gstack’s /land-and-deploy now waits for you to type it, and /runaway-stopper catches runaway agents

1 day ago

gstack 1.95.5.0 to 1.96.0.0: /land-and-deploy and /modernize now wait for you to type them in Claude Code and Cursor, the new /runaway-stopper skill catches runaway agent threads, and /ship's reviewers stop cloning other repositories. Changelog: https://github.com/garrytan/gstack/blob/main/CHANGELOG.md 1.96.0.0: https://github.com/garrytan/gstack/blob/main/CHANGELOG.md#19600---2026-11-05 1.95.5.0: https://github.com/garrytan/gstack/blob/main/CHANGELOG.md#19550---2026-11-04 gstack: https://github.com/garrytan/gstack

Ask about this presentation

Chapters

  1. 0:00Typed commands only
  2. 0:1302 Runaways
  3. 0:15/runaway-stopper
  4. 0:28Rotate, stop, warn
  5. 0:38100M cap
  6. 0:5203 Reviews
  7. 0:53Denials
  8. 1:04Can’t verify
  9. 1:16Fetch rule
  10. 1:28Fewer stops
  11. 1:40/review helps
  12. 1:5504 Browser and design
  13. 1:57Taps land
  14. 2:10Ease out
  15. 2:2205 Setup
  16. 2:23/setup-gbrain
  17. 2:34Secret scanner
  18. 2:47Platforms
  19. 3:00Upgrade
Show transcript
  1. In gstack, Claude Code and Cursor no longer start /land-and-deploy or /modernize on their own. Say land it, and you get a pointer to the command instead of a mistaken merge.

    Slide text
    gstack logo

    gstack’s /land-and-deploy now waits for you to type it, and /runaway-stopper catches runaway agents

    /land-and-deploy /modernize
    “Land it” now gets a pointer to the command. /ship, /careful, /freeze, /guard and /unfreeze still start from plain words.
    Commands Runaways Reviews Design Setup
    garrytan/gstack · CHANGELOG.md · 1.96.0.0
  2. Runaway agents.

    Slide text
    02
    Runaways
    A new skill for agent threads that never stop
  3. The new /runaway-stopper skill finds agent threads that have run away and stops them at a clean point. One Founderstar thread used nearly seven billion tokens in thirteen days.

    Slide text

    New: /runaway-stopper finds runaway agent threads and stops them at a clean point

    6.9B tokens one Founderstar thread, in 13 days
    about 4B its own context, re-read on every step
    Commands Runaways Reviews Design Setup
    garrytan/gstack · CHANGELOG.md · 1.96.0.0
  4. A thread past five hundred million tokens, or active more than three days, hands off to a successor. Captains are only reported, and a thread mid-ship waits.

    Slide text

    A thread past 500M tokens or three days hands off to a fresh successor

    rotate own tokens past 500M · active past 3 days
    stop a subagent past 100M
    warn empty wakes past 25% · over 2 review rounds a PR
    Captains are reported, never rotated. Threads mid-/ship or mid-merge are deferred.
    Commands Runaways Reviews Design Setup
    garrytan/gstack · CHANGELOG.md · 1.96.0.0
  5. A /ship subagent that meters its own cost now stops at a hundred million tokens. A normal run takes about sixty-four million. One subagent reused for five ships reached 485 million.

    Slide text

    A /ship subagent now stops at 100M tokens by default

    one /ship run, median
    64M
    the new default cap
    100M
    one subagent reused for five /ship runs
    485M
    For a subagent whose brief names its title. The cap can be turned off.
    Commands Runaways Reviews Design Setup
    garrytan/gstack · CHANGELOG.md · 1.96.0.0
  6. Reviews.

    Slide text
    03
    Reviews
    What /ship counts as a finding
  7. A reviewer's line that only denied a severity used to become a blocker in /ship, costing a Captain's exception. Now it collects nothing. A file and line still blocks.

    Slide text

    “I found nothing CRITICAL” no longer counts as a CRITICAL finding

    “Neither finding is CRITICAL.”
    before blocker
    now nothing collected
    A file:line, an “except…” clause or a Severity field still blocks.
    Commands Runaways Reviews Design Setup
    garrytan/gstack · CHANGELOG.md · 1.96.0.0
  8. A Codex outside review that ended with can't verify was printed as clean, with no findings. Now it reads as unverified, and a numbered finding without a priority tag still counts.

    Slide text

    A Codex review that couldn’t verify the change no longer reads as a pass

    Verdict: Can’t verify (four required suites blocked)
    VERDICT: clean
    VERDICT: unverified
    A numbered finding with no P0 to P3 tag now counts as a finding.
    Commands Runaways Reviews Design Setup
    garrytan/gstack · CHANGELOG.md · 1.96.0.0
  9. One reviewer cloned a 4.2 gigabyte repository of sealed test conversations. Briefs now forbid fetching, and a stage that downloads one is thrown out and rerun once.

    Slide text

    /ship’s reviewers no longer clone other repositories

    4.2 GB public repo one reader cloned, with sealed held-out eval conversations
    rerun once a review stage that downloads a repo is thrown out
    Caught so far: repos with a network remote and files of 100 MB or more.
    Commands Runaways Reviews Design Setup
    garrytan/gstack · CHANGELOG.md · 1.96.0.0
  10. A timeout-only test fix, one rerun of a flaky check, and a small docs fix after review are now checked and recorded, and the run keeps going. Everything else still goes to the owner.

    Slide text

    Routine /ship stops no longer wait for a Captain

    a test fix that only raises a timeout
    one rerun of a flaky pregate case
    a small docs-audit fix after reviewone scoped reader, not a review round
    Each is checked and recorded in the receipt. Every other stop still goes to the owner.
    Commands Runaways Reviews Design Setup
    garrytan/gstack · CHANGELOG.md · 1.96.0.0
  11. On a development bug suite, /review with its stop hook caught about seven times as many blocking bugs as Opus 5.5 on its own. Nearly all of those bugs came from gstack reviews, so it makes no release claim.

    Slide text

    /review with its Stop hook finds blocking bugs plain Opus 5.5 misses

    /review with the Stop hook · blocking recall
    26.2%
    Opus 5.5, no gstack
    3.6%
    41 of the 42 bugs came from gstack-family reviews. A development suite, so no release claim.
    Commands Runaways Reviews Design Setup
    garrytan/gstack · CHANGELOG.md · 1.96.0.0
  12. The browser, and design.

    Slide text
    04
    Browser and design
    Touch input and motion
  13. On a phone-emulated tab, a tap right after loading was scaled to the desktop width and hit the wrong element. Every tap, swipe and pinch now waits for the page to draw first.

    Slide text

    A tap right after a page loads now lands where you aimed

    $B goto <url> && $B tap 240 440
    landed at (96, 176), on the zoom area
    waits for the first drawn frame
    Under CPU load: 45 of 60 missed before, 0 of 60 after.
    Commands Runaways Reviews Design Setup
    garrytan/gstack · CHANGELOG.md · 1.96.0.0
  14. /design-consultation wrote ease in for exits, and /design-review checked for it. Ease in starts slow while you watch, so a closing menu feels laggy. Both now use ease out.

    Slide text

    /design-consultation now writes exits that ease out

    exit(ease-in) exit(ease-out)
    /design-review flags plain ease-in. Ease-out to enter and exit, ease-in-out to move.
    Commands Runaways Reviews Design Setup
    garrytan/gstack · CHANGELOG.md · 1.96.0.0
  15. Setup.

    Slide text
    05
    Setup
    Fewer false alarms on your machine
  16. One failing check in another agent's skills folder stopped /setup-gbrain, and the agent then offered to edit that workspace. Now only brain checks count.

    Slide text

    /setup-gbrain no longer stops on a working brain

    gbrain doctor also graded other agents’ skills folders, such as ~/.openclaw/workspace.
    now brain checks only
    A failed connection, schema_version or pgvector check still stops setup.
    Commands Runaways Reviews Design Setup
    garrytan/gstack · CHANGELOG.md · 1.96.0.0
  17. The secret scanner stops reading a commit hash as a phone number or a Bitcoin address, and a gstack version number as an IP address. Real phone numbers and addresses still report.

    Slide text

    The secret scanner stops flagging three things that aren’t secrets

    an all-digit sha after review_loop_grant read as a phone number
    a git sha that starts with bc1 read as a Bitcoin address
    gstack 1.94.4.0 in a lane reportread as a public IP address
    Commands Runaways Reviews Design Setup
    garrytan/gstack · CHANGELOG.md · 1.96.0.0
  18. On Windows, the question log hook stops killing a slow first run. On a Mac, gstack doctor works with the built-in bash again. And pregate no longer crashes on very large diffs.

    Slide text

    Fixes for Windows, macOS’s built-in bash and very large diffs

    the AskUserQuestion log hook stops killing a slow first runWindows
    gstack-doctor finds the launch probe againmacOS /bin/bash 3.2
    pregate no longer crashes on large diffsseen on a 335-file diff
    Commands Runaways Reviews Design Setup
    garrytan/gstack · CHANGELOG.md · 1.96.0.0
  19. If Claude Code runs /ship for you unattended, I would upgrade for the fetch rule alone, knowing small downloads still get past it. Details are in the gstack changelog.

    Slide text

    If /ship runs unattended, upgrade for the reviewer fetch rule

    Smaller downloads and extracted archives aren’t caught yet.
    garrytan/gstack · CHANGELOG.md
    Setup details for every change are in gstack’s changelog.
    garrytan/gstack · CHANGELOG.md · 1.96.0.0

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