gstack's plan reviews now check a plan's facts before its taste

1 day ago

gstack 1.91.77.0 to 1.91.78.0 (2026-10-11), for Claude Code users: plan reviews (CEO review, eng review, autoplan) now check a plan's facts against the code before judging it, /review sweeps every sibling of a finding, /investigate gets recipes for flaky tests, polluters, races, eval bisects and backlog triage, and /ship runs a two-stage pre-gate before anything paid. https://github.com/garrytan/gstack/blob/main/CHANGELOG.md#191770---2026-10-11 https://github.com/garrytan/gstack/blob/main/CHANGELOG.md#191780---2026-10-11 https://github.com/garrytan/gstack/blob/main/docs/pregate.md

Ask

Ask about this presentation

Answers are generated from this presentation.

Chapters

  1. 0:00Reality
  2. 0:13Sideways
  3. 0:27Investigate
  4. 0:43Pre-gate
  5. 0:56/ship fixes
  6. 1:12Still waiting
Show transcript

Reality

gstack logo

gstack’s plan reviews now check a plan’s facts before its taste

REALITY: <row> pass|finding|n/a <summary> <file:line>
Unattended run, a row missingverdict=incomplete
RealitySidewaysInvestigatePre-gate/ship fixesStill waiting
garrytan/gstack · CHANGELOG.md 1.91.78.0

gstack's CEO review, eng review and autoplan now check a plan's facts before its taste: every claim cites a file and line. In an unattended run, a missing row marks the review incomplete.

Sideways

/review now sweeps every sibling of a finding

Same pattern · every other call site, command and ingress
Severity · the worst sibling’s
Always CRITICAL · data loss, trust-boundary bypass, uncapped spend
RealitySidewaysInvestigatePre-gate/ship fixesStill waiting
garrytan/gstack · CHANGELOG.md 1.91.78.0

/review now searches for each finding's pattern at every other call site and rates it by the worst sibling. Data loss, trust boundary bypass and uncapped spend are always critical. /cso runs the same sweep.

Investigate

/investigate has a recipe for each kind of hard bug

Flake · reproduce on main, never loosen the test
Polluter · bisect the CI test order
Race · force the interleaving
Eval bisect · cheap cached scorer per revision
Backlog · still open only with a repro run
RealitySidewaysInvestigatePre-gate/ship fixesStill waiting
garrytan/gstack · CHANGELOG.md 1.91.78.0

/investigate now ships recipes for flaky tests, test polluters, races, eval bisects and backlog triage. A flaky test gets reproduced on main in a worktree first, and the fix never loosens the test.

Pre-gate

/ship runs a pre-gate before anything paid

Stage 1 · regen, secrets, strays, literalspreflight
Stage 2 · the test lanes your files selectlanes
The tree moves after the runPREGATE_STALE
RealitySidewaysInvestigatePre-gate/ship fixesStill waiting
garrytan/gstack · CHANGELOG.md 1.91.77.0

/ship now runs a two-stage pre-gate before anything paid: quick checks for secrets and stray files, then the test lanes your changes select. If the tree moves, those results go stale.

/ship fixes

/ship closes an issue only when a commit says so

Fixes #NFixes, Closes, Resolves trailer
Refs #Nbare mention, partial work
Generated files · rebuilt in a scratch worktree and diffed
Background lanes · a failure piped into tail now fails
RealitySidewaysInvestigatePre-gate/ship fixesStill waiting
garrytan/gstack · CHANGELOG.md 1.91.77.0

/ship now writes Fixes only when a commit trailer says Fixes, Closes or Resolves; a bare mention becomes Refs. Generated files get rebuilt and compared, and a lane failure hidden behind tail now fails.

Still waiting

Autoplan’s submit step doesn’t run the reality check yet

A parent loop runs the check before submit
Database, performance and search rows load only with --scope
Pre-gate lint, links and guards come in a follow-up
docs/pregate.mdthe pre-gate guide
RealitySidewaysInvestigatePre-gate/ship fixesStill waiting
garrytan/gstack · CHANGELOG.md 1.91.77.0 · 1.91.78.0

Autoplan's submit step doesn't run the reality check yet, so a parent loop runs it first. Database, performance and search rows load only on request, and more pre-gate checks follow. The pre-gate guide lists them.