[rawphp]

Closes #74

Contract

• done: a MessageRejectedError on a forwarded comment posts the dropped-comment notice with the reason "T3 refused it", then writes comment-dropped <id>; the comment is never sent again. A refused body edit posts the fixed notice ("The factory did not send the edited issue body to the build session, because T3 refused it. Post the change as a new comment and the factory will send it.") then writes body-edit-dropped <edited_at>, which the ledger reader treats like body-edit-forwarded; a newer edit is still sent. A refusal does not stop the call. Refused items are left out of ForwardOutcome, so an all-refused call returns {comments: 0, bodyEdit: false} and the pass does not hold the gate. A forward-wait <k> under which nothing was sent is cancelled with forward-wait-cancelled <k>; the reader clears waitingAfter for that forward marker and restores lastWait, so advance does not answer running. Transport and timeout failures (SessionsError) and a gone thread keep today's behaviour.
• out: the pass's notes wording; the reminder and gate send-back paths (they keep their own refusal handling); the tracker ever producing "refused"; softfac's docs.
• risk: none
• test: vp test run apps/server/src/factory/runner/forward.test.ts apps/server/src/factory/runner/progress.test.ts → pass (117). At the test: commit's tests against the base product code: 13 failed, 104 passed (red confirmed before ship).

Acceptance criteria

• A refused comment is dropped, the notice posted with "T3 refused it", then comment-dropped <id>; never sent again; posted once — forward.test.ts "a send T3 refuses is dropped with a notice, and the call goes on"
• A refused body edit is dropped with the fixed notice, then body-edit-dropped <edited_at>; not sent or reported again; a newer edit still forwarded — "a refused body edit is dropped with a notice and its step, and a newer edit is still sent"; progress.test.ts "keeps the time of every body-edit-dropped step with the forwarded ones"
• A refusal does not stop the call; later comments and the edit go in order — "a refusal between two sendable comments stops nothing", "a refused comment followed by a sendable edit sends the edit and keeps the wait"
• A refused item counts as not sent; an all-refused call does not hold the gate; its forward-wait does not keep advance at running — "a call whose every send was refused holds no gate and leaves no wait" (outcome {0,false}, forward-wait-cancelled 1, advance → ready-for-gate, then a fresh marker on the next send); progress.test.ts "a cancel for the idle marker clears the wait and restores the last wait", "a cancel does not undo a marker a send used"
• Other send failures behave as today — "a failed send (transport) stops the call, and the rest is sent in order under the same ids", "a transport failure after a refusal still stops the call, and the refused item is told next time"; the status-read thread-gone tests are unchanged
• The two ported retry-on-refusal tests are changed to the new rule and say they depart from softfac on purpose — both carry a "Departs from softfac on purpose (A forward that T3 refuses is dropped with a notice instead of retried forever #74)" comment
• New unit tests: refused comment; refused body edit (+ newer edit); refusal between two sendable comments; all-refused pass; crash between notice and step — "a crash before the notice says it on the next call, and the step comes after the comment"; plus "a marker left idle by a crash is cancelled on the next call" and "a refused send under a marker another message used keeps the wait"

Gate

• factory/scripts/check.sh → exit 0 (check-seams; vp test run apps/server/src/factory: 41 files, 1666 tests passed; vp run --filter t3 typecheck; vp run --filter @t3tools/web typecheck). Log: .factory/gate-0.log (not committed).

Models

• maker: claude-sonnet-5-5 (via host; no models block in .dw/config.yml, so Tom's roster)
• checker: claude-opus-5-5 (via host)
• orchestrator: this session, claude-fable-5-1

Review

• cycles: 1
• roster: correctness, reliability (read-only), testing (12 named mutants, all reached, 11 killed at 3aeeda02a)
• fixed: testing-1 (blocker; surviving mutant: deleting !editSent from the cancel condition) and testing-2 (nit) → e353b8d51, a test on an idle session that pins the marker stays when the edit went in; the mutant now fails that test
• leftovers (non-blockers, dropped on purpose):
  • correctness-1 / reliability-1 (nit): if a call crashes between a forward-wait and its cancel, and the refused comment then disappears before the next pass (deleted, or now reported as a tracker drop), the nothing-to-send early return never cancels the idle marker and the run waits until the time limit. Both reviewers note the same stale wait exists today for a crash between the marker and the first send. Out of this issue's scope; a follow-up could cancel seen.idleForwardWait on the early-return path when the session is still idle at that ordinal.
  • reliability-2 (nit): while GitHub comment posts fail, each pass re-sends every not-yet-noticed refused item once (T3 answers the same refusal), writes a marker/cancel pair, and the failed forward holds the gate for the outage. Bounded by the number of refused items and self-healing; acceptable as is.
• not verified live: the change was exercised through the runner's fakes only. No live T3 session or GitHub issue was driven to a refusal in this run.

Lessons

• Read the worktree's branch at Isolate; the harness snapshot can name another. Symptom: the maker's brief said the branch was t3/c19a1b2a (from the session's git status snapshot) while the factory's worktree was on t3/factory-issue-74; the maker noticed and stayed put. Cause: the factory names its worktree branch after the issue, and the snapshot was taken before that. Default: run git branch --show-current at Isolate and put that name in every packet and push.
• In a pnpm monorepo, prove a test: commit red in place, not in a detached worktree. Symptom: a detached worktree needs its own node_modules and a fresh install to run vitest. Default: with the maker and reviewers idle and git status clean, git checkout <base> -- <the product files> (name each path; zsh does not split an unquoted $VAR), run the contract's test files, then git checkout HEAD -- <the same paths> and run once more for green. Record both counts on the receipt.

Commits

• 413e6231e test(factory): pin that a forward T3 refuses is dropped with a notice — Make slice (red: 13 failing)
• 3aeeda02a fix(factory): drop a forward T3 refuses with a notice instead of retrying forever — Make slice (green)
• e353b8d51 test(factory): a sent edit keeps the forward-wait its refused comments shared — Check cycle 1

Files

• apps/server/src/factory/runner/building.ts — forward catches a refusal per send, goes on, counts only what was sent, cancels an idle forward-wait, and tells the issue through one tellIssue helper (replaces reportDrops)
• apps/server/src/factory/runner/progress.ts — BODY_EDIT_DROPPED, FORWARD_WAIT_CANCELLED, Seen.idleForwardWait; the reader treats body-edit-dropped like body-edit-forwarded and applies the cancel
• apps/server/src/factory/runner/messages.ts — the "T3 refused it" reason text and REFUSED_BODY_EDIT
• apps/server/src/factory/tracker/github.ts — DropReason gains "refused" (the runner's own; the tracker never produces it)
• apps/server/src/factory/runner/forward.test.ts — rig can refuse a command id; two pinned softfac tests moved to the new rule; new describe "forward: a send T3 refuses"
• apps/server/src/factory/runner/progress.test.ts — reader tests for the two new steps and idleForwardWait

Run notes

• The worktree was fast-forwarded from 214c0e092 to origin/main (5df529b18) at Isolate before any read or write.
• Repo labels: all seven factory:* labels already existed; none created.

[rawphp]

Factory gate: passed (attempt 1, run by the factory in a fresh checkout of the pull request's head e353b8d merged with the base branch tip 5df529b)

Command	Exit	Time
factory/scripts/check.sh	0	142.4 s