b9f51836e6
* refactor(#3180): one owner for completion ratio, a prompt-layer drift guard, and a written behavior contract The 2026-08-08 coverage audit on #3180 found the epic's copy counts were a lower bound for the third consecutive time, and that two derivation families had never been named at all. ADR-3180 gains Decision 7 — a normative behavior contract that says what the right answer IS for each derivation, not merely who owns it. A reviewer with no written rule can only ask "does this look like the others", which is how a fifth copy passes review. Decision 4 gains (d) scan surface is every authored surface and an owner FILE is never exempt, only its named functions; and (e) a surface that cannot be consolidated today ships ratcheted, never unguarded. Completion ratio: `clampPercent` sat exported and unused beside six hand-inlined copies of its own body across five modules. All six now route through it; `clampPercentFromFraction` is added for the one caller that already held a fraction. Every migration is behaviour-identical — clampPercent's first line IS the `total > 0 ? … : 0` ternary each copy carried. Guarded by lint-completion-ratio-drift.cjs, which reports zero re-derivations with no file-level exemption. Prompt layer: workflow markdown re-derives live-plan counting in raw shell (#1762), invisible to every `src/`-scoped guard. lint-planning-prompt-drift.cjs scans it with a shrink-only baseline of the 7 sites that exist today — new sites fail, and a baseline entry that stops firing fails too, so an acknowledgment can never outlive the thing it describes. lint-milestone-window-drift.cjs stops exempting its owner file wholesale; only the four named canonical functions are exempt now. The blanket exemption was pointed at the one file most likely to grow the next copy, and it had. Refs #3180 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * docs(#3180): link Phases 6-8 sub-issues (#3216, #3217, #3218) from ADR-3180 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix(#3180): address orthogonal review — consumer-output identity tests, count-keyed ratchet, property coverage Five findings from the two orthogonal review passes, all fixed. Decision 4(c) breach: the completion-ratio identity test asserted at the OWNER, which is exactly the bypass that decision exists to close — a consumer can call clampPercent and then post-process locally, leaving both the lint and an owner-level test green. It now drives `roadmap analyze`, `query progress` and `stats` and asserts on their own output, over a fixture containing a `status: superseded` plan so a consumer that re-counted raw files would report 60 where the owner reports 75. Decision 4(e) breach: ratchet entries named the epic (#3180) rather than the issue that removes them. They name Phase 8 (#3218) now. The ratchet keyed on (file, text) alone, so plan-phase.md's two byte-identical sites were one indistinguishable key and migrating either would have left the guard green with the other alive. Entries carry an occurrence count; fewer than acknowledged fails as a partial migration, more fails as a new copy. Adds the missing MAX_REGEX_LITERAL_LEN boundary coverage the sibling guard's test already had, and the fast-check property tests CONTRIBUTING requires for clamp/budget-limit functions. Refs #3180 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * test: stop wrapping a nested double-spawn in a 15s wall-clock budget (bug #641 probes) `tests/ci-test-scope.test.cjs`'s `bug #641` block spawned `run-tests.cjs` under PROBE_TIMEOUT_MS=15000; that child then spawned a nested `node --test`. A fixed wall-clock budget around a double spawn, running inside a container that is concurrently executing the full ~31k-test suite, fails by construction under load. Confirmed against three full matrix runs. Every failure was shaped `null !== 0` — the child was KILLED, never an assertion about the thing under test. One captured probe had already printed the correct resolution (`suite="all" files=2: a.test.cjs b.test.cjs`) and was killed anyway. It reproduces on `next` alone: 5 failures on linux-node22, 0 on linux-node24. The victim subset varies by run and by lane. What these tests are actually about is suite-token RESOLUTION — `unit` as a bare token in --files/--files-from. Executing the seeded trivial files is incidental and is the entire timeout surface, so the assertions move in-process against the same functions `main()` calls, in the same order. `parseArgs`, `selectExplicitFiles`, `selectFiles` and `walkTestFiles` are exported for that; no behavior, signature or logic changed. No coverage lost: `tests/run-tests-harness.test.cjs` already spawns the harness for real and asserts exit codes end to end, on a 120s budget. Pre-existing on `next`, fixed here rather than deferred. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * test: delete the three elapsed-time assertions CLAUDE.md forbids asserting on wall-clock time. Three assertions did, and all three are load-sensitive: on a saturated bench each can fail while the code under test is correct. In every case the load-bearing assertion sits on the line above and the timing line adds no discrimination. run-with-timeout: the stated worry — "was this 124 the cap firing or the 30s harness backstop?" — is already answered by the assertion above it. A backstop kills by signal, which surfaces as status null, never 124. Observed directly this session: three matrix runs produced exactly that null shape from killed children. normalize-test-command and context-predicates: both bounded a ReDoS check. A threshold only ever separates "fast" from "slightly slow", which is bench load, not correctness — catastrophic backtracking on 800 KB of input does not take 251ms, it does not finish at all. A real regression therefore shows up as the suite being killed on that test, which is louder and more reliable than a number. The structural assertions (returned unchanged; cleanly rejected) are what actually carry those tests, and they stay. The sweep now reports zero elapsed-time assertions in tests/. The remaining Date.now() uses are unique-path suffixes, barrier deadlines, fixture timestamps and fake mtimes — none of them assertions. Pre-existing on `next`, fixed here rather than deferred. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * chore(#3180): backfill changeset PR number (#3223) Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix(#3180): key the prompt-drift ratchet on POSIX paths so it works on Windows The baseline keys on (file, trimmed text). `file` came from scanTree's `path.relative()`, which uses NATIVE separators, while the committed baseline stores POSIX. On Windows every violation was therefore unmatched — reported as FRESH — and every baseline entry matched nothing — reported as STALE. The guard failed 100% of the time there, on both CI shards: ✖ scanRepo(repoRoot) matches the baseline exactly: zero fresh AND zero stale + { file: 'gsd-core\\workflows\\execute-plan.md', ... } The remote runner this repo gates on is Linux-only and cannot see this class at all; the GitHub Actions Windows lane is what caught it. Normalization is unconditional — never gated on process.platform. A platform-conditional normalizer makes the POSIX path the special case and leaves the Windows branch unexercised on every other OS, which is the same blind spot in a different place. It is applied at one seam inside findPromptDrift, which builds `file` on every returned violation, so the baseline key, the --update writer, the stderr report and the tests all consume one normalized value. The regression tests drive a Windows-shaped relPath directly and run on every OS rather than skipping off-Windows — a test that only runs on the platform where the bug lives is why this escaped. They include a sanity check that un-normalized input does NOT match, so the assertion cannot pass vacuously. Audited the three sibling guards: none keys against a committed cross-platform baseline, and their exemption keys are path.join-built, so producer and consumer share the native convention. Left correct code alone rather than making them look alike. scripts/lib/drift-scan.cjs is untouched — normalizing there would break those three on Windows. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> --------- Co-authored-by: sim <sim@local> Co-authored-by: Claude Opus 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.