Why (in your own words)

gstack-gate --help still gives the reply grammar as it shipped in #3115: all · <id><option> tokens · all except <tokens>.

Since 1.91.107.0 (#3190) the parser leads with rec, keeps all as an alias and adds as-is. Everything else already says so:

• the rendered gate list (lib/gate-list.ts:117)
• the GATE_REPLY_UNPARSED fix text in RESULT_CODES (lib/result-codes.ts:117)
• docs/unattended.md

A parent agent that learns the grammar from --help never sees rec or as-is, while the list it relays to the owner leads with rec. Nothing parses differently (all except d1b and rec except d1b resolve to the same choices); the help text was left behind.

The same old grammar also sat in the Fix. paragraph of docs/troubleshooting.md#gate-reply-unparsed. That copy now matches the result-code row.

Fixes #3282.

Why this shape

• The help line is the reporter's expected text, word for word: "Reply grammar: rec (alias all) · rec except <tokens> · <id><option> tokens (d3b uc1a) · as-is. Bare yes/no is rejected."
• The troubleshooting copy isn't in the issue. The docs rule (fix stale copies of the same sentence in the same commit) brings it in, in the order the RESULT_CODES row uses.
• The test doesn't pin the new wording.
  • It reads the reply forms from the rendered list's "Reply with exactly one of:" line, and requires each one in --help's "Reply grammar:" line.
  • The next grammar change then fails the test until --help catches up, and the test never needs the words. That keeps the file's rule: "Prose in the rendering is not pinned."
  • It is red on upstream/main (no `rec`) and green here.
• CONTRIBUTING's wave rule: if two PRs fix the same thing, keep the smaller one. This is 1 commit in 3 files (+12/−4), on current main, and fine to absorb into a wave.

What it does

• bin/gstack-gate: the Reply grammar: line of --help.
• docs/troubleshooting.md: the GATE_REPLY_UNPARSED Fix. paragraph.
• test/gate-list.test.ts: one test in describe('bin/gstack-gate').

What it deliberately does not do

• No parser or render change. all, approve all and the other aliases keep working, and the rendered Reply with line is untouched.
• No single-source refactor (building --help from the render's grammar). The test guards the drift instead. Happy to do the refactor if you'd rather have it.
• Historical copies left alone: docs/designs/* and test/fixtures/multi-agent-wave/* keep the original B9 grammar, as records of the plan.
• Release files and docs: VERSION / CHANGELOG / README / ETHOS untouched, and no generated files. /ship not run (the wave stamps the version).

AI was used for assistance.

Live evidence

Setup. A synthetic one-item gate list in a temp dir. gstack-gate writes nothing.

Before (upstream/main at afac8dfd8): --help and the rendered list disagree.

$ gstack-gate --help | grep "Reply grammar"
Reply grammar: `all` · `<id><option>` tokens (`d3b uc1a`) · `all except <tokens>`. Bare yes/no is rejected.
$ gstack-gate render gate.json | grep "Reply with"
Reply with exactly one of: `rec` (default: take every recommendation, plan changes included) · `rec except <tokens>` · `<id><option>` tokens, e.g. `d1b` · `as-is` (the plan as written).

After (this branch, afff2b927): they agree, and the all alias still parses the same as rec.

$ gstack-gate --help | grep "Reply grammar"
Reply grammar: `rec` (alias `all`) · `rec except <tokens>` · `<id><option>` tokens (`d3b uc1a`) · `as-is`. Bare yes/no is rejected.
$ gstack-gate render gate.json | grep "Reply with"
Reply with exactly one of: `rec` (default: take every recommendation, plan changes included) · `rec except <tokens>` · `<id><option>` tokens, e.g. `d1b` · `as-is` (the plan as written).
$ gstack-gate parse gate.json --reply "all except d1b" --gate-rev 1 --json
{"valid":true,"gate_rev":1,"reply":"all except d1b","choices":[{"id":"d1","chosen":"b","kind":"auto","how":"override"}],"pending":[],"unparsed":[],"mode":"rec"}
$ gstack-gate parse gate.json --reply "rec except d1b" --gate-rev 1 --json
{"valid":true,"gate_rev":1,"reply":"rec except d1b","choices":[{"id":"d1","chosen":"b","kind":"auto","how":"override"}],"pending":[],"unparsed":[],"mode":"rec"}

Scope

• Changed: bin/gstack-gate (+1/−1), docs/troubleshooting.md (+3/−3), test/gate-list.test.ts (+8). No VERSION, no CHANGELOG, no generated files.
• Verified live by:
  • The new test is red on upstream/main, green here.
    • On afac8dfd8 it fails at toContain('`rec`'), receiving the old Reply grammar: line.
    • Here, all 15 tests in test/gate-list.test.ts pass.
  • Every file pregate selects for this diff. origin is my fork, so I pointed a temporary origin/pregate-base ref at upstream/main (afac8dfd8):
    • pregate selects 17 files, and I ran each alone with the runner's --timeout=30000.
    • 16 pass.
    • autoplan-gate-artifacts fails 3 tests here and the same 3 on upstream/main: gstack-autoplan next refuses with "bun 1.4.0 is below the supported minimum 1.4.2". This box has Bun 1.4.0; CI pins 1.4.2.
  • Typecheck: bun run typecheck exits 0.
  • Open PRs on the same files: it merges cleanly with main and with the four open PRs that touch docs/troubleshooting.md (fix(pregate): Dart import graph and test lanes #3285, feat(pregate): user-owned policy file and local lanes #3286, v1.91.153.0 fix(outside-review): no-findings counts only as the review's conclusion; add regressions (#3294) #3304, v1.91.161.0 fix(pregate): Django, Dart, asset and tool-config lanes; init drafts from the repo (#3266-#3269) #3306).
  • Full bun run test: not green on this Linux box, on upstream/main either.
    • This run on afff2b927 failed 365 tests in 69 files.
    • The main causes:
      • Bun 1.4.0 here is below gstack's 1.4.2 floor.
      • Chromium can't launch here (no Xvfb).
      • The /cso suites need bun run build:cso and Docker.
    • pregate's import graph links one of the 69 files to this diff: autoplan-gate-artifacts, above.
    • Against my previous full run on this box (another branch, on 2d42e4ece), no file fails now that didn't fail then, and every common file has the same count.
    • The serial shard 17 (bootstrap-retention) passed its 36 tests but was marked failed by the home-containment check: ~/.claude/sessions/<id>.json changed while it ran, and that file belongs to a live Claude Code session on this machine. No new ~/.gstack/projects entries.
    • CI's Linux free suite is the gate.
  • slop-scan: 706 findings on both trees, none in the changed files.
  • Paid evals: bun run eval:select --base upstream/main picks the whole PR gate (108 behaviors, 24 judges), because every gate case reaches bin/gstack-gate. Fork PRs get no eval secrets, so I couldn't run it. The change is the --help text only: render, parse and decisions output is byte-identical (the two Reply with lines above), so a case that relays or answers a gate is unaffected.
  • Redaction: bin/gstack-redact --repo-visibility public on the diff reports no findings. git diff --check is clean.
• Did NOT test: Windows. CI's Windows Free Tests covers it.

Liveness proof (required for external contributors)

Checklist

• Liveness screenshot attached (GSTACK PR typed live into a real surface) or PR author is @garrytan (owner exemption)
• This is not a generated-file-only diff (I edited the source/template and regenerated)
• No ETHOS.md edits, and no changes to voice / founder perspective / YC references
• New public command / external service / host adapter has an accepted issue linked (or N/A)
• Linked issue or reproduction: gstack-gate --help still documents the pre-1.91.107.0 reply grammar (all), without rec and as-is #3282

Test plan

• The new test is red on upstream/main and green here
• Every file pregate selects for this diff (above)
• Before/After evidence for --help, the rendered list and the alias
• Full bun run test: the same environment failures as upstream/main; none comes from the changed files
• CI free tests on this PR (Free Tests and Windows Free Tests)

Fork PRs do not get eval secrets; the free files above are the gate.

Fixes #3282