CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Quick Start

npm install                    # Install deps (auto-generates .env from .env.example)
npm run dev                    # Dev server at http://localhost:20128
npm run build                  # Production build (Next.js 16 standalone)
npm run lint                   # ESLint (0 errors expected; warnings are pre-existing)
npm run typecheck:core         # TypeScript check (should be clean)
npm run typecheck:noimplicit:core  # Strict check (no implicit any)
npm run test:coverage          # Unit tests + coverage gate (60/60/60/60 — statements/lines/functions/branches)
npm run check                  # lint + test combined
npm run check:cycles           # Detect circular dependencies

Running Tests

# Single test file (Node.js native test runner — most tests)
node --import tsx/esm --test tests/unit/your-file.test.ts

# Vitest (MCP server, autoCombo, cache)
npm run test:vitest

# All suites
npm run test:all

For full test matrix, see CONTRIBUTING.md → "Running Tests". For deep architecture, see AGENTS.md.

Project at a Glance

Nerve — unified AI proxy/router. One endpoint, 290 LLM providers, auto-fallback.

Monorepo: src/ (Next.js 16 app), open-sse/ (streaming engine workspace), electron/ (desktop app), tests/, bin/ (CLI entry point).

Request Pipeline

Client → /v1/chat/completions (Next.js route)
  → CORS → Zod validation → auth? → policy check → prompt injection guard
  → handleChatCore() [open-sse/handlers/chatCore.ts]
    → cache check → rate limit → combo routing?
      → resolveComboTargets() → handleSingleModel() per target
    → translateRequest() → getExecutor() → executor.execute()
      → fetch() upstream → retry w/ backoff
    → response translation → SSE stream or JSON
    → If Responses API: responsesTransformer.ts TransformStream

API routes follow a consistent pattern: Route → CORS preflight → Zod body validation → Optional auth (extractApiKey/isValidApiKey) → API key policy enforcement → Handler delegation (open-sse). No global Next.js middleware — interception is route-specific.

Combo routing (open-sse/services/combo.ts): 19 strategies (priority, weighted, fill-first, round-robin, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, context-relay, cache-optimized, fusion, pipeline). Each target calls handleSingleModel() which wraps handleChatCore() with per-target error handling and circuit breaker checks. The fusion strategy is the exception: it fans out to a panel of models in parallel, then a judge model synthesizes one final answer (open-sse/services/fusion.ts). See docs/routing/AUTO-COMBO.md for the 12-factor Auto-Combo scoring + the full strategy table and docs/architecture/RESILIENCE_GUIDE.md for the 3 resilience layers.

Resilience Runtime State

Nerve has three related but distinct temporary-failure mechanisms. Keep their scope separate when debugging routing behavior. See the 3-layer resilience diagram (source: docs/diagrams/resilience-3layers.mmd) for an at-a-glance map.

Provider Circuit Breaker

Scope: whole provider, e.g. glm, openai, anthropic.

Purpose: stop sending traffic to a provider that is repeatedly failing at the upstream/service level, so one unhealthy provider does not slow down every request.

Implementation:

• Core class: src/shared/utils/circuitBreaker.ts
• Chat gate/execution wiring: src/sse/handlers/chatHelpers.ts, src/sse/handlers/chat.ts
• Runtime status API: src/app/api/monitoring/health/route.ts
• Shared wrappers: open-sse/services/accountFallback.ts
• Persisted state table: domain_circuit_breakers

States:

• CLOSED: normal traffic is allowed.
• OPEN: provider is temporarily blocked; callers get a provider-circuit-open response or combo routing skips to another target.
• HALF_OPEN: reset timeout has elapsed; allow a probe request. Success closes the breaker, failure opens it again.

Defaults (open-sse/config/constants.ts):

• OAuth providers: threshold 3, reset timeout 60s.
• API-key providers: threshold 5, reset timeout 30s.
• Local providers: threshold 2, reset timeout 15s.

Only provider-level failure statuses should trip the provider breaker:

(408, 500, 502, 503, 504);

Do not trip the whole-provider breaker for normal account/key/model errors like most 401, 403, or 429 cases. Those usually belong to connection cooldown or model lockout. A generic API-key provider 403 should be recoverable unless it is classified as a terminal provider/account error.

The breaker uses lazy recovery, not a background timer. When OPEN expires, reads such as getStatus(), canExecute(), and getRetryAfterMs() refresh the state to HALF_OPEN, so dashboards and combo candidate builders do not keep excluding an expired provider forever.

Connection Cooldown

Scope: one provider connection/account/key.

Purpose: temporarily skip one bad key/account while allowing other connections for the same provider to continue serving requests.

Implementation:

• Write/update path: src/sse/services/auth.ts::markAccountUnavailable()
• Account selection/filtering: src/sse/services/auth.ts::getProviderCredentials...
• Cooldown calculation: open-sse/services/accountFallback.ts::checkFallbackError()
• Settings: src/lib/resilience/settings.ts

Important fields on provider connections:

rateLimitedUntil;
testStatus: "unavailable";
lastError;
lastErrorType;
errorCode;
backoffLevel;

During account selection, a connection is skipped while:

new Date(rateLimitedUntil).getTime() > Date.now();

Cooldowns are also lazy: when rateLimitedUntil is in the past, the connection becomes eligible again. On successful use, clearAccountError() clears testStatus, rateLimitedUntil, error fields, and backoffLevel.

Default connection cooldown behavior:

• OAuth base cooldown: 5s.
• API-key base cooldown: 3s.
• API-key 429 should prefer upstream retry hints (Retry-After, reset headers, or parseable reset text) when available.
• Repeated recoverable failures use exponential backoff:

baseCooldownMs * 2 ** failureIndex;

The anti-thundering-herd guard prevents concurrent failures on the same connection from repeatedly extending the cooldown or double-incrementing backoffLevel.

Terminal states are not cooldowns. banned, expired, and credits_exhausted are intended to stay unavailable until credentials/settings change or an operator resets them. Do not overwrite terminal states with transient cooldown state.

Model Lockout

Scope: provider + connection + model.

Purpose: avoid disabling a whole connection when only one model is unavailable or quota-limited for that connection.

Examples:

• Per-model quota providers returning 429.
• Local providers returning 404 for one missing model.
• Provider-specific mode/model permission failures such as selected Grok modes.

Model lockout lives in open-sse/services/accountFallback.ts and lets the same connection continue serving other models.

Debugging Guidance

• If all keys for a provider are skipped, inspect both provider breaker state and each connection's rateLimitedUntil/testStatus.
• If a provider appears permanently excluded after the reset window, check whether code is reading raw state instead of using getStatus()/canExecute().
• If one provider key fails but others should work, prefer connection cooldown over provider breaker.
• If only one model fails, prefer model lockout over connection cooldown.
• If a state should self-recover, it should have a future timestamp/reset timeout and a read path that refreshes expired state. Permanent statuses require manual credential or config changes.

Key Conventions

Code Style

• 2 spaces, semicolons, double quotes, 100 char width, es5 trailing commas (enforced by lint-staged via Prettier)
• Imports: external → internal (@/, @nerve/open-sse) → relative
• Naming: files=camelCase/kebab, components=PascalCase, constants=UPPER_SNAKE
• ESLint: no-eval, no-implied-eval, no-new-func = error everywhere; no-explicit-any = error in open-sse/ and tests/ (since #6218 — pre-existing violations are frozen in config/quality/eslint-suppressions.json, new ones must be fixed; npm run lint applies the suppressions and is what CI runs)
• TypeScript: strict: false, target ES2022, module esnext, resolution bundler. Prefer explicit types.

Database

• Always go through src/lib/db/ domain modules — never write raw SQL in routes or handlers
• Never add logic to src/lib/localDb.ts (re-export layer only)
• Never barrel-import from localDb.ts — import specific db/ modules instead
• DB singleton: getDbInstance() from src/lib/db/core.ts (WAL journaling)
• Migrations: src/lib/db/migrations/ — versioned SQL files, idempotent, run in transactions

Error Handling

• try/catch with specific error types, log with pino context
• Never swallow errors in SSE streams — use abort signals for cleanup
• Return proper HTTP status codes (4xx/5xx)

Security

• Never use eval(), new Function(), or implied eval
• Validate all inputs with Zod schemas
• Encrypt credentials at rest (AES-256-GCM)
• Upstream header denylist: src/shared/constants/upstreamHeaders.ts — keep sanitize, Zod schemas, and unit tests aligned when editing
• Public upstream credentials (Gemini/Antigravity/Windsurf-style OAuth client_id/secret + Firebase Web keys extracted from public CLIs): MUST be embedded via resolvePublicCred() from open-sse/utils/publicCreds.ts — never as string literals. See docs/security/PUBLIC_CREDS.md for the mandatory pattern.
• Error responses (HTTP / SSE / executor / MCP handler): MUST route through buildErrorBody() or sanitizeErrorMessage() from open-sse/utils/error.ts — never put raw err.stack or err.message in a response body. See docs/security/ERROR_SANITIZATION.md.
• Shell commands built from variables: when calling exec()/spawn() with a script that needs runtime values, pass them via the env option (shell-escaped automatically) — never string-interpolate untrusted/external paths into the script body. Reference: src/mitm/cert/install.ts::updateNssDatabases.
• Secure-by-default libraries (tldrsec/awesome-secure-defaults): prefer Helmet.js, DOMPurify, ssrf-req-filter, safe-regex, Google Tink over custom implementations whenever adding new security-sensitive surfaces.

Common Modification Scenarios

Adding a New Provider

1. Register in src/shared/constants/providers.ts (Zod-validated at load)
2. Add executor in open-sse/executors/ if custom logic needed (extend BaseExecutor)
3. Add translator in open-sse/translator/ if non-OpenAI format
4. Add OAuth config in src/lib/oauth/constants/oauth.ts if OAuth-based — if the upstream CLI ships a public client_id/secret, embed via resolvePublicCred() (see docs/security/PUBLIC_CREDS.md), never as a literal
5. Register models in open-sse/config/providerRegistry.ts
6. Write tests in tests/unit/ (include the publicCreds shape assertion if you added a new embedded default)

Adding a New API Route

1. Create directory under src/app/api/v1/your-route/
2. Create route.ts with GET/POST handlers
3. Follow pattern: CORS → Zod body validation → optional auth → handler delegation
4. Handler goes in open-sse/handlers/ (import from there, not inline)
5. Error responses use buildErrorBody() / errorResponse() from open-sse/utils/error.ts (auto-sanitized — never put err.stack or err.message raw in the body). See docs/security/ERROR_SANITIZATION.md.
6. Add tests — including at least one assertion that error responses do not leak stack traces (!body.error.message.includes("at /"))

Adding a New DB Module

1. Create src/lib/db/yourModule.ts — import getDbInstance from ./core.ts
2. Export CRUD functions for your domain table(s)
3. Add migration in src/lib/db/migrations/ if new tables needed
4. Re-export from src/lib/localDb.ts (add to the re-export list only)
5. Write tests

Adding a New MCP Tool

1. Add tool definition in open-sse/mcp-server/tools/ with Zod input schema + async handler
2. Register in tool set (wired by createMcpServer())
3. Assign to appropriate scope(s)
4. Write tests (tool invocation logged to mcp_audit table)

Adding a New A2A Skill

1. Create skill in src/lib/a2a/skills/ (5 already exist: smart-routing, quota-management, provider-discovery, cost-analysis, health-report)
2. Skill receives task context (messages, metadata) → returns structured result
3. Register in A2A_SKILL_HANDLERS in src/lib/a2a/taskExecution.ts
4. Expose in src/app/.well-known/agent.json/route.ts (Agent Card)
5. Write tests in tests/unit/
6. Document in docs/frameworks/A2A-SERVER.md skill table

Adding a New Cloud Agent

1. Create agent class in src/lib/cloudAgent/agents/ extending CloudAgentBase (3 already exist: codex-cloud, devin, jules)
2. Implement createTask, getStatus, approvePlan, sendMessage, listSources
3. Register in src/lib/cloudAgent/registry.ts
4. Add OAuth/credentials handling if needed (src/lib/oauth/providers/)
5. Tests + document in docs/frameworks/CLOUD_AGENT.md

Adding a New Embedded Service

1. Create installer in src/lib/services/installers/{name}.ts modeled on ninerouter.ts (use runNpm from installers/utils.ts — no shell interpolation, hard rule #13).
2. Register the service in src/lib/services/bootstrap.ts (add to SERVICES[] array and extend buildSpawnArgsFactory()).
3. Add a DB seed row for the new service in src/lib/db/migrations/ (version_manager table, status='not_installed', auto_start=0).
4. Create 7 API endpoints under src/app/api/services/{name}/ (_lib.ts, install, start, stop, restart, update, status, auto-start). All delegate errors through createErrorResponse(). The shared logs endpoint is already wired via [name]/logs/route.ts.
5. Verify /api/services/ is in LOCAL_ONLY_API_PREFIXES in src/server/authz/routeGuard.ts; add a test asserting isLocalOnlyPath() returns true for the new prefix if you add one (hard rule #17).
6. Add a UI tab in src/app/(dashboard)/dashboard/providers/services/tabs/ reusing ServiceStatusCard, ServiceLifecycleButtons, ServiceLogsPanel.
7. Document in docs/frameworks/EMBEDDED-SERVICES.md (update §1 service table + §4 API reference) and docs/openapi.yaml.
8. Write tests: unit (tests/unit/services/), integration (tests/integration/services/, gated by RUN_SERVICES_INT=1), and update docs/ops/RELEASE_CHECKLIST.md smoke section.

Adding a New Guardrail / Eval / Skill / Webhook event

• Guardrail: src/lib/guardrails/ → docs: docs/security/GUARDRAILS.md
• Eval suite: src/lib/evals/ → docs: docs/frameworks/EVALS.md
• Skill (sandbox): src/lib/skills/ → docs: docs/frameworks/SKILLS.md
• Webhook event: src/lib/webhookDispatcher.ts → docs: docs/frameworks/WEBHOOKS.md

Reference Documentation

For any non-trivial change, read the matching deep-dive first:

Testing

PR rule: If you change production code in src/, open-sse/, electron/, or bin/, you must include or update tests in the same PR.

Test layer preference: unit first → integration (multi-module or DB state) → e2e (UI/workflow only). Encode bug reproductions as automated tests before or alongside the fix.

Both test runners must pass: npm run test:unit (Node native — most tests) AND npm run test:vitest (MCP server, autoCombo, cache) cover non-overlapping files. Both are wired in CI (jobs test-unit and test-vitest) and must be green before merging. A PR where only one suite passes may silently ship broken MCP tools or routing regressions.

Bug fix / issue triage protocol (Hard Rule #18): Every fix for a reported issue must be validated by one of the following — no exceptions:

1. TDD (preferred) — write a failing test reproducing the bug → fix it → confirm the test passes. The test becomes the permanent regression guard. Touch only the files the test proves need changing; nothing more.
2. Real-environment test (when TDD is not possible) — deploy to the production VPS (root@192.168.0.15) and run a documented live test. Record the exact command + result in the PR description. Applies to: OAuth upstream flows, Cloudflare/WS upstream behavior, UI-only regressions, hardware-dependent behavior.
3. "It worked locally without a test" does not count. A fix without a test or a VPS validation record is not a fix — it is a guess.

Why this matters: fixing bug A while opening bug B is worse than not fixing at all. The TDD/VPS gate enforces surgical scope — you touch only what the failing test proves is broken. Examples where this paid off: #3090 (claude-web 403), #3113 (WS HTTP fallback), #3052 (heap-guard auto-calibration).

Copilot coverage policy: When a PR changes production code and coverage is below 60% (statements/lines/functions/branches), do not just report — add or update tests, rerun the coverage gate, then ask for confirmation. Include commands run, changed test files, and final coverage result in the PR report.

Planning & Research Artifacts (superpowers, deep-research)

_tasks/ is a separate, isolated git repository that is gitignored by the main repo (.gitignore → _tasks/). It is the canonical home for working artifacts — plans, specs/designs, research, hand-offs — so they stay versioned in their own repo instead of polluting the main Nerve tree.

Hard rule — never write superpowers / planning / research output under docs/ or the repo root. The superpowers skills ship with defaults that point at docs/… (writing-plans → docs/superpowers/plans/, brainstorming → docs/superpowers/specs/). Those defaults are overridden here. Whenever you invoke superpowers (or any plan/spec/research generator) in this project, save to _tasks/ instead, using the same filename convention:

When a superpowers skill announces a path like "saved to docs/superpowers/plans/…", rewrite it to the _tasks/… equivalent before writing. Commit those artifacts inside the _tasks/ repo (git -C _tasks …), never in the main repo.

Git Workflow

# Never commit directly to main
git checkout -b feat/your-feature
git commit -m "feat: describe your change"
git push -u origin feat/your-feature

Branch prefixes: feat/, fix/, refactor/, docs/, test/, chore/

Commit format (Conventional Commits): feat(db): add circuit breaker — scopes: db, sse, oauth, dashboard, api, cli, docker, ci, mcp, a2a, memory, skills

Husky hooks:

• pre-commit: lint-staged + check-docs-sync + check:any-budget:t11 + check:tracked-artifacts
• pre-push: intentionally light (PATH/npm sanity only). any-budget + tracked-artifacts already run on pre-commit; re-running them on every push was pure double-pay. CI still enforces both. (Was Fase 6A.12 full pre-push gate; folded into pre-commit in #6716.)

Worktree isolation (MANDATORY for every development task)

Multiple sessions/agents work this repo in parallel. The main checkout is shared, so a git checkout/branch switch in it silently discards another session's uncommitted work and yanks the branch out from under whatever else is running (incidents: 2026-06-05, 2026-06-13).

Rule: never develop on the shared main checkout. Every task gets its own git worktree on its own dedicated branch, and you MUST confirm the base branch with the operator before creating it.

1. Ask first — which base branch? Before creating anything, ask the operator (via AskUserQuestion, unless they already told you) from which branch the new worktree/branch should be cut. Do NOT assume main or "whatever I'm on" — the answer is usually the active release/vX.Y.Z, but it can be another feature/release branch. Get the base explicitly.

2. Create an isolated worktree + branch off that base (never reuse the main checkout). 🔴 MANDATORY PATH: every worktree lives under .claude/worktrees/ — and nowhere else. This is the single canonical location (the same dir the native EnterWorktree tool uses). It is gitignored AND in the tsconfig.json / .dockerignore excludes, so worktrees never leak into the build scope. Never use .worktrees/, repo-root, or any other path — a worktree outside .claude/worktrees/ (a) escapes the build-scope excludes and poisons next build (the tsconfig include: **/* globs ~70× the codebase → OOM; incident 2026-06-25) and (b) scatters worktrees across two dirs.

   BASE_BRANCH="release/vX.Y.Z"          # ← the branch the operator confirmed in step 1
   TASK="feat/your-feature"               # feat/ fix/ refactor/ docs/ test/ chore/
   git fetch origin "$BASE_BRANCH"
   git worktree add ".claude/worktrees/${TASK##*/}" -b "$TASK" "origin/$BASE_BRANCH"
   cd ".claude/worktrees/${TASK##*/}"
   # symlink node_modules from the main checkout to skip a per-worktree npm install:
   ln -s "$(git -C <main_checkout> rev-parse --show-toplevel)/node_modules" node_modules

   In Claude Code prefer the native EnterWorktree tool (it already creates worktrees under .claude/worktrees/): create the worktree with the command above, then call EnterWorktree with its path.

3. Work, commit, push, open the PR — all from inside the worktree. Never git checkout a different branch inside a worktree another session might share.
4. Tear down only your own worktree + branch when done, from the main checkout: git worktree remove .claude/worktrees/<dir> then git branch -D <task>. Never blanket-delete fix/*/feat/* — other sessions keep their own; delete only the branches you created, by name.
5. Never touch another session's worktree, branch, or uncommitted changes. If git worktree list shows worktrees you didn't create, leave them alone. End every session with the main checkout back on the branch it started on (the active release/vX.Y.Z, never main).

Environment

• Runtime: Node.js ≥22.0.0 <23 || ≥24.0.0 <27, ES Modules. This is the only supported runtime for the published nerve CLI, the server, and the test suites (node:test + vitest) — engines.node is authoritative and end users never need Bun. A best-effort bun:sqlite compatibility path exists so a global Bun install (bun install -g nerve) can start without better-sqlite3 (driver adapter + Bun-aware process spawning); it is not a supported runtime — no support guarantees — and every Bun-specific runtime change MUST preserve the Node driver/fallback chain and ship a Bun test (test:bun:db) or an explicit reason why the path is Node-only.
• Bun (build/dev script runner + compatibility smoke only): Bun 1.3.14 is pinned as an exact devDependency (provisioned through the existing npm ci via the lockfile's @oven/bun-* platform binaries — no setup-bun/ad-hoc install). It is used only to execute a small, allow-listed set of TypeScript gate/generator scripts (replacing node --import tsx for startup speed): the CI checks check:provider-consistency, check:compression-budget, check:known-symbols, and the non-CI gen:provider-reference, bench:compression — plus the focused test:bun:db compatibility smoke suite for the best-effort bun:sqlite path. Do NOT widen Bun to npm install, the build (build:cli*), check:pack-artifact, the supported published runtime, or the main test runners — those stay on Node. Any new Bun-invoking gate/generator script must be validated byte-identical against its node --import tsx output first. After pulling the lockfile change, run npm install so bun resolves locally (a stale node_modules will fail those scripts with bun: not found).
• TypeScript: 6.0+, target ES2022, module esnext, resolution bundler
• Path aliases: @/* → src/, @nerve/open-sse → open-sse/, @nerve/open-sse/* → open-sse/*
• Default port: 20128 (API + dashboard on same port)
• Data directory: DATA_DIR env var, defaults to ~/.nerve/
• Key env vars: PORT, JWT_SECRET, API_KEY_SECRET, INITIAL_PASSWORD, REQUIRE_API_KEY, APP_LOG_LEVEL
• Setup: cp .env.example .env then generate JWT_SECRET (openssl rand -base64 48) and API_KEY_SECRET (openssl rand -hex 32)

Quality Gates & Ratchets

Nerve has ~48 quality-gate scripts (scripts/check/ + scripts/quality/) wired across 9 gate-running jobs in .github/workflows/ci.yml (lint, quality-gate, quality-extended, docs-sync-strict, i18n-ui-coverage, i18n, pr-test-policy, test-vitest, sonarqube), plus the quality.yml fast-gates job (PR→release/**) and 3 nightly workflows (nightly-property, nightly-resilience, nightly-llm-security; nightly-mutation once merged). Full inventory, per-job breakdown, and operational procedures are in docs/architecture/QUALITY_GATES.md.

Quick reference:

• Gates in jobs lint + docs-sync-strict: pass/fail policy gates — fix the violation or add an allowlist entry with a justification comment + tracking issue.
• Gates in job quality-gate: ratchet — metrics (ESLint warnings, code coverage, duplication, complexity) must not regress vs quality-baseline.json. Update via npm run quality:ratchet -- --update when a metric genuinely improves.
• Job test-vitest runs npm run test:vitest (MCP tools, autoCombo, cache) — blocking. test:vitest:ui is advisory until UI component tests are triaged.

Allowlist policy (short form): Fix the cause; use the allowlist only for pre-existing violations you cannot fix in the same PR. Add a comment with justification + issue number. Stale allowlist entries (suppressing a violation that no longer exists) will be caught by the stale-enforcement added in Fase 6A.3.

Hard Rules

1. Never commit secrets or credentials
2. Never add logic to localDb.ts
3. Never use eval() / new Function() / implied eval
4. Never commit directly to main
5. Never write raw SQL in routes — use src/lib/db/ modules
6. Never silently swallow errors in SSE streams
7. Always validate inputs with Zod schemas
8. Always include tests when changing production code
9. Coverage must not regress below the baseline frozen in quality-baseline.json (ratchet); absolute floor is 60% (statements/lines/functions/branches). Update the baseline via npm run quality:ratchet -- --update only when coverage genuinely improves. See docs/architecture/QUALITY_GATES.md.
10. Never bypass Husky hooks (--no-verify, --no-gpg-sign) without explicit operator approval.
11. Never embed public upstream OAuth client_id/secret or Firebase Web keys as string literals — always go through resolvePublicCred() (open-sse/utils/publicCreds.ts). See docs/security/PUBLIC_CREDS.md.
12. Never return raw err.stack / err.message in HTTP / SSE / executor responses — always route through buildErrorBody() or sanitizeErrorMessage() (open-sse/utils/error.ts). See docs/security/ERROR_SANITIZATION.md.
13. Never string-interpolate external paths or runtime values into shell scripts passed to exec()/spawn() — pass via the env option instead. Reference: src/mitm/cert/install.ts::updateNssDatabases.
14. Never dismiss a CodeQL / Secret-Scanning alert without (a) first checking the pattern docs above to see if the helper applies, and (b) recording the technical justification in the dismissal comment. Precedent: js/stack-trace-exposure raised on callsites that already route through sanitizeErrorMessage() is a known CodeQL limitation (custom sanitizers not recognized) — dismiss as false positive referencing docs/security/ERROR_SANITIZATION.md.
15. Never expose routes that spawn child processes (/api/mcp/, /api/cli-tools/runtime/) without isLocalOnlyPath() classification in src/server/authz/routeGuard.ts. Loopback enforcement happens unconditionally before any auth check — leaked JWT via tunnel cannot trigger process spawning. See docs/security/ROUTE_GUARD_TIERS.md.
16. Never credit or advertise an AI assistant, LLM, or automation account in any commit/PR metadata. Two forbidden forms, both equivalent — they route attribution to a bot account (or advertise AI authorship) and hide the real author (diegosouzapw): (a) Co-Authored-By trailers naming an AI/bot (e.g. names containing "Claude", "GPT", "Copilot", "Bot"; emails at anthropic.com / openai.com / bot-owned noreply.github.com addresses); (b) AI-generation footers or descriptions anywhere in a commit message, PR title/body, or CHANGELOG — e.g. 🤖 Generated with [Claude Code], "Generated with Claude Code", "Made with <AI tool>", or any Co-authored-by: Claude/GPT/Copilot line. This overrides any harness, template, or tool default that auto-appends such a footer (e.g. the Claude Code PR-body/commit default) — strip it before pushing; do not let it reach a commit, PR, or CHANGELOG. Human collaborators — including upstream PR authors and issue reporters being ported into Nerve — MAY and SHOULD be credited with standard Co-authored-by: Name <email> trailers; the upstream-port workflows (/port-upstream-features, /port-upstream-issues) depend on this.
17. Never expose routes under /api/services/ or /dashboard/providers/services/*/embed/ without isLocalOnlyPath() classification in src/server/authz/routeGuard.ts. These routes can spawn child processes (npm install, node). Loopback enforcement happens unconditionally before any auth check — a leaked JWT via tunnel cannot trigger process spawning. See docs/security/ROUTE_GUARD_TIERS.md.
18. Every bug fix must be validated before shipping: a failing-then-passing unit/integration test (TDD) OR a documented live test on the production VPS (192.168.0.15). A fix without either is not merged. See Testing → "Bug fix / issue triage protocol" for the full decision tree.
19. Never develop on the shared main checkout. Every development task runs in its own git worktree on its own dedicated branch, and you MUST confirm the base branch with the operator (e.g. via AskUserQuestion) before creating the worktree/branch — never assume main or the currently checked-out branch. A git checkout in the shared checkout silently destroys other sessions' uncommitted work. Tear down only the worktrees/branches you created (by name, never fix/*/feat/* wildcards), leave other sessions' worktrees untouched, and end on the branch you started on (the active release/vX.Y.Z, never main). See Git Workflow → "Worktree isolation".
20. PII redaction/sanitization is opt-in — never on by default. Nerve proxies for self-hosted/local LLMs where the operator owns the data, so mutating request/response payloads by default would silently corrupt legitimate traffic. The two data-mutating PII feature flags MUST keep defaultValue: "false" in src/shared/constants/featureFlagDefinitions.ts: PII_REDACTION_ENABLED (request-side) and PII_RESPONSE_SANITIZATION (response + streaming). All three application points — src/lib/guardrails/piiMasker.ts (request guardrail), src/lib/piiSanitizer.ts (response), src/lib/streamingPiiTransform.ts (SSE) — are gated on these flags; with both off the pii-masker guardrail still runs but never mutates payloads (data passes through untouched). Flipping either default to "true" requires explicit operator approval. The regression guard is tests/unit/pii-opt-in-default.test.ts (asserts both definition defaults + behavioral pass-through). Opt-in is per-operator via env or the settings/DB override (src/lib/db/featureFlags.ts), never a silent default. See docs/security/GUARDRAILS.md.
21. Release-freeze — the FROZEN release branch belongs to the release captain; development does NOT stop (parallel-cycle model, 2026-07-04). /generate-release opens a marker issue labeled release-freeze at the start of reconciliation (Phase 0a), immediately cuts the next cycle's branch release/vX+1 from the frozen tip (Phase 0a.0b — bump + living release PR + re-home of open PRs), and closes the freeze once the release PR squash-merges to main. Before merging any PR, every campaign workflow (/review-prs, /review-group-prs, /merge-prs, /triage-fix-bugs, /implement-fix-bugs, /triage-features, /implement-features, /green-prs, /port-upstream-*) MUST check gh issue list --repo diegosouzapw/Nerve --label release-freeze --state open — if a freeze is active: NEVER merge into the frozen release/vX.Y.Z named in the freeze title; instead resolve the ACTIVE development branch (the highest release/v* by semver — normally release/vX+1, announced in a freeze-issue comment) and retarget the PR there (gh pr edit <N> --base release/vX+1, then VERIFY with gh pr view <N> --json baseRefName — the edit fails silently) and merge normally. HOLD only when the highest release/v* branch IS the frozen one (the short window before 0a.0b completes, or a pre-parallel-cycle release) — in that case leave the PR ready and open, tell the operator, and resume when the next branch appears or the freeze lifts. The just-shipped fixes reach release/vX+1 via the Phase 5 sync-back (scripts/release/sync-next-cycle.mjs); do not try to sync mid-release. This is a coordination signal, not a permission lock: the release captain and the campaign sessions share the diegosouzapw identity, so a GitHub branch-protection lock cannot distinguish them — only this honored marker prevents the mid-release commit races that forced full CHANGELOG re-reconciliation in v3.8.40/v3.8.41 (a parallel campaign advanced release/vX.Y.Z by 34 commits mid-run). The release captain's own reconciliation/cycle-open pushes are exempt — they are the release. Fixes that must land during a freeze (a homologation finding) follow the post-merge read-only rule: land on main first via fix/release-vX.Y.Z-*. ⛔ ONLY /generate-release may raise a release-freeze, and ONLY at its Phase 0a (start of generating a new version) — lifted at Phase 12c after the squash-merge to main. No campaign, session, or agent may open a release-freeze marker at any other time — a freeze is never a mid-development coordination tool. If a session ever believes a freeze is genuinely, unavoidably necessary outside the /generate-release flow, it MUST first ask the operator (diegosouzapw) in chat, explicitly alert "estou criando um freeze" and get an explicit yes — never open, extend, or re-open a release-freeze autonomously. Conversely, do not close/lift an active /generate-release freeze to unblock campaign merges: it protects the captain's single clean CI run and auto-lifts at Phase 12c — closing it early re-triggers the exact commit race it prevents. Verify a freeze is legitimate before acting on it: an open release-freeze whose title/body references an OPEN release PR (gh pr view <N> --json state) is the authorized captain freeze — hold, don't touch.
22. Cross-session safety — this repo is worked by MANY parallel sessions/agents at once; never step on another's in-flight work. Two absolute bans, both recurring incidents (this rule exists because they keep happening):
    • (a) Never git stash / git stash pop — ANYWHERE in this repo, including inside an isolated worktree, and including inside any subagent you dispatch. git stash operates on the shared repository object store, not the per-worktree working tree — so a stash pushed or popped in one session can silently clobber or resurrect another parallel session's uncommitted changes. This is not hypothetical: 2026-07-02 a #5923 quotaCache change leaked into the unrelated #2296 worktree via a global stash pop, and the same class reincided through a subagent. To compare working changes against a base ref without stashing, use git show <ref>:<path> or git diff <ref> -- <path>; to confirm a typecheck/lint error is pre-existing on the base, inspect the base ref directly (git show origin/release/vX.Y.Z:<path>) — never stash your tree away to "get it clean". Put this ban verbatim in the prompt of every subagent that touches git (agents don't inherit this file's context — the recurrence was a subagent).
    • (b) Never merge, push, rebase, or force-push a PR / branch / worktree that another session is actively working. An open PR whose head is a live fix worktree in .claude/worktrees/ you did not create (e.g. fix-5852/fix-5923 carrying fresh commits, even when they share your diegosouzapw identity), or any branch another session owns, is off-limits — HOLD, and let the owning session merge it. Before merging or pushing to any PR you did not create this session, run git worktree list to check for a matching in-flight worktree and re-check gh pr view <N> --json state,headRefOid. Only the owning session merges its own in-flight PR; mid-flight merges race the owner and re-trigger the exact commit/CHANGELOG races Rule #19 and Rule #21 guard against. (Reinforces Rule #19.)

PII & Stream Sanitization Learnings

1. Regex Security (ReDoS)

All regex patterns matching variable-length strings (e.g. IPv6 address, credit cards) must use strictly bounded, non-overlapping sequences (e.g., limit occurrences with bounded ranges {1,7}) to prevent catastrophic backtracking when processing untrusted inputs.

2. SSE Snapshot Handling

When parsing streaming LLM responses (e.g. Responses API), check if a chunk represents a final snapshot (done or completed events). Snapshot text must be sanitized directly as a standalone string (bypassing rolling delta buffers) to prevent text duplication at the end of the stream.

3. Database Handles in Tests

Ensure that any unit tests that trigger database migrations or establish SQLite connections call resetDbInstance() and properly clean up/close all DB handles in a test.after(...) hook. Failure to release database connection handles will cause Node's native test runner to hang indefinitely.