path: root/tasks.md

# Dispatch — tasks (live progress)

> **Live status + roadmap only.** Completed milestones are summarized, not
> narrated. Old blow-by-blow history is pruned — it lives in git (`git log`).
> Keep this lean and current; do not let it re-accrete a step-by-step changelog.

## Status (current)
`tsc -b` EXIT 0 · biome clean · **1730 vitest** pass (+6 sshd-integration skipped). (worktree `feature/ssh-support`;
merged `dev` — brings retry-with-backoff (`provider-retry` AgentEvent) + the LSP-dead-server fix alongside the
SSH waves below.)

## Retry with backoff on retryable provider errors (DONE — from dev)
When the upstream LLM API returns a retryable error (HTTP 429 / 5xx "overloaded"),
the kernel now retries `provider.stream()` with a stepped backoff, visibly, until
the 8h cumulative-sleep budget is exhausted — then emits the final error and
seals the turn. Retries fire ONLY when no content was emitted yet this step (the
safety invariant — never duplicate partial output). Plan:
`notes/retry-with-backoff-plan.md`; report: `reports/retry-with-backoff.md`.
- **Architecture (kernel hook + shell policy/I/O):** kernel provides the hook
  (`RetryStrategy` contract + the retry loop in `runTurn`); the shell
  (session-orchestrator) provides the policy (the schedule) + the I/O (an
  abortable `setTimeout` sleep). Kernel imports no timer. `retry?` is optional
  → omit = no retry (backward-compatible).
- **New transient `AgentEvent` variant** `provider-retry` (`@dispatch/wire`),
  emitted once per scheduled retry BEFORE the sleep so the UI can show
  "⚠ retrying in Ns…" immediately; NOT persisted to model history (never
  pollutes the prompt). Final failure is still a persisted `error` + seal.
- **Schedule:** `5s,10s,30s,60s,5m,10m,15m,30m`, then repeat 30m until 8h of
  cumulative scheduled sleep → ~21 retries then give up. Pure `delayFor(attempt)`.
- **Retry trigger:** emitted `error` with `retryable===true` → retry;
  `retryable` false/absent → give up; a THROWN error → retryable-by-default
  ONLY when pre-content. All gated on `!hadContent` (text/reasoning/tool-call/usage).
- **Frontend handoff (5d3f, separate repo `../dispatch-web`):** render
  `provider-retry` as a yellow warning system-message bubble showing `message`
  (+`code`) with the `delayMs` countdown.

## SSH support — transparent remote execution (DONE — waves 0-5c)
Plan: `notes/ssh-support-plan.md` (decisions locked in §0.5/§13). Orchestrated in
waves (ORCHESTRATOR.md §2a — pre-author the contract seam, then parallel
owner-agents on disjoint packages).
- [x] **Wave 0** (orchestrator): kernel contract seam — `computerId` on
      `ToolExecuteContext` + `RunTurnInput` (additive optional; backward
      compatible). `tsc -b` EXIT 0.
- [x] **Wave 1** (parallel): `wire` (Computer/defaultComputerId types) +
      `exec-backend` (NEW pkg: ExecBackend contract + LocalExecBackend + handle +
      resolver) + `kernel` runtime (thread computerId through dispatch/run-turn) +
      `conversation-store` (contract fan-out: defaultComputerId + getEffectiveComputer
      + per-conv computerId get/set/clear). `tsc -b` EXIT 0, biome clean, **1592 vitest**
      (was 1549, +43).
- [x] **Wave 2** (parallel): refactor `tool-shell`/`read-file`/`write-file`/
      `edit-file` behind `ExecBackend` (local-only; spawn.ts deleted — logic moved
      to exec-backend; edit_file gains forward-compatible remote-diagnostics skip).
      `tsc -b` EXIT 0, biome clean, **1599 vitest** (was 1592).
- [x] **Wave 3** (parallel): `session-orchestrator` (thread computerId end-to-end
      + remote tool-drop filter: drops `lsp` + `__`-namespaced MCP tools when
      remote) + `transport-contract` (ChatRequest.computerId + computer endpoint
      API types). `tsc -b` EXIT 0, biome clean, **1620 vitest** (was 1599).
- [x] **Wave 4** (parallel): `transport-http` (computer endpoints + `/chat`
      threading + the `ComputerService` seam the ssh package will provide) +
      `transport-ws` (computerId through chat.send/queue) + `mcp` (CR-1: preserve
      computerId in filter). `tsc -b` EXIT 0, biome clean, **1641 vitest** (was 1620).
- [x] **Wave 5a**: `exec-backend` — remote-backend factory handle (lazy lookup;
      computerId set -> SshExecBackend via factory; absent -> clear error). +24 tests.
- [x] **Wave 5b**: `ssh` package (NEW) — SshConnectionPool (per-alias ssh2.Client,
      lazy connect, keep-alive, idle reap), SshExecBackend (ssh2 exec+sftp, node:fs
      .code error mapping), ~/.ssh/config reader (ssh-config), known_hosts
      auto-trust-and-pin, key-only auth from ~/.ssh. LOAD-BEARING: ssh2 verified
      under Bun (connected to local sshd :22, exec OK) — decision #1 confirmed.
      Provides remoteExecBackendFactoryHandle + computerServiceHandle. +45 tests
      (6 sshd integration tests skipped). tsc -b EXIT 0, biome clean, **1690 vitest**
      (was 1641).
- [x] **Wave 5c**: host-bin — register exec-backend + ssh extensions in
      CORE_EXTENSIONS (correct DAG order); transport-http CR-5 barrel re-export of
      computerServiceHandle. orchestrator added missing @dispatch/exec-backend dep to
      host-bin + bun install. **LIVE-VERIFIED**: server boots clean ("Dispatch booted",
      no disabled extensions). tsc -b EXIT 0, biome clean, 1690 vitest (+6 sshd skipped).
- [x] **Merge dev**: brought retry-with-backoff (`provider-retry` AgentEvent — what
      the FE consumes) + LSP-dead-server fix into the SSH branch. All code files
      auto-merged cleanly; only `tasks.md` conflicted (orchestrator-resolved).
- [x] **FE handoff #3 (provider-retry merge) — RESOLVED**: FE re-synced both pinned
      file: deps (`@dispatch/wire` + `@dispatch/transport-contract`) against merged
      `feature/ssh-support`; both resolve `TurnProviderRetryEvent`. The 11 provider-
      retry svelte-check errors cleared with ZERO further FE code changes (consumer
      already complete + tested). FE full suite green: typecheck 0/0, 795/795 tests,
      biome clean, vite build OK. Earlier SSH handoffs (#1 wire types, #2 computer
      HTTP API) now also typecheck-clean against the merged wire. Nothing further
      needed from backend on this.
- [x] **FE final sync check — GREEN, all three handoffs + cross-cutting verified**:
      FE confirmed whole-tree green (typecheck 0/0, 795/795 tests, biome clean, build
      OK, git clean). (1) provider-retry (§2c): TurnProviderRetryEvent resolves;
      assertAgentEventExhaustive covers it (typecheck-green = exhaustive); ChatView
      renders yellow alert-warning bubble w/ attemptLabel + delayLabel (delayMs via
      viewProviderRetry/formatRetryDelay) + code badge, gated {#if providerRetry}.
      (2) SSH handoff #1: Workspace.defaultComputerId + Computer/ComputerEntry resolve;
      2 Workspace literals supply defaultComputerId: null; catalog flows through
      store.computers. (3) SSH handoff #2: full src/features/computer/ (ComputerField
      w/ per-conv selector + connection-status badge + Test-connection polling;
      ComputerSelect reusable; store computerId/refreshComputer/setComputer + computers
      catalog on boot + computerStatus/testComputer; WorkspaceCard default-computer
      selector via setDefaultComputer) — 20 view-model tests, typecheck-clean, chat.send
      unchanged. CROSS-CUTTING (key integration question): GREEN, no collision —
      provider-retry is WS-stream → TranscriptState.providerRetry → ChatView (transcript,
      keyed activeConversationId); computer is HTTP-ONLY (imports NO AgentEvent/chunks/
      TranscriptState) → AppStore.computerId (per-conv persisted) → ComputerField (sidebar,
      keyed currentConversationId). Disjoint state, disjoint channels (WS vs HTTP),
      disjoint regions (transcript vs sidebar), disjoint mount keys. The conversation-
      switch lifecycle is the only shared touchpoint and is correct + independent.
      assertAgentEventExhaustive confirms computer is NOT an AgentEvent (HTTP-only).
      We're done — nothing further needed from either side.
- [ ] **DEFERRED — CR-6 usageCount**: `listComputers()` returns `usageCount: 0` until a
      conversation-store count-by-alias helper + host-bin wiring is added (non-blocking —
      discovery/connect/execute all work; only the count badge shows 0). Follow-up.
- [ ] **DEFERRED — cache-warming**: computerId threading intentionally NOT done
      (user-deferred — cache-warming is not needed right now). Known limitation:
      a warm probe on a remote turn assembles the tool set WITHOUT the remote-drop
      → a potential prompt-cache miss (performance-only, not correctness). Revisit
      when cache-warming is re-enabled.
Key decisions: ssh2 + ssh-config (project-local deps); key-only auth from
`~/.ssh`; auto-trust-and-pin host keys; computers discovered read-only from
`~/.ssh/config` (no CRUD entity); computerId persisted per-conversation; LSP/MCP
silently dropped on remote turns; edit_file works w/o diagnostics remotely.

## Per-edit LSP diagnostics auto-append (DONE)
After a successful `edit_file`, the extension now calls LSP `getDiagnostics` on the
post-edit buffer and appends any errors/warnings (severity ≤ 2) to the tool result —
so the model sees lint/diagnostics feedback inline without a separate round-trip.
Multi-server aggregation queries ALL connected servers matching the file's extension
(not just the first), merging diagnostics tagged by source (`[steep]`, `[ruby-lsp]`, etc.).
Incremental sync (`textDocument/didChange`) captures each server's `change` kind during
`initialize` and computes prefix/suffix diff ranges for `change:2` servers, full content
for `change:1`. New pure `diff.ts` (`computeChangeRange` + `offsetToPosition`, O(n)).
60s timeout; slow warning if >10s; graceful degradation when no LSP available. Generic
— works for any LSP. `languageId` mapping extended (`.rb`/`.rbs`/`.c`/`.cpp`/etc.).
- [x] Wave 1 — `packages/lsp/` (single unit): diff.ts, client, tool, diagnostics, language, types, extension. 15 new diff tests + multi-server tool test.
- [x] Wave 2 — `packages/tool-edit-file/`: optional dep on `@dispatch/lsp` via `host.getService()` (not manifest `dependsOn`); appends diagnostics after successful edit.
- [x] Verified: `tsc -b` EXIT 0, biome clean, **1468 vitest** pass (was 1453, +15).
- [x] **LIVE-VERIFIED** (production dispatch-server :24991): edit_file now surfaces LSP diagnostics inline — a deliberate type error (`const x: number = "not a number"`) in a .ts file produces `[TypeScript Language Server] ERROR (2322) L3:9: Type 'string' is not assignable to type 'number'` appended to the edit result. Required a lazy LSP service lookup fix (commits e03a96e + d4ff45c) — tool-edit-file activates at position 5 in CORE_EXTENSIONS while lsp activates at position 20, so getService always threw at activation time.

## MCP (Model Context Protocol) integration (DONE)
Dispatch is now an MCP host. A new `mcp` standard extension (`packages/mcp/`) spawns
configured MCP servers (stdio child processes), performs the MCP handshake, discovers
tools via `tools/list`, and registers each as a first-class Dispatch `ToolContract` via
`host.defineTool`. When the model calls an MCP tool, the extension proxies to `tools/call`
on the MCP server and returns the flattened result. Config: `.dispatch/mcp.json` (servers
key) → `opencode.json` mcp key fallback, resolved per-cwd (mirrors LSP). Tool names namespaced
as `<serverId>__<toolName>`. A `toolsFilter` drops tools from disconnected servers. Phase 1:
stdio only, Tools only (no Resources/Prompts/HTTP/sampling). Hand-rolled JSON-RPC (zero deps).
- **Design:** `notes/mcp-design.md` + `PLAN-mcp.md`.
- [x] Wave 1 — `packages/mcp/` (agent via dispatch CLI): 12 source + 8 test files, 69 tests.
- [x] Wave 2 — orchestrator: root tsconfig ref, host-bin CORE_EXTENSIONS registration, bun install.
- [x] Verified: `tsc -b` EXIT 0, biome clean, **1537 vitest** pass (was 1468, +69).
- [x] **LIVE-VERIFIED** (production dispatch-server :24991): a minimal test MCP server (stdio,
  one `ping` tool) configured in `.dispatch/mcp.json` → model discovered `test__ping`,
  called it with `{"msg":"hello"}`, received `pong` — full turn lifecycle (tool-call →
  tool-result → done). Tool name namespacing (`<serverId>__<toolName>`) confirmed on the wire.
- **Bug found + fixed during live-verify:** `edit_file` tool was missing from the toolset
  because the per-edit diagnostics change called `host.getService(lspServiceHandle)` at
  activation time, but `tool-edit-file` activates BEFORE `lsp` in CORE_EXTENSIONS → getService
  threw → activate crashed → tool never registered. Fix: lazy lookup at edit time (commits
  e03a96e, d4ff45c).

## Broken-chat self-repair (read-time reconcile) (DONE)
Conversation `77574596` broke unrecoverably: `reconcile()` only repaired orphaned
tool-calls, not (a) a trailing assistant message whose only chunk is `error`
(serializes to empty content → uncontinuable) and (b) a `tool-call` whose `input`
is a raw malformed-JSON string (re-sent as OpenAI `arguments` → provider 400s on
every continuation). `load()` also had no try/catch on `JSON.parse` (one corrupt
row would brick a chat). Fix = read-time repair so broken chats auto-heal on next
open — NO DB surgery (append-only preserved; repair is a turn-path transform on
`load()`). Full diagnosis + plan: `broken-chat-repair-handoff.md` +
`reports/broken-chat-repair-diagnosis.md`.
- **Layer 1 — `conversation-store` `reconcile.ts` (protects ALL providers):**
  `reconcileWithReport` now (1) strips `error` chunks from assistant messages, (2)
  drops any assistant message left with no `text`/`tool-call` (the emptied error-only
  msg — safe: never followed by a `tool` msg), (3) keeps orphaned-tool-call synthesis
  unchanged. `ReconcileReport` +2 additive counts (`strippedErrorChunks`,
  `droppedEmptyMessages`) for the repair span. `loadSince` (FE reads) intentionally
  NOT reconciled — the user still SEES the error while the provider gets clean history.
  **Hardening:** `store.ts` `load()` wraps per-chunk `JSON.parse` in try/catch →
  corrupt row skipped (log + continue), reconcile runs on the rest. +6 reconcile/store
  tests.
- **Layer 2 — `openai-stream` `convert-messages.ts` (per-provider args safety):** new
  pure `serializeToolArguments` — object→stringify; valid-string→parse+restringify;
  malformed-string→fallback `{ _malformed_arguments: <truncated 200> }`. Output ALWAYS
  `JSON.parse`s → provider stops 400ing on stored malformed args. +4 tests.
- **Layer 2 (equiv) — `../claude` `provider-anthropic` `convert.ts`:** `safeJson` now
  returns a valid object fallback (`{ _malformed_arguments: s.slice(0,200) }`) on
  parse failure, not the raw string (`tool_use.input` must be an object for Anthropic).
  Exported for direct testing. +3 tests. (Separate repo, separate agent.)
- **Wave 1+2 (parallel, disjoint):** conversation-store + openai-stream (arch-rewrite)
  + provider-anthropic (`../claude`). All in-lane; zero internal mocks; no contract/type
  change. Reports: `reports/conversation-store.md`, `reports/openai-stream.md`,
  `../claude/reports/provider-anthropic.md`.
- [x] Verified: arch-rewrite `tsc -b` EXIT 0, biome clean, **1453 vitest** (was 1443);
  `../claude` `tsc -b` EXIT 0, 71 vitest, biome clean. Both pure-core units zero
  internal mocks.
- [x] **LIVE-VERIFIED** (dev stack `bin/up` :24203): reproduced 77574596's REAL broken
  tail (the actual malformed-args tool-call + trailing error chunk) in the dev DB;
  `POST /chat` continued it cleanly (`text-delta:"OK"` → `done` reason `"stop"`, no
  400) — the provider accepted the reconciled history (error stripped, args sanitized).
  The historical error chunk remains in storage by design (read-time repair only); no
  new error was appended. Cleaned up the test conversation after.

## LSP — broken-server recovery + config source attribution (DONE)
Handoff from an agent running in raylib-jamstack (configuring ruby-lsp under the
installed Dispatch harness `/usr/bin/dispatch-server`): two issues found by
decompiling the running binary. (Previous orchestrator session 77574596 did the
investigation + Wave 0 + wrote the prompt; its chat broke mid-summon — resumed.)
- **Issue 2 (blocker):** a failed LSP server was `broken` FOREVER — the manager's
  `broken` set (keyed `${serverId}:${root}`) was cleared ONLY in `shutdownAll()`, so a
  server that failed (bad env, missing binary, OR a since-fixed bad config) stayed
  `state:"error"` for the whole process. For an agent running *inside* dispatch the
  only recovery (server restart) kills its own session.
- **Issue 1:** `.dispatch/lsp.json` (read first) silently shadowed `opencode.json`'s
  `lsp` key — a broken entry won with no warning, and the caller couldn't tell which
  config source a server came from (`status()` was its only visibility).
- **Wave 0 (orchestrator, contracts):** additive `readonly configSource?: string` on
  `LspServerInfo` (`@dispatch/transport-contract` `0.20.0→0.21.0`) + a type-test
  assertion (8→9). tsc/biome/vitest clean.
- **Wave 1 — `lsp` extension:** (a) broken-server now self-heals when its *resolved
  config changes* since it was marked broken (a config edit is a discrete event → no
  retry storm; bounded backoff for transient failures); (b) `configSource?` mirrored on
  `LspServerStatus` + populated in `status()` (`.dispatch/lsp.json` / `opencode.json` /
  `built-in`); (c) shadow warning via `host.logger` when both configs declare lsp; (d)
  spawn-failure `error` strings now name the config source. 6 required named tests +
  extras. Report: (agent cut off before writing `reports/lsp.md`; work independently
  verified — 50 lsp tests, tsc EXIT 0, biome clean).
- **Wave 1 CR (transport-http):** the `GET /conversations/:id/lsp` handler mapped
  `LspServerStatus`→`LspServerInfo` field-by-field and DROPPED `configSource` (never
  reached the wire). Summoned the transport-http owner for the one-line conditional-spread
  pass-through (mirrors `error`, honors `exactOptionalPropertyTypes`) + a named pass-through
  test (present + undefined-omitted). Report: `reports/transport-http.md`.
- [x] Verified: `tsc -b` EXIT 0, biome clean, **1443 vitest** pass; all agents in-lane
  (only packages/lsp + transport-contract + transport-http touched; pre-existing
  uncommitted WIP in kernel/tool-shell left untouched). Zero internal mocks.
- [x] **LIVE-VERIFIED** (dev stack `bin/up` on :24203, new code via `--watch`):
  (A) `configSource` reaches the wire — built-in TS server reports
  `configSource:"built-in"`, `state:"connected"` (Wave 0 + transport-http pass-through
  confirmed end-to-end); (B) a broken server (`.dispatch/lsp.json` → nonexistent binary)
  reports `state:"error"` + `configSource:".dispatch/lsp.json"` + a source-named error
  string (`broken-ts [from .dispatch/lsp.json]: Executable not found in $PATH: …`);
  (C) **recovery without restart** (the blocker) — same conversation/process went
  `error`→`connected` after the config was fixed (config change clears the broken key →
  re-spawn → connects); (D) no retry storm — repeated `status()` with no config change
  stays `error`; (E) shadow warning logged via `host.logger` (`extensionId:"lsp"`,
  level `warn`) when both `.dispatch/lsp.json` and `opencode.json` declare lsp.

## Per-conversation model persistence (DONE)
Bug: a chat's selected provider + model was NOT persisted per conversation.
Opening the same chat in a new browser session defaulted to the server's
default model rather than recalling the originally selected one.
- **Wave 0 (orchestrator, contracts):** `@dispatch/transport-contract`
  `0.19.0→0.20.0` — additive `ModelResponse` + `SetModelRequest` types for
  `GET/PUT /conversations/:id/model`.
- **Wave 1 — `conversation-store`:** `getModel`/`setModel` (`model:<id>` key,
  mirrors `getReasoningEffort`/`setReasoningEffort`); `forkHistory` copies model;
  empty string clears (idempotent). +13 tests.
- **Wave 2 (parallel):** `session-orchestrator` (resolve model from persisted
  store when no per-turn override → `resolveModel`; persist the resolved model
  so it sticks; warm path parity; `resolveModelName` pure helper; +4 tests) +
  `transport-http` (`GET/PUT /conversations/:id/model` with validation +
  `parseModelBody` pure validator; +10 tests).
- [x] Verified: `tsc -b` EXIT 0, biome clean, **1433 vitest** pass; all in-lane.

## System-prompt stale on cwd change (DONE)
Bug: the system-prompt service constructed the resolved prompt once on the first
turn and reused it via `get()` on subsequent turns (cache-safe design). But the
prompt is cwd-sensitive (`[file:AGENTS.md]`, `[prompt:cwd]` variables). When a
conversation's cwd changed after the first turn, the cached prompt was stale —
referenced files from the new cwd were not loaded.
- **Wave 1 — `system-prompt`:** added `getWithMeta(conversationId)` returning
  `{ prompt, cwd }` — reads both `resolved:<id>` and a new `resolved-cwd:<id>`
  sibling key. `construct()` now also stores the cwd. All additive, no existing
  method signature/behavior changed. +5 tests.
- **Wave 2 — `session-orchestrator`:** subsequent turns call `getWithMeta`,
  compare stored cwd vs `effectiveCwd ?? process.cwd()`, and `construct` if they
  differ (or if no stored prompt exists). Compaction path (always constructs)
  and warm path (no system prompt) unaffected. +1 test.
- [x] Verified: `tsc -b` EXIT 0, biome clean, **1411 vitest** pass; both in-lane.
- No FE handoff needed (backend-only fix; no contract version bump).

## Workspace tab issue — conversation.open drops workspaceId (DONE)
Cross-repo additive fix: `conversation.open` / `conversation.statusChanged` WS
broadcasts now carry the conversation's persisted workspace id, so a frontend
opens/focuses a tab in the correct workspace instead of the viewer's current
workspace (`activeWorkspaceId`). CLI `dispatch <model> --open --workspace my-ws`
now opens only in `my-ws`.
- **Wave 0 (orchestrator, contracts):** `@dispatch/transport-contract`
  `0.18.0→0.19.0` — additive `readonly workspaceId: string` on
  `ConversationOpenMessage` and `ConversationStatusChangedMessage`.
- **Wave 1 (parallel):** `session-orchestrator` (add `workspaceId` to
  `ConversationOpenedPayload`/`ConversationStatusChangedPayload`; resolve from
  `conversationStore.getWorkspaceId` at all status-change emit sites) +
  `transport-ws` (thread `workspaceId` from hook payload into WS broadcasts) —
  disjoint packages.
- **Wave 2:** `transport-http` — `POST /conversations/:id/open` now awaits
  `getWorkspaceId(conversationId)` and emits `conversationOpened` with it.
- [x] Verified: `tsc -b` EXIT 0, biome clean, **1405 vitest** green; all agents in-lane.
- [x] **FE courier** to `29ae`: `frontend-workspace-open-handoff.md` — parse/use
  `workspaceId` from `conversation.open` and `conversation.statusChanged`;
  re-pin `@dispatch/transport-contract` `0.19.0`; re-mirror reference.md.

## LSP cwd resolution — server-default fallthrough + workspace assignment (DONE)
Bug: `GET /conversations/:id/lsp` called `getEffectiveCwd` directly, which falls through
to `serverDefaultCwd` (`process.cwd()`) when no conversation cwd is set — the LSP
connected on the wrong dir. Additionally, a new conversation's workspace isn't assigned
until the first `chat.send`, so `getEffectiveCwd` resolved against `"default"` (not the
intended workspace) when the FE set the cwd before the first turn.
- **Wave 0 (orchestrator, contracts):** `@dispatch/transport-contract` `0.16.0→0.17.0` —
  additive `SetCwdRequest.workspaceId?: string` + updated `LspStatusResponse.cwd` comment
  ("resolved working directory the LSP connects on, or null when no cwd is set").
- **Wave 1 — transport-http:** `GET /conversations/:id/lsp` now gates on `getCwd`
  (persisted) first — returns `{ cwd: null, servers: [] }` when no cwd set (LSP does NOT
  connect); only calls `getEffectiveCwd` + `lspService.status()` when a persisted cwd
  exists. `PUT /conversations/:id/cwd` now accepts optional `workspaceId` — validates
  with `isValidWorkspaceSlug`, then `ensureWorkspace` → `setWorkspaceId` → `setCwd`
  (assigns workspace before persisting cwd). 5 new tests + 1 assertion updated.
  Report: `reports/transport-http.md`.
- [x] Verified: `tsc -b` EXIT 0, biome clean, **1332 vitest** pass; agent in-lane.
- [x] **FE courier** sent to FE agent `ffe3`: `frontend-lsp-cwd-workspace-handoff.md`
  — send `workspaceId` on `PUT /conversations/:id/cwd`; `GET /conversations/:id/lsp`
  now returns `cwd: null` + empty `servers` when no working dir is set.

## Workspace cwd fallthrough + relative resolution (DONE)
FE courier in: bug report + behavior change (`workspace defaultCwd` not used at turn start when
a conversation has no explicit cwd; plus per-conversation cwd should be **relative to the workspace
`defaultCwd`** unless absolute). Resolution is backend-owned (the FE omits `cwd` on `chat.send`).
- **Scope:** single unit — `conversation-store` owns `getEffectiveCwd` (already consumed unchanged
  by `session-orchestrator` turn/warm + `transport-http` `GET /conversations/:id/lsp`), so no
  cross-package surface change and no fan-out. `GET /conversations/:id/cwd` uses `getCwd` (raw
  explicit cwd) — unchanged.
- [x] **conversation-store** — added injectable `serverDefaultCwd` (default `process.cwd()`) to
  `createConversationStore`; rewrote `getEffectiveCwd` with the new algorithm: explicit conversation
  cwd null → `workspaceCwd ?? serverDefaultCwd` (bug fix: was returning null, skipping the workspace
  default); absolute (starts `/`) → overrides; relative → `path.resolve(workspaceCwd ??
  serverDefaultCwd, conversationCwd)`. Public signature `(conversationId) => Promise<string | null>`
  unchanged. 8 regression tests. Report: `reports/conversation-store-workspace-cwd.md`.
- [x] Verified: `tsc -b` EXIT 0, biome clean, **1289 vitest** pass; agent in-lane; zero internal mocks.

## Per-turn cwd override not resolved relative to workspace (CURRENT — live-found)
Live investigation (dev stack, tab 4ef4 in workspace `test` with `defaultCwd=/home/tradam/projects/
dispatch`): `getEffectiveCwd` resolves a persisted relative cwd correctly (LSP endpoint + a chat
**omitting** `cwd` both return `/home/tradam/projects/dispatch/arch-rewrite`). BUT a per-turn `cwd`
sent on `chat.send` is used **as-is** by `session-orchestrator` (`cwd !== undefined ?
Promise.resolve(cwd)`, orchestrator.ts:360), bypassing `getEffectiveCwd`. So raw `arch-rewrite`
reaches `run_shell` → `resolve("arch-rewrite")` = `<process.cwd>/arch-rewrite` (nonexistent) → `pwd`
broken; `./` → `resolve("./")` = `process.cwd()` (valid) → "works". The FE sends the CwdField value
as a per-turn `cwd` (transport-ws threads it: router.ts:173 → extension.ts:277).
- **Fix (2 waves):** add an optional `overrideCwd?: string` to `ConversationStore.getEffectiveCwd`
  (resolve the override if provided, else the persisted `getCwd` — same relative algorithm), then
  `session-orchestrator` passes the per-turn `cwd` (turn start + warm `opts.cwd`) as the override.
- [x] **Wave 1 — conversation-store:** added `overrideCwd?` param + impl + tests.
- [x] **Wave 2 — session-orchestrator:** pass per-turn cwd as override (turn start + warm) + tests.
- [x] Verified: `tsc -b` EXIT 0, biome clean, **1298 vitest** pass; both agents in-lane; zero
  internal mocks.
- [x] **LIVE-VERIFIED** (dev stack, workspace `test` defaultCwd `/home/tradam/projects/dispatch`):
  a per-turn `cwd:"arch-rewrite"` on an existing conversation (assigned to `test`) → `pwd`
  returns `/home/tradam/projects/dispatch/arch-rewrite` (resolved, not broken). Both the
  omit-cwd path (Wave 0) and the per-turn-cwd path (Wave 2) confirmed working.
- **Known edge case (pre-existing, not a regression):** a brand-NEW conversation's FIRST turn runs
  `getEffectiveCwd` *before* the workspace is assigned (orchestrator.ts assigns it later in the
  IIFE), so a relative per-turn cwd resolves against the "default" workspace (server default)
  instead of the intended one. Uncommon (CwdField typically set after the first message). Deferred.
- **Note (separate pre-existing bug, not touched):** `DELETE /conversations/:id/cwd` returns
  `cwd:null` but does NOT clear the persisted cwd (transport-http app.ts:538 — the route is a stub).

## Cwd edge cases — timing + DELETE stub (DONE)
Two pre-existing bugs surfaced during live-verify of the relative-cwd fix:
- **Edge 1 (timing):** a NEW conversation's first turn ran `getEffectiveCwd` BEFORE the workspace
  was assigned, so a relative per-turn cwd resolved against `"default"` (server default) not the
  intended workspace. **Fix:** session-orchestrator now assigns the workspace (for new
  conversations, detected via `getConversationMeta === null`) BEFORE resolving the effective cwd;
  removed the duplicate assignment site. 3 tests.
- **Edge 2 (DELETE stub):** `DELETE /conversations/:id/cwd` returned `{cwd:null}` but did NOT
  clear the persisted cwd (no `clearCwd` on the store). **Fix:** conversation-store added
  `clearCwd(id)` (`storage.delete(cwdKey)`, idempotent) + tests; transport-http DELETE handler now
  `await clearCwd` for real.
- [x] **Wave A (parallel):** conversation-store (clearCwd) + session-orchestrator (timing) — disjoint.
- [x] **Wave B:** transport-http (DELETE handler uses clearCwd).
- [x] Verified: `tsc -b` EXIT 0, biome clean, **1311 vitest** pass; all in-lane; zero internal mocks.
- [x] **LIVE-VERIFIED** (dev stack): Edge 2 — PUT→GET(`/tmp/test`)→DELETE→GET(`null`) actually
  cleared. Edge 1 — NEW conversation, workspace `test`, per-turn `cwd:"arch-rewrite"` → `pwd`
  returns `/home/tradam/projects/dispatch/arch-rewrite` (resolved against workspace default, not
  broken).
- [x] **FE courier handoff** written + sent: `frontend-cwd-resolution-handoff.md` couriered to FE
  orchestrator conversation `b18a` via `dispatch send b18a --queue` (turn started). Behavior-only
  — no `@dispatch/wire`/`transport-contract`/`ui-contract` version bumps; no FE contract change
  needed. Notes: `DELETE /conversations/:id/cwd` now actually clears; per-turn `cwd` on `chat.send`
  resolved relative to workspace `defaultCwd`; FE MAY omit `cwd` on `chat.send` (backend resolves
  persisted).

Built and verified live (full-fidelity: every feature is a manifest-loaded
extension through the host):
- **kernel** — contracts (ABI), bus, `runTurn` turn loop, extension host.
- **core extensions** — storage-sqlite, auth-apikey, provider-openai-compat
  (OpenCode Go), conversation-store, session-orchestrator, transport-http,
  credential-store; tool extensions `read_file` (files + directory listing), `run_shell`,
  `edit_file`, `write_file`.
- **observability** — structured Logger/Span ABI + journal-sink → out-of-process
  collector → trace-store (`bun:sqlite`); host-bin supervises the collector;
  nested turn→step→{prompt, provider.request, ttft, decode} spans; D5 verbatim
  provider capture (self-redacted); `trace-replay` record/replay lib + fixtures.
- **CLI** — one-shot HTTP client (`bun packages/cli/src/main.ts`); `GET /models`,
  `--cwd`, `--conversation`.
- **web frontend** — SEPARATE repo `../dispatch-web`. Slice 1 (surface system)
  shipped via `ui-contract` + `surface-registry` + `transport-ws` +
  `surface-loaded-extensions`. Slice 2 (browser chat) in progress there.

## How to run
```bash
# .env auto-loads DISPATCH_API_KEY (do NOT re-export) and pins BACKEND_PORT (beats PORT).
# Private probe instance: override the port + ISOLATE data paths (ORCHESTRATOR §8):
BACKEND_PORT=4567 SURFACE_WS_PORT=4569 DISPATCH_DB=/tmp/opencode/probe/dispatch.db \
  DISPATCH_TRACE_DB=/tmp/opencode/probe/traces.db DISPATCH_JOURNAL=/tmp/opencode/probe/app.ndjson \
  bun packages/host-bin/src/main.ts   # boots app + collector
curl -s -X POST localhost:4567/chat -H 'content-type: application/json' \
  -d '{"conversationId":"c1","message":"Say hello in 3 words."}'        # field = conversationId
```
Process cleanup uses the `[x]` bracket trick (ORCHESTRATOR §8) — leaked
server/collector procs poison the next run's counts.

**Two stacks:** `bin/up` = dev (live-reload backend, ports 24203/24205/24204).
`../bin/up2` = a **stable, no-watch** second stack on **25203/25205/25204** with
ISOLATED data (`./.dispatch-data/up2/`, `./.dispatch/journal/up2/`) — runs ALONGSIDE
`bin/up`, edit backend code freely without restarting it; Ctrl-C stops only itself.
Enabled by a new env knob **`SURFACE_WS_PORT`** → `surfaceWsPort` config
(`host-bin/config.ts`; default 24205 when unset, so dev is unchanged).

## Foundation (done — summarized; details in git)
- **MVP + multi-turn:** curl → transport-http → session-orchestrator →
  host/registry → provider → OpenCode Go → AgentEvents → NDJSON;
  `conversationId` threads history.
- **Post-MVP:** auth→provider seam; `read_file` tool (live tool-dispatch loop);
  `getHostAPI()` hygiene; `tabId → conversationId` rename.
- **Observability Phase A/B:** the substrate + collector/store + supervision +
  replay fixtures (see bullet list above).
- **CLI MVP:** credential-store + transport-contract + cli; model catalog; cwd
  threading; multi-turn.
- **FE Slice 1:** the surface system across both repos (live WS probe verified).
- **FE Slice 2 backend prereqs:** `@dispatch/wire` split; per-chunk `seq` cursor;
  read endpoint `GET /conversations/:id?sinceSeq=`; WS chat-deltas (transport-ws);
  turn-lifecycle events (`turn-start`/`done`/`turn-sealed`); step grouping
  (`stepId` on tool chunks/events); live stream metrics (`step-complete` +
  `usage`/`done` token/timing — "Pass 1"); CORS.

## Metrics — token + timing (current milestone)
- [x] **Pass 1 — live stream metrics** (done): `step-complete` event +
  `usage`(stepId) + `done`(durationMs + aggregate usage).
- [x] **Observability spans** (done): turn & step span-close stamp all four
  `Usage` fields (added cacheRead/cacheWrite; normalized `usage_*` → `usage.*`).
- [x] **Pass 2 — persisted replay metrics** (done, was deferred): `StepMetrics`/
  `TurnMetrics` wire types; conversation-store `appendMetrics`/`loadMetrics`
  (separate key space, turn-append order); session-orchestrator accumulates
  per-step+turn metrics from the event stream and persists after seal;
  transport-http `GET /conversations/:id/metrics` → `ConversationMetricsResponse`.
  `@dispatch/wire` + `@dispatch/transport-contract` → `0.4.0`. Commit `6db12ff`.
- [x] **Live-verified end-to-end** (against flash): live `step-complete`/`done`
  metrics ↔ persisted `GET /conversations/:id/metrics` byte-match (aggregate +
  per-step `stepId` + ttft/decode/genTotal + durationMs); journal turn/step spans
  carry dotted `usage.*` incl. `usage.cacheReadTokens` (the #2 fix).
- [x] **FE courier handoff** written: `frontend-metrics-pass2-handoff.md` (in
  this repo; user couriers to `../dispatch-web`; ORCHESTRATOR §7).

## dedup / storage growth (DONE)
Design `notes/observability-design.md` §12. User-gated calls: extend existing
pipeline (no new ext); scope = **de-dup + retention/rotation** (D9 roll-ups
deferred); dedup = **content-addressed bodies** (body-hash, NOT fingerprint-gated).
- [x] **Wave 1 — `trace-store`**: content-addressed `bodies` table (SHA-256),
  at-rest gzip (>1 KiB), `prune(policy)` (age + drop-oldest byte-cap + orphan GC) /
  `RetentionPolicy` / `PruneSummary` / `DEFAULT_RETENTION` (7d/256MiB); reads
  transparent.
- [x] **Wave 2 — `observability-collector`**: pure `shouldPrune` cadence helper;
  `main.ts` calls `store.prune(DEFAULT_RETENTION)` on a coarse cadence
  (`--prune-interval-ms`, default 60s; host-bin-overridable), log-and-continue on
  error.
- [x] Glossary: added content-addressed body, trace retention, prefix fingerprint,
  warm vs real.
- [x] **Migration bug** (found by live boot, fixed): Wave 1 created the
  `idx_records_bodyHash` index BEFORE running `migrateOldBodies`, so opening a
  pre-existing OLD-schema `traces.db` crashed the collector
  (`no such column: bodyHash`, crash-looped). Fix = reorder migration before the
  index + 3 regression tests that seed a real old-schema DB. bun 106→109.
- Tests: bun 89→109. typecheck/biome clean. **Live-verified** against a real
  old-schema `traces.db`: 0 crashes, collector stays up, schema migrates
  (bodyHash + content-addressed bodies), real-data dedup (318 body refs → 270
  stored bodies), prune cadence fires cleanly (14× `prune completed`). Optional
  follow-up: host-bin env-override for the retention policy.

## Standard tools — fs + shell (DONE)
User-gated calls: **one tool per extension** (matches `tool-read-file` precedent); tools are
**standard** tier (a turn completes with `tools:[]`, §2.6/§2.8). **Zero ABI change** — the
`ToolContract`/`ToolExecuteContext` already carry `signal`/`onOutput`/`cwd`/`log`.
- **Wave 1 (parallel, disjoint pkgs, kernel-only dep) — all green:**
  - [x] `tool-read-file` — EXTENDED `read_file` to list directory contents (sorted, `/`-suffixed
    subdirs; files unchanged). 41 tests.
  - [x] `tool-shell` (new) — `run_shell`: foreground, streamed via `ctx.onOutput`, `ctx.signal`
    cancel, `ctx.cwd`, timeout + output cap, `concurrencySafe:false`; injected `spawn`. 31 tests.
  - [x] `tool-edit-file` (new) — `edit_file`: `oldString`/`newString`/`replaceAll`; errors on
    absent/non-unique/identical; workdir-contained; `concurrencySafe:false`. 38 tests.
  - [x] `tool-write-file` (new) — `write_file`: explicit `overwrite` flag (absent+unset→create;
    exists+unset→error; exists+true→overwrite; absent+true→error); no parent auto-create. 33 tests.
- **Wave 2 (done):** orchestrator added 3 root tsconfig refs + `bun install`; host-bin owner
  registered the 3 new extensions in `CORE_EXTENSIONS` (same pattern as `read_file`).
- **Live-verified:** clean boot (`Dispatch booted`, collector up, no activation/capability-gate
  error — the new `shell` capability is accepted); full-graph `tsc -b` EXIT 0, biome clean.
- **Recovery notes (scar tissue):** `tool-write-file` first returned plan-only (§5a) → re-summoned
  with "IMPLEMENT NOW". `tool-edit-file` hung vitest at collection — `computeReplacement` infinite-
  looped on empty `oldString` (`"".indexOf("") === 0`, index never advances) invoked at a test's
  `describe` scope; fixed with an early empty-string guard + validation. One agent deleted
  `ORCHESTRATOR.md` out-of-lane → caught by post-wave `git status`, restored from git.
- Deferred (not selected): `glob`, `grep`/`search_code`, background shells.

## Skill system + load_skill tool (DONE)
User-gated calls: skills list lives in the **`load_skill` tool definition** (NOT the system prompt),
refreshed **per new turn** (cache-stable across steps), **live file read** on execute. One `skills`
standard extension (loader + filter + tool). Skill = md in `.skills/`; discovered from `~/.skills` +
`<cwd>/.skills` (cwd shadows home); name = filename w/o `.md`. Format: line1 = summary,
line2 = `---`, body = line3+; on load the first two lines are stripped; malformed (no `---`) =
no summary but still loadable. Glossary: added `skill`, `skill summary`, `tools filter`.
- **Mechanism — the per-turn `tools` filter chain** (first concrete use of the §3.2 context-assembly
  chain; reusable for persona/agents later):
  - [x] **kernel** — exposed `HostAPI.applyFilters` (delegates to the bus's existing `applyFilters`).
  - [x] **session-orchestrator** — defines+exports `toolsFilter`/`ToolAssembly`; applies it ONCE per
    turn (injected `applyToolsFilter` dep) before `runTurn`, threading `cwd`+`conversationId`.
  - [x] **skills** (new ext, `dependsOn session-orchestrator`) — pure parse/merge/render +
    `load_skill` tool (live read, strips first two lines, path-contained) + a `toolsFilter` filter
    that rewrites `load_skill`'s description + `name` enum with the per-cwd catalog. 42 tests.
  - [x] **host-bin** — registered `skills` in `CORE_EXTENSIONS`.
  - [x] **Fan-out (§5.3):** `applyFilters` was a required `HostAPI` addition → broke one consumer
    (transport-http `server.bun.test.ts` inline HostAPI stub) → fixed by its owner.
- **Live-verified:** clean boot (`skills` activates, filter registered, no crash); full-graph
  `tsc -b` EXIT 0, biome clean. (End-to-end load_skill via a real LLM turn not yet exercised —
  unit/integration tests cover the filter rewrite + live read.)

## Cache warming (core DONE; control surface PARTIAL)
User-gated calls: target the external **Claude** provider (`../claude` provider-anthropic, loaded via
`DISPATCH_EXTERNAL_EXTENSIONS`); warm-assembly lives in **session-orchestrator** (`warm()` reuses the
real turn's assembly → byte-identical prefix, provider-agnostic); **surface system** for controls;
**per-conversation** controls; interval default 4 min, free value. Old-code invariants honored
(primary-model/full-prefix via reuse; refuse mid-turn; never persist/emit; in-flight invalidation;
arm-on-settle/cancel-on-start; `pct = round(clamp(cacheRead/input,0,1)*100)`).
- **Mechanism (2nd use of bus hooks; first event-hook emit):**
  - [x] **kernel** — exposed `HostAPI.emit` (delegates to bus.emit), counterpart of `on`.
  - [x] **session-orchestrator** — `turnStarted`/`turnSettled` event hooks (carry conversationId/cwd/
    modelName) emitted per turn; `warm()` service (`cacheWarmHandle`) reusing assembly, refusing
    mid-turn, never persisting/emitting; returns Usage.
  - [x] **cache-warming** (new ext) — per-conversation timers (arm/cancel/in-flight token),
    calls `warm()`, computes `lastPct`, persists `{enabled,intervalMs}` (default on/240s) in
    host.storage; registers a controls Surface. 19 tests.
  - [x] **host-bin** — registered cache-warming; **transport-http** HostAPI stub fixed for `emit`.
- **Manual trigger endpoint:** `POST /chat/warm {conversationId, model?, cwd?}` → `WarmResponse`
  `{inputTokens,outputTokens,cacheReadTokens,cacheWriteTokens,cachePct}` (409 if generating). Powers a
  FE "warm now" button + fast tests. Types in `@dispatch/transport-contract`; route in transport-http.
- **LIVE-VERIFIED against Claude haiku:** automatic timer warm → journal `warm complete pct:100`;
  manual `POST /chat/warm` → `cacheReadTokens:6799, cachePct:100` (100% hit), HTTP 200. The external
  `../claude` provider-anthropic is loaded via `bin/up` (`DISPATCH_EXTERNAL_EXTENSIONS`).
- **Cache-metric fix + retention metric:** `provider-anthropic` (in `../claude`, commit `0e9d118`)
  now reports `Usage.inputTokens` as the TOTAL prompt (was the uncached remainder → the cache rate
  inflated/clamped to 100% on Claude). So `cacheRead/inputTokens` is now the true rate (live: a turn
  adding new content reads 61%, not 100%). Added **`expectedCacheRate`** = `cacheRead/(cacheRead+
  cacheWrite)` (retention/health, ~100% when warm, 0% when the cache expired) to `WarmResponse` +
  `POST /chat/warm` + the cache-warming surface (a "cache retention" stat). Live-verified: warm
  within TTL → 100%; warm after >5 min idle → 0% (cache expired). FE handoff updated with both
  metrics + the cross-turn real-turn `expectedCache = cacheRead_N/(cacheRead_{N-1}+cacheWrite_{N-1})`.
- **Surface framework extended (DONE):** added `NumberField` to `ui-contract` + per-conversation
  surface scoping (optional `conversationId` on subscribe/unsubscribe/invoke + surface/update; new
  `SurfaceContext` on `SurfaceProvider.getSpec/invoke`; transport-ws keys subscriptions by
  `(surfaceId, conversationId)` and tags updates). cache-warming now serves a PER-CONVERSATION
  surface: `Toggle`(enabled) · `Number`(interval, seconds, `cache-warming/set-interval`) ·
  `Stat`(last cache %). All backward-compatible (global surfaces like `surface-loaded-extensions`
  unchanged). **FE courier:** `frontend-cache-warming-handoff.md` (this repo) — the web must render
  the `number` field kind + send/handle `conversationId` on the surface WS protocol.

## Cache warming — FE CR-3 (DONE)
FE asked (dispatch-web `backend-handoff-cache-warming-timer.md`): expose next/last-warm timestamps +
make a manual warm reset the timer/refresh the surface. Done via an **inversion** (commit `bfbad3a`):
session-orchestrator `warm()` (the single chokepoint for manual `/chat/warm` AND the auto timer) emits
a `warmCompleted` bus event; cache-warming subscribes and does all post-warm handling — so manual
warms re-arm the timer + push a surface update with **no transport-http change** (core can't depend on
the standard cache-warming ext). Added `nextWarmAt`/`lastWarmAt` state + a `custom`
`rendererId:"cache-warming-timer"` surface field (no ui-contract bump). Caught + fixed a wiring bug
(`createWarmService` missed the `emit` dep → `deps.emit?.` silently no-oped; made it required).
Live-verified vs claude haiku (manual warm logs `warm complete` ~2s after the turn, not the 4-min
timer). FE handoff updated. (FE CR-1 table + CR-2 catalog `scope` flag still open, not requested.)

## LSP integration + per-conversation CWD (DONE)
Design: `notes/lsp-design.md`. FE courier: `frontend-lsp-cwd-handoff.md`. Decisions
(locked): **single `lsp` extension**; **hand-rolled pure JSON-RPC codec** (zero dep,
injected-stream tested); **diagnostics-on-write deferred** (on-demand `lsp` tool
only); **cwd persisted in `conversation-store`**; config = **built-in TypeScript +
`<cwd>/.dispatch/lsp.json` + `<cwd>/opencode.json` `lsp` fallback** (Roblox works
with its existing config). Glossary: added LSP, language server, diagnostics,
workspace root, working directory.
- **The bug we fixed** (opencode root cause, confirmed): opencode's
  `client/registerCapability` ignores all but `textDocument/diagnostic`, so
  `workspace/didChangeWatchedFiles` registrations are dropped + no real fs watcher
  → stale `sourcemap.json` → "Unknown require" mid-session. Fix = honor the
  registration + real fs watcher + forward `didChangeWatchedFiles` + auto-spawn
  `rojo sourcemap --watch` sidecar when `luau-lsp.sourcemap.autogenerate`. Covered
  by a regression test in `packages/lsp/src/client.test.ts`.
- **`lsp` extension** (new, bundled core): hand-rolled LSP client (framing + rpc +
  watched-files + diagnostics + config + root + tool + manager), zero external deps.
  Lazy-spawn one server per `(serverID, root)`; config resolved **per cwd**;
  `lspServiceHandle.status(cwd)` lazy-connects + reports state; `deactivate` kills
  all child procs (host-bin shutdown now calls `host.deactivate()`).
- **CWD:** `conversation-store.getCwd/setCwd`; `session-orchestrator` defaults a
  turn's cwd from the store; endpoints `GET`/`PUT /conversations/:id/cwd` +
  `GET /conversations/:id/lsp` in transport-http; wire types in
  `@dispatch/transport-contract` (→ `0.5.0`).
- **LIVE-VERIFIED:** this repo (`typescript`) → `connected`; `/home/tradam/projects/
  roblox` (`luau-lsp`) → `connected` (via the project's own `opencode.json` + rojo
  sidecar); cwd PUT/GET round-trip 200. Op note: LSP binaries must be on the server
  process PATH (`~/.local/bin` daemon-PATH caveat for `typescript-language-server`).
- **Recovery (scar tissue):** the `lsp` agent stalled on the final stretch (1 hung
  test + ~40 biome `!`/dot-key findings) → at the user's request the orchestrator
  finished it directly; also fixed a real design bug the agent missed: the manager
  read config statically instead of per-cwd (would have broken Roblox).

## Context size — current context-window usage (DONE)
User-gated decisions: term = **context size** (current usage; reserve "context window" for the
model's max LIMIT, a later feature); definition = the turn's **FINAL step `inputTokens +
outputTokens`** (NOT the aggregate `usage`, which sums per-step prompts and overcounts a
multi-step turn); delivery = a backend-computed field on BOTH the live `done` event and the
persisted `TurnMetrics`.
- [x] **Contract (orchestrator):** optional `contextSize?: number` added to `TurnDoneEvent` +
  `TurnMetrics` in `@dispatch/wire` (`0.4.0→0.5.0`); `@dispatch/transport-contract`
  `0.5.0→0.6.0` (re-exports both — no other change). Glossary: added **context size**.
- [x] **Wave (parallel, disjoint pkgs):**
  - [x] **kernel** — `run-turn.ts` tracks the last step's `Usage`; `doneEvent()` stamps
    `done.contextSize = lastStep.input + lastStep.output` (omitted when no usage). +3 tests.
  - [x] **session-orchestrator** — `metrics.ts build()` stamps `TurnMetrics.contextSize` from
    the final per-step metrics (same definition; equals the live value). +5 tests.
- [x] Verified: `tsc -b` EXIT 0, biome clean, 881 vitest pass; both owners stayed in-lane.
  `conversation-store` (JSON passthrough) + `transport-http` (forwards/serves) unchanged.
- [x] **LIVE-VERIFIED against flash** (`deepseek-v4-flash`): turn 1 → live `done.contextSize`
  1255 == persisted `turns[-1].contextSize` 1255 == final-step `1206 in + 49 out` (NOT the
  aggregate); turn 2 (same conversation) → 1286 (grew cumulatively), live == persisted. Both
  carriers agree; "current" = latest turn's value.
- [x] **FE courier handoff:** `frontend-context-size-handoff.md` (user couriers to
  `../dispatch-web`).

## Turn continuity — detached turns + multi-client live view (DONE)
Design: `notes/turn-continuity-design.md`. FE courier: `frontend-turn-continuity-handoff.md`.
Problem (code-traced): a turn's lifetime was bound to the WS connection — `transport-ws` aborted
the in-flight turn on socket close, so a backgrounded/reloaded mobile browser killed generation.
Principle enforced: **the FE is only a control interface; the AI runs independent of it**, and
**multiple clients may watch the same conversation** (multi-device handoff).
- **Decisions (locked):** broadcast hub lives in the CORE (`session-orchestrator`), not a
  transport; additive `SessionOrchestrator` handle (keep `handleMessage`); persist-at-seal kept,
  per-step R1 deferred; late-join served by an in-memory in-flight buffer; subscribers persist
  per-conversation independent of turns; no concurrent-send arbitration; no explicit stop op.
- **Contract (orchestrator):** `@dispatch/transport-contract` `0.6.0→0.7.0` — additive WS ops
  `chat.subscribe`/`chat.unsubscribe` on `WsClientMessage` (events still arrive as `chat.delta`).
- **Wave 1 — `session-orchestrator`:** detached per-conversation turn ownership + broadcast;
  `startTurn`/`subscribe`/`isActive` added to the handle; `handleMessage` → convenience wrapper
  (dropped `signal`). **Two-map model** (`subscribers` persistent + `activeTurns` buffer) — the
  fix for the live-found bug where pre-turn subscribers were dropped. 63 tests.
- **Wave 2 (parallel) — `transport-ws`** (fan-out: per-connection chat-subscription map;
  `chat.send` auto-subscribes sender + `startTurn`; new ops in pure `router.ts`; `close` drops
  subs but NEVER aborts a turn; removed the turn `AbortController`) + **`transport-http`** (only
  test fakes updated for the 3 new methods; runtime unchanged). host-bin untouched.
- **LIVE-VERIFIED against flash** (2-client WS test, `/tmp/ws_multi.ts`): (S1) two clients both
  stream a turn; closing the SENDER mid-turn → the other keeps receiving through `done` and the
  turn persists (1197 chars) — AI kept going independent of the interface; (S2) a client joining
  mid-turn gets `turn-start` replayed + the rest live. `RESULT OVERALL: OK`.
- **Recovery (scar tissue):** first Wave-1 impl stored listeners INSIDE the per-turn hub and
  `startTurn` made a fresh empty-listener hub → every pre-turn subscriber dropped; live test got
  zero deltas though the turn ran+persisted. Caught by live-verify (unit test had subscribed
  AFTER start, masking it). Fixed via the persistent-subscribers / per-turn-buffer split.

## Turn continuity — CR-3: user prompt on the event stream (DONE)
FE bug (multi-client): a pure watcher (subscribed, not the sender) couldn't see the USER prompt until
seal — the user message was passed to the provider + persisted only at seal, never on the turn's
outward stream/buffer. FE courier: `frontend-cr3-user-message-handoff.md`.
- **Contract:** `@dispatch/wire` `0.5.0→0.6.0` — additive `TurnInputEvent`
  `{ type:"user-message"; conversationId; turnId; text }` on the `AgentEvent` union (kernel barrels
  re-export it). `@dispatch/transport-contract` `0.7.0→0.8.0` (re-export only). Widening broke NO
  exhaustive switch (typecheck clean) — zero consumer fan-out.
- **session-orchestrator:** `emitToHub({type:"user-message",…})` as the FIRST event of `runTurnDetached`
  (before `runTurn`) → buffered + broadcast to all subscribers (live + late-join); HTTP path covered via
  `handleMessage`'s buffer replay. Persistence + metrics unchanged. +3 tests; 3 Wave-1 tests updated
  (user-message now precedes turn-start).
- **LIVE-VERIFIED vs flash:** a watcher that never sent receives `user-message` (correct text) as its
  FIRST `chat.delta`, before `turn-sealed`, then the streaming reply. `RESULT: OK`.
- **Process note:** implemented directly by the orchestrator as a one-off (user-approved at the
  time). SUPERSEDED — the user has since confirmed the ORCHESTRATOR.md model governs: the
  orchestrator summons owner-agents and does not write feature code itself.

## Cache warming — FE CR-4 lifecycle + CR-1 extensions table + CR-2 catalog scope (DONE)
FE courier in: `../dispatch-web/backend-handoff-cache-warming.md` (+ CR-1/CR-2 from their living
`backend-handoff.md`). Courier out: `frontend-cache-warming-lifecycle-handoff.md`. Full report:
`reports/cr4-cache-warming-lifecycle.md`.
- **CR-4a:** warming defaults OFF (opt-in per conversation) — `parseSettings` + `DEFAULT_STATE`;
  re-enabling now restores the persisted interval. Known gap (pre-existing, fail-safe): no boot
  hydration of persisted opt-in across server restarts.
- **CR-4b:** post-warm surface updates now carry the FUTURE `nextWarmAt` (re-arm BEFORE notify);
  `turnSettled`/`turnStarted` also push (fresh schedule after seal / `null` while generating).
- **CR-4c:** new `POST /conversations/:id/close` (tab close ≠ disconnect): aborts the in-flight
  turn via a per-turn `AbortController` → kernel `runTurn` `signal` (partial persist + normal seal,
  `done.reason:"aborted"`), and emits new typed hook `conversationClosed` → cache-warming disables
  sync + persists OFF. Disconnect/`chat.unsubscribe` semantics unchanged.
- **CR-4d:** no change — initial `surface` echo already at HEAD (FE probed a stale up2 boot).
- **CR-1:** loaded-extensions emits count stat + ONE `custom`/`rendererId:"table"` field
  (`TablePayload` exported); columns Name|Version|Trust|Activation, all trust tiers.
- **CR-2:** `SurfaceCatalogEntry.scope?: "global"|"conversation"` (`ui-contract` `0.1.0→0.2.0`);
  set on both surfaces. `transport-contract` `0.8.0→0.9.0` (additive `CloseConversationResponse`).
- 907 tests pass (+13 new); typecheck + biome clean. **LIVE-VERIFIED vs `bin/up`:** default-off,
  2 automatic warms @5s each pushing future `nextWarmAt`, mid-turn close → `abortedTurn:true` +
  `done.reason:"aborted"` + warming disabled, catalog scopes + table field present, echo present.

## History windowing — FE CR-5 (DONE)
FE courier in: `../dispatch-web/backend-handoff-chat-limit.md` (+ living `backend-handoff.md` §2
CR-5). Courier out: `frontend-history-windowing-handoff.md`. User-gated call: ask #3 shipped as
the INVARIANT option (no new field) — seq is contractually **1-based, monotonic, gap-free**; FE
derives `hasOlder` from `chunks[0].seq > 1`.
- **Wave 0 (orchestrator, contracts):** `limit`/`beforeSeq` query-param semantics + validation +
  `latestSeq` windowed-read caveat documented on `ConversationHistoryResponse`
  (`@dispatch/transport-contract` `0.9.0→0.10.0`); 1-based seq guarantee codified on
  `StoredChunk` (`@dispatch/wire` `0.6.0→0.6.1`, doc-only).
- **Wave 1 — `conversation-store`:** additive `loadSince(id, sinceSeq?, window?: { beforeSeq?,
  limit? })` — selection `sinceSeq < seq < beforeSeq`, newest-`limit` window, result stays
  ascending; garbage-in treated as absent (transport validates upstream). +8 tests.
- **Wave 2 — `transport-http`:** parses + validates the params (positive integers; malformed/
  zero/negative → 400 `{ error }`, store never called with an invalid window); two-arg call
  shape preserved when no params (regression-guarded). +20 tests.
- 935 vitest + 112 bun tests, typecheck + biome clean. **LIVE-VERIFIED** (isolated boot, real
  flash turns): firstSeq=1; `limit=2`→`[5,6]` ascending w/ correct `latestSeq`; `limit=9999`→
  full log; `beforeSeq=3`→`[1,2]`; `beforeSeq=3&limit=1`→`[2]`; `limit=0`/`beforeSeq=0`/
  `limit=abc`→400×3. `RESULT: OK` ×6.
- **Scar tissue (process):** (1) probing with a PRIVATE boot was overkill — the windowing checks
  are read-only GETs and the dev stack was running; prefer probing `bin/up`/`up2` or asking the
  user (ORCHESTRATOR §8 updated). (2) The §8 boot recipe was stale (`DISPATCH_API_KEY_OPENCODE1`
  doesn't exist; an empty re-export OVERRIDES `.env` → "No providers registered"; `.env`'s
  `BACKEND_PORT` beats `PORT`; un-isolated data paths spawn a duplicate collector on the dev
  DB) — recipe fixed in §8 + above. (3) Violated the bracket trick once (`pkill -f 'cr5-data'`
  self-matched → killed parent shell, timeout-with-no-output); the existing §8 rule stands.

## Reasoning effort (current milestone)
User-gated calls: canonical term **reasoning effort** (GLOSSARY); ladder `low|medium|high|xhigh|max`
(Anthropic-driven, includes xhigh/max); scope = **(c)** persisted per-conversation + per-turn
`ChatRequest.reasoningEffort` override; resolution default **`high`**; provider picks sensible
budget_tokens; `../claude` orchestrated DIRECTLY (mode A); CLI `--effort` now.
- [x] **Wave 0 (orchestrator, contracts):** `ReasoningEffort` in `@dispatch/wire` (`0.6.1→0.7.0`);
  `ProviderStreamOptions.reasoningEffort` (kernel contract; runtime untouched — providerOpts is
  forwarded verbatim); `ChatRequest.reasoningEffort` + `ReasoningEffortResponse`/
  `SetReasoningEffortRequest` GET/PUT types (`@dispatch/transport-contract` `0.10.0→0.11.0`);
  glossary entry. typecheck + biome clean.
- [x] **Wave 1 (parallel ×3, disjoint):** `conversation-store` get/setReasoningEffort (own key
  space, mirrors cwd; +12 tests); `provider-anthropic` (../claude commit `c0835a4`, mode A summon
  with `--dir ../claude`, contract excerpt INLINED per the cross-`--dir` hang rule) —
  `REASONING_EFFORT_BUDGETS` 4096/10240/16384/32768/65536, raises max_tokens above budget, strips
  temperature when thinking on, absent → byte-stable body (+12 tests); `cli` `--effort` flag,
  parse-validated, body key omitted when unset (+8 tests).
- [x] **Wave 2:** `session-orchestrator` — exported pure `resolveReasoningEffort` (override →
  stored → `"high"`), additive `StartTurnInput.reasoningEffort`, providerOpts always stamped,
  **warm() parity** (same resolved effort as a real turn — prompt-cache safe), own fakes fixed
  (+9 tests).
- [x] **Wave 3 (parallel ×2):** `transport-http` — `/chat` validation (400 names valid levels,
  orchestrator never sees bad input), threads to startTurn, GET/PUT
  `/conversations/:id/reasoning-effort` mirroring cwd endpoints, own fakes fixed; `transport-ws` —
  `chat.send` threading + validation (+3 tests).
- [x] Verified: `tsc -b` EXIT 0, biome clean, **993 vitest + 189 bun** green; all agents in-lane.
  Commits: arch-rewrite `35197ed` (contracts) + `020e051` (impl); ../claude `c0835a4`.
- [x] Live-verified vs claude (thinking deltas streamed at xhigh; persisted PUT honored next turn).
- [x] FE courier handoff written: `frontend-reasoning-effort-handoff.md` (user couriers to
  `../dispatch-web`): ChatRequest/chat.send field + GET/PUT endpoints + ladder + default-`high`
  semantics + cache note.

## Message queue + steering injection (DONE)
Design: this file's roadmap item 3 (now implemented). User-gated calls: a **separate
`message-queue` standard extension** (dependsOn `surface-registry`) owns the queue STATE +
a per-conversation `custom` surface; the **session-orchestrator** owns delivery (drain →
inject → carry) + emits the `steering` event (it owns the chat hub — no `chatEmit` service
needed); the **kernel** gets a generic `drainSteering` callback. Glossary: added
**message queue**, **steering**, **queued message**. Enqueue when idle **starts a turn**
(user choice; `chat.queue` degrades to `chat.send`). Steering text rendered live via a new
additive `steering` `AgentEvent`; queue state via the surface (NOT the chat stream).
- **Wave 0 (orchestrator, contracts):** `RunTurnInput.drainSteering?: () => readonly
  ChatMessage[]` (kernel contract — generic, kernel stays pure); `QueuedMessage` +
  `QueuePayload` + `TurnSteeringEvent` (type `"steering"`, additive to `AgentEvent`) in
  `@dispatch/wire` (`0.7.0→0.8.0`); `POST /conversations/:id/queue` + WS `chat.queue` op +
  `QueueRequest`/`QueueResponse` in `@dispatch/transport-contract` (`0.11.0→0.12.0`). typecheck
  clean except the expected transport-ws exhaustive-switch fan-out (fixed in Wave 3).
- **Wave 1 (parallel ×2, disjoint):** `kernel` runtime — calls `drainSteering` at the
  tool-result boundary only when continuing to a next step (gated; no drain on max-steps),
  +6 pure tests (65 total); `message-queue` (NEW ext) — pure queue core (enqueue/getQueue/
  drain/combine) + `MessageQueueService`/`messageQueueHandle` + per-conversation `custom`
  surface (`rendererId:"message-queue"`, `QueuePayload`), 12 tests. (The message-queue agent
  DIED mid-task after writing all src+tests but before verifying/reporting; orchestrator
  recovered by running `bun install` + root tsconfig ref + verifying directly — tsc/vitest/
  biome clean, 12 tests pass; no hand-fixing of impl.)
- **Wave 2:** `session-orchestrator` — added `enqueue` facade (idle→`startTurn`,
  active→queue.enqueue) + `resolveQueue?` dep (self-wired lazily in `activate` via
  `host.getService(messageQueueHandle)` — host-bin does NOT wire it) + `drainSteering` wrapper
  (drain → emit `steering` → return one combined user `ChatMessage`) + post-seal carry
  (non-empty queue → new turn), +8 tests (85 total). `message-queue` is an OPTIONAL dep
  (feature degrades off if absent).
- **Wave 3 (parallel ×3):** `host-bin` — registered `message-queue` in `CORE_EXTENSIONS`
  (+dep+ref), 28 tests; `transport-http` — `POST /conversations/:id/queue` route + validation,
  145 tests; `transport-ws` — `chat.queue` op + fixed the Wave-0 exhaustive-switch fan-out,
  29 vitest + 20 bun.
- Verified: `tsc -b` EXIT 0, biome clean (280 files), **1043 vitest + 199 transport bun** pass;
  all agents in-lane. **Boot smoke:** private instance boots clean with `message-queue`
  registered (no activation crash).
- [x] FE courier handoff written: `frontend-message-queue-handoff.md` (user couriers to
  `../dispatch-web`): surface (`rendererId:"message-queue"`), `chat.queue` WS op, `steering`
  event, HTTP `POST /queue`, auto-start-when-idle, carry semantics, version bumps.

## Umans AI Coding Plan provider (DONE)
User-gated calls: a new **`provider-umans`** standard extension wrapping the Umans
OpenAI-compatible backend (`https://api.code.umans.ai/v1`). Built via the **full-refactor
path**: first extract a generic `@dispatch/openai-stream` library from
`provider-openai-compat`, then build `provider-umans` on top. Self-contained (reads
`UMANS_API_KEY` from env directly — no `auth-apikey` dep).
- **Wave 1 — `@dispatch/openai-stream` lib (NEW package):** extracted the generic OpenAI
  functions (convert-messages, convert-tools, parse-sse, listModels, stream, provider)
  from `provider-openai-compat` into a pure library package. `createOpenAICompatProvider`
  parameterized: `id: string` (was hardcoded `"openai-compat"`) + `transformBody?: (body,
  opts) => Record<string,unknown>` hook (for provider-specific body fields). Refactored
  `provider-openai-compat` to import from the lib (thin extension.ts, backward-compat
  re-exports, manifest unchanged, byte-identical behavior). Full tsc EXIT 0, 66 vitest,
  biome clean. Report: `reports/provider-umans-wave1-openai-stream.md`.
- **Wave 2 — `provider-umans` (NEW ext):** imports `createOpenAICompatProvider` from the
  lib; registers provider id `"umans"`; `transformBody` maps Dispatch `reasoningEffort`
  (`low|medium|high|xhigh|max`) → Umans `reasoning_effort` (`none|low|medium|high`,
  capping `xhigh`/`max`→`high`); dynamic `listModels` (GET /v1/models); default model
  `umans-coder` (env `UMANS_MODEL` or config `provider.umans.model`); baseURL env
  `UMANS_BASE_URL`; absent key → warn + skip registration (graceful). Pure core:
  `mapReasoningEffort` + `resolveUmansConfig` (factored out for direct unit testing).
  12 tests. Report: `reports/provider-umans.md`.
- **Wave 3 — host-bin wiring:** registered `provider-umans` in `CORE_EXTENSIONS` + added
  `@dispatch/provider-umans` dep + root tsconfig ref. No credential-store entry needed
  (self-contained — reads env directly, doesn't go through `auth-apikey`). 28 host-bin
  tests.
- Verified: full-graph `tsc -b` EXIT 0, biome clean (293 files), **1059 vitest** pass.
  **Boot smoke:** without `UMANS_API_KEY` → `"provider-umans: no UMANS_API_KEY. Provider
  not registered."` (graceful skip); with `UMANS_API_KEY=sk-test` → `"provider-umans:
  registered (model=umans-coder)"`.
- [x] **LIVE-VERIFIED against the real Umans API:** the dev stack (umans-glm-5.2) called
  `web_search` (Firecrawl) in a real turn — first live Umans API call, clean response.

## web_search tool — Firecrawl (DONE)
Standard tool extension `tool-web-search` backed by a self-hosted Firecrawl instance
(`http://100.102.55.49:31329/v1`, Tailscale, no API key). One tool `web_search` with 4
modes: search, scrape, crawl (polls status URL), map — mirroring the proven opencode tool.
Pure core: `validateArgs` (discriminated union by mode) + `format*` functions + `truncateOutput`.
Injected edge: `FirecrawlClient` (injectable `fetchFn` + `sleep` + `now`), `AbortSignal.any`
for per-request timeout + caller cancellation. `concurrencySafe: true`, `capabilities: { network: true }`.
38 tests. Report: `reports/tool-web-search.md`.
- **LIVE-VERIFIED:** the dev stack (umans-glm-5.2) called `web_search` → Firecrawl returned
  real results (Paris, France) — first live Umans API call too.

## todo tool — per-conversation task list + surface (DONE)
Standard tool extension with a single `todo_write` tool (opencode `todowrite` pattern:
full-list replace, returns JSON, no business-rule enforcement — the description guides
the model). Per-conversation in-memory state (`Map<conversationId, TodoItem[]>`). Per-
conversation surface (`rendererId: "todo"`, `scope: "conversation"`) via subscriber-notify
(message-queue pattern). `concurrencySafe: false` (mutates shared state).
- **Wave 0 (orchestrator, kernel contract):** added `conversationId?: string` to
  `ToolExecuteContext` (additive, backward-compatible). Wired in `dispatch.ts` — the
  kernel already had `conversationId` as a parameter, just wasn't passing it through to
  the tool context. 170 kernel tests pass.
- **Wave 1 (todo extension):** pure core (`validateTodos` — shape only; `getTodos`/
  `setTodos`/`clearTodos` — fresh array copies; `buildTodoSpec`; `formatTodoResult` →
  `JSON.stringify`). Shell: `createTodoWriteTool({ state, notify })` + surface provider.
  26 tests. Report: `reports/todo.md`.
- **Wave 2 (host-bin wiring):** registered `todo` in `CORE_EXTENSIONS` + dep + root tsconfig
  ref. 28 host-bin tests.
- Verified: full-graph `tsc -b` EXIT 0, biome clean (314 files), **1123 vitest** pass.
  **Boot smoke:** `"todo: registered"` + activated.
- [x] Live-verified (model uses `todo_write` in a real turn).

## youtube_transcript tool (DONE)
Standard tool extension `tool-youtube-transcript` backed by a self-hosted transcriber
service (`http://100.102.55.49:41090`, Tailscale, no API key). One tool
`youtube_transcript` — takes a YouTube URL, fetches the transcript (completed → full
text + timestamped segments; queued/processing → position + ETA + `.youtube_subtitles_pending`
retry convention; failed → error). Pure core: `validateUrl` + `format*` functions +
`truncateOutput`. Injected edge: `TranscriptClient` (injectable `fetchFn`, `AbortSignal.any`
for cancellation). `concurrencySafe: true`, `capabilities: { network: true }`. 30 tests.
Report: `reports/tool-youtube-transcript.md`.

## CLI — cross-client messaging + open tab (DONE)
Roadmap items 2 + 4. The CLI can now list conversations, read the last AI message
(blocking), send messages (blocking or `--queue`), and signal the frontend to open a
conversation tab. Short-ID prefix resolution (4+ chars → full ID via `GET /conversations?q=`).
- **Wave 0 (orchestrator, contracts):** `ConversationMeta` in `@dispatch/wire`
  (`0.8.0→0.9.0`); `ConversationListResponse`, `LastMessageResponse`,
  `OpenConversationResponse`, `SetTitleRequest`, `TitleResponse`, WS
  `conversation.open` in `@dispatch/transport-contract` (`0.12.0→0.13.0`);
  `listConversations()`/`getConversationMeta()`/`setConversationTitle()` on
  `ConversationStore`; new routes declared in transport-http manifest;
  `conversationOpened` hook in session-orchestrator.
- **Wave 1 (conversation-store):** metadata tracking (createdAt on first write,
  lastActivityAt on every append, title from first user message truncated 80 chars);
  `conv-index` key tracks all conversation IDs; `extractTitle` pure helper. 21 new
  tests (81 total).
- **Wave 2 (parallel, transport-http + transport-ws):** `GET /conversations` (list
  with `?q=` prefix filter), `GET /conversations/:id/last` (blocks until turn settles
  via subscribe-then-checkIsActive, returns last assistant text via pure
  `extractLastAssistantText`), `POST /conversations/:id/open` (emits
  `conversationOpened` hook), `PUT /conversations/:id/title`; `emit` threaded from
  `host.emit` → `createApp`. transport-ws subscribes to `conversationOpened` +
  broadcasts `ConversationOpenMessage` to all connected WS clients. 21+2 new tests.
- **Wave 3 (CLI):** `dispatch list` (table: short ID + title + activity),
  `dispatch read <id>` (blocking, prints last AI message), `dispatch send <id> --text`
  (blocking by default; `--queue` for non-blocking enqueue; `--open` signals FE).
  Short-ID resolution (4+ chars → prefix search; 32+ chars = full UUID). 48 new
  tests (108 total).
- Verified: full-graph `tsc -b` EXIT 0, biome clean (327 files), **1240 vitest** pass.
  **Boot smoke + endpoint smoke:** `GET /conversations` → `[]`, `GET /conversations/:id/last`
  → `{content:""}`, `POST /conversations/:id/open` → `{conversationId}`.
- [x] Live-verified end-to-end (CLI → real conversation → FE tab open).

## Workspaces (DONE)
Cross-repo design ask from `../dispatch-web` (`backend-handoff-workspaces.md`).
Outbound courier: `frontend-workspaces-handoff.md` (final shapes + Q1–Q8).
- **Boundary decision:** workspaces live inside `conversation-store` (metadata +
  cwd persistence owner); no new extension. Single owner-agent for all workspace
  storage + service methods.
- **Versions:** `@dispatch/wire` `0.11.0→0.12.0`, `@dispatch/transport-contract`
  `0.15.0→0.16.0`, `@dispatch/ui-contract` unchanged. Kernel re-exports
  `Workspace`/`WorkspaceEntry`.
- **Key decisions:** `DELETE /workspaces/:id` closes all conversations (status→
  "closed") + reassigns to "default" + deletes workspace; auto-create workspace on
  turn start if missing; `PUT /workspaces/:id` create-on-miss with optional
  `title`/`defaultCwd`; `DELETE /conversations/:id/cwd` to clear explicit cwd;
  `GET /conversations/:id/lsp` roots at effective cwd; WS lifecycle push deferred.
- **Waves:**
  - **Wave 0 (orchestrator):** contracts (wire `0.12.0` + transport-contract
    `0.16.0` + kernel re-exports). tsc + biome clean.
  - **Wave 1 (conversation-store):** workspace persistence + service methods
    (`getWorkspace`, `ensureWorkspace`, `setWorkspaceTitle`, `setWorkspaceDefaultCwd`,
    `deleteWorkspace`, `listWorkspaces`, `getWorkspaceId`, `setWorkspaceId`,
    `getEffectiveCwd`, `isValidWorkspaceSlug`); `listConversations` filter;
    `forkHistory`/`replaceHistory` preserve `workspaceId`. 111 bun tests. CRs
    (kernel re-exports, `bun install`) resolved by orchestrator.
  - **Wave 2 (session-orchestrator):** `workspaceId` on `StartTurnInput`/
    `EnqueueInput`; effective cwd resolution (`getCwd` → `getEffectiveCwd`); auto-
    create workspace on turn start; warm parity. 93 vitest (+8).
  - **Wave 3 (parallel):** `transport-http` (workspace routes, `workspaceId`
    threading, `?workspaceId=` filter, `DELETE /conversations/:id/cwd`, effective
    cwd for LSP, slug validation; 166 tests), `transport-ws` (`workspaceId` on
    `chat.send`/`chat.queue`; 32 tests), `cli` (`--workspace`/`-w` flag; 123 tests).
  - FE handoff sent to agent 4091 via `dispatch send --queue` (non-blocking).
- Verified: full-graph `tsc -b` EXIT 0, biome clean (328 files), **1283 vitest +
  199 transport bun** pass (1 pre-existing `tool-shell` failure unrelated).
- **LIVE-VERIFIED** against dev stack (`bin/up`): 11/11 workspace checks pass —
  create-on-miss, rename, set default-cwd, invalid-slug 400, unknown 404, delete-
  default 409, chat with workspaceId stamps conversation, workspace filter, cwd
  inheritance (null = inheriting), delete cascade (closedCount:1, workspace→404).
- `dist/` rebuilt for FE (wire + transport-contract + kernel .d.ts contain Workspace
  types). FE agent 4091 notified twice (handoff + dist-ready).

## Open items
- **`prefix.fingerprint` / `warm|real` cache-bust attributes (deferred):** decoupled
  from dedup by the content-addressed decision; also gated on cache-warming being
  built (not yet) so `warm|real` can't be honestly stamped. Later cache-bust-debug
  milestone (`notes/observability-design.md` §3.1, §12).
- **D9 analytics roll-ups (deferred):** rollup table shape + `GROUP BY` indexes +
  retention asymmetry + periodic rollup job (`notes/observability-design.md` §2 D9,
  §12). The scheduler mechanism (`host.scheduler.register`) already exists.
- **D8 `prompt.assembly` segments:** deferred-by-design (await the context-filter
  chain).
- **In-memory state persistence (message queue + todo list):** both the message
  queue and the todo list are in-memory only (`Map<conversationId, …>` in the
  extension's `activate`). Neither persists across server restarts. If persistence
  is needed later, both would write through `host.storage` (the conversation-store
  pattern: separate key space per feature, append/write per conversation).

## Roadmap
1. **Web frontend** (in progress, SEPARATE repo `../dispatch-web`; Svelte +
   DaisyUI, same methodology). Slice 2 = browser chat MVP consuming the
   wire/transport-contract + metrics. Cross-repo contract changes are couriered
   via the user (ORCHESTRATOR §7); `lsp references` does not span repos.
2. **Message queue — close-with-queued-messages (deferred product decision):**
   if a client closes a conversation (`POST /conversations/:id/close`) while the
   queue is non-empty, the carry currently still fires (starts a new turn on the
   closed conversation). Decide: does closing discard pending steering, or honor
   it? If "discard," gate the carry on `finishReason !== "aborted"` in
   session-orchestrator (one-line). No FE action either way.
3. **FE: consume `GET /conversations/:id/status` for crash-recovery re-sync.**
   Backend endpoint shipped: returns `{ conversationId, isActive, status }` where
   `isActive` is the orchestrator's in-memory truth and `status` is the persisted
   lifecycle status. On reconnect (WS re-establish or page reload), the FE should
   call this for any tab it believes is "generating"; if `isActive: false`,
   override the local spinner to idle regardless of the persisted `status`
   (defense-in-depth against status drift the boot-sweep didn't catch).

(Done and dropped from the list: CLI; dedup / storage growth; message queue +
steering injection; CLI open-tab handoff; `todo` tool; `web_search` tool; tab
persistence across devices; conversation compacting; live-verify steering flow.)

## Stop generation must abort a hanging tool + not brick the conversation (DONE)
FE courier in: "Stop generation doesn't abort a hanging tool call." When the user clicks Stop during
a tool that hangs (e.g. `run_shell` with a blocking/grandchild-holding process), the turn never
sealed → the FE spinner spun forever AND the conversation was bricked (next `chat.send` rejected as
`"already-active"` because `activeTurns` was never cleared).
- **Root cause:** the kernel's `executeToolCall` awaited `tool.execute(...)` with **no race against
  the abort signal** — a tool that ignored `ctx.signal` (or blocked on something it couldn't
  interrupt) blocked `drain` → `runTurn` never returned → session-orchestrator's `finally` (which
  clears `activeTurns`) never ran. (The `/stop` endpoint, `stopTurn`, and the `finally` cleanup were
  already correct — they just needed `runTurn` to return.) Secondary: `realSpawn` resolved on
  `child.on("close")` (waits for stdio) and killed only the immediate child, so a grandchild holding
  the pipes could stall the spawn promise + leak.
- [x] **kernel** — `executeToolCall` now **races** `tool.execute` against `signal` via `Promise.race`;
  on abort it **resolves** (not rejects) `{ content: "Aborted", isError: true }` so the step completes
  normally → kernel's existing `signal.aborted → finishReason "aborted"` path runs → turn seals
  cleanly (`done` + `turn-sealed`) → `finally` clears `activeTurns` → **conversation freed, next
  message accepted**. Late rejections from the orphaned tool promise are swallowed. 11 tests incl.
  the durability test (hanging tool `new Promise(() => {})` + abort → `runTurn` returns
  `finishReason "aborted"`, doesn't hang). Report: `reports/kernel-abort-race.md`.
- [x] **tool-shell** — `realSpawn` spawns `detached: true` (own process group); on abort **and**
  timeout kills the **group** (`process.kill(-pgid, "SIGKILL")`) AND resolves immediately (no
  `close`-dependency) so a grandchild holding the pipes can't stall the spawn or leak. 4 tests
  (grandchild abort, grandchild timeout, normal-completion stdout capture, simple abort). Report:
  `reports/tool-shell-process-group-kill.md`.
- [x] Verified: `tsc -b` EXIT 0, biome clean, **1326 vitest** pass; both in-lane; kernel zero
  internal mocks.
- [x] **Live-verified** (fresh `bin/up`): start a hanging tool (`run_shell` sleep/grandchild),
  Stop, then send a NEW message → it must be ACCEPTED (conversation not bricked) and the spinner
  clears.

## System prompt builder — template-based system context (DONE)
Design: `notes/system-prompt-design.md`. FE courier: `frontend-system-prompt-handoff.md`.
Problem: no system prompt was sent to the provider for regular turns (the messages array
started with the user message; `providerOpts.systemPrompt` was never set). This adds a
template-based system prompt builder with variable placeholders (`[type:name]`) and
conditionals (`[if]`/`[else]`/`[endif]`).
- **Cache constraint (critical):** the system prompt is constructed ONCE (first turn of
  a new conversation) and persisted. Reused on all subsequent turns (no reconstruction —
  cache-safe). Reconstructed only on **compaction** (fresh variable resolution + compaction
  instructions appended).
- **Variable types:** `system:time/date/os/hostname`, `prompt:cwd/model/conversation_id`,
  `git:branch/status`, `file:<path>` (dynamic — any path).
- **Wave 0 (orchestrator, contracts):** `@dispatch/transport-contract` `0.17.0→0.18.0` —
  `SystemPromptTemplateResponse`, `SetSystemPromptTemplateRequest`, `SystemPromptVariable`,
  `SystemPromptVariablesResponse`.
- **Wave 1 — `system-prompt` (NEW ext):** pure parser (29 tests) + variable resolver
  (injected adapters, 12 tests) + catalog (3 tests) + service handle (`construct` +
  `get` + `getTemplate` + `setTemplate`, 8 tests). 52 tests total. Default template:
  persona + AGENTS.md if exists + cwd.
- **Wave 2 (parallel):** `session-orchestrator` (wire service: construct on first turn,
  get on subsequent, construct+append on compaction; 12 tests) + `transport-http`
  (GET/PUT `/system-prompt`, GET `/system-prompt/variables`; 6 tests).
- **Wave 3 — host-bin:** registered `system-prompt` in `CORE_EXTENSIONS`.
- [x] Verified: `tsc -b` EXIT 0, biome clean, **1396 vitest** pass.
- [x] Live-verified (boot smoke: extension activates, `GET /system-prompt` returns default
  template, `GET /system-prompt/variables` returns catalog).
- [x] **FE courier** sent to FE agent `ffe3`: `frontend-system-prompt-handoff.md`.