gstack's DevEx reviews now grade every error message for the AI agent that meets it first

1 hour ago

gstack 1.91.114.0 to 1.91.115.0: the DevEx reviews score every error message for an unattended AI agent and run CLI errors in a sealed capture, /autoplan stops at its round cap and spend budget and exports a public copy of a run, /eval-plan refuses unknown or outdated models before money moves, and a reviewer subagent installs only gstack's read-only helpers. Release notes 1.91.114.0: https://github.com/garrytan/gstack/blob/main/CHANGELOG.md#1911140---2026-10-17 Release notes 1.91.115.0: https://github.com/garrytan/gstack/blob/main/CHANGELOG.md#1911150---2026-10-17 Agent-ready errors: https://github.com/garrytan/gstack/blob/main/docs/agent-ready-errors.md Commits: https://github.com/garrytan/gstack/commits/main

Ask

Ask about this presentation

Answers are generated from this presentation.

Chapters

  1. 0:00Errors
  2. 0:15Capture
  3. 0:31Cap
  4. 0:43Export
  5. 0:55Models
  6. 1:08Reviewer
  7. 1:22Trends
Show transcript

Errors

gstack logo

gstack's DevEx reviews now grade every error message for the AI agent that meets it first

what · fix · recovery · code · context · trace · redaction · channel
Leaked secret, failure reported as success, unsafe fix, block on a terminalError Messages capped at 4/10
Missing fix, recovery, context or tracecapped at 6/10
ErrorsCaptureCapExportModelsReviewerTrends
garrytan/gstack · CHANGELOG.md 1.91.114.0

gstack's DevEx reviews now score every error message for an AI agent running unattended. A leaked secret or a fake success caps the score at four out of ten. A missing fix caps it at six.

Capture

/devex-review now triggers your CLI's errors in a sealed, read-only run

No terminal, stdin closed, temporary cwd and HOME, scrubbed environment, a deadlinegstack-error-capture
Exit code, deadline kill, terminal verdict, secret findingsone redacted JSON record
Credentials, config files or a real directoryneeds --approved
ErrorsCaptureCapExportModelsReviewerTrends
garrytan/gstack · CHANGELOG.md 1.91.114.0

/devex-review now triggers your CLI's errors in a sealed run: no terminal, closed input, a scrubbed environment and a deadline. It returns one redacted record. Passing in real credentials needs your approval.

Cap

Autoplan now stops at its round cap and its spend budget

Round cap reached with open findingsstarted another round
Round cap reached with open findingsaccepted with findings, deferred to the build
Spend budgetenforced the same way
ErrorsCaptureCapExportModelsReviewerTrends
garrytan/gstack · CHANGELOG.md 1.91.115.0

/autoplan now stops when it said it would. A run that reaches its round cap with open findings closes, and hands those findings to the build. The spend budget holds the same way.

Export

Autoplan exports a shareable copy of a run, with your machine paths rewritten

$ gstack-autoplan export --out <run> --public <dir>
A machine path that survives, as in a code spanreported as ARTIFACT_MACHINE_PATH
A target folder holding unrelated filesrefused
ErrorsCaptureCapExportModelsReviewerTrends
garrytan/gstack · CHANGELOG.md 1.91.115.0

Autoplan can now export a public copy of a run, with your machine's paths rewritten. A path it cannot rewrite gets reported. And it refuses a target folder that holds other files.

Models

/eval-plan refuses an unknown or outdated model before any money is reserved

.gstack/eval-models.json, read from the base branchgates every arm
Unknown, denied or older-generation modelrefused before money moves
Preregistered diagnostics now run before the freeze, and a miscount is reported.
ErrorsCaptureCapExportModelsReviewerTrends
garrytan/gstack · CHANGELOG.md 1.91.115.0

/eval-plan now checks every arm's model against your repo's rules. An unknown, denied or older model is refused before any money is reserved. Declared diagnostics now get tested first.

Reviewer

A reviewer subagent installs only gstack's read-only helpers, in seconds

$ --for review-voice
No ./setup, no host profilesBun at gstack's own minimum
A full workflow given that installrefused, with a --for <skill> fix
Autoplan reviewer promptsrun it and relay its gstack: line
ErrorsCaptureCapExportModelsReviewerTrends
garrytan/gstack · CHANGELOG.md 1.91.115.0

A reviewer subagent can now install just gstack's read-only helpers in seconds, without running setup. A full workflow refuses that install and names the fix. Autoplan's reviewers now use it.

Trends

DX trends stop comparing scores across rubric versions

Trend Check and Boomerang print "rubric changed, not comparable" instead of a misleading delta1.91.114.0
AI coding agent is a persona option for CLI tools and Claude Code skills1.91.114.0
Durable autoplan checkpoints are opt-in1.91.115.0
docs/agent-ready-errors.mdThe error contract, in full
ErrorsCaptureCapExportModelsReviewerTrends
garrytan/gstack · CHANGELOG.md 1.91.114.0 · 1.91.115.0

Old and new DX scores are no longer compared: Trend Check and Boomerang say the rubric changed, instead of showing a misleading delta. AI coding agent is now a persona option. And durable autoplan checkpoints are opt in.