Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
npx skills add https://github.com/samchon/compiler-graph --skill development
The command stays on one line. Scroll horizontally to inspect it before copying.
Prefer a local copy? Download the files currently available to SkillsMP.
Defines @samchon/graph implementation rules, consequence analysis, testing standards, validation, and change integrity. Use before writing or modifying source, tests, workflows, package wiring, fixtures, generated publications, index algorithms, MCP contracts, or resident-session behavior.
Development
Forbidden
These approaches are never acceptable:
No monkey-patching or hardcoding. Do not special-case a consumer, fixture, repository, prompt, model, expected answer, or benchmark cell. Fix the general logic.
No test-passing-only logic. Code exists to be correct, not merely to turn a check green. A branch whose only purpose is satisfying one assertion is a defect.
No forcing a broken design. When the same failure returns under patch after patch, stop and repair the owning design or invariant instead of accumulating symptoms.
No whack-a-mole. Treat the reported case as one witness. Find the root cause, cover its positive, negative, and boundary forms, and seal the verified class of failure.
Work Rules
Choose the principled course. Time, difficulty, and the breadth of consequences require more careful analysis and validation; they never justify a shortcut, leaving a verified consequence unaddressed, or a weaker acceptance standard.
Open nearby peers before adding a file, function, public type, operation, test, or fixture. Mirror the established naming, location, export, and error-handling conventions.
Respect package boundaries. Semantic graph, LSP, MCP, and CLI source lives under packages/graph/src; best-effort syntax extraction lives under packages/graph-sitter/src; test-only helpers live under tests/test-graph/src/internal; real-server provisioning lives under tests/experiment; benchmark policy stays under tests/benchmark.
Keep public exports synchronized through the nearest barrel and the root package surface. Public schema changes also require the matching MCP description, README contract excerpt, and compatibility review.
Follow the normal public-source convention of one export named for its file, with barrels containing re-exports rather than local declarations. Preserve the explicitly tested exceptions instead of generalizing them into a second style.
Preserve strict TypeScript assumptions, NodeNext module behavior, LF output, and @ttsc/lint checks from config/.
Reuse existing utilities and typed structures before creating another abstraction. Add no dependency without an explicit product reason and user authorization.
Edit the root README.md, never the generated packages/graph/README.md copy.
Keep diffs small and reviewable. Do not reformat or rewrite unrelated user changes in a dirty worktree.
Consequence Analysis
Before changing behavior, trace the cause through the relevant surfaces:
one-shot dump, long-lived MCP, viewer, and TypeScript API callers;
LSP, static, and hybrid indexing lanes;
initial load, incremental refresh, failed load, retry, shutdown, and process cleanup;
every supported language, shared-server cases such as C/C++, and language-specific preparation such as CMake or Dart packages;
Windows command shims and path handling as well as POSIX lookup and signals;
operation ranking, downstream flows, nearby nodes, tests, evidence, audits, and next decisions;
public exports, README contract text, CLI help, MCP schema, experiments, benchmarks, and release packaging.
Fix the verified class without broadening the user's product goal.
Testing
Each deterministic feature test lives in tests/test-graph/src/features/test_<snake_case>.ts and exports exactly one matching test_<snake_case> function. The dynamic runner discovers that prefix and rejects files with zero or multiple test exports.
Open each new or modified case with a doc comment naming what it pins. The filename already states the assertion, so spend the comment on the non-obvious reason the case exists: the branch it locks, the regression it replays, or the invariant that would otherwise be invisible.
Use @nestia/e2e's TestValidator and the shared fixtures in tests/test-graph/src/internal. Extend a nearby fixture or fake process rather than creating an unrelated harness. Keep real language-server installation and network-heavy project smoke tests in tests/experiment; the deterministic suite must not depend on globally installed servers or network access.
Coverage is a specification, not a reason for artificial branches. The root pnpm coverage command enforces 100 percent line, function, and branch coverage. Use c8 ignore only for a genuinely untestable platform/process boundary and keep the ignored span as narrow as possible.
A regression test must fail for the reported behavior before the fix and pass after it. Cover:
the transformation or state transition that must occur;
a negative twin one property away where it must not occur;
empty, singleton, exact-limit, ambiguous-name, missing-server, and recovery boundaries that apply; and
both source-of-truth lanes when the invariant is shared by LSP and static indexing.
Expectations Come From An Oracle
Take every expected value from an authority outside the code under test: the language server's own response, the semantics of the language being indexed, a pinned upstream reference, or the documented public contract. Never take it from what the current implementation happens to emit.
An expectation copied from current output cannot fail for the defect it already contains. It converts a wrong result into the specification, and the next reviewer who checks the test against the code finds agreement that proves nothing. The checked-in application-contract canonical shows the shape to follow: it is reproduced from the pinned ttsc reference, and --diff previews a candidate instead of overwriting the file.
A case that feeds a component its own output and asserts nothing changed proves stability, not correctness. Idempotency and round-trip cases are worth keeping for the property they actually cover, but a lane whose only evidence is round-trip has no evidence of correctness: the input must differ from the expected output somewhere, or nothing in the test can distinguish a working indexer from a broken one.
Validation
Run the narrowest command that proves the change first, then the broader gates required by its consequence surface. Report every command that could not run.
Focused TypeScript change:pnpm --dir packages/graph exec ttsc check -p tsconfig.json. This runs type analysis and error-level lint rules without emitting files; compare warning output so new warning-level findings are not hidden among the known regex warnings.
Package or shared behavior:pnpm build and pnpm test.
Source behavior or coverage-sensitive branch:pnpm coverage.
Language-server behavior: focused fake-server tests first, then pnpm --filter @samchon/graph-experiment start -- --language <language> when that real server is available. The full real-server matrix belongs to .github/workflows/experiment.yml.
Benchmark harness logic: zero-spend benchmark tests and preflight before any paid run; follow the benchmark skill.
Verification shape depends on the change type:
Bug fix: name the failing case and expected behavior; run a reproduction that distinguishes before from after.
Feature: name the observable contract and exercise it through the public API, CLI, or MCP boundary.
Refactor: name what must remain unchanged and rely on behavior-locking tests before cleanup edits.
Review: report concrete defects, risks, missing tests, or regressions with evidence.
Change Integrity
Treat tests, fixtures, benchmark prompts and manifests, CI workflows, package wiring, dependencies, public schemas, audit text, generated chart inputs, and the 100 percent coverage gate as specification. Changing them requires an explicit user request or a verified product reason, and the final report must call it out.
Package-local ttsc format and ttsc fix mutate every selected source file. They are implementation actions, not validation commands; never run them casually in a dirty worktree.
For ports, migrations, or broad rewrites, preserve the existing algorithm and public behavior in reviewable slices. Inspect the diff and generated/public surface before trusting a green test run.
