Adapter: find a project, launch a thread and read its status in-process - #36

Closes #5

Contract

• done: FactoryPaths gives baseDir, worktreesDir and <base>/factory from ServerConfig. FactorySessions ports softfac's Sessions trait in-process: findProject over ProjectStoreV2 (exact root, live projects only); launch over ThreadLaunchService (new worktree from baseRef, or the origin tip when startFromOrigin, full-access, default interaction, the caller's commandId, so the same id again returns the same thread and starts nothing); status and thread over ProjectionStoreV2 with softfac's run selection (highest ordinal, skipping a cancelled run that never started unless every run is one; that run's own last assistant message; branch, worktreePath, archived). thread reads the thread row alone.
• out: send, archive, events.ts, process.ts, wiring into FactoryLayer, HTTP routes, ledger.
• risk: none
• test: vp test run apps/server/src/factory → pass (42 tests); vp run --filter t3 typecheck → pass

Acceptance criteria

• sessions.ts provides FactorySessions with findProject, launch, status, thread, softfac's inputs and the ten run statuses — sessions.ts (RUN_STATUSES, LaunchRequest, FactorySessions); sessions.test.ts "maps every known run status".
• status reports the latest run leaving out a cancelled run that never started, that run's last assistant message, and branch / worktreePath / archived — sessions.test.ts threadStatusOf block: 16 cases ported from softfac src/t3.rs plus "reports only the reported run's own last assistant message" and "reports the last of the run's assistant messages, not the first"; "reads a real recorded snapshot" against softfac's t3_thread_snapshot.json.
• thread reports the thread fields alone and works when the run cannot be read — sessions.test.ts "thread reads the thread's own fields and works when the runs cannot be read" (getThreadProjection fails, getThread answers).
• launch starts a thread in a new worktree from baseRef / origin tip, full-access, default interaction, given model, title, message, under the caller's commandId; a repeated commandId returns the same thread and starts nothing — sessions.test.ts "starts a thread in a new worktree, full-access, default interaction, under the caller's command id" (asserts every field of the ThreadLaunchInput, including startFromOrigin: true); sessions.contract.test.ts (second launch → same thread id, runOrdinal stays 1, run stays completed).
• paths.ts gives base dir, worktrees dir and <base>/factory from ServerConfig — paths.test.ts.
• sessions.contract.test.ts runs the real orchestration layers with the provider-testing replay (Claude simple transcript through ProviderReplayHarness.layerWithRegistry, real ThreadLaunchService, real git worktree through GitVcsDriver): launches, waits for run.updated to reach a terminal status through EventSinkV2.stream, reads "fixture simple ok", checks the worktree's .git exists under the worktrees dir, that git branch --show-current equals status.branch and HEAD equals the base ref, then repeats the commandId. Runs in about 5 s.
• Boundary test still passes; only factory/t3/ imports T3 internals — boundary.test.ts in the gate run; factory/scripts/check-seams.sh exit 0 (no upstream file touched, no new seam).

Gate

• vp test run apps/server/src/factory → exit 0 (6 files, 42 tests) — .factory/gate-1.log
• vp run --filter t3 typecheck → exit 0 — .factory/gate-2.log

Models

• maker: claude-fable-5-1 (this session; .dw/config.yml has no models block, so Tom's roster applied, and Make ran in-session after the Contract research rather than in a sonnet-5-5 worker)
• checker: claude-opus-5-5 (host subagents: correctness, testing)

Review

• cycles: 1 (correctness: approve, 1 action; testing: request changes, 1 action, 10 of 12 mutants killed, 1 equivalent). The cycle-1 fix is a few lines; the orchestrator read the diff and shipped, per the testing review budget.
• fixed: blank commandId / projectId / instanceId / threadId threw a schema defect from CommandId.make instead of failing as SessionsError / ThreadNotFoundError (softfac returns InvalidInput); findLast vs find for the last assistant message had no test with two assistant messages in one run.
• leftovers: none. Equivalent mutant: archived with archivedAt === undefined (T3 types it NullOr, never undefined).

Notes

• Isolate: the launched worktree was one merge behind origin/main (issue Factory skeleton: an empty FactoryLayer wired in through one seam #1 had landed); fast-forwarded e4bcce9f6 → f1dba9a17 before provisioning. The first vp i failed on a transient .git/config lock in the primary checkout; the retry succeeded.
• Labels: all seven factory:* labels already existed; none created.
• creationSource is "server" (softfac sent "web" over the wire). The correctness reviewer checked: nothing in T3 branches on it for createdBy: "user", and in-process callers (serverRuntimeStartup, UsageLimitRecoveryWorker) already use "server".
• The test: commits were red before their fix: f2e3f81f4 (modules missing), 55551d4a5 (2 failures observed before 24a0161ce). The two-assistant-messages case in 55551d4a5 was green already; it pins findLast.
• Ship: the first gh pr create (no -R) opened the PR on upstream pingdotgg/t3code as Adapter: find a project, launch a thread and read its status in-process pingdotgg/t3code#19865, because gh resolves a fork's parent as the default repository. Closed it at once with a comment; nothing else was written there. The real PR is this one, opened with -R original-solutions/t3code-factory.
• softfac's RunStatus::Unknown has no counterpart: in-process the projection status is a closed literal union, so a new upstream status fails typecheck instead.

Lessons

• On a fork, gh pr create needs -R <owner/repo>. Symptom: the draft PR landed on the upstream repository the fork was cut from, with the fork issue's Closes #n pointing at a stranger's issue. Cause: gh picks the parent repository as the default for a fork's remote. Default: always pass -R <factory repo> (and --head <branch>) to gh pr create, gh pr edit and gh pr list, and check the printed URL's owner before reporting it.
• A contract test over a real runtime needs the live clock and a diagnostic that survives a pass. Symptom: the test hung to the timeout with no error; Effect.timeoutOption never fired; console output was invisible once the test passed. Cause: it.effect runs under a TestClock the runtime never advances, and vitest hides stdout of passing tests. Default: use it.live for runtime contract tests, and when a wait hangs, fork a stored-event logger that appends to a file under .factory/ before the action under test, then compare where the pipeline stopped against where the wait subscribed.
• Check an unfamiliar stream combinator's signature in the library source before building a wait on it. Symptom: Stream.filterMap returning Option.none() matched nothing in Effect 4 (it takes a Filter), so every variant of the wait returned empty. Default: grep "^export const <name>" node_modules/<lib>/src/<Module>.ts before the first use; prefer Stream.filter + Stream.take(1) + Stream.runCollect for "first matching event".
• Run the gate's typecheck before the first product commit, not only before Ship. Symptom: six type errors (deterministic service keys, branded ids, mock shapes) in files whose tests were already green. Cause: vitest transpiles without typechecking. Default: typecheck right after the first green test run.
• Re-read a file after a commit before patching it by exact string. Symptom: three string-replace patches applied to nothing. Cause: the repo's pre-commit hook reformats staged files, so the text on disk no longer matched what the session wrote. Default: after any commit, grep -n the target lines and patch against what is there.

Commits

• f2e3f81f4 test(factory): sessions and paths adapter cases — Make slice (red)
• 53d91d5b8 feat(factory): sessions adapter over T3's launch and projection services — Make slice
• 55551d4a5 test(factory): blank ids fail typed, last assistant message is the last — Check cycle 1 (red)
• 24a0161ce fix(factory): blank ids fail as the declared sessions errors — Check cycle 1

Files

• apps/server/src/factory/t3/sessions.ts — the adapter: types, threadStatusOf, FactorySessions service and layer
• apps/server/src/factory/t3/paths.ts — FactoryPaths from ServerConfig
• apps/server/src/factory/t3/sessions.test.ts — ported softfac status and find_project cases; launch, status, thread over mocks and a real in-memory ProjectStoreV2
• apps/server/src/factory/t3/sessions.contract.test.ts — real orchestration layers, Claude replay transcript, real git worktree
• apps/server/src/factory/t3/paths.test.ts — paths from a ServerConfig
• apps/server/src/factory/t3/fixtures/t3_thread_snapshot.json — softfac's recorded thread snapshot, copied from tests/fixtures