feat(sessions): FTS session search across transcripts + member scope; session links resolve (v0.461.0)

CHANGELOG.md

@@ -8,6 +8,29 @@ new version heading in the same commit.
 
 ## [Unreleased]
 
+## [0.461.0] - 2026-10-13
+### Added
+- **`session_history` is a real keyword search, and can find a run on another agent.** A query is now
+  ranked full-text search (FTS5, bm25, stemmed) over each run's title, seed task and report summary AND
+  over what was said in it: prompts, replies, and the files/URLs each action touched. The transcript half
+  is a per-conversation index filled off the scheduler tick in 150ms slices, capped at 32KB per
+  conversation and sampled evenly, so the last topic of a days-long run survives. Every-word matches rank
+  above any-word ones, the resumed runs of one conversation collapse to a single hit, and each hit shows
+  a snippet of where it matched. A new `scope:'member'` searches every session the person the run acts for
+  can see in the console (the console's own visibility rule, audited `session.searched`), and
+  `session_open` opens those too (audited `session.opened`). The motivating case: asked to find "the
+  session where we fixed the PayPal webhook", an agent found the old tool could not see work buried late in
+  a run titled after its first job, so it read the tenant DB with `node:sqlite` from Bash and grepped other
+  agents' transcripts instead.
+  **For users:** ask any agent to find a past session by what happened in it (for example "the session
+  where we fixed the PayPal webhook"). It searches what was said in the run, not just its title, and can
+  find sessions you were in with other agents.
+### Fixed
+- **Session links from agents open the session.** The console's session route expects the tmux name
+  (`#/sessions/aos-ses_…`), but agents built `#/sessions/ses_…` from the bare id, which opened a blank
+  pane. The console now accepts the bare-id form and rewrites the address bar. `session_history` and
+  `session_open` also return the canonical `url`, and accept a bare id, `aos-<id>` or a full link.
+
 ## [0.460.1] - 2026-10-12
 ### Removed
 - **The sidebar's "N ended" toggle.** The sidebar session list now shows only live sessions (plus the

CLAUDE.md

@@ -389,9 +389,15 @@ Key modules:
   + 13 conditional (chat-reply / egress / media / multi-org `github_token`, each exposed only when its
   env flag is set; full list in `docs/agent-mcp-tools.md`). Memory: `recall`/`remember`/`revise`/`forget` (recall returns each memory's id, the
   handle for revise/forget). Episodic self-query (the run-history companion to semantic memory):
-  `session_history` lists the agent's OWN past sessions (id/title/status/rating, own-scoped, "have I done
-  this before?") and `session_open` reopens any one — the friendly transcript timeline (`readConversation`)
-  or a `summary:true` recap — gated server-side to the caller's own agent. KB:
+  `session_history` searches past sessions — with a `query`, ranked FTS5/bm25 over each run's title/task/
+  report summary (`term_sessions_fts`, trigger-kept) AND what was said in it (`session_text_fts`, one row per
+  conversation, filled off the scheduler tick by `indexSessionText` — see `src/edge/session-search.ts`).
+  Scope `agent` (default) = own runs; `member` = everything the run-as human can see in the console
+  (`canReadRow`, audited `session.searched`). `session_open` reopens one — the friendly transcript timeline
+  (`readConversation`) or a `summary:true` recap — own-agent, or member-visible (audited `session.opened`).
+  Both return the console `url` (`TerminalManager.sessionUrl`): the sessions route's detail is the TMUX name
+  `aos-<id>`, so an agent-built `#/sessions/ses_…` used to land on a blank pane (the console now rewrites
+  that form too). KB:
   `kb_search`/`kb_read`/`kb_write`/`kb_history`/`kb_revert`. Operator/inbox:
   `ask`/`check_inbox`/`report`/`update`/`notify`/`publish`/`library_list` (session cards are
   **owner-scoped** — addressed to the run's `run_as`/spawner via the `sessionOwner` audience — so the

docs/agent-mcp-tools.md

@@ -79,8 +79,8 @@ scope/parity properties, but not the cost.
 | `skill_find` | `GET /api/skills/discover` | `TerminalManager.requestableSkills` (+ `searchSkillsh` when `q`) | R | the caller's installed library (each flagged `active` for this agent) + the bundled catalog — what's installable to ask for; with a `query` also returns matching **community** skills from the skills.sh directory, each with its `source` (`owner/repo`) |
 | `skill_request` | `POST /api/skills/request` | `TerminalManager.requestSkill` + messages | W | **asks** an owner/admin to install a skill (never installs itself). `source` omitted ⇒ the bundled catalog (validated against it); `source: 'owner/repo'` ⇒ a **remote GitHub repo** (`browseRepo` resolves the skill + its path at request time so a typo fails fast). Dedupes an open request for the same skill+source, posts a `skill.request` card to owner/admins; audited `skill.requested`. Human installs via `POST /api/skills/requests/:id/approve` (owner/admin; catalog → `SkillsStore.install`, remote → `fetchSkill` + `installFiles`; optional `scope:'agent'`) or dismisses via `POST /api/skills/requests/:id/dismiss`. On approve, `TerminalManager.refreshAgentSkills` delivers same-session to the requester's LIVE interactive sessions (materialise into the watched `.claude/skills` + inject `/reload-skills`, claude ≥2.1.152); headless/next-launch otherwise |
 | `library_list` | `GET /api/agent/artifacts` | `ArtifactStore.list` + `ArtifactStore.folders` | R | list scoped to the agent's own deliverables in the Library; also returns the tenant-wide Library folders so a `publish` files into the existing tree (discovery) |
-| `session_history` | `GET /api/agent/sessions` | `TerminalManager.sessionsForAgent` | R | **episodic self-query** — the agent's OWN past runs ("have I done this before, how did it go?"), the run-history companion to the semantic memory plane. Own-scoped server-side to the caller's agent id (never a sibling's); excludes the current run; metadata only (id/title/status/rating/dates), not transcripts; optional `query` substring over title/task |
-| `session_open` | `GET /api/agent/session` | `readConversation` / `summarizeConversation` (`src/edge`) | R | opens ONE of the caller's own past sessions and returns the friendly transcript **timeline** (`readConversation` → user/assistant/activity turns, tail-capped at 60), or with `summary:true` a throwaway-claude **recap** of the whole run. Own-scoped: the target must belong to the caller's agent (`sessionsForAgent` lookup) else **403**. Falls back to the raw `session-<id>.log` pane tail for a headless run with no structured transcript; summary audited `session.summarized` (`principal: agent:<id>`) |
+| `session_history` | `GET /api/agent/sessions` | `TerminalManager.searchSessions` / `sessionsForAgent` / `recentSessionsForMember` | R | **episodic search** — past runs ("have I done this before?", "find the session where we fixed X"). With a `query`: ranked FTS5/bm25 over the row's title/task/report summary (`term_sessions_fts`) AND the run's own words (`session_text_fts`: prompts, replies, action hints — filled off the tick by `indexSessionText`, ≤32KB per conversation, sampled evenly so a long run's LAST topic survives); all-words matches rank first; runs of one conversation collapse to one hit; each hit carries a match `snippet`. `scope:'agent'` (default) = the caller's own runs; `scope:'member'` (+ optional `agent`) = every run the run-as human can see in the console (`canReadRow`), audited `session.searched`; a run acting for nobody gets `no_member`. Excludes the current run. Every row carries `url` — the console link, which agents must hand over verbatim |
+| `session_open` | `GET /api/agent/session` | `readConversation` / `summarizeConversation` (`src/edge`) | R | opens ONE past session and returns the friendly transcript **timeline** (`readConversation` → user/assistant/activity turns, tail-capped at 60), or with `summary:true` a throwaway-claude **recap** of the whole run. Takes a bare id, `aos-<id>`, or a console URL (`normalizeSessionRef`). Allowed for the caller's own agent's runs, or a run the run-as human can see (`memberCanViewSession`, audited `session.opened`) — else **403**. Meta includes `agent` + `url`. Falls back to the raw `session-<id>.log` pane tail for a headless run with no structured transcript; summary audited `session.summarized` (`principal: agent:<id>`) |
 | `schedule` | `POST /api/agent/schedule` | `Automations.schedule` (`type:'once'`) | W | one-shot deferred self-run; same agent + run-as; **resumes the scheduling conversation by default** (`resume_claude_id` → `--resume` the transcript; `resume:false` = fresh session); bounded 1 min–30 days, ≤25 pending/agent |
 | `unschedule` | `POST /api/agent/schedule/cancel` | `Automations.cancelScheduled` | W | cancel a pending one, scoped to the agent |
 | `stop` | `POST /api/agent/stop` | `TerminalManager.stopSession` (`by` = agent id) | W | the session ends ITSELF — same halt as the console kill (kills tmux, cancels its pending questions/approvals, blocks auto-resume, writes a `stopped` episode). Server acks then halts ~150ms later so the reply flushes first; audited `session.stopped` with the agent id + optional `reason` |