gstack

Upstream: github.com/garrytan/gstack — MIT, by Garry Tan (YC). Pinned at the single-branch depth-1 clone taken on 2026-04-15; skill manifests declare version: 1.0.0, the repo’s VERSION file reports 0.17.0.0, and package.json reports 0.16.2.0. Runtime stack: Claude Code + Bun v1.0+ (plus Node on Windows). Marketed as “Garry’s Stack — Claude Code skills + fast headless browser. One repo, one install, entire AI engineering workflow.”

What it is

gstack is a skill pack — 37 directories under config/gstack-t1/skills/ — that frames Claude Code as a simulated product team: CEO (/plan-ceo-review), eng manager (/plan-eng-review), designer (/plan-design-review, /design-review), reviewer (/review), QA (/qa, /qa-only), security officer (/cso), release engineer (/ship, /land-and-deploy, /canary), plus debugging (/investigate), planning (/autoplan, /office-hours), retros, scope locks (/freeze, /guard), and a headless-browser binary. The installer (./setup --no-prefix) drops each skill into $CLAUDE_CONFIG_DIR/skills/, where Claude Code reads them as top-level slash commands.

Every skill begins with a long # Preamble (run first) Bash block that probes ~/.gstack/ for telemetry, proactive-suggest, routing, and vendoring markers, optionally appends a “Skill routing” block to the project CLAUDE.md, and (when proactive is on) auto-invokes peer skills from conversational triggers — "why is this broken" routes to /investigate, "ship it" to /ship, "architecture review" to /plan-eng-review, and so on. Each skill’s description: frontmatter explicitly names its trigger phrases; that prose is what Claude Code matches against the user turn.

Benchmark configuration

• Config dir: config/gstack-t1/. settings.json contains only {"skipDangerousModePermissionPrompt": true}; no MCP, no hooks, no allow-list. All behaviour comes from the skill tree. Install shim in scripts/setup-tool-config.sh (gstack case, ~L176) clones upstream into $TOOL_CONFIG/skills/gstack and runs HOME=$TOOL_CONFIG ./setup --no-prefix so ~/.gstack/ writes stay inside the trial.
• Launch prompt (scripts/manual-bench.sh, ~L158): feature and refactor run the raw $SHARED_TASK with the suffix “When implementation is complete and committed, run /ship to review the diff and finalize.”; bugfix prepends /investigate $SHARED_TASK with the same /ship tail. /investigate is the only gstack entry point fired by command — everywhere else, activation is left to the skill’s prose triggers.
• Plan mode is deliberately off for gstack (pipeline.md §1; manual-bench.sh L47-54). The rationale in the script comment: “gstack deliberately excluded — its /ship workflow already runs an eng review gate; we evaluate gstack on its native surface without forcing Claude’s plan mode.” Only pure and mindful get --permission-mode plan; every tool with its own planning surface (bmad, omc, ecc, compound, gstack, …) runs without it so results reflect the setup’s own gate, not Claude’s.

How the entry points behave

/investigate is a four-phase debugger enforcing what the skill calls the Iron Law: no fixes without root cause. Phase 1 is root-cause investigation with a regression-diff check; Phase 2 is pattern analysis with an optional sanitized web search; Phase 3 is explicit hypothesis confirmation via temporary logs/assertions before any edit; Phase 4 is the minimal fix; Phase 5 writes a regression test and a capture-learnings entry to ~/.gstack/projects/<slug>/learnings.jsonl. PreToolUse hooks on Edit/Write call freeze/bin/check-freeze.sh to enforce a scope lock.

/ship is where gstack’s “eng-review gate” lives — the reason gstack is excluded from plan mode. It runs: Step 0 platform detection → Step 1 pre-flight + Review Readiness Dashboard (tallies prior /plan-ceo-review, /codex review, /plan-eng-review, /plan-design-review, /plan-devex-review runs from ~/.gstack/ logs; verdict is NO REVIEWS YET until they’ve run) → Step 2 merge base branch → Step 2.5 test-framework bootstrap → Step 3 tests, with a Test Failure Ownership Triage that classifies failures as in-branch vs pre-existing and, on collaborative repos, can open and assign a GitHub issue via gh. Only after review+tests+eval-suites clear does it bump VERSION, update CHANGELOG, commit, push, and open the PR.

Results

Rank 4 / 9 overall, z̄ = +0.071, top-4 tie with bmad / pure / ecc. Per-task: feature rank 2 (z = +0.158, mean 123.65), bugfix rank 5 (z = +0.165), refactor rank 8 (z = −0.111). sigma_round on feature is 3.34, the highest single-task round-to-round spread in the cohort. Self-preference is negative on feature (−3.22) and strongly positive on bugfix (+9.0) and refactor (+8.5) — gstack under-rates its own feature work relative to the panel. Auto-metrics: feature 11 files / +772 / −4 / 94 tests pass / 10 ESLint errors; bugfix 2 files / +185 / 0 / 15 test failures / 10 ESLint errors; refactor 9 files / +121 / −71 / 0 test failures / 7 ESLint errors. The feature session shows Claude auto-routing to /autoplan from the task prose before any user-typed slash command (50 Bash / 18 Read / 7 Edit / 5 Write / 1 Skill invocation across 239 log lines).

Where it fits

gstack trades blank-prompt flexibility for a fixed organisational metaphor. The ceiling is high when the task lines up with a role gstack already has — greenfield feature work, PR shipping — and lower when the bottleneck is exploratory reasoning the preamble didn’t anticipate. On this benchmark the eng-review gate was enough to tie with the leaders on feature work but didn’t rescue the refactor run.

Observed in trial timelines

Skill activations land at mean 0.5 (feature), 1.0 (bugfix), 1.0 (refactor) — the auto-route to /autoplan from task prose is visible in feature t1 but does not always fire. Subagent dispatch is light (mean 2.2 / 1.0 / 1.0) and Bash dominates on refactor (mean 36) — the /ship test+lint+gate sequence is Bash-heavy.

Detail: see the per-trial timeline files linked below.

Trial timelines

Per-trial event timelines auto-extracted from session-logs/*.jsonl — skill activations, plugin/skill file reads, subagents dispatched, code mutations, Bash usage:

• feature
• bugfix
• refactor