e57e3b5d6e
Phase 3 of #3464, following #3465 and #3466. Those phases cut the exemption ceiling 305 -> 278 by removing vestigial markers and rewriting real source-greps behaviorally. This one addresses why the ratchet was weak in the first place: it counts markers, and markers correlated only loosely with violations, because the rule's implementation was far narrower than its intent. Four gaps, each measured against 833 files under tests/ before any code was written: A TEXT_METHODS omitted matchAll, split and replace, and never handled the regex-side form re.test(tracked) / /lit/.test(tracked) where the tracked value is the ARGUMENT rather than the callee object. B The extension test was /\.(?:cjs|js|ts)/, which does not match .cts, .mts or .mjs. Under ADR-457 this repo's production modules live in src/**/*.cts, so the rule has been structurally blind to the entire TypeScript source surface since that migration. Highest-value fix here. C Tracking stopped at one hop, so an intermediate transform (const b = strip(a); b.match(...)) escaped. D Variables were tracked by NAME in a flat Set<string>, with no scope resolution, so a name reused across describe/test blocks was conflated. D removes one verified false positive where an outer `const src = readFileSync(...)` was cross-attributed to a shadowed arrow-function parameter of the same name. Deliberately NOT implemented: flagging reads whose path cannot be statically resolved. Measured at 4255 sites across 255 files, 66 of them newly red, with a 6-of-6 false-positive rate in spot-checking -- every sampled site read a markdown workflow or fixture doc through a path variable, not JS source. The heuristic "no JS extension literal present" inverts to "not a source file" in this codebase. Shipping it would have manufactured exactly the marker-spam dynamic this epic exists to stop. The principled version needs real static resolution (constant-folding path.join and template literals) and is left to a future phase. The widening surfaced two genuinely-invisible violations, handled on their merits rather than uniformly: tests/verifier-behavior-unverified.test.cjs read src/verification.cts and regexed it for the VERIFIER_STATUSES array. Fixed BEHAVIORALLY with no marker: that constant is already exported, so the test now asserts the real runtime value -- strictly stronger, and immune to source formatting. Mutation-checked: injecting present_behavior_unverified into the exported array turns it red, restoring turns it green. tests/adr-index-gate.test.cjs scans src/plan-drift-guard.cts for docs/adr/<name>.md citations and asserts each cited ADR exists. Those citations live in COMMENTS, erased at compile time: no exported value, no runtime observable, and making it behavioral would mean contorting production code into exporting its own documentation citations. Irreducible, so it takes one marker, cited to #3502, stating exactly why. Documenting a real exemption beats leaving the violation invisible, which was the status quo. Two defects in this branch's own work, both found by review and fixed here rather than shipped: FALSE POSITIVE (adversarial review). Hop propagation walked every Identifier in a declarator init and treated any reference to a tracked variable as derivation, regardless of whether the derived VALUE still carried source text. So `const len = raw.length; /^\d+$/.test(len)` was reported as a source-grep. Propagation is now value-shape aware: it follows identity, string-returning string methods, split/join, template embedding, string concatenation, conditional branches and call arguments; it stops at .length, numeric methods (indexOf/search/charCodeAt), boolean methods (includes/startsWith/test), comparisons, negation, typeof, and Number/parseInt/Boolean coercions. Unrecognized shapes still propagate -- the conservative default for a linter is a rarer false positive over a silent false negative, and that choice is documented inline. Seven RuleTester rows now cover this axis, which was previously untested. SUPER-QUADRATIC SCAN (security review). resolveVariable() resolved each identifier with two linear Array.find passes over scope.references and scope.variables, once per identifier walked -- O(vars-in-scope) per lookup. On a synthetic single-scope file of N consts each referencing ~20 priors: 4.76s at N=3000 and 24.32s at N=6000 (~5.1x for 2x N). Replaced with a Map<IdentifierNode, Variable> built once per file, lazily, from the scope manager: 0.19s and 0.37s for the same inputs (~1.95x for 2x N, linear). Semantics unchanged. No measurable effect on the real repo either way, but this is exactly the bug class ADR-3212 / local/no-unbounded-quantifier exists to catch, and this repo has a prior incident where a rule written to catch complexity bugs shipped with one of its own. The per-widening measurement had a flaw worth recording: each widening was measured in isolation, so a site needing TWO at once appeared in neither column, and the UNION column was dominated by the rejected dynamic-path noise and never inspected for interactions. adr-index-gate needs both B (.cts) and A (.matchAll) and was missed for exactly that reason. Real newly-red count was 2, not the 1 predicted. Corrected on #3502 rather than quietly amended. Marker-bearing files 277 -> 278 against an unchanged ceiling of 278. The ceiling is NOT raised: the sole addition is the cited irreducible exemption. 27 RuleTester rows cover the change. The valid rows carry as much weight as the invalid ones -- shadowed same-name bindings, sibling block scopes, .md and .json literal reads, dynamic path variables, non-textual derivations, reads never text-searched, and require() of a .cjs must all stay valid. A widening that flagged those would be worse than the status quo, because it would push contributors toward adding markers to silence noise. Known limit, documented rather than papered over: a tracked value round-tripped through an array or object literal and read back via destructuring is still not tracked. Pre-existing, not introduced here, and deliberately not widened for -- closing it means tracking member identity, with its own false-positive surface. Closes #3502 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.