515191f07d
* chore(#3677): checkpoint design artifacts (gitignored, dev-only) * test(#3677): add failing regression test for the crash-window duplicate-dispatch gap (RED) Independently re-traces resume-mode.md/planner-wave.md/worktree-dispatch.md/ merge-wave.md and src/quick-batch.cts's resumeBatch (lines 894-899) and confirms the prior research pass's Open Question 1: a coordinator crash between Step 6 (executor commits, SUMMARY.md written) and Step 7 (merge) leaves BATCH.json at "pending" with no STATE.md row yet (only written in Step 9), so --resume's eligibility re-derivation would dispatch a second executor into a new worktree for the same item, orphaning the first. This test asserts worktree-dispatch.md's Step 6 excludes an item whose SUMMARY.md already exists from the spawn set, mirroring planner-wave.md's existing PLAN.md-existence check one layer earlier. Fails against the current worktree-dispatch.md, which has no such guard. See .gsd/phase/feat-3677-quick-batch-hardening-acceptance/40-design.md §1 for the full trace and fix-location rationale. * fix(#3677): guard worktree-dispatch.md against re-dispatching an already-executed item (GREEN) worktree-dispatch.md's Step 6 re-derives eligibility every dispatch round via the same quick-batch resume call resume-mode.md uses, but had no check for "did this item already finish executing" the way planner-wave.md already checks "did this item already get planned" (PLAN.md existence) before re-planning. A coordinator crash between Step 6 (executor commits, SUMMARY.md written) and Step 7 (merge) left the item eligible for a second dispatch on --resume, orphaning the first worktree's real, already- committed work and silently losing it once the second executor's SUMMARY.md write clobbered the first at the same item_dir path. Adds a SUMMARY.md-existence exclusion before spawn-plan is computed, symmetric to planner-wave.md's PLAN.md check. The excluded item is not lost: merge-wave.md's own mergeable-wave criterion (status=pending, SUMMARY.md on disk, not yet merged) already picks it up independently of this eligible/spawn list. Workflow-prose-only fix — touches no already-merged/reviewed .cts module. See .gsd/phase/feat-3677-quick-batch-hardening-acceptance/40-design.md §1 for the fix-location rationale (why not resumeBatch itself). * test(#3677): add real-git coverage for worktree-ownership tampering, scope drift, and submodules Closes the three coverage gaps identified in 40-design.md §2/§3 (#3677, epic #3344 Phase 5's own AC bullets: "arbitrary-worktree ownership attempts", "scope drift", "submodules"): - Arbitrary-worktree ownership tampering: a manifest entry naming a non-agent branch is silently dropped at normalization before any git subprocess runs; a manifest entry naming a plausible agent-branch that was never actually created by this repo's own worktree.create (a genuinely foreign repo/branch) is blocked via base_mismatch. Both leave the foreign location and repoRoot's HEAD provably untouched. - Advisory scope drift: a committed path outside declared files_modified still merges successfully (advisory, never blocking) while surfacing a scope_out_of_declared warning naming the drifted path; an exact declared-scope match produces zero warnings (boundary case). - Real .gitmodules submodule integration: a repo containing a real local git submodule merges cleanly through executeWorktreeWaveCleanupPlan for an unrelated plan; a real gitlink pointer bump (declared) merges cleanly with the superproject tree reflecting the new pinned commit; an undeclared bump is advisory-only and surfaces a scope warning naming vendor/sub, same as any other undeclared modification. No src/*.cts changes — all three gaps were coverage-only; the underlying primitives already behaved correctly (independently verified against real git subprocess output before writing each assertion). * docs(#3677): document how to diagnose a preserved quick-batch worktree Extends the one-sentence "worktree is preserved (never deleted)" mention into a concrete diagnosis procedure: where the preserved directory is, how to read the executor's real commits/diff against the plan's declared files_modified, how to read the item's own SUMMARY.md independent of merge outcome, how to manually merge-and-clean-up or discard, and how to re-run --resume afterward. Also documents that a SUMMARY.md-written-but-still- pending item (the crash-window case fixed in this same PR) needs no manual intervention — --resume routes it straight to the merge step. * chore(#3677): checkpoint final acceptance-evidence mapping (gitignored, dev-only) * fix(#3677): make crash-window duplicate-dispatch guard behaviorally provable and durably recoverable Orthogonal review (Spec finding): the crash-window regression test added earlier this phase only asserted readStep('worktree-dispatch.md') + regex matches against the markdown prose — proving the DOCUMENTATION says the right thing, never that the runtime condition (pending status + on-disk SUMMARY.md + absent STATE row) is actually handled correctly. #3677's own "Alternatives considered" explicitly rejects "document recovery without fault injection" for exactly this reason. Extracts the filtering decision into a pure, independently testable function, filterAlreadyExecuted(eligibleIds, executedIds) in src/quick-batch-dispatch.cts, wired to a new `quick-batch filter-executed` CLI verb (src/quick-batch-command-router.cts) — the same pure-decision- then-CLI-wired pattern computeSpawnPlan/computeMergeOrder already establish. worktree-dispatch.md now calls this verb explicitly instead of only describing the decision in prose. A genuine fixture-based test in tests/quick-batch.test.cjs constructs a REAL BATCH.json (createBatch), writes a REAL SUMMARY.md on disk at the item's real item_dir, calls the REAL resumeBatch, and proves both that resumeBatch alone still reports the item eligible AND that filterAlreadyExecuted (fed a real filesystem check) correctly excludes it. The prior prose-assertion tests are kept — they now prove the workflow markdown is correctly WIRED to the verb — but are no longer the only proof. Self-discovered defect while building that fixture (fixed inline, not deferred): tracing merge-wave.md against /gsd:quick's own prior art (QUICK_WORKTREE_MANIFEST=$(mktemp ...), quick.md:415) showed $QUICK_BATCH_WORKTREE_MANIFEST is a fresh PER-PROCESS temp file. A resumed coordinator correctly does not re-dispatch an already-executed item (this fix), but nothing durably recorded that item's worktree_path/branch/base either — Step 7 in the resumed process would have had no data to build its cleanup-wave entry from. Adds dispatched_worktree/dispatched_branch/ dispatched_base to QuickBatchItem (src/quick-batch.cts) — deliberately NOT a reuse of the pre-existing `worktree` field, whose loadBatch validation requires the path to exist on disk (verified empirically: reusing it made the batch permanently unloadable the moment a legitimately-merged worktree was removed). worktree-dispatch.md persists the triple once a worktree is created; merge-wave.md falls back to it when the ephemeral manifest lacks an entry, clears it after a successful merge, and fails closed rather than guessing if no record exists anywhere. See .gsd/phase/feat-3677-quick-batch-hardening-acceptance/40-design.md §9.1 and §9.3 for the full trace, empirical verification notes, and rejected alternatives (reusing `worktree` directly). * test(#3677): prove the arbitrary-worktree-ownership boundary against two real sibling worktrees Orthogonal review (Security finding): the two existing ownership-tampering tests didn't test ownership — one was trivially rejected by WORKTREE_AGENT_BRANCH_RE's shape check before any git call (proves branch- NAME filtering, not ownership), the other pointed at a wholly separate, never-linked foreign repo, so merge-base failed immediately because the branch didn't exist as a ref at all. Neither exercised the real scenario: a manifest entry whose worktree_path/branch are swapped to point at a DIFFERENT, GENUINELY-REGISTERED sibling worktree of the SAME repoRoot, with a branch name passing the shape check and a base in allowed_bases. Investigated executeWorktreeWaveCleanupPlan (src/worktree-safety.cts) directly: this is NOT a reachable gap. Git enforces branch-per-worktree uniqueness, so a swapped-in entry.branch can only match worktree_path's ACTUAL checked-out branch if it names that sibling's own real, uniquely- generated branch name — which manifest tampering confined to one batch's own record has no way to know (branch names are agent-<quick_id>[-<timestamp>]-shaped, and quick_id allocation is collision-checked GLOBALLY across every existing quick task and batch, not merely within one batch). Adds a stronger test that empirically proves this: two REAL, concurrently- alive sibling worktrees of the same repo (both via real `git worktree add`, both WORKTREE_AGENT_BRANCH_RE-passing, both sharing one merge-base), with worktree_path/branch swapped between them in both directions. Both attempts are blocked via branch_mismatch; both real worktrees, their branches, and one sibling's real uncommitted-to-main commit survive completely untouched. Supplements (does not replace) the original two tests, which still prove distinct, real boundaries. See .gsd/phase/feat-3677-quick-batch-hardening-acceptance/40-design.md §9.2 for the full trace, including the one explicitly-documented (not fixed) trust boundary this investigation surfaced: the primitive defends against fabricated data, not a caller bug that misattributes a real-but-wrong item's own triple to a different item. * chore(#3677): checkpoint design-doc addendum for review pass 2 findings (gitignored, dev-only) * docs(#3677): add changeset for PR 4240 Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> --------- Co-authored-by: sim <sim@local> Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
GSD Core
Git. Ship. Done.
English · Português · 简体中文 · 日本語 · 한국어
A light-weight meta-prompting, context engineering, and spec-driven development system for Claude Code, OpenCode, Antigravity CLI, Kimi CLI, Kilo, Codex, Copilot, Cursor, Windsurf, and more.
What is GSD Core
GSD Core is a context-engineering and spec-driven development framework that drives AI coding agents (Claude Code, Codex, Antigravity CLI, Kimi CLI, Copilot, Cursor, and more) through a disciplined phase loop. It solves context rot — the quality degradation that accumulates as an AI fills its context window — by running all heavy research, planning, and execution work in fresh-context subagents while keeping your main session lean.
How it works
Each milestone repeats the same five-step loop, one phase at a time:
- Discuss — capture implementation decisions before anything is planned
- Plan — research, decompose, and verify the plan fits a fresh context window
- Execute — run plans in parallel waves; each executor starts with a clean 200k-token context
- Verify — walk through what was built; diagnose and fix before declaring done
- Ship — create the PR, archive the phase, repeat for the next one
Quickstart
npx @opengsd/gsd-core@latest
The installer prompts for your runtime (Claude Code, OpenCode, Antigravity CLI, Kimi CLI, Kilo, Codex, Copilot, Cursor, Windsurf, and more) and whether to install globally or locally. The installer is required for cross-runtime compatibility — do not copy files from agents/ or commands/ directly.
On another runtime or without Node.js? See Install on your runtime.
Once installed, start a new project or onboard an existing repo:
/gsd-new-project # greenfield project
/gsd-onboard # existing codebase
New here? Follow Your first project for a guided walkthrough from install to first shipped phase, or Onboarding an existing codebase for brownfield setup.
Documentation
What's new in 1.7.0 → docs/whats-new-1.7.0.md
Tutorials — learning by doing:
How-to guides — task-focused recipes:
Reference — authoritative facts:
Explanation — concepts and design decisions:
Full index: docs/README.md. Other languages: 日本語 · 한국어 · Português · 简体中文.
Why it works
Most AI-coding setups fail at scale because context bloat silently degrades output quality, there is no shared memory between sessions, and nothing verifies that code actually works. GSD Core solves all three: heavy work runs in fresh subagents, structured artifacts like STATE.md and CONTEXT.md survive session boundaries, and the verify step walks through what was built and generates fix plans before a phase is declared done. See docs/explanation/context-engineering.md for the full reasoning.
Troubleshooting? See docs/how-to/recover-and-troubleshoot.md.
Community
| Project | Platform |
|---|---|
| gsd-opencode | Original OpenCode port |
| Discord | Community support |
Star History
License
MIT License. See LICENSE for details.
Claude Code is powerful. GSD Core makes it reliable.