gstack 1.98.0.0 drops /investigate and /plan-eng-review, and new installs go stable

2 days ago

gstack 1.98.0.0 retires /investigate and /plan-eng-review after evals found Opus 5.5 did as well without them, adds an opt-in gate that keeps /autoplan running until every phase reports, moves new installs to the stable channel, and stops draining GitHub's GraphQL quota. Changelog: https://github.com/garrytan/gstack/blob/main/CHANGELOG.md#19800---2026-11-07 Skill gates: https://github.com/garrytan/gstack/blob/main/docs/skill-gates.md#the-autoplan-gate Commits: https://github.com/garrytan/gstack/commits/main gstack: https://github.com/garrytan/gstack

Ask about this presentation

Chapters

  1. 0:00Two skills retired
  2. 0:13/investigate eval
  3. 0:27Bugs fixed directly
  4. 0:39What stays
  5. 0:54/plan-eng-review
  6. 1:04Why it went
  7. 1:1402 /autoplan
  8. 1:16/autoplan gate
  9. 1:30Gate limits
  10. 1:38Transcript wait
  11. 1:5303 Installs
  12. 1:54Stable channel
  13. 2:06GitHub quota
  14. 2:19Lit templates
  15. 2:3404 Security
  16. 2:36Local only
  17. 2:52Redaction
  18. 3:07Turn the gate on
Show transcript
  1. gstack 1.98 retires two skills, /investigate and /plan-eng-review. Setup stops installing them on every host, and your next setup or upgrade removes the copies already there.

    Slide text
    gstack logo

    gstack 1.98.0.0 drops /investigate and /plan-eng-review, and new installs go stable

    /investigate · removed /plan-eng-review · removed
    Setup stops installing both, and the next setup or upgrade removes the copies you have.
    Retired skills /autoplan Installs Security
    garrytan/gstack · CHANGELOG.md · 1.98.0.0
  2. On real bug fixes, Opus 5.5 with /investigate passed about as often as Opus with no skill at all. The skill made each passing fix cost about 43 percent more.

    Slide text

    Opus fixed bugs about as often without /investigate

    Opus 5.5 told to use /investigate
    50.0%
    Opus 5.5 with no skill
    47.2%
    18 real bug-fix tasks, 20 trials each. A passing fix cost about 43% more with the skill.
    Retired skills /autoplan Installs Security
    garrytan/gstack · CHANGELOG.md · 1.98.0.0
  3. Bug reports now go straight to a fix. The agent checks the cause once without editing, writes a failing test if the cause is still unclear, then fixes the root cause and reruns.

    Slide text

    Bug reports now go straight to a fix

    1. One read-only check of the cause
    2. Still unclear? A test that fails on the bug, before any edit
    3. The root-cause fix, then a rerun
    Retired skills /autoplan Installs Security
    garrytan/gstack · CHANGELOG.md · 1.98.0.0
  4. An investigate skill file you edited moves to a backup folder instead of being deleted. Last release's investigate gate goes with it. One review checklist line still points timing flakes at investigate, for now.

    Slide text

    Your own edits survive the removal

    A SKILL.md you edited moves to a backup folder
    The /investigate gate and its four Claude Code hooks are removed
    /review still sends timing flakes to /investigate, for now
    Retired skills /autoplan Installs Security
    garrytan/gstack · CHANGELOG.md · 1.98.0.0
  5. /plan-eng-review goes too, but its method stays, unchanged, as the third phase of /autoplan. Ask for an architecture review now and gstack sends you there.

    Slide text

    /plan-eng-review lives on as /autoplan’s Phase 3

    /plan-eng-review → /autoplan · Phase 3
    Its methodology is kept byte for byte. Architecture and engineering-review requests now route to /autoplan.
    Retired skills /autoplan Installs Security
    garrytan/gstack · CHANGELOG.md · 1.98.0.0
  6. Opus 5.5 with no gstack at all caught nearly every planted engineering flaw. A session told to use the skill cost about seven times as much.

    Slide text

    Plain Opus already caught the planted flaws

    15 of 16 reviews by Opus 5.5 with no gstack found the planted engineering flaws
    7.1× the cost of a plain session, for one told to use the skill
    Retired skills /autoplan Installs Security
    garrytan/gstack · CHANGELOG.md · 1.98.0.0
  7. Autoplan learns to finish.

    Slide text
    02
    /autoplan
    A gate that holds the session until every phase is done
    garrytan/gstack · CHANGELOG.md · 1.98.0.0
  8. Told to run /autoplan on its own, Opus 5.5 showed none of its phases in 23 sessions. With the new opt-in gate, the session can't end until every phase has published its report.

    Slide text

    /autoplan can now refuse to stop until every phase reports

    0 of 23 sessions showed autoplan’s phases
    Phase 0 scope → each phase report → Final Approval Gate
    Retired skills /autoplan Installs Security
    garrytan/gstack · CHANGELOG.md · 1.98.0.0
  9. A stop that asks you a question always goes through, and after three blocks the gate lets the session end. It ships turned off.

    Slide text

    The gate steps aside when it should

    A stop that asks you a question goes through
    After 3 blocks, the next stop goes through
    Off by default. Un-gated runs are unchanged
    Retired skills /autoplan Installs Security
    garrytan/gstack · CHANGELOG.md · 1.98.0.0
  10. Claude Code can fire its stop event a fifth of a second before it writes the last lines of the transcript. The CEO review gate and the contract hook now wait up to two seconds, so finished work isn't held.

    Slide text

    Stop gates now wait for Claude Code to finish writing

    Stop fires ~200 ms before the last transcript lines land
    The /plan-ceo-review gate and the skill contract hook re-read the transcript for up to 2 seconds.
    Retired skills /autoplan Installs Security
    garrytan/gstack · CHANGELOG.md · 1.98.0.0
  11. Installs.

    Slide text
    03
    Installs
    Which gstack you get, and how often it asks GitHub
    garrytan/gstack · CHANGELOG.md · 1.98.0.0
  12. A fresh clone and setup now follows stable, the tested release that ships twice a week. Every install set up before this keeps following main, so nothing switches under you.

    Slide text

    New installs follow the stable channel

    New install stable: the tested release, twice a week
    Existing install keeps following main. So do contributor checkouts and Capy crews
    Retired skills /autoplan Installs Security
    garrytan/gstack · CHANGELOG.md · 1.98.0.0
  13. Watching one pull request used to spend about 360 GitHub GraphQL calls an hour. gstack now polls over REST, and every wait between GitHub reads lasts at least a minute.

    Slide text

    Waiting on a PR stops draining GitHub’s GraphQL quota

    ~360 GraphQL calls/hour REST, with a 60 s floor
    gstack-pr-wait reads GraphQL only when the PR changed, or every 10 minutes.
    Retired skills /autoplan Installs Security
    garrytan/gstack · CHANGELOG.md · 1.98.0.0
  14. A screen written as a Lit or htm template inside a TypeScript file used to read as backend, so /ship skipped its design check. Now it counts as frontend. A rendered design review still doesn't run inside ship.

    Slide text

    /ship and /autoplan now see UI built from Lit and htm templates

    personas.ts: html`<div class="row">…`
    These files now count as frontend. A rendered /design-review still doesn’t run inside /ship.
    Retired skills /autoplan Installs Security
    garrytan/gstack · CHANGELOG.md · 1.98.0.0
  15. Security.

    Slide text
    04
    Security
    Tighter local daemons and more secrets caught
    garrytan/gstack · CHANGELOG.md · 1.98.0.0
  16. Agents paired over the browse tunnel can now open only web addresses. The design board and iOS QA daemons answer only local clients, and make PDF reads images only from the document's own folder.

    Slide text

    Three local services stop reaching past their job

    Browse tunnel agents open only web URLs and can’t touch the daemon’s disk
    Design board and iOS QA daemons serve only local clients
    make-pdf reads images only from the document’s own folder
    Retired skills /autoplan Installs Security
    garrytan/gstack · CHANGELOG.md · 1.98.0.0
  17. The redaction gate now blocks OpenRouter keys, GitHub App tokens, Stripe restricted keys, Supabase keys and temporary AWS key IDs. And gbrain setup stops putting your database password on the command line.

    Slide text

    The redaction gate catches more credential formats

    sk-or-v1- ghu_ ghr_ rk_live_ sb_secret_ sbp_ ASIA…
    /setup-gbrain also stops putting the gbrain database URL, password included, on the command line.
    Retired skills /autoplan Installs Security
    garrytan/gstack · CHANGELOG.md · 1.98.0.0
  18. If you run /autoplan without watching it, I'd turn the new gate on, because without it Opus can stop after any phase. The full list is in the gstack changelog.

    Slide text

    Run /autoplan unattended? Turn its gate on

    It ships off by default.
    garrytan/gstack · CHANGELOG.md
    Every change in 1.98.0.0, and the gate’s setup in docs/skill-gates.md.
    garrytan/gstack · CHANGELOG.md · 1.98.0.0

What happened in AI today, in your inbox

One short email each evening. Read it and you’re caught up.

Built with ❤️ in New York