isolate work in a worktree when it is parallel to in-progress work, risky/experimental, a hotfix interrupting other work, or splits into independent units — just create when clear, propose in one line when ambiguous, skip quick/small single-stream work. Branch from a stable base (default branch); worktrees persist (never auto-remove or auto-merge); they isolate code state, not runtime (ports/DBs/services still collide). Full heuristic: invoke `using-worktrees`.
+
 ## Code Reviewer
 ---
 temperature: 0.6

data/skills/boocode/committing-changes/SKILL.md

@@ -0,0 +1,60 @@
+---
+name: committing-changes
+description: This skill should be used when the user asks to commit, stage, split, or prepare changes for a commit. Examples: "commit this", "stage these", "split this into commits", "help me commit", "prepare a commit", "make a commit for the dcp fix".
+---
+
+# Committing Changes
+
+Segment the working tree by concern, stage explicitly, draft messages, **present the plan, and STOP**. Commit only on the user's explicit command for this turn. Never push — the user pushes manually (Gitea + GitHub mirror).
+
+**The default is to prepare and propose, not to commit.** A request to "commit X" is a request to get X *ready* and show the plan, unless the user has, in this turn, told you to actually run the commit. When in doubt, present and wait.
+
+## Workflow
+
+1. **Inspect.** `git status` then `git diff` (and `git diff --staged` if anything is already staged). Read what actually changed — do not commit from memory of what you wrote.
+2. **Segment by concern.** Group the changes into buckets, one per coherent concern. State the grouping in plain language before staging anything (e.g. "two concerns: (a) the SSE fix in opencode-server.ts, (b) an unrelated typo in README").
+3. **Safety scan.** Before staging, scan the diff for: secrets / keys / tokens, debug code, stray `console.log`/`print`/`dbg!`, commented-out experiments, and edits to files the user did not ask you to touch (their in-progress work). Flag anything found; do not silently stage it.
+4. **Stage explicitly, per bucket.** Stage named files (`git add path/a path/b`) or hunks (`git add -p`). **Never `git add -A`, `git add .`, or `git add -u`** — those sweep up unrelated work. If `-p` can't cleanly split adjacent hunks, hand-edit the patch (`git add -e`) or revert the unrelated hunk in the working tree first.
+5. **Draft messages.** One message per bucket, in the repo's scope-prefix style (see `references/message-style.md`). Explain *why* for anything non-obvious — the diff already shows *what*. Imperative mood. No emojis. Do not impose Conventional-Commits ceremony (type enums, `BREAKING CHANGE:` footers) unless the user asks.
+6. **Present the plan + STOP.** Show: the buckets, the files in each, the drafted message for each, and the current staged state. Then wait. **Do not run `git commit`.**
+7. **On the user's command**, execute the agreed `git add` / `git commit` exactly as presented, using the identity below. Then report the resulting hashes. Still do not push.
+
+## Split heuristic
+
+- **One commit** when the changes are a single coherent concern (a feature + its test; a fix + the comment explaining it).
+- **Multiple commits** when concerns are independently revertable or reviewable — a bug fix and an unrelated refactor that happen to share the working tree should be two commits even if they touch the same file.
+- A migration/schema change and the code that uses it are usually *one* concern (they're not independently revertable). A doc/changelog update alongside code is usually a *separate* concern.
+
+## Identity (always)
+
+Commit as:
+
+```
+user.name  = indifferentketchup
+user.email = sam@indifferentketchup.com
+```
+
+Never use a personal Gmail or the host's default git identity. If unsure the repo config is right, pass it inline: `git -c user.name=indifferentketchup -c user.email=sam@indifferentketchup.com commit -m "..."`.
+
+## DO-NOT
+
+- **Never push.** No `git push` under any circumstances — that is the user's manual step (dual remote: Gitea + GitHub mirror).
+- **Never auto-commit.** Preparing ≠ committing. Commit only when told to, this turn.
+- **Never `git add -A` / `git add .` / `git add -u`.** Stage by name or by hunk.
+- **Never commit the user's unrelated/in-progress files.** If a file changed that the task didn't touch, leave it; surface it.
+- **No emojis** in messages.
+- **No amending or rebasing** published or shared commits without an explicit instruction.
+
+## Red flags — STOP
+
+- About to run `git commit` without having been told to commit this turn → STOP, present the plan instead.
+- About to `git add -A` "to save time" → STOP, stage by concern.
+- About to `git push` "to finish the job" → STOP, that is never part of this skill.
+- A secret or debug line is in the diff and you're staging anyway → STOP, surface it.
+
+## Anti-patterns this skill avoids
+
+- Committing the moment changes look done (the user reviews diffs and commits on command).
+- Collapsing several concerns into one "WIP" commit because staging separately is tedious.
+- Pushing after committing because the work "feels finished."
+- Reformatting the message into strict Conventional Commits when the repo uses freeform scope-prefixes.

data/skills/boocode/committing-changes/eval.yaml

@@ -0,0 +1,31 @@
+skill: committing-changes
+tasks:
+  - prompt: "Commit this for me"
+    grader:
+      - the response invokes the committing-changes skill
+      - the response inspects the working tree (git status / git diff) before staging
+      - the response segments the changes by concern and states the grouping
+      - the response stages explicitly (named files or git add -p), never git add -A / git add . / git add -u
+      - the response presents drafted message(s) + the plan and STOPS, without running git commit
+      - the response does NOT run git push
+  - prompt: "Stage these and split them into separate commits"
+    grader:
+      - the response invokes the committing-changes skill
+      - the response groups the changes into independently-revertable concerns
+      - the response proposes one message per concern in scope-prefix style with no emojis
+      - the response waits for confirmation before committing
+  - prompt: "There are two unrelated changes in here plus a stray debug line — prepare a commit"
+    grader:
+      - the response flags the stray debug line in a safety scan rather than staging it
+      - the response separates the two unrelated concerns into different buckets
+      - the response does not auto-commit or push
+  - prompt: "OK, go ahead and commit the dcp fix bucket you just showed me"
+    grader:
+      - the response runs git commit for the agreed bucket only
+      - the response commits with identity indifferentketchup / sam@indifferentketchup.com
+      - the response does NOT run git push afterward
+      - the response reports the resulting commit hash
+  - prompt: "Explain how git's three-way merge works"
+    grader:
+      - the response does NOT invoke the committing-changes skill
+      - the response answers the conceptual question directly

data/skills/boocode/committing-changes/references/message-style.md

@@ -0,0 +1,43 @@
+# Commit message style
+
+Freeform **scope-prefix** messages. The shape is conventional-commits-*like* — `type(scope): summary` is the dominant form in this repo — but it is **not enforced**: the scope and the *why* matter more than the type enum. Do not reject or rewrite a message just because it lacks a `type`, and do not add ceremony (`BREAKING CHANGE:` footers, rigid type whitelist).
+
+## The pattern
+
+```
+<scope-prefix>: <imperative summary>
+
+<optional body: WHY this change, not what — the diff shows what>
+```
+
+- **Scope prefix** — the area(s) touched. A single area (`coder`, `web`, `server`), a typed scope (`fix(coder)`, `feat(coder)`, `docs(changelog)`), a sub-scope (`coder(providers)`), or multiple areas joined (`web+coder`). Pick whatever names the blast radius honestly.
+- **Imperative summary** — "strip dcp tags", not "stripped" / "strips". One line, no trailing period needed.
+- **Body** — only when the *why* isn't obvious from the summary. Explain the reason, the failure it fixes, or the constraint it satisfies. Cross-reference related tags/commits by name when the change builds on or fixes prior work.
+- **No emojis.** Anywhere — summary, body, or trailers.
+
+## Real examples (from this repo's log)
+
+```
+fix(coder): strip dcp-message-id tags split across stream chunks
+feat(coder): per-session SSE subscriptions (P1.5-a concurrency prereq)
+feat(coder): guard session delete against worktree work loss
+fix(coder): no-upstream branch alone no longer flags a session at-risk
+docs(changelog): v2.6.2-delete-guard-and-sse
+chore(coder): untrack live coder-providers.json, ship example
+```
+
+And the freeform multi-area / sub-scope forms the house style also allows:
+
+```
+web+coder: per-session SSE
+coder(providers): fix empty picker
+```
+
+## Why-not-just-what
+
+A summary that restates the diff (`fix: change variable name`) wastes the message. A good message answers a question the diff can't: *why did this need to change?* Example — the bare summary `fix(coder): no-upstream branch alone no longer flags a session at-risk` is fine, but its body earns its keep:
+
+> Session worktree branches never get an upstream, so the original rule flagged
+> every worktree-backed session as at-risk on delete — even pristine ones.
+
+That sentence is the part a future reader (or `git blame`) actually needs.

data/skills/boocode/using-worktrees/SKILL.md

@@ -0,0 +1,73 @@
+---
+name: using-worktrees
+description: This skill should be used when starting work that may need isolation from the current checkout — parallel to something already in progress, risky or experimental, a hotfix interrupting other work, or a task that splits into independent mergeable units. Also when the user explicitly asks for a worktree. Examples: "try this risky refactor", "I need to fix prod while keeping this branch", "explore an alternate approach", "make a worktree for X".
+---
+
+# Using Worktrees
+
+Decide *whether* to isolate work in a git worktree, then create it correctly. The judgment — "does this need its own worktree?" — is the point of this skill; the mechanics are routine.
+
+**Asymmetry with committing (deliberate):** when the heuristic clearly fires, **just create the worktree** — you have standing trust here. When it's ambiguous, **propose it in one line and wait**. This is unlike committing, which is always command-gated. Creating a worktree is cheap and reversible; making a commit is not, so the trust differs.
+
+## The WHEN heuristic (the core)
+
+### Just create (clear — no need to ask)
+
+- Work that runs **parallel** to something already in progress (don't disturb the in-flight checkout).
+- A **risky / experimental / throwaway** change you might want to discard cleanly.
+- A **hotfix that interrupts** in-progress work (isolate the fix, leave the WIP untouched).
+- Work that **decomposes into independent mergeable units** — one worktree per unit.
+- Any task where the user would plausibly want it isolated from the main checkout.
+
+### Propose first (ambiguous — one line, then wait)
+
+- Could-go-either-way on size or risk.
+- Unsure whether the user wants isolation at all.
+- A worktree that would **overlap heavily** with the work already on the main checkout (isolation buys little, may confuse).
+
+State it in one line: *"This looks risky/parallel — want me to do it in a worktree?"* Then wait.
+
+### Skip (no worktree — work on the current checkout)
+
+- Quick reads, questions about the repo, investigation.
+- Small single-stream fixes with nothing to run in parallel.
+- Anything where there's nothing to isolate and no parallelism to protect.
+
+```
+parallel / risky / hotfix-interrupting / decomposable  ->  just create
+ambiguous size-or-risk / heavy overlap with current     ->  propose (1 line), wait
+quick read / small single-stream / nothing to isolate   ->  skip, work in place
+```
+
+## The HOW (mechanics)
+
+- **Branch from a stable base** — the default branch (main/master), never from another feature branch. A worktree off a half-done branch inherits its instability.
+- **Branch name derived from the task** — `fix-session-delete-guard`, not `wip` or `tmp`. No emojis.
+- **Collision-safe path** — a unique dir outside the main checkout (e.g. a per-task or per-branch path), so two worktrees never share a directory.
+- **Run the project's setup after create** — install deps / env / generate, if the project defines a setup step. A fresh worktree has the code but not the installed/generated state. (Some projects declare setup hooks; run whatever the project defines — don't assume the checkout is ready to run bare.)
+
+## Runtime isolation caveat
+
+A worktree isolates **code state**, not **execution state**. Ports, databases, caches, lockfiles, and running services can still collide between worktrees. Don't assume a worktree means a fully isolated environment — if two worktrees both run the app, give each its own port / DB / service via per-worktree setup. Code isolation ≠ runtime isolation.
+
+## Lifecycle
+
+- Worktrees **persist** — they are not auto-reaped. Leaving one around is fine; it's not litter.
+- **Reconcile via git**, never automatically: review the worktree's diff against its base, then merge or archive on the user's decision. Do not auto-merge.
+- **Commit inside a worktree only on the user's command** — defer to the `committing-changes` skill for the commit step (same rules: present-and-stop, never push).
+
+## DO-NOT
+
+- **Never branch from a non-stable base** (another feature branch). Stable base only.
+- **Never auto-merge or auto-reconcile** a worktree back. That's a reviewed decision.
+- **Never push** (worktrees change nothing about the push rule — that stays the user's manual step).
+- **Never `git worktree remove`** without the user's say. Worktrees persist; removing one can discard uncommitted work.
+- **No emojis** in branch names.
+
+## Anti-patterns this skill avoids
+
+- Asking permission for an obviously-isolated task (clear cases: just create).
+- Creating a worktree for a quick read or a one-line fix (nothing to isolate).
+- Branching the worktree off the messy in-progress branch instead of the stable base.
+- Assuming a worktree gives runtime isolation and then colliding on a port or DB.
+- Auto-removing or auto-merging a worktree the user hasn't reconciled.

data/skills/boocode/using-worktrees/eval.yaml

@@ -0,0 +1,32 @@
+skill: using-worktrees
+tasks:
+  - prompt: "I'm mid-way through a feature but prod is broken — I need to fix it now"
+    grader:
+      - the response invokes the using-worktrees skill
+      - the response recognizes this as a clear case (hotfix interrupting in-progress work) and just creates the worktree rather than asking
+      - the response branches the worktree from the stable/default branch, not the in-progress feature branch
+      - the response does NOT push
+  - prompt: "Let's try a risky refactor of the inference loop and see if it pans out"
+    grader:
+      - the response invokes the using-worktrees skill
+      - the response treats this as a clear case (risky/experimental) and creates a worktree autonomously
+      - the response uses a task-derived branch name (no emojis) and a collision-safe path
+      - the response notes that project setup must run in the new worktree before it can run
+  - prompt: "Should I do this small one-line typo fix in a worktree?"
+    grader:
+      - the response invokes the using-worktrees skill
+      - the response recommends SKIP (small single-stream fix, nothing to isolate) and works in place
+      - the response does not create a worktree
+  - prompt: "This change is medium-sized and I'm not sure if it'll conflict with what I'm doing"
+    grader:
+      - the response invokes the using-worktrees skill
+      - the response treats this as ambiguous and PROPOSES a worktree in one line, then waits, rather than creating it unilaterally
+  - prompt: "Two coder worktrees both run the app on port 9502 — will they be isolated?"
+    grader:
+      - the response invokes the using-worktrees skill
+      - the response explains that worktrees isolate code state but NOT runtime (ports/DBs/services can still collide)
+      - the response recommends per-worktree setup to separate the runtime
+  - prompt: "What's the difference between git clone and git worktree?"
+    grader:
+      - the response does NOT invoke the using-worktrees skill
+      - the response answers the conceptual question directly