gstack

Upstream: github.com/garrytan/gstack — MIT, by Garry Tan (YC). Pinned at SHA 443bde05, VERSION 1.28.0.0. 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, ~L208): feature and refactor fire bare /autoplan (no task argument — the skill auto-discovers the PRD from the repo) with the suffix “When implementation is complete and committed, run /ship to review the diff and finalize.”; bugfix prepends /investigate $SHARED_TASK with the suffix “When the fix is committed, run /ship to review the diff and finalize.” Every task now fires an explicit entry skill per the 2026-04-22 “explicit skill activation” harness change — previously feature and refactor ran with raw $SHARED_TASK and relied on prose triggers; that form produced zero or near-zero Skill tool calls on some runs, so the rerun pins /autoplan (feature/refactor) and /investigate (bugfix) as the primary entry points. The 2026-04-23 follow-up dropped $SHARED_TASK from the /autoplan line after confirming the skill locates the PRD on its own — passing the task text inline was redundant and risked steering the skill away from its own discovery path.

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.

What the transcripts show (session audit)

Numbers below are mean across 3 trials per task (session-audit run-time subsystem is n=3; the score table above is the full n=5 cohort), from scripts/audit-sessions.py. Per-trial JSON: results/gstack/t<N>/session-audit.json (feature), results/{bugfix,refactor}/gstack/t<N>/session-audit.json. Cohort summary: results/_audits/session-audit.md.

Task	wall min	main turns	sidechain turns	sub-agent disp.	files read (config / target)	files edited	cache hit
feature	41.0	155.7	31.0	2.7	21.0 (0 / 21.0)	13.3	0.97
bugfix	5.9	72.0	0.0	0.0	12.7 (0 / 12.7)	1.3	0.97
refactor	46.3	258.7	42.7	2.7	40.3 (0 / 40.3)	32.7	0.98

Three skills drive the runs: autoplan (456 turns), ship (336), investigate (216). The plan→ship pair fires every feature/refactor trial; investigate fires only on bugfix as intended.
Slash commands typed: /autoplan (7), /ship (7), /investigate (3) — /autoplan+/ship fire on every feature/refactor trial (6 trials, with one double-fire on each), /investigate fires on every bugfix trial.
Sub-agent types dispatched: general-purpose (14), Explore (2). Notably ≈ 0 dispatch on bugfix — /investigate’s “no fixes without root cause” phases run as sequential turns in the main thread.
Refactor main-turn count is the cohort’s highest (258.7) — gstack’s preamble + skill-routing chatter adds noticeable per-skill turn overhead, especially when /autoplan and /ship both fire.

Results

Task	Weighted mean /200	within_σ	between_σ	Rank
feature	131.98	18.35	19.41	8 / 8
bugfix	159.97	9.13	16.12	8 / 8
refactor	144.92	58.43	12.52	8 / 8

gstack is the cohort’s weakest performer overall — rank-8 on all three tasks. On feature it sits ≈ 7.5 weighted pts below rank-5 omc (139.49) — the only setup cleanly below the main cluster on any task. within_σ of 18.35 on feature is the highest in the cohort, indicating noticeable run-to-run instability in addition to between-judge spread. The refactor within_σ of 58.43 is a severe outlier — an order of magnitude above the 4–13 range every other tool sits in on refactor — signalling extreme run-to-run variance on that task, which drags gstack’s refactor weighted mean (144.92) far below the rest of the cohort and into rank-8.

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, and lower when the bottleneck is exploratory reasoning the preamble didn’t anticipate. The eng-review gate enforces process cleanliness but does not catch all code-quality issues that judges weight heavily; on this corpus that translates to rank-8 on every task.