GitNexus/GUARDRAILS.md
Gergő Magyar 0a4b31b3c5
docs: optimize context files for LLM accuracy and token efficiency (#857)
* docs: optimize context files for LLM accuracy and token efficiency

Fix factual errors across all five root context files and optimize
for LLM context window efficiency.

Corrections:
- Web UI: "runs entirely in WASM" -> thin client backed by HTTP API
- Pre-commit hook: "typecheck + tests" -> formatting + typecheck only
- MCP tools: 7 -> 16 (added api_impact, route_map, tool_map,
  shape_check, group_list/query/sync/contracts/status)
- Default serve port: 3741 -> 4747
- E2E tests: "5 tests" -> 7 spec files
- ESLint: "no config" -> eslint.config.mjs exists with TS/React rules
- npm test: "vitest run test/unit" -> "vitest run" (full suite)
- Removed nonexistent test:all script
- ci-quality.yml: added missing format + lint job descriptions
- Pipeline phase deps: added missing structure dep on mro/communities/processes
- Ingestion entry: added missing run-analyze.ts intermediate orchestrator
- Tools Quick Reference: added missing list_repos
- Group tool examples: fixed param name (group -> name)
- Removed stale vite-plugin-wasm gotcha
- Added gitnexus-shared to repository layout tables

New documentation:
- ARCHITECTURE.md: language-agnostic graph feeding (provider pattern,
  unified capture tags, import resolution tiers, chunked parse, MRO)
- ARCHITECTURE.md: full analysis flow (10 stages with progress %)
- ARCHITECTURE.md: storage layout, LadybugDB schema, embeddings, search
- ARCHITECTURE.md: DAG runner internals (Kahn's sort, dep isolation, error handling)

Token optimization:
- Removed filler prose, compressed descriptions into dense tables
- Front-loaded key facts in every section
- Eliminated redundancy between sections
- AGENTS.md: 219 -> 201 lines. ARCHITECTURE.md: 192 -> 298 lines
  (more info in fewer tokens via tables and structure)

* docs: optimize GUARDRAILS.md for LLM context efficiency

Tighten prose without losing information:
- Compressed intro, scope section, and Signs format labels
- Shortened Sign headers (removed "Sign:" prefix)
- Replaced verbose "Instruction/Reason" labels with "Do/Why"
- Removed trailing whitespace and redundant emphasis
2026-04-16 08:43:11 +01:00

85 lines
3.8 KiB
Markdown

# Guardrails — GitNexus
Rules for **human contributors** and **AI agents**. Complements `AGENTS.md` (workflows) and `CONTRIBUTING.md` (PR process).
## Scope (least privilege)
- **Read:** Source, tests, docs, public config as needed.
- **Write:** Only files required for the fix or feature; no unrelated formatting or refactors.
- **Execute:** Tests, typecheck, documented CLI commands. No destructive commands on user data without approval.
- **Off-limits:** Other people's machines, production deployments you don't own, credentials you lack permission to use.
Maintainer may widen scope per task.
---
## Non-negotiables
1. **Never commit secrets** — API keys, tokens, real `.env` values, private URLs, session cookies. Use `.env.example` with placeholders.
2. **Never rename with find-and-replace** in GitNexus-indexed projects — use `rename` MCP tool with `dry_run: true` first, review `graph` vs `text_search` edits. No separate `gitnexus rename` CLI exists.
3. **Run impact analysis before editing shared symbols**`impact` (upstream) for functions/classes/methods others call. Do not ignore HIGH/CRITICAL without maintainer sign-off.
4. **Run `detect_changes` before commit** — confirm diffs map to expected symbols/processes when the graph is available.
5. **Preserve embeddings** — if `.gitnexus/meta.json` shows embeddings, use `npx gitnexus analyze --embeddings`; plain `analyze` drops them.
---
## Signs (recurring failure patterns)
Format: **Trigger → Instruction → Reason**. Append new Signs when the same mistake repeats.
### Stale graph after edits
- **Trigger:** MCP warns index is behind `HEAD`, or search doesn't match latest commit.
- **Do:** `npx gitnexus analyze` (plus `--embeddings` if used).
- **Why:** Tools query LadybugDB from last analyze; git changes are invisible until re-indexed.
### Embeddings vanished after analyze
- **Trigger:** Semantic search quality drops; `stats.embeddings` in `meta.json` is 0 after refresh.
- **Do:** `npx gitnexus analyze --embeddings`, confirm `meta.json` reflects stored embeddings.
- **Why:** Embedding generation is opt-in; analyze without the flag does not preserve prior vectors.
### MCP lists no repos
- **Trigger:** MCP stderr says no indexed repos.
- **Do:** `npx gitnexus analyze` in the target repo; verify `npx gitnexus list` shows it.
- **Why:** MCP discovers repos via `~/.gitnexus/registry.json`, populated by analyze.
### Wrong repo in multi-repo setups
- **Trigger:** Query/impact results belong to another project.
- **Do:** Call `list_repos`, then pass `repo` on subsequent tools.
- **Why:** Default target is ambiguous when multiple repos are registered.
### LadybugDB lock / "database busy"
- **Trigger:** Errors opening `.gitnexus/lbug` while MCP and analyze both run.
- **Do:** Stop overlapping processes (one writer at a time). Retry analyze or restart MCP.
- **Why:** Embedded DB expects single-process ownership.
---
## Publishing & supply chain
- **npm:** Do not publish from unreviewed automation. Bump version intentionally; tag releases to match `package.json`.
- **Dependencies:** Minimal, auditable `package.json` changes; run tests and CI after lockfile updates.
- **License:** PolyForm Noncommercial 1.0.0 — do not relicense without maintainer approval.
---
## Escalation
Stop and ask a **human maintainer** when:
- Impact analysis shows HIGH/CRITICAL risk and the task still requires the change.
- You need to alter CI, release, or security-sensitive config.
- Requirements conflict (e.g. "speed up analyze" vs "must keep all embeddings on huge repo").
- You are unsure whether data loss is acceptable (`clean`, forced migrations, schema changes).
---
## Related docs
- [ARCHITECTURE.md](ARCHITECTURE.md) — components and data flow
- [RUNBOOK.md](RUNBOOK.md) — commands for recovery
- [CONTRIBUTING.md](CONTRIBUTING.md) — PR and commit expectations