@specforge/canary-cli 0.2.16 → 0.2.17

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@specforge/canary-cli",
-  "version": "0.2.16",
+  "version": "0.2.17",
   "description": "MCP server for SpecForge - AI agent integration",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",
@@ -19,9 +19,9 @@
     "ora": "^8.2.0",
     "zod": "^3.25.0",
     "@specforge/api-types": "0.4.17",
-    "@specforge/report-types": "0.4.9",
+    "@specforge/spec-types": "0.4.17",
     "@specforge/session-types": "0.4.14",
-    "@specforge/spec-types": "0.4.17"
+    "@specforge/report-types": "0.4.9"
   },
   "devDependencies": {
     "@types/inquirer": "^9.0.9",
@@ -33,7 +33,7 @@
     "tsx": "^4.7.0",
     "typescript": "^5.0.0",
     "vitest": "4.0.18",
-    "@specforge/lifecycle": "0.3.22"
+    "@specforge/lifecycle": "0.3.23"
   },
   "engines": {
     "node": ">=18.0.0"
@@ -61,7 +61,7 @@
     "@specforge/session-types",
     "@specforge/report-types"
   ],
-  "gitHead": "5f09988626cf09d1d0f400bef277fac50c962344",
+  "gitHead": "64fe3cf4657914ccecc2c26c67c09d6890597a0a",
   "scripts": {
     "build": "tsup && tsc --emitDeclarationOnly --outDir dist",
     "typecheck": "tsc --noEmit",

package/src/cli/templates/agents/content/core/sfag-orchestrator.ts

@@ -103,9 +103,17 @@ This is how implementation runs. Internalize it before dispatching anything.
 - **Each worker is fully autonomous.** It picks up its ticket, runs the whole SWS → action_work_session
   → complete_work_session loop, records every dimension through the assay, commits, and finalizes
   \`active → done\` with no human touch. You do not step inside a worker's loop.
-- **Isolate the workers.** Give each worker its own git worktree/branch (use the \`branchPrefix\` from
-  config, e.g. \`ticket/<ref>\`) so concurrent sessions don't collide on the worktree. SWS enforces a
-  clean worktree per session.
+- **Isolate the workers — one worktree each, and hand each worker its path.** Before dispatching a
+  worker, create a fresh git worktree/branch for its ticket (\`git worktree add\`, using the
+  \`branchPrefix\` from config, e.g. \`ticket/<ref>\`). Then **tell the worker its worktree path** and
+  instruct it to pass that path as \`worktree\` on \`start_work_session\` and \`complete_work_session\`.
+  This is not optional for a parallel fleet: your workers all run under **one shared MCP server**
+  (one process, one working directory), so the MCP-local git evidence — clean status (at
+  start_work_session) and commit hash (at complete_work_session) — is read from THAT one directory
+  unless each worker names its own worktree. Two
+  concurrent sessions that end up in the same worktree are **denied** (\`concurrent_worktree\`); that
+  denial is the signal a worker was launched without its own worktree/path. SWS also enforces a clean
+  worktree per session.
 - **Respect the DAG.** Only \`ready\` (dependency-free) tickets are dispatchable. When a worker completes
   a ticket, the readiness cascade unblocks its dependents (\`pending → ready\`); you then dispatch the
   newly-ready ones. Never dispatch a ticket whose dependencies aren't \`done\`.
@@ -152,7 +160,9 @@ get_next_actionable_tickets({ specificationId, limit }) // the ready tickets to
 get_implementation_status({ projectId, specificationId, status: "active" }) // who is already running
 \`\`\`
 Then dispatch:
-- Launch one \`sfag-ticket-implementer\` per \`ready\` ticket, each in its own worktree/branch.
+- Launch one \`sfag-ticket-implementer\` per \`ready\` ticket. For each, first \`git worktree add\` a
+  fresh worktree/branch, then hand the worker its worktree path (it passes it to SWS/CWS as
+  \`worktree\`). A single non-parallel worker can run in the current directory with no worktree arg.
 - Bound concurrency by the config: at most \`maxParallelEpics\` epics in flight and \`maxTicketsPerTeam\`
   tickets per epic team. If the strategy is \`single\`, run one worker at a time; \`parallel\` runs
   independent epics concurrently; \`phased\` runs the DAG in dependency-ordered phases; \`auto\` picks
@@ -258,7 +268,9 @@ sfag-ticket-implementer (ticket C, worktree C) ─┘   poll get_implementation_
 - ❌ Don't launch a worker without a spec. If no spec, HALT and hand planning to the main agent.
 - ❌ Don't try to create a spec, and don't delegate spec creation to any subagent. Planning is main-conversation-only.
 - ❌ Don't dispatch a ticket out of dependency order. Only \`ready\` (dependency-free) tickets are dispatchable.
-- ❌ Don't run workers in the same worktree. Give each its own worktree/branch or SWS collides on git-clean.
+- ❌ Don't run workers in the same worktree, and don't dispatch a worker without giving it its own
+  worktree path to pass to SWS/CWS. Concurrent sessions sharing a worktree are DENIED
+  (\`concurrent_worktree\`) and their git evidence would otherwise be the launcher's, not the worker's.
 - ❌ Don't create the ImplementationSession yourself. The first worker's SWS creates it (first-write-wins).
 - ❌ Don't wait for a review/approval step — there isn't one. Workers self-complete through the CWS gates.
 - ❌ Don't resolve or unblock a discovery yourself. Hand it to \`sfag-work-resolver\`; the human unblocks in the web app.

package/src/cli/templates/agents/content/core/sfag-ticket-implementer.ts

@@ -98,6 +98,21 @@ So, before every \`start_work_session\`:
 (If you are not inside a git worktree, SWS still runs — the precondition treats a
 non-git caller as clean.)
 
+### Pass your \`worktree\` when you run in parallel
+
+The MCP-local git reads (clean status, commit hash) resolve in the MCP server's current
+directory. When several workers run **concurrently under one orchestrator, they share one
+MCP server** — so each worker MUST tell SWS/CWS which worktree it lives in, or every
+worker's git evidence would come from the launcher's directory instead of its own.
+
+- If the orchestrator gave you a worktree path, pass it as \`worktree\` on **every**
+  \`start_work_session\` and \`complete_work_session\` call (use the same value for both).
+- Two active sessions in the **same** worktree are denied (\`concurrent_worktree\`) — that
+  denial means you are missing (or sharing) a worktree; create your own (\`git worktree
+  add\`) and retry with its path.
+- Running a **single** session in the current directory? Omit \`worktree\` — the current
+  directory is used, exactly as before.
+
 ## Step 1: Pick a ticket (autonomous)
 
 Pick the next actionable ticket yourself — no need to ask anyone:
@@ -111,7 +126,8 @@ Keep going through ready tickets until none remain.
 
 Ensure the worktree is clean (see the git-clean precondition), then:
 \`\`\`
-start_work_session({ ticketId })
+start_work_session({ ticketId })                       // single session, current directory
+start_work_session({ ticketId, worktree: "/abs/path/to/your/worktree" })  // parallel worker
 // ready → active; returns the full orientation + workSessionId (the sessionId).
 \`\`\`
 Read the orientation in full. It IS your brief:
@@ -124,7 +140,9 @@ Read the orientation in full. It IS your brief:
 
 If SWS comes back \`outcome: "denied"\`, read the guidance and act on it: \`git_dirty\` →
 commit and retry; \`unfinished_dependencies\` → a dependency is not \`done\`, pick a
-different ready ticket; an aborted implementation → stop, this spec is off-limits.
+different ready ticket; \`concurrent_worktree\` → another active session is in your
+worktree, create your own (\`git worktree add\`) and retry with its \`worktree\` path; an
+aborted implementation → stop, this spec is off-limits.
 
 ## Step 3: Implement + record — the assay-driven order
 
@@ -241,7 +259,10 @@ complete_work_session({
   ticketId,
   summary: "Implemented user registration: service, endpoint, validation, tests",
   filesCreated: [...], filesModified: [...],
-  validation: { tests: { passed: 8, failed: 0 } }
+  validation: { tests: { passed: 8, failed: 0 } },
+  // If you passed a worktree to start_work_session, pass the SAME one here so the
+  // commit hash is read from YOUR worktree (omit it for a single session).
+  // worktree: "/abs/path/to/your/worktree"
 })
 // active → done; finalizes the WorkSession; the cascade unblocks dependents.
 // When every spec ticket is done, the spec → done and the ImplementationSession finalizes.
@@ -330,6 +351,7 @@ ticket — never a reset.
 ## Quality Gate (self-check before complete_work_session)
 
 - [ ] Worktree was clean before start_work_session (no \`git_dirty\` denial)
+- [ ] If running in parallel: passed your \`worktree\` path to start_work_session AND complete_work_session
 - [ ] All implementation steps marked done via mark_implementation_step_completion
 - [ ] Tests run and recorded via record_test_result; every skip/failure justified
 - [ ] Every touched/divergent file recorded via record_file_change; divergences justified