Comparison

A · CLAUDE.md · diegosouzapw/OmniRoute B · AGENTS.md · diegosouzapw/OmniRoute

What each file covers, counted
Dimension	Shared	Only in A	Only in B	Overlap
Sections	4	36	49	4%
Commands	14	26	15	25%
Section tags	12	1	4	71%

What each file covers

Sections

4 shared · 36 only in A · 49 only in B

− CLAUDE.md
− Quick Start
− Single test file (Node.js native test runner — most tests)
− Vitest (MCP server, autoCombo, cache)
− All suites
− Project at a Glance
− Request Pipeline
− Resilience Runtime State
− Provider Circuit Breaker
− Connection Cooldown
− Model Lockout
− Debugging Guidance
− Key Conventions
− Code Style
− Database
− Common Modification Scenarios
− Adding a New API Route
− Adding a New DB Module
− Adding a New MCP Tool
− Adding a New A2A Skill
− Adding a New Cloud Agent
− Adding a New Embedded Service
− Adding a New Guardrail / Eval / Skill / Webhook event
− Reference Documentation
− Testing
− Planning & Research Artifacts (superpowers, deep-research)
− Git Workflow
− Never commit directly to main
− Worktree isolation (MANDATORY for every development task)
− Environment
− Quality Gates & Ratchets
− Hard Rules
− PII & Stream Sanitization Learnings
− 1. Regex Security (ReDoS)
− 2. SSE Snapshot Handling
− 3. Database Handles in Tests
+ omniroute — Agent Guidelines
+ Project
+ Doc Accuracy Discipline (read before writing any doc)
+ Stack
+ Build, Lint, and Test Commands
+ All tests (unit + vitest + ecosystem + e2e)
+ Single test file (Node.js native test runner — most tests use this)
+ Integration tests
+ Vitest (MCP server, autoCombo)
+ E2E with Playwright
+ Protocol clients E2E (MCP transports, A2A)
+ Ecosystem compatibility tests
+ Coverage (see CONTRIBUTING.md)
+ Code Style Guidelines
+ Formatting (Prettier — enforced via lint-staged)
+ TypeScript
+ ESLint Rules
+ Naming
+ Imports
+ Architecture
+ Data Layer (`src/lib/db/`)
+ API Route Layer (`src/app/api/v1/`)
+ Request Pipeline (`open-sse/`)
+ Provider Categories
+ Executors (`open-sse/executors/`)
+ Translator (`open-sse/translator/`)
+ Transformer (`open-sse/transformer/`)
+ Services (`open-sse/services/`)
+ Domain Layer (`src/domain/`)
+ MCP Server (`open-sse/mcp-server/`)
+ A2A Server (`src/lib/a2a/`)
+ ACP Module (`src/lib/acp/`)
+ Memory System (`src/lib/memory/`)
+ Skills System (`src/lib/skills/`)
+ Compliance (`src/lib/compliance/`)
+ MITM Proxy (`src/mitm/`)
+ Middleware (`src/middleware/`)
+ Guardrails (`src/lib/guardrails/`)
+ Cloud Agents (`src/lib/cloudAgent/`)
+ Evals (`src/lib/evals/`)
+ Webhooks (`src/lib/webhookDispatcher.ts`)
+ Authorization Pipeline (`src/server/authz/`)
+ Reasoning Replay (`src/lib/db/reasoningCache.ts` + `open-sse/services/reasoningCache.ts`)
+ Tunnels (`src/lib/{cloudflaredTunnel,ngrokTunnel}.ts` + `src/app/api/tunnels/`)
+ Subdirectory AGENTS.md Files
+ Reference Documentation (docs/)
+ Fork / Upstream Workflow
+ the default branch is the active release line, e.g. release/v3.8.49
+ Review Focus
Running Tests
Error Handling
Security
Adding a New Provider

Commands

14 shared · 26 only in A · 15 only in B

− npm install
− git checkout -b feat/your-feature
− git commit -m "feat: describe your change"
− git push -u origin feat/your-feature
− git fetch origin "$BASE_BRANCH"
− git worktree add ".claude/worktrees/${TASK##*/}" -b "$TASK" "origin/$BASE_BRANCH"
− npm run test:unit
− node --import tsx/esm --test tests/unit/file.test.ts
− npm run coverage:report
− git -C _tasks …
− docker
− git checkout
− git worktree remove .claude/worktrees/<dir>
− git branch -D <task>
− node:test
− bun:sqlite
− bun install -g omniroute
− npm ci
− node --import tsx
− bun
− bun: not found
− npm run quality:ratchet -- --update
− node
− gh issue list --repo diegosouzapw/OmniRoute --label release-freeze --state open
− gh pr edit <N> --base release/vX+1
− gh pr view <N> --json baseRefName
+ node --import tsx/esm --test tests/unit/plan3-p0.test.ts
+ node --import tsx/esm --test tests/unit/fixes-p1.test.ts
+ node --import tsx/esm --test tests/unit/security-fase01.test.ts
+ node --import tsx/esm --test tests/integration/*.test.ts
+ git fetch upstream
+ git switch -c <branch-name> upstream/release/vX.Y.Z
+ npm run check:docs-all
+ npm run check:fabricated-docs
+ npm run build:release
+ npm run start
+ npm run build:cli
+ npm run electron:dev
+ npm run electron:build
+ prettier --write
+ npm run check:docs-counts
npm run dev
npm run build
npm run lint
npm run typecheck:core
npm run typecheck:noimplicit:core
npm run test:coverage
npm run check
npm run check:cycles
node --import tsx/esm --test tests/unit/your-file.test.ts
npm run test:vitest
npm run test:all
npm run test:e2e
npm run test:protocols:e2e
npm run test:ecosystem

Section tags

12 shared · 1 only in A · 4 only in B

− setup
+ types
+ testing-strategy
+ performance
+ monorepo
build
test
lint-format
code-style
git-pr
security
database
api
deployment
do-not
agent-behaviour
docs

Line diff

+458 added−430 removed143 unchanged23.8% identical

diegosouzapw/OmniRoute · CLAUDE.md

@@ −1 @@

1# CLAUDE.md

3This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

5## Quick Start

7```bash

8npm install # Install deps (auto-generates .env from .env.example)

9npm run dev # Dev server at http://localhost:20128

10npm run build # Production build (Next.js 16 standalone)

11npm run lint # ESLint (0 errors expected; warnings are pre-existing)

12npm run typecheck:core # TypeScript check (should be clean)

13npm run typecheck:noimplicit:core # Strict check (no implicit any)

14npm run test:coverage # Unit tests + coverage gate (60/60/60/60 — statements/lines/functions/branches)

15npm run check # lint + test combined

16npm run check:cycles # Detect circular dependencies

17```

19### Running Tests

21```bash

22# Single test file (Node.js native test runner — most tests)

23node --import tsx/esm --test tests/unit/your-file.test.ts

25# Vitest (MCP server, autoCombo, cache)

26npm run test:vitest

28# All suites

29npm run test:all

30```

32For full test matrix, see `CONTRIBUTING.md` → "Running Tests". For deep architecture, see `AGENTS.md`.

34---

36## Project at a Glance

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

40| Layer | Location | Purpose |

41| ------------- | ----------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------- |

42| API Routes | `src/app/api/v1/` | Next.js App Router — entry points |

43| Handlers | `open-sse/handlers/` | Request processing (chat, embeddings, etc) |

44| Executors | `open-sse/executors/` | Provider-specific HTTP dispatch |

45| Translators | `open-sse/translator/` | Format conversion (OpenAI↔Claude↔Gemini) |

46| Transformer | `open-sse/transformer/` | Responses API ↔ Chat Completions |

47| Services | `open-sse/services/` | Combo routing, rate limits, caching, etc |

48| Database | `src/lib/db/` | SQLite domain modules (130 migrations) |

49| Domain/Policy | `src/domain/` | Policy engine, cost rules, fallback logic |

50| MCP Server | `open-sse/mcp-server/` | 104 tools (42 base + memory/skill/agentSkill/pool/notion/obsidian/gamification/plugin modules), 3 transports (stdio / SSE / Streamable HTTP), 31 scopes |

51| A2A Server | `src/lib/a2a/` | JSON-RPC 2.0 agent protocol |

52| Skills | `src/lib/skills/` | Extensible skill framework |

53| Memory | `src/lib/memory/` | Persistent conversational memory |

55Monorepo: `src/` (Next.js 16 app), `open-sse/` (streaming engine workspace), `electron/` (desktop app), `tests/`, `bin/` (CLI entry point).

57---

59## Request Pipeline

61```

62Client → /v1/chat/completions (Next.js route)

63 → CORS → Zod validation → auth? → policy check → prompt injection guard

64 → handleChatCore() [open-sse/handlers/chatCore.ts]

65 → cache check → rate limit → combo routing?

66 → resolveComboTargets() → handleSingleModel() per target

67 → translateRequest() → getExecutor() → executor.execute()

68 → fetch() upstream → retry w/ backoff

69 → response translation → SSE stream or JSON

70 → If Responses API: responsesTransformer.ts TransformStream

71```

73API 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.

75**Combo routing** (`open-sse/services/combo.ts`): 19 public strategies (priority, weighted, fill-first, round-robin, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, cache-optimized, context-relay, 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 13-factor Auto-Combo scoring + the full strategy table and `docs/architecture/RESILIENCE_GUIDE.md` for the 3 resilience layers.

77---

79## Resilience Runtime State

81OmniRoute has three related but distinct temporary-failure mechanisms. Keep their

82scope separate when debugging routing behavior. See the

83[3-layer resilience diagram](./docs/diagrams/exported/resilience-3layers.svg)

84(source: [docs/diagrams/resilience-3layers.mmd](./docs/diagrams/resilience-3layers.mmd))

85for an at-a-glance map.

87### Provider Circuit Breaker

89**Scope**: whole provider, e.g. `glm`, `openai`, `anthropic`.

91**Purpose**: stop sending traffic to a provider that is repeatedly failing at the

92upstream/service level, so one unhealthy provider does not slow down every request.

94**Implementation**:

96- Core class: `src/shared/utils/circuitBreaker.ts`

97- Chat gate/execution wiring: `src/sse/handlers/chatHelpers.ts`, `src/sse/handlers/chat.ts`

98- Runtime status API: `src/app/api/monitoring/health/route.ts`

99- Shared wrappers: `open-sse/services/accountFallback.ts`

100- Persisted state table: `domain_circuit_breakers`

101

102**States**:

103

104- `CLOSED`: normal traffic is allowed.

105- `OPEN`: provider is temporarily blocked; callers get a provider-circuit-open response

106 or combo routing skips to another target.

107- `HALF_OPEN`: reset timeout has elapsed; allow a probe request. Success closes the

108 breaker, failure opens it again.

109

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

111

112- OAuth providers: threshold `3`, reset timeout `60s`.

113- API-key providers: threshold `5`, reset timeout `30s`.

114- Local providers: threshold `2`, reset timeout `15s`.

115

116Only provider-level failure statuses should trip the provider breaker:

117

118```ts

119(408, 500, 502, 503, 504);

120```

121

122Do not trip the whole-provider breaker for normal account/key/model errors like most

123`401`, `403`, or `429` cases. Those usually belong to connection cooldown or model

124lockout. A generic API-key provider `403` should be recoverable unless it is classified

125as a terminal provider/account error.

126

127The breaker uses lazy recovery, not a background timer. When `OPEN` expires, reads such

128as `getStatus()`, `canExecute()`, and `getRetryAfterMs()` refresh the state to

129`HALF_OPEN`, so dashboards and combo candidate builders do not keep excluding an

130expired provider forever.

131

132### Connection Cooldown

133

134**Scope**: one provider connection/account/key.

135

136**Purpose**: temporarily skip one bad key/account while allowing other connections for

137the same provider to continue serving requests.

138

139**Implementation**:

140

141- Write/update path: `src/sse/services/auth.ts::markAccountUnavailable()`

142- Account selection/filtering: `src/sse/services/auth.ts::getProviderCredentials...`

143- Cooldown calculation: `open-sse/services/accountFallback.ts::checkFallbackError()`

144- Settings: `src/lib/resilience/settings.ts`

145

146Important fields on provider connections:

147

148```ts

149rateLimitedUntil;

150testStatus: "unavailable";

151lastError;

152lastErrorType;

153errorCode;

154backoffLevel;

155```

156

157During account selection, a connection is skipped while:

158

159```ts

160new Date(rateLimitedUntil).getTime() > Date.now();

161```

162

163Cooldowns are also lazy: when `rateLimitedUntil` is in the past, the connection becomes

164eligible again. On successful use, `clearAccountError()` clears `testStatus`,

165`rateLimitedUntil`, error fields, and `backoffLevel`.

166

167Default connection cooldown behavior:

168

169- OAuth base cooldown: `5s`.

170- API-key base cooldown: `3s`.

171- API-key `429` should prefer upstream retry hints (`Retry-After`, reset headers, or

172 parseable reset text) when available.

173- Repeated recoverable failures use exponential backoff:

174

175```ts

176baseCooldownMs * 2 ** failureIndex;

177```

178

179The anti-thundering-herd guard prevents concurrent failures on the same connection from

180repeatedly extending the cooldown or double-incrementing `backoffLevel`.

181

182Terminal states are not cooldowns. `banned`, `expired`, and `credits_exhausted` are

183intended to stay unavailable until credentials/settings change or an operator resets

184them. Do not overwrite terminal states with transient cooldown state.

185

186### Model Lockout

187

188**Scope**: provider + connection + model.

189

190**Purpose**: avoid disabling a whole connection when only one model is unavailable or

191quota-limited for that connection.

192

193Examples:

194

195- Per-model quota providers returning `429`.

196- Local providers returning `404` for one missing model.

197- Provider-specific mode/model permission failures such as selected Grok modes.

198

199Model lockout lives in `open-sse/services/accountFallback.ts` and lets the same

200connection continue serving other models.

201

202### Debugging Guidance

203

204- If all keys for a provider are skipped, inspect both provider breaker state and each

205 connection's `rateLimitedUntil`/`testStatus`.

206- If a provider appears permanently excluded after the reset window, check whether code

207 is reading raw `state` instead of using `getStatus()`/`canExecute()`.

208- If one provider key fails but others should work, prefer connection cooldown over

209 provider breaker.

210- If only one model fails, prefer model lockout over connection cooldown.

211- If a state should self-recover, it should have a future timestamp/reset timeout and a

212 read path that refreshes expired state. Permanent statuses require manual credential

213 or config changes.

214

215---

216

217## Key Conventions

218

219### Code Style

220

221- **2 spaces**, semicolons, double quotes, 100 char width, es5 trailing commas (enforced by lint-staged via Prettier)

222- **Imports**: external → internal (`@/`, `@omniroute/open-sse`) → relative

223- **Naming**: files=camelCase/kebab, components=PascalCase, constants=UPPER_SNAKE

224- **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)

225- **TypeScript**: `strict: false`, target ES2022, module esnext, resolution bundler. Prefer explicit types.

226

227### Database

228

229- **Always** go through `src/lib/db/` domain modules — **never** write raw SQL in routes or handlers

230- **Never** add logic to `src/lib/localDb.ts` (re-export layer only)

231- **Never** barrel-import from `localDb.ts` — import specific `db/` modules instead

232- DB singleton: `getDbInstance()` from `src/lib/db/core.ts` (WAL journaling)

233- Migrations: `src/lib/db/migrations/` — versioned SQL files, idempotent, run in transactions

234

235### Error Handling

236

237- try/catch with specific error types, log with pino context

238- Never swallow errors in SSE streams — use abort signals for cleanup

239- Return proper HTTP status codes (4xx/5xx)

240

241### Security

242

243- **Never** use `eval()`, `new Function()`, or implied eval

244- Validate all inputs with Zod schemas

245- Encrypt credentials at rest (AES-256-GCM)

246- Upstream header denylist: `src/shared/constants/upstreamHeaders.ts` — keep sanitize, Zod schemas, and unit tests aligned when editing

247- **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.

248- **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`.

249- **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`.

250- **Secure-by-default libraries** ([tldrsec/awesome-secure-defaults](https://github.com/tldrsec/awesome-secure-defaults)): prefer Helmet.js, DOMPurify, ssrf-req-filter, safe-regex, Google Tink over custom implementations whenever adding new security-sensitive surfaces.

251

252---

253

254## Common Modification Scenarios

255

256### Adding a New Provider

257

2581. Register in `src/shared/constants/providers.ts` (Zod-validated at load)

2592. Add executor in `open-sse/executors/` if custom logic needed (extend `BaseExecutor`)

2603. Add translator in `open-sse/translator/` if non-OpenAI format

2614. 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

2625. Register models in `open-sse/config/providerRegistry.ts`

2636. Write tests in `tests/unit/` (include the publicCreds shape assertion if you added a new embedded default)

264

265### Adding a New API Route

266

2671. Create directory under `src/app/api/v1/your-route/`

2682. Create `route.ts` with `GET`/`POST` handlers

2693. Follow pattern: CORS → Zod body validation → optional auth → handler delegation

2704. Handler goes in `open-sse/handlers/` (import from there, not inline)

2715. 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`.

2726. Add tests — including at least one assertion that error responses do not leak stack traces (`!body.error.message.includes("at /")`)

273

274### Adding a New DB Module

275

2761. Create `src/lib/db/yourModule.ts` — import `getDbInstance` from `./core.ts`

2772. Export CRUD functions for your domain table(s)

2783. Add migration in `src/lib/db/migrations/` if new tables needed

2794. Re-export from `src/lib/localDb.ts` (add to the re-export list only)

2805. Write tests

281

282### Adding a New MCP Tool

283

2841. Add tool definition in `open-sse/mcp-server/tools/` with Zod input schema + async handler

2852. Register in tool set (wired by `createMcpServer()`)

2863. Assign to appropriate scope(s)

2874. Write tests (tool invocation logged to `mcp_audit` table)

288

289### Adding a New A2A Skill

290

2911. Create skill in `src/lib/a2a/skills/` (5 already exist: smart-routing, quota-management, provider-discovery, cost-analysis, health-report)

2922. Skill receives task context (messages, metadata) → returns structured result

2933. Register in `A2A_SKILL_HANDLERS` in `src/lib/a2a/taskExecution.ts`

2944. Expose in `src/app/.well-known/agent.json/route.ts` (Agent Card)

2955. Write tests in `tests/unit/`

2966. Document in `docs/frameworks/A2A-SERVER.md` skill table

297

298### Adding a New Cloud Agent

299

3001. Create agent class in `src/lib/cloudAgent/agents/` extending `CloudAgentBase` (3 already exist: codex-cloud, devin, jules)

3012. Implement `createTask`, `getStatus`, `approvePlan`, `sendMessage`, `listSources`

3023. Register in `src/lib/cloudAgent/registry.ts`

3034. Add OAuth/credentials handling if needed (`src/lib/oauth/providers/`)

3045. Tests + document in `docs/frameworks/CLOUD_AGENT.md`

305

306### Adding a New Embedded Service

307

3081. 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).

3092. Register the service in `src/lib/services/bootstrap.ts` (add to `SERVICES[]` array and extend `buildSpawnArgsFactory()`).

3103. Add a DB seed row for the new service in `src/lib/db/migrations/` (`version_manager` table, `status='not_installed'`, `auto_start=0`).

3114. 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`.

3125. 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).

3136. Add a UI tab in `src/app/(dashboard)/dashboard/providers/services/tabs/` reusing `ServiceStatusCard`, `ServiceLifecycleButtons`, `ServiceLogsPanel`.

3147. Document in `docs/frameworks/EMBEDDED-SERVICES.md` (update §1 service table + §4 API reference) and `docs/openapi.yaml`.

3158. 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.

316

317### Adding a New Guardrail / Eval / Skill / Webhook event

318

319- Guardrail: `src/lib/guardrails/` → docs: `docs/security/GUARDRAILS.md`

320- Eval suite: `src/lib/evals/` → docs: `docs/frameworks/EVALS.md`

321- Skill (sandbox): `src/lib/skills/` → docs: `docs/frameworks/SKILLS.md`

322- Webhook event: `src/lib/webhookDispatcher.ts` → docs: `docs/frameworks/WEBHOOKS.md`

323

324---

325

326## Reference Documentation

327

328For any non-trivial change, read the matching deep-dive first:

329

330| Area | Doc |

331| --------------------------------------------- | ------------------------------------------------------- |

332| Repo navigation | `docs/architecture/REPOSITORY_MAP.md` |

333| Architecture | `docs/architecture/ARCHITECTURE.md` |

334| Engineering reference | `docs/architecture/CODEBASE_DOCUMENTATION.md` |

335| Auto-Combo (13-factor scoring, 19 strategies) | `docs/routing/AUTO-COMBO.md` |

336| Resilience (3 mechanisms) | `docs/architecture/RESILIENCE_GUIDE.md` |

337| Reasoning replay | `docs/routing/REASONING_REPLAY.md` |

338| Skills framework | `docs/frameworks/SKILLS.md` |

339| Memory system (FTS5 + Qdrant) | `docs/frameworks/MEMORY.md` |

340| Cloud agents | `docs/frameworks/CLOUD_AGENT.md` |

341| Guardrails (PII / injection / vision) | `docs/security/GUARDRAILS.md` |

342| Public upstream credentials (Gemini/etc.) | `docs/security/PUBLIC_CREDS.md` |

343| Error message sanitization | `docs/security/ERROR_SANITIZATION.md` |

344| Evals | `docs/frameworks/EVALS.md` |

345| Compliance / audit | `docs/security/COMPLIANCE.md` |

346| Webhooks | `docs/frameworks/WEBHOOKS.md` |

347| Authorization pipeline | `docs/architecture/AUTHZ_GUIDE.md` |

348| Stealth (TLS / fingerprint) | `docs/security/STEALTH_GUIDE.md` |

349| Agent protocols (A2A / ACP / Cloud) | `docs/frameworks/AGENT_PROTOCOLS_GUIDE.md` |

350| MCP server | `docs/frameworks/MCP-SERVER.md` |

351| A2A server | `docs/frameworks/A2A-SERVER.md` |

352| API reference + OpenAPI | `docs/reference/API_REFERENCE.md` + `docs/openapi.yaml` |

353| Provider catalog (auto-generated) | `docs/reference/PROVIDER_REFERENCE.md` |

354| Release flow | `docs/ops/RELEASE_CHECKLIST.md` |

355| Embedded services | `docs/frameworks/EMBEDDED-SERVICES.md` |

356| Quality gates (~48 scripts, allowlist policy) | `docs/architecture/QUALITY_GATES.md` |

357

358---

359

360## Testing

361

362| What | Command |

363| ----------------------- | --------------------------------------------------------------------------- |

364| Unit tests | `npm run test:unit` |

365| Single file | `node --import tsx/esm --test tests/unit/file.test.ts` |

366| Vitest (MCP, autoCombo) | `npm run test:vitest` |

367| E2E (Playwright) | `npm run test:e2e` |

368| Protocol E2E (MCP+A2A) | `npm run test:protocols:e2e` |

369| Ecosystem | `npm run test:ecosystem` |

370| Coverage gate | `npm run test:coverage` (60/60/60/60 — statements/lines/functions/branches) |

371| Coverage report | `npm run coverage:report` |

372

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

374

375**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.

376

377**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.

378

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

380

3811. **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.

3822. **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.

3833. "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.

384

385Why 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).

386

387**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.

388

389---

390

391## Planning & Research Artifacts (superpowers, deep-research)

392

393`_tasks/` is a **separate, isolated git repository** that is gitignored by the main

394repo (`.gitignore` → `_tasks/`). It is the canonical home for working artifacts —

395plans, specs/designs, research, hand-offs — so they stay **versioned in their own

396repo** instead of polluting the main OmniRoute tree.

397

398**Hard rule — never write superpowers / planning / research output under `docs/` or

399the repo root.** The superpowers skills ship with defaults that point at `docs/…`

400(`writing-plans` → `docs/superpowers/plans/`, `brainstorming` → `docs/superpowers/specs/`).

401Those defaults are **overridden here**. Whenever you invoke superpowers (or any

402plan/spec/research generator) in this project, save to `_tasks/` instead, using the

403same filename convention:

404

405| Artifact (skill) | Default (do NOT use) | Save here instead |

406| ---------------------------------- | ------------------------- | ------------------------------------------------------------- |

407| Plans (`writing-plans`) | `docs/superpowers/plans/` | `_tasks/superpowers/plans/YYYY-MM-DD-<feature>.md` |

408| Specs / design (`brainstorming`) | `docs/superpowers/specs/` | `_tasks/superpowers/specs/YYYY-MM-DD-<topic>-design.md` |

409| Research (`deep-research`, ad-hoc) | `docs/research/` | `_tasks/research/…` |

410| Hand-offs (`/handoff`) | — | `_tasks/hands-off/<YYYY-MM-DD>_<branch>_v<versão>_sess-<id>/` |

411

412When a superpowers skill announces a path like "saved to `docs/superpowers/plans/…`",

413rewrite it to the `_tasks/…` equivalent before writing. Commit those artifacts inside

414the `_tasks/` repo (`git -C _tasks …`), never in the main repo.

415

416## Git Workflow

417

418```bash

419# Never commit directly to main

420git checkout -b feat/your-feature

421git commit -m "feat: describe your change"

422git push -u origin feat/your-feature

423```

424

425**Branch prefixes**: `feat/`, `fix/`, `refactor/`, `docs/`, `test/`, `chore/`

426

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

428

429**Husky hooks**:

430

431- **pre-commit**: lint-staged + `check-docs-sync` + `check:any-budget:t11` + `check:tracked-artifacts`

432- **pre-push**: intentionally light (PATH/npm sanity only). `any-budget` + `tracked-artifacts`

433 already run on pre-commit; re-running them on every push was pure double-pay. CI still

434 enforces both. (Was Fase 6A.12 full pre-push gate; folded into pre-commit in #6716.)

435

436### Worktree isolation (MANDATORY for every development task)

437

438Multiple sessions/agents work this repo in parallel. The main checkout is **shared**, so a

439`git checkout`/branch switch in it silently discards another session's uncommitted work and

440yanks the branch out from under whatever else is running (incidents: 2026-06-05, 2026-06-13).

441

442**Rule: never develop on the shared main checkout. Every task gets its own git worktree on its

443own dedicated branch, and you MUST confirm the base branch with the operator before creating it.**

444

4451. **Ask first — which base branch?** Before creating anything, ask the operator (via

446 `AskUserQuestion`, unless they already told you) from which branch the new worktree/branch

447 should be cut. Do NOT assume `main` or "whatever I'm on" — the answer is usually the active

448 `release/vX.Y.Z`, but it can be another feature/release branch. Get the base explicitly.

4492. **Create an isolated worktree + branch off that base** (never reuse the main checkout).

450 **🔴 MANDATORY PATH: every worktree lives under `.claude/worktrees/` — and nowhere else.**

451 This is the single canonical location (the same dir the native `EnterWorktree` tool uses). It

452 is gitignored AND in the `tsconfig.json` / `.dockerignore` excludes, so worktrees never leak

453 into the build scope. **Never** use `.worktrees/`, repo-root, or any other path — a worktree

454 outside `.claude/worktrees/` (a) escapes the build-scope excludes and poisons `next build` (the

455 `tsconfig` `include: **/*` globs ~70× the codebase → OOM; incident 2026-06-25) and (b) scatters

456 worktrees across two dirs.

457

458 ```bash

459 BASE_BRANCH="release/vX.Y.Z" # ← the branch the operator confirmed in step 1

460 TASK="feat/your-feature" # feat/ fix/ refactor/ docs/ test/ chore/

461 git fetch origin "$BASE_BRANCH"

462 git worktree add ".claude/worktrees/${TASK##*/}" -b "$TASK" "origin/$BASE_BRANCH"

463 cd ".claude/worktrees/${TASK##*/}"

464 # Reuse the main checkout's node_modules to skip a per-worktree npm install.

465 # HARD LINKS (`cp -al`), never a symlink: ~5s for the whole tree and near-zero extra

466 # disk (the inodes are shared), and unlike a symlink it does not break the dev server.

467 cp -al "$(git -C <main_checkout> rev-parse --show-toplevel)/node_modules" node_modules

468 ```

469

470 **Never `ln -s` node_modules.** Turbopack rejects a symlink that resolves outside the

471 project root, so `npm run dev` dies with a FATAL panic (`Symlink [project]/node_modules

472 is invalid, it points out of the filesystem root`) while typecheck, lint and the test

473 runners all keep passing — the error names "filesystem root", not the worktree, so it

474 reads like a Next/build bug and costs real time to trace (incident 2026-07-31, #9043).

475

476 In Claude Code prefer the native `EnterWorktree` tool (it already creates worktrees under

477 `.claude/worktrees/`): create the worktree with the command above, then call `EnterWorktree`

478 with its `path`.

479

4803. **Work, commit, push, open the PR — all from inside the worktree.** Never `git checkout` a

481 different branch inside a worktree another session might share.

4824. **Tear down only your own** worktree + branch when done, from the main checkout:

483 `git worktree remove .claude/worktrees/<dir>` then `git branch -D <task>`. Never blanket-delete

484 `fix/*`/`feat/*` — other sessions keep their own; delete only the branches you created, by name.

4855. **Never touch another session's worktree, branch, or uncommitted changes.** If `git worktree

486list` shows worktrees you didn't create, leave them alone. End every session with the main

487 checkout back on the branch it started on (the active `release/vX.Y.Z`, never `main`).

488

489---

490

491## Environment

492

493- **Runtime**: Node.js ≥22.0.0 <23 || ≥24.0.0 <27, ES Modules. This is the **only supported** runtime for the published `omniroute` 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 omniroute`) 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.

494- **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`).

495- **TypeScript**: 6.0+, target ES2022, module esnext, resolution bundler

496- **Path aliases**: `@/*` → `src/`, `@omniroute/open-sse` → `open-sse/`, `@omniroute/open-sse/*` → `open-sse/*`

497- **Default port**: 20128 (API + dashboard on same port)

498- **Data directory**: `DATA_DIR` env var, defaults to `~/.omniroute/`

499- **Key env vars**: `PORT`, `JWT_SECRET`, `API_KEY_SECRET`, `INITIAL_PASSWORD`, `REQUIRE_API_KEY`, `APP_LOG_LEVEL`

500- Setup: `cp .env.example .env` then generate `JWT_SECRET` (`openssl rand -base64 48`) and `API_KEY_SECRET` (`openssl rand -hex 32`)

501

502---

503

504## Quality Gates & Ratchets

505

506OmniRoute has **~48 quality-gate scripts** (`scripts/check/` + `scripts/quality/`) wired

507across **9 gate-running jobs** in `.github/workflows/ci.yml` (`lint`, `quality-gate`,

508`quality-extended`, `docs-sync-strict`, `i18n-ui-coverage`, `i18n`, `pr-test-policy`,

509`test-vitest`, `sonarqube`), plus the `quality.yml` fast-gates job (PR→`release/**`) and

5103 nightly workflows (`nightly-property`, `nightly-resilience`, `nightly-llm-security`;

511`nightly-mutation` once merged). Full inventory, per-job breakdown, and operational

512procedures are in [`docs/architecture/QUALITY_GATES.md`](docs/architecture/QUALITY_GATES.md).

513

514**Quick reference:**

515

516- Gates in jobs `lint` + `docs-sync-strict`: pass/fail policy gates —

517 fix the violation or add an allowlist entry with a justification comment + tracking issue.

518- Gates in job `quality-gate`: ratchet — metrics (ESLint warnings, code coverage, duplication,

519 complexity) must not regress vs `quality-baseline.json`. Update via

520 `npm run quality:ratchet -- --update` when a metric genuinely improves.

521- Job `test-vitest` runs `npm run test:vitest` (MCP tools, autoCombo, cache) — blocking.

522 `test:vitest:ui` is advisory until UI component tests are triaged.

523

524**Allowlist policy (short form):** Fix the cause; use the allowlist only for pre-existing

525violations you cannot fix in the same PR. Add a comment with justification + issue number.

526Stale allowlist entries (suppressing a violation that no longer exists) will be caught by

527the stale-enforcement added in Fase 6A.3.

528

529---

530

531## Hard Rules

532

5331. Never commit secrets or credentials

5342. Never add logic to `localDb.ts`

5353. Never use `eval()` / `new Function()` / implied eval

5364. Never commit directly to `main`

5375. Never write raw SQL in routes — use `src/lib/db/` modules

5386. Never silently swallow errors in SSE streams

5397. Always validate inputs with Zod schemas

5408. Always include tests when changing production code

5419. 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`.

54210. Never bypass Husky hooks (`--no-verify`, `--no-gpg-sign`) without explicit operator approval.

54311. 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`.

54412. 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`.

54513. 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`.

54614. 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`.

54715. 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`.

54816. 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 OmniRoute — 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.

54917. 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`.

55018. 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.

55119. 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".

55220. PII redaction/sanitization is **opt-in — never on by default**. OmniRoute 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`.

55321. **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/OmniRoute --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.

55422. **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):

555 - **(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).

556 - **(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.)

557

558---

559

560## PII & Stream Sanitization Learnings

561

562### 1. Regex Security (ReDoS)

563

564All 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.

565

566### 2. SSE Snapshot Handling

567

568When 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.

569

570### 3. Database Handles in Tests

571

572Ensure 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.

573

diegosouzapw/OmniRoute · AGENTS.md

@@ +1 @@

1# omniroute — Agent Guidelines

3## Project

5Unified AI proxy/router — route any LLM through one endpoint. Multi-provider support

6with **290 provider entries** (OpenAI, Anthropic, Gemini, DeepSeek, Groq, xAI, Mistral, Fireworks,

7Cohere, NVIDIA, Cerebras, Pollinations, Puter, Cloudflare AI, HuggingFace, DeepInfra,

8SambaNova, Meta Llama API, Moonshot AI, AI21 Labs, Databricks, Snowflake, and many more)

9with **MCP Server** (104 tools), **A2A v0.3 Protocol**, and **Electron desktop app**.

11> **Live counts (v3.8.49)**: providers 290 · MCP tools 104 · MCP scopes 30 · A2A skills 6 ·

12> open-sse services 134 · routing strategies 17 · auto-combo scoring factors 12 ·

13> DB modules 95 · DB migrations 110 · base tables 17 · search providers 11 ·

14> i18n locales 42. **Refresh with `npm run check:docs-all`.**

16## Doc Accuracy Discipline (read before writing any doc)

18> **If `grep -rn "name" src/ open-sse/ bin/` returns nothing, the name does not exist. Do not document it.**

20The recurring failure mode in AI-generated docs is _plausible-but-unverified specifics_.

21Every claim in a `.md` file under `docs/` should be verifiable against the source.

23**Rules (enforced by `npm run check:fabricated-docs`):**

251. **Never state an API name, endpoint, path, CLI command, or env var without grepping for it first.**

26 ```bash

27 grep -rn "theName" src/ open-sse/ bin/

28 # 0 hits → do not document

29 ```

302. **Never write a line count, file size, migration count, provider count, or strategy count from memory.**

31 ```bash

32 wc -l <file> # exact line count

33 ls <dir>/*.ts | wc -l # file count

34 ```

353. **Every code example should be copy-pasted from real usage or actually run** — not synthesized.

36 Link to a real call site (`path:line`) instead of inventing a signature.

374. **Prefer citing real source (`file.ts:line`) over paraphrasing behavior** — verifiable and self-correcting.

385. **A shorter doc that is 100% accurate beats a comprehensive one with fabrications.**

39 Wrong docs cost more than missing docs, because people trust and act on them.

41The script `scripts/check/check-fabricated-docs.mjs` extracts every route path, env var, hook

42name, function name, and file reference from `docs/**/*.md` and verifies each one against the

43codebase. Run it locally before pushing docs; it runs in CI via `npm run check:docs-all`.

45## Stack

47- **Runtime**: Next.js 16 (App Router), Node.js `>=22.0.0 <23 || >=24.0.0 <27`, ES Modules (`"type": "module"`)

48- **Language**: TypeScript 6.0 (`src/`) + JavaScript (`open-sse/`, `electron/`)

49- **Database**: better-sqlite3 (SQLite) — `DATA_DIR` configurable, default `~/.omniroute/`

50- **Streaming**: SSE via `open-sse` internal workspace package

51- **Styling**: Tailwind CSS v4

52- **i18n**: next-intl with 42 locales (`src/i18n/messages/`) — refresh with `ls src/i18n/messages/*.json | wc -l`

53- **Desktop**: Electron (cross-platform: Windows, macOS, Linux)

54- **Schemas**: Zod v4 for all API / MCP input validation

56---

58## Build, Lint, and Test Commands

60| Command | Description |

61| ----------------------------------- | ------------------------------------------------------------------ |

62| `npm run dev` | Start Next.js dev server |

63| `npm run build` | Production build: `next build` → `.build/next/` + assemble `dist/` |

64| `npm run build:release` | Clean rebuild + HEAD sentinel (`dist/BUILD_SHA`) — use for deploy |

65| `npm run start` | Run production build |

66| `npm run build:cli` | Build CLI package |

67| `npm run lint` | ESLint on all source files |

68| `npm run typecheck:core` | TypeScript core type checking |

69| `npm run typecheck:noimplicit:core` | Strict checking (no implicit any) |

70| `npm run check` | Run lint + test |

71| `npm run check:cycles` | Check for circular dependencies |

72| `npm run electron:dev` | Run Electron app in dev mode |

73| `npm run electron:build` | Build Electron app for current OS |

75**Build output layout:**

77| Directory | Purpose | Gitignored |

78| --------- | -------------------------------------------------- | ---------- |

79| `src/` | Application source (TypeScript / TSX) | No |

80| `.build/` | Build intermediates (`distDir = .build/next`) | Yes |

81| `dist/` | Shippable bundle assembled by `assembleStandalone` | Yes |

83The pipeline is a single `next build` pass — intermediates land in `.build/next/`, the

84assembled bundle in `dist/`. VPS deploys rsync `dist/` into the remote

85`/usr/lib/node_modules/omniroute/app/` directory (VPS image path is unchanged).

87### Running Tests

89```bash

90# All tests (unit + vitest + ecosystem + e2e)

91npm run test:all

93# Single test file (Node.js native test runner — most tests use this)

94node --import tsx/esm --test tests/unit/your-file.test.ts

95node --import tsx/esm --test tests/unit/plan3-p0.test.ts

96node --import tsx/esm --test tests/unit/fixes-p1.test.ts

97node --import tsx/esm --test tests/unit/security-fase01.test.ts

99# Integration tests

100node --import tsx/esm --test tests/integration/*.test.ts

101

102# Vitest (MCP server, autoCombo)

103npm run test:vitest

104

105# E2E with Playwright

106npm run test:e2e

107

108# Protocol clients E2E (MCP transports, A2A)

109npm run test:protocols:e2e

110

111# Ecosystem compatibility tests

112npm run test:ecosystem

113

114# Coverage (see CONTRIBUTING.md)

115npm run test:coverage

116```

117

118**For authoritative coverage requirements, test execution, and PR gates, see [`CONTRIBUTING.md`](CONTRIBUTING.md#running-tests).**

119

120---

121

122## Code Style Guidelines

123

124### Formatting (Prettier — enforced via lint-staged)

125

1262 spaces · semicolons required · double quotes (`"`) · 100 char width · es5 trailing commas.

127Always run `prettier --write` on changed files.

128

129### TypeScript

130

131- **Target**: ES2022 · **Module**: `esnext` · **Resolution**: `bundler`

132- `strict: false` — prefer explicit types, don't rely on inference

133- Path aliases: `@/*` → `src/`, `@omniroute/open-sse` → `open-sse/`, `@omniroute/open-sse/*` → `open-sse/*`

134

135### ESLint Rules

136

137- **Security (error, everywhere)**: `no-eval`, `no-implied-eval`, `no-new-func`

138- **Relaxed in `open-sse/` and `tests/`**: `@typescript-eslint/no-explicit-any` = warn

139- React hooks rules and `@next/next/no-assign-module-variable` disabled in `open-sse/` and `tests/`

140

141### Naming

142

143| Element | Convention | Example |

144| ------------------- | -------------------------------- | ------------------------------------ |

145| Files | camelCase / kebab-case | `chatCore.ts`, `tokenHealthCheck.ts` |

146| React components | PascalCase | `Dashboard.tsx`, `ProviderCard.tsx` |

147| Functions/variables | camelCase | `getHealth()`, `switchCombo()` |

148| Constants | UPPER_SNAKE | `MAX_RETRIES`, `DEFAULT_TIMEOUT` |

149| Interfaces | PascalCase (`I` prefix optional) | `ProviderConfig` |

150| Enums | PascalCase (members too) | `LogLevel.Error` |

151

152### Imports

153

154- **Order**: external → internal (`@/`, `@omniroute/open-sse`) → relative (`./`, `../`)

155- **No barrel imports** from `localDb.ts` — import from the specific `db/` module instead

156

157### Error Handling

158

159- try/catch with specific error types; always log with context (pino logger)

160- Never silently swallow errors in SSE streams — use abort signals for cleanup

161- Return proper HTTP status codes (4xx client, 5xx server)

162

163### Security

164

165- **NEVER** commit API keys, secrets, or credentials

166- Validate all user inputs with Zod schemas

167- Auth middleware required on all API routes

168- Never log SQLite encryption keys

169- Sanitize user content (dompurify for HTML)

170- **Public upstream OAuth identifiers** (Gemini / Antigravity / Windsurf-style client_id/secret + Firebase Web keys extracted from public CLIs): use `resolvePublicCred()` from `open-sse/utils/publicCreds.ts`, **never** as string literals. Full pattern in `docs/security/PUBLIC_CREDS.md`.

171- **Error responses** (HTTP / SSE / executor / MCP): use `buildErrorBody()` or `sanitizeErrorMessage()` from `open-sse/utils/error.ts`, **never** put raw `err.stack` / `err.message` in a Response body. Full pattern in `docs/security/ERROR_SANITIZATION.md`.

172- **`exec()` / `spawn()` with runtime values**: pass via the `env` option, **never** string-interpolate paths/values into the script body. Reference: `src/mitm/cert/install.ts::updateNssDatabases`.

173- Prefer secure-by-default libraries when available — see [tldrsec/awesome-secure-defaults](https://github.com/tldrsec/awesome-secure-defaults) for the curated list (Helmet.js, DOMPurify, ssrf-req-filter, safe-regex, Google Tink, etc.).

174

175---

176

177## Architecture

178

179### Data Layer (`src/lib/db/`)

180

181All persistence uses SQLite through **95 domain-specific modules** in `src/lib/db/`. Top modules:

182

183- Core: `core.ts`, `migrationRunner.ts`, `encryption.ts`, `stateReset.ts`

184- Providers / catalog: `providers.ts`, `models.ts`, `providerLimits.ts`, `compressionAnalytics.ts`

185- Routing: `combos.ts`, `modelComboMappings.ts`, `domainState.ts`, `commandCodeAuth.ts`

186- Auth: `apiKeys.ts`, `secrets.ts`, `registeredKeys.ts`, `sessionAccountAffinity.ts`

187- Usage / billing: `quotaSnapshots.ts`, `creditBalance.ts`, `usage*.ts`, `compressionCacheStats.ts`

188- Storage: `backup.ts`, `cleanup.ts`, `jsonMigration.ts`, `healthCheck.ts`, `databaseSettings.ts`

189- Extension modules: `evals.ts`, `webhooks.ts`, `reasoningCache.ts`, `readCache.ts`, `tierConfig.ts`, `compressionCombos.ts`, `compressionScheduler.ts`, `batches.ts`, `files.ts`, `syncTokens.ts`, `proxies.ts`, `oneproxy.ts`, `upstreamProxy.ts`, `versionManager.ts`, `cliToolState.ts`, `prompts.ts`, `detailedLogs.ts`, `contextHandoffs.ts`, `compression.ts`, `stats.ts`

190

191Live count: `ls src/lib/db/*.ts | wc -l` (currently 95). Drift detection: `npm run check:docs-counts`.

192Schema migrations live in `db/migrations/` (**110 files** as of v3.8.43) and run via `migrationRunner.ts`.

193`src/lib/localDb.ts` is a **re-export layer only** — never add logic there.

194

195#### DB Internals

196

197- **`core.ts`**: `getDbInstance()` returns a singleton `better-sqlite3` instance with WAL

198 journaling. `SCHEMA_SQL` defines **17 base tables** (verify with `grep -c "CREATE TABLE" src/lib/db/core.ts` minus 1 for the bookkeeping `_omniroute_migrations` table). Helpers: `rowToCamel`, `encryptConnectionFields`.

199- **`migrationRunner.ts`**: Applies versioned SQL files from `db/migrations/` inside transactions.

200 Tracks applied migrations in `_omniroute_migrations` table.

201- **Migrations**: 110 files (`001_initial_schema.sql` → `110_*.sql`).

202 Each migration is idempotent and runs in a transaction. Live count: `ls src/lib/db/migrations/*.sql | wc -l`.

203- **Domain modules** import `getDbInstance()` from `core.ts` for all CRUD operations.

204 Each module owns a specific table/set of tables (e.g., `providers.ts` → `provider_connections`,

205 `combos.ts` → `combos`). Encryption helpers protect sensitive fields at rest.

206- **`localDb.ts`** re-exports all domain modules — consumers import from here for convenience.

207

208### API Route Layer (`src/app/api/v1/`)

209

210Next.js App Router routes — each follows a consistent pattern:

211

212```

213Route → CORS preflight → Body validation (Zod) → Optional auth (extractApiKey/isValidApiKey)

214 → API key policy enforcement (enforceApiKeyPolicy) → Handler delegation (open-sse)

215```

216

217| Route | Handler | Notes |

218| ------------------------------- | ------------------------- | ------------------------------------------------------------- |

219| `chat/completions/route.ts` | `handleChat()` | + prompt injection guard (clones request) |

220| `responses/route.ts` | `handleChat()` (unified) | Responses API format |

221| `embeddings/route.ts` | `handleEmbedding()` | Model listing + creation |

222| `images/generations/route.ts` | `handleImageGeneration()` | Model listing + creation |

223| `audio/transcriptions/route.ts` | audio handler | Multipart form data |

224| `audio/speech/route.ts` | TTS handler | Binary audio response |

225| `videos/generations/route.ts` | video handler | ComfyUI/SD WebUI |

226| `music/generations/route.ts` | music handler | ComfyUI workflows |

227| `moderations/route.ts` | moderation handler | Content safety |

228| `rerank/route.ts` | rerank handler | Document relevance |

229| `search/route.ts` | search handler | Web search (12 providers per `open-sse/handlers/search.ts:6`) |

230

231**No global Next.js middleware file** — interception is route-specific. Auth is optional

232(controlled by `REQUIRE_API_KEY` env). Prompt injection guard is unique to chat completions.

233

234### Request Pipeline (`open-sse/`)

235

236The `open-sse/` workspace is the core streaming engine. Full request flow:

237

238```

239Client Request

240 → src/app/api/v1/.../route.ts (Next.js route)

241 → open-sse/handlers/chatCore.ts::handleChatCore()

242 → Semantic/signature cache check

243 → Rate limit check (rateLimitManager)

244 → Combo routing? → open-sse/services/combo.ts::handleComboChat()

245 → resolveComboTargets() → ordered ResolvedComboTarget[]

246 → For each target: handleSingleModel() (wraps chatCore)

247 → translateRequest() (open-sse/translator/)

248 → Convert source format (e.g., OpenAI) → target format (e.g., Claude)

249 → getExecutor() → provider-specific executor instance

250 → executor.execute() (BaseExecutor → DefaultExecutor or provider-specific)

251 → buildUrl() + buildHeaders() + transformRequest()

252 → fetch() to upstream provider

253 → Retry logic with exponential backoff

254 → Response translation back to client format

255 → If Responses API: responsesTransformer.ts TransformStream

256 → SSE stream or JSON response to client

257```

258

259**Handlers** (`open-sse/handlers/`): `chatCore.ts`, `responsesHandler.ts`, `embeddings.ts`,

260`imageGeneration.ts`, `videoGeneration.ts`, `musicGeneration.ts`, `audioSpeech.ts`,

261`audioTranscription.ts`, `moderations.ts`, `rerank.ts`, `search.ts`.

262

263**Upstream headers**: merged after default auth; same header name replaces executor value.

264**T5 intra-family fallback** recomputes headers using only the fallback model id.

265Forbidden header names: `src/shared/constants/upstreamHeaders.ts` — keep sanitize,

266Zod schemas, and unit tests aligned when editing.

267

268### Provider Categories

269

270- **Free** (2): Qoder AI, Kiro AI

271- **OAuth** (13): Claude Code, Antigravity, Codex, GitHub Copilot, Cursor, Kimi Coding, Kilo Code, Cline, Kiro, Qoder, Gemini, Windsurf (v3.8), GitLab Duo (v3.8)

272- **API Key** (120+): OpenAI, Anthropic, Gemini, DeepSeek, Groq, xAI, Mistral, Perplexity,

273 Together, Fireworks, Cerebras, Cohere, NVIDIA, Nebius, SiliconFlow, Hyperbolic,

274 HuggingFace, OpenRouter, Vertex AI, Cloudflare AI, Scaleway, AI/ML API, Pollinations,

275 Puter, Longcat, Alibaba, Kimi, Minimax, Blackbox, Synthetic, Kilo Gateway,

276 Z.AI, GLM, Deepgram, AssemblyAI, ElevenLabs, Cartesia, PlayHT, Inworld,

277 NanoBanana, SD WebUI, ComfyUI, Ollama Cloud, Perplexity Search, Serper, Brave, Exa,

278 Tavily, OpenCode Zen/Go, Bailian Coding Plan, DeepInfra, Vercel AI Gateway,

279 Lambda AI, SambaNova, nScale, OVHcloud AI, Baseten, PublicAI, Moonshot AI,

280 Meta Llama API, v0 (Vercel), Morph, Featherless AI, FriendliAI, LlamaGate,

281 Galadriel, Weights & Biases Inference, Volcengine, AI21 Labs, Venice.ai,

282 Codestral, Upstage, Maritalk, Xiaomi MiMo, Inference.net, NanoGPT, Predibase,

283 Bytez, Heroku AI, Databricks, Snowflake Cortex, GigaChat (Sber), CrofAI,

284 AgentRouter, ChatGPT Web, Baidu Qianfan, AWS Polly, RunwayML, GitLab Duo,

285 Amazon Q, Empower, Poe, and many more.

286- **Self-Hosted** (8+): LM Studio, vLLM, Lemonade, Llamafile, Triton, Docker Model Runner, Xinference, Oobabooga

287- **Custom**: OpenAI-compatible (`openai-compatible-*`) and Anthropic-compatible (`anthropic-compatible-*`) prefixes

288

289Providers are registered in `src/shared/constants/providers.ts` with Zod validation at module load.

290

291### Executors (`open-sse/executors/`)

292

293Provider-specific request executors: `base.ts`, `default.ts`, `cursor.ts`, `codex.ts`,

294`antigravity.ts`, `github.ts`, `kiro.ts`, `qoder.ts`, `vertex.ts`,

295`cloudflare-ai.ts`, `opencode.ts`, `pollinations.ts`, `puter.ts`.

296

297#### Executor Internals

298

299- **`base.ts`** (`BaseExecutor`): Abstract base with `buildUrl()`, `buildHeaders()`,

300 `transformRequest()`, retry logic (exponential backoff), and `execute()`. Subclasses

301 override URL/header/transform methods for provider-specific behavior.

302- **`default.ts`** (`DefaultExecutor extends BaseExecutor`): Handles most OpenAI-compatible

303 providers. Reads provider config from `providerRegistry.ts` to resolve base URL, auth

304 header format, and request transformations.

305- **`getExecutor()`** (`executors/index.ts`): Factory that returns the correct executor

306 instance based on provider ID. Provider-specific executors (Cursor, Codex, Vertex, etc.)

307 override only what differs from the default.

308

309### Translator (`open-sse/translator/`)

310

311Translates between API formats (OpenAI-format ↔ Anthropic, Gemini, etc.).

312Includes request/response translators with helpers for image handling.

313

314#### Translator Internals

315

316- **`translator/index.ts`**: Exports `translateRequest()` and format constants. Called by

317 `chatCore.ts` before executor dispatch.

318- **Flow**: `translateRequest(body, sourceFormat, targetFormat)` → detects source format

319 (OpenAI, Anthropic, Gemini) → applies the matching translator module → returns

320 transformed body ready for the target provider.

321- **Response translation** runs in reverse after upstream response, converting back to

322 the client's expected format.

323

324### Transformer (`open-sse/transformer/`)

325

326`responsesTransformer.ts` — transforms Responses API format to/from Chat Completions format.

327

328#### Transformer Internals

329

330- **`createResponsesApiTransformStream()`**: Returns a `TransformStream` that converts

331 Chat Completions SSE chunks (`data: {"choices":[...]}`) into Responses API SSE events

332 (`response.output_item.added`, `response.output_text.delta`, etc.).

333- Used when the client sends a Responses API request: the request is internally converted

334 to Chat Completions format, dispatched normally, and the response is piped through this

335 transform stream before reaching the client.

336

337### Services (`open-sse/services/`)

338

339134 service modules in `open-sse/services/` (top-level only; more including sub-dirs like `autoCombo/` and `compression/`). Refresh: `ls open-sse/services/*.ts | wc -l`. Key modules:

340`combo.ts` (routing engine), `usage.ts`, `tokenRefresh.ts`,

341`rateLimitManager.ts`, `accountFallback.ts`, `sessionManager.ts`, `wildcardRouter.ts`,

342`autoCombo/`, `intentClassifier.ts`, `taskAwareRouter.ts`, `thinkingBudget.ts`,

343`contextManager.ts`, `modelDeprecation.ts`, `modelFamilyFallback.ts`,

344`emergencyFallback.ts`, `workflowFSM.ts`, `backgroundTaskDetector.ts`, `ipFilter.ts`,

345`signatureCache.ts`, `volumeDetector.ts`, `contextHandoff.ts`, `compression/` (prompt

346compression pipeline), and more.

347

348#### Prompt Compression Pipeline (`compression/`)

349

350Modular prompt compression that runs proactively before the existing reactive context manager.

351

352- **`strategySelector.ts`**: Selects compression mode based on config, compression combo assignments,

353 combo overrides, auto-trigger thresholds, and defaults. Priority: assigned compression combo >

354 combo override > auto-trigger > default mode > off.

355- **`lite.ts`**: 5 lite-mode techniques: `collapseWhitespace`, `dedupSystemPrompt`,

356 `compressToolResults`, `removeRedundantContent`, `replaceImageUrls`. Target: 10-15% savings at

357 <1ms latency.

358- **`caveman.ts` / `cavemanRules.ts`**: Caveman-style semantic condensation backed by built-in

359 rules plus file-loaded language packs under `compression/rules/`.

360- **`engines/rtk/`**: Rule-based terminal/tool-output compression inspired by RTK patterns. Detects

361 command output classes, applies JSON filter packs, deduplicates repeated lines, strips ANSI/code

362 noise, and preserves errors/actionable context. The RTK JSON DSL supports replace,

363 match-output short-circuit, strip/keep, per-line truncation, head/tail/max-line truncation,

364 inline tests, trust-gated project/global custom filters, and optional redacted raw-output

365 retention for authenticated recovery.

366- **`engines/registry.ts`**: Registers engines (`caveman`, `rtk`) and powers stacked pipelines.

367- **`stats.ts`**: Per-request compression stats tracking (original tokens, compressed tokens,

368 savings %, techniques used, engine breakdown, compression combo id).

369- **`types.ts`**: `CompressionMode` (off/lite/standard/aggressive/ultra/rtk/stacked),

370 `CompressionConfig`, `CompressionStats`, `CompressionResult`.

371- DB settings in `src/lib/db/compression.ts`, compression combos in

372 `src/lib/db/compressionCombos.ts`, API routes under `src/app/api/settings/compression/`,

373 `src/app/api/context/*`, and preview/language-pack routes under `src/app/api/compression/*`.

374

375#### Combo Routing Engine (`combo.ts`)

376

377- **`handleComboChat()`**: Entry point for combo-routed requests. Receives the combo config

378 and iterates through targets in order until one succeeds or all fail.

379- **`resolveComboTargets()`**: Expands a combo configuration into an ordered array of

380 `ResolvedComboTarget[]`, each specifying provider + model + account + credentials.

381- **Strategies** (17): priority, weighted, fill-first, round-robin, P2C, random, least-used, reset-aware (v3.8),

382 reset-window, cost-optimized, strict-random, auto, lkgp, context-optimized, context-relay, headroom, fusion. Source: `ROUTING_STRATEGY_VALUES` in `src/shared/constants/routingStrategies.ts`.

383- Each target calls **`handleSingleModel()`** which wraps `handleChatCore()` with

384 per-target error handling and circuit breaker checks.

385

386### Domain Layer (`src/domain/`)

387

388Policy engine modules: `policyEngine.ts`, `comboResolver.ts`, `costRules.ts`,

389`degradation.ts`, `fallbackPolicy.ts`, `lockoutPolicy.ts`, `modelAvailability.ts`,

390`providerExpiration.ts`, `quotaCache.ts`, `responses.ts`, `configAudit.ts`.

391

392### MCP Server (`open-sse/mcp-server/`)

393

394**104 tools** total (`TOTAL_MCP_TOOL_COUNT`, `open-sse/mcp-server/server.ts`): a 42-entry base registry (`MCP_TOOLS` in `schemas/tools.ts`, bundling the core / cache / compression / 1proxy / advanced tools) **plus** standalone module sets — memory (3), skill (4), agentSkill (3), pool (6), gamification (8), plugin (8), notion (6), obsidian (22). 3 transports (stdio / SSE / Streamable HTTP). Scoped auth (31 scopes — see `OMNIROUTE_MCP_SCOPES`), Zod schemas. See [`docs/frameworks/MCP-SERVER.md`](docs/frameworks/MCP-SERVER.md).

395

396**Core tools** (20): get_health, list_combos, get_combo_metrics, switch_combo, check_quota,

397route_request, cost_report, list_models_catalog, web_search, simulate_route, set_budget_guard,

398set_routing_strategy, set_resilience_profile, test_combo, get_provider_metrics,

399best_combo_for_task, explain_route, get_session_snapshot, db_health_check, sync_pricing.

400

401**Cache tools** (2): cache_stats, cache_flush.

402

403**Compression tools** (5): compression_status, compression_configure, set_compression_engine,

404list_compression_combos, compression_combo_stats.

405

406**1proxy tools** (3): oneproxy_fetch, oneproxy_rotate, oneproxy_stats.

407

408**Memory tools** (3): memory_search, memory_add, memory_clear.

409

410**Skill tools** (4): skills_list, skills_enable, skills_execute, skills_executions.

411

412**Agent-skill tools** (3): A2A skill discovery / invocation bridges.

413

414**Gamification tools** (8): levels, badges, leaderboard, and community-federation queries.

415

416**Plugin tools** (8): plugin marketplace listing, install/enable/disable, and runtime inspection.

417

418**Notion tools** (6) + **Obsidian tools** (22): knowledge-base read/write integrations (the largest tool family — vault search, note CRUD, WebDAV-backed file ops).

419

420#### MCP Internals

421

422- **Tool registration**: Each tool is an object with `{ name, description, inputSchema: ZodSchema,

423handler: async (args) => {...} }`. Zod validates inputs before the handler fires.

424- **`createMcpServer()`** and **`startMcpStdio()`** exported from `mcp-server/index.ts`.

425 `createMcpServer()` wires all tool sets; `startMcpStdio()` launches the stdio transport.

426- **Transports**: stdio (CLI `omniroute --mcp`), SSE (`/api/mcp/sse`), Streamable HTTP

427 (`/api/mcp/stream`). All share the same tool/scope engine.

428- **Scopes** (30): Control which tool categories an API key can access. Enforcement happens

429 before handler dispatch.

430- **Audit**: Every tool invocation is logged to SQLite (`mcp_audit` table) with tool name,

431 args, success/failure, API key attribution, and timestamp.

432

433### A2A Server (`src/lib/a2a/`)

434

435JSON-RPC 2.0, SSE streaming, Task Manager with TTL cleanup.

436Agent Card at `/.well-known/agent.json`.

437Skills (6): `smartRouting.ts`, `quotaManagement.ts`, `providerDiscovery.ts`, `costAnalysis.ts`, `healthReport.ts`, `listCapabilities.ts`.

438

439#### A2A Internals

440

441- **`taskManager.ts`**: State machine lifecycle for tasks: `submitted → working →

442completed | failed | canceled`. Tasks have TTL and are cleaned up automatically.

443- **JSON-RPC methods**: `message/send` (sync), `message/stream` (SSE), `tasks/get`,

444 `tasks/cancel`. Dispatched via `POST /a2a`.

445- **Skills**: Registered in a DB-backed registry. Each skill receives task context

446 (messages, metadata) and returns structured results. `quotaManagement.ts` summarizes

447 quota; `smartRouting.ts` recommends routing decisions.

448- **Agent Card**: `/.well-known/agent.json` exposes capabilities, skills, and metadata

449 for client auto-discovery.

450

451### ACP Module (`src/lib/acp/`)

452

453Agent Communication Protocol registry and manager.

454

455### Memory System (`src/lib/memory/`)

456

457Extraction, injection, retrieval, summarization, and store modules for persistent

458conversational memory across sessions.

459

460### Skills System (`src/lib/skills/`)

461

462Extensible skill framework: registry, executor, sandbox, built-in skills,

463custom skill support, interception, and injection.

464

465#### Skills Internals

466

467- **`registry.ts`**: DB-backed skill registration and discovery. Skills have metadata

468 (name, description, version, enabled status) stored in SQLite.

469- **`executor.ts`**: Execution engine with configurable timeout and retry logic.

470 Receives skill name + input, looks up the skill, runs it in the sandbox.

471- **`sandbox.ts`**: Isolation layer for custom (user-provided) skills. Limits resource

472 access and execution time.

473- **Built-in skills**: Ship with OmniRoute (e.g., quota management, routing). Located

474 alongside the registry.

475- **Interception/Injection**: Skills can intercept requests in the pipeline (pre/post

476 processing) or inject context into prompts.

477

478### Compliance (`src/lib/compliance/`)

479

480Policy index for compliance enforcement.

481

482### MITM Proxy (`src/mitm/`)

483

484MITM proxy capability with certificate management, DNS handling, and target routing.

485

486### Middleware (`src/middleware/`)

487

488Request middleware including `promptInjectionGuard.ts`.

489

490### Guardrails (`src/lib/guardrails/`)

491

492Hot-reloadable guardrails framework (3 built-in: pii-masker, prompt-injection, vision-bridge). Fail-open. The `pii-masker` guardrail is registered and runs on every request, but its data-mutating logic is **opt-in** and OFF by default — it only redacts when `PII_REDACTION_ENABLED` (request) / `PII_RESPONSE_SANITIZATION` (response + streaming) are enabled (both `defaultValue: "false"`); with them off, payloads pass through untouched. A request can additionally opt OUT of any guardrail via header (`x-omniroute-disabled-guardrails`). Never make PII default-on (Hard Rule #20). See [`docs/security/GUARDRAILS.md`](docs/security/GUARDRAILS.md).

493

494### Cloud Agents (`src/lib/cloudAgent/`)

495

496`CloudAgentBase` abstract class + 3 agents (codex-cloud, devin, jules). Tasks persisted in `cloud_agent_tasks`; management auth required. See [`docs/frameworks/CLOUD_AGENT.md`](docs/frameworks/CLOUD_AGENT.md).

497

498### Evals (`src/lib/evals/`)

499

500Generic eval framework: `evalRunner.ts`, `runtime.ts`. Targets: combo / model / suite-default. See [`docs/frameworks/EVALS.md`](docs/frameworks/EVALS.md).

501

502### Webhooks (`src/lib/webhookDispatcher.ts`)

503

504HMAC-signed delivery, exponential backoff, auto-disable after 10 failures. 7 event types. See [`docs/frameworks/WEBHOOKS.md`](docs/frameworks/WEBHOOKS.md).

505

506### Authorization Pipeline (`src/server/authz/`)

507

508`classify → policies → enforce`. 3 route classes (PUBLIC / CLIENT_API / MANAGEMENT). See [`docs/architecture/AUTHZ_GUIDE.md`](docs/architecture/AUTHZ_GUIDE.md).

509

510### Reasoning Replay (`src/lib/db/reasoningCache.ts` + `open-sse/services/reasoningCache.ts`)

511

512Hybrid in-memory + SQLite cache for `reasoning_content`. Re-injects on multi-turn for strict providers (DeepSeek V4, Kimi K2, Qwen-Thinking, GLM, xiaomi-mimo). See [`docs/routing/REASONING_REPLAY.md`](docs/routing/REASONING_REPLAY.md).

513

514### Tunnels (`src/lib/{cloudflaredTunnel,ngrokTunnel}.ts` + `src/app/api/tunnels/`)

515

516Cloudflare Quick/Named, ngrok, Tailscale Funnel. See [`docs/ops/TUNNELS_GUIDE.md`](docs/ops/TUNNELS_GUIDE.md).

517

518### Adding a New Provider

519

5201. Register in `src/shared/constants/providers.ts`

5212. Add executor in `open-sse/executors/` (if custom logic needed)

5223. Add translator in `open-sse/translator/` (if non-OpenAI format)

5234. Add OAuth config in `src/lib/oauth/constants/oauth.ts` (if OAuth-based)

5245. Add models in `open-sse/config/providerRegistry.ts`

525

526---

527

528## Subdirectory AGENTS.md Files

529

530- **[`src/lib/db/AGENTS.md`](src/lib/db/AGENTS.md)** — SQLite persistence, domain modules, migrations

531- **[`open-sse/services/AGENTS.md`](open-sse/services/AGENTS.md)** — Routing engine, combo resolution, strategy selection

532

533## Reference Documentation (docs/)

534

535For any non-trivial change, read the matching deep-dive first:

536

537| Area | Doc |

538| ------------------------------------------ | --------------------------------------------------------------------------------------------------------------- |

539| Repo navigation | [`docs/architecture/REPOSITORY_MAP.md`](docs/architecture/REPOSITORY_MAP.md) |

540| Architecture | [`docs/architecture/ARCHITECTURE.md`](docs/architecture/ARCHITECTURE.md) |

541| Engineering reference | [`docs/architecture/CODEBASE_DOCUMENTATION.md`](docs/architecture/CODEBASE_DOCUMENTATION.md) |

542| Auto-Combo (12-factor, 18 strategies) | [`docs/routing/AUTO-COMBO.md`](docs/routing/AUTO-COMBO.md) |

543| Resilience (3 layers) | [`docs/architecture/RESILIENCE_GUIDE.md`](docs/architecture/RESILIENCE_GUIDE.md) |

544| Skills | [`docs/frameworks/SKILLS.md`](docs/frameworks/SKILLS.md) |

545| Memory | [`docs/frameworks/MEMORY.md`](docs/frameworks/MEMORY.md) |

546| Cloud agents | [`docs/frameworks/CLOUD_AGENT.md`](docs/frameworks/CLOUD_AGENT.md) |

547| Guardrails | [`docs/security/GUARDRAILS.md`](docs/security/GUARDRAILS.md) |

548| Evals | [`docs/frameworks/EVALS.md`](docs/frameworks/EVALS.md) |

549| Compliance | [`docs/security/COMPLIANCE.md`](docs/security/COMPLIANCE.md) |

550| Webhooks | [`docs/frameworks/WEBHOOKS.md`](docs/frameworks/WEBHOOKS.md) |

551| Authz | [`docs/architecture/AUTHZ_GUIDE.md`](docs/architecture/AUTHZ_GUIDE.md) |

552| Stealth | [`docs/security/STEALTH_GUIDE.md`](docs/security/STEALTH_GUIDE.md) |

553| Reasoning replay | [`docs/routing/REASONING_REPLAY.md`](docs/routing/REASONING_REPLAY.md) |

554| Agent protocols (A2A / ACP / Cloud) | [`docs/frameworks/AGENT_PROTOCOLS_GUIDE.md`](docs/frameworks/AGENT_PROTOCOLS_GUIDE.md) |

555| MCP server | [`docs/frameworks/MCP-SERVER.md`](docs/frameworks/MCP-SERVER.md) |

556| A2A server | [`docs/frameworks/A2A-SERVER.md`](docs/frameworks/A2A-SERVER.md) |

557| API reference | [`docs/reference/API_REFERENCE.md`](docs/reference/API_REFERENCE.md) + [`docs/openapi.yaml`](docs/openapi.yaml) |

558| Provider catalog (auto-generated) | [`docs/reference/PROVIDER_REFERENCE.md`](docs/reference/PROVIDER_REFERENCE.md) |

559| Tunnels | [`docs/ops/TUNNELS_GUIDE.md`](docs/ops/TUNNELS_GUIDE.md) |

560| Electron desktop | [`docs/guides/ELECTRON_GUIDE.md`](docs/guides/ELECTRON_GUIDE.md) |

561| Release flow | [`docs/ops/RELEASE_CHECKLIST.md`](docs/ops/RELEASE_CHECKLIST.md) |

562| Quality gates (35 gates, allowlist policy) | [`docs/architecture/QUALITY_GATES.md`](docs/architecture/QUALITY_GATES.md) |

563| Cluster opt-in profiles (memory, bifrost) | [`docs/architecture/cluster-decisions.md`](docs/architecture/cluster-decisions.md) |

564

565---

566

567## Fork / Upstream Workflow

568

569This repository is a fork of `diegosouzapw/OmniRoute`. Keep fork-only operational

570changes (for example GHCR image publishing, personal deployment workflows, or local

571automation) out of upstream contribution PRs.

572

573When preparing a PR for upstream, always start the work branch from the upstream

574**default branch** — the active `release/vX.Y.Z` line (today `release/v3.8.49`).

575Never branch from `main`: `main` only receives release squash-merges, so a branch

576cut there is weeks behind and produces conflict-heavy PRs

577(see `CONTRIBUTING.md` and `docs/ops/BRANCHING_MODEL.md`):

578

579```bash

580git fetch upstream

581# the default branch is the active release line, e.g. release/v3.8.49

582git switch -c <branch-name> upstream/release/vX.Y.Z

583```

584

585Only cherry-pick or reapply the changes intended for the upstream PR.

586

587---

588

589## Review Focus

590

591- **DB ops** go through `src/lib/db/` modules, never raw SQL in routes

592- **Provider requests** flow through `open-sse/handlers/`

593- **MCP/A2A pages** are tabs inside `/dashboard/endpoint`, not standalone routes

594- **No memory leaks** in SSE streams (abort signals, cleanup)

595- **Rate limit headers** must be parsed correctly

596- All API inputs validated with **Zod schemas**

597- **Provider constants** validated at module load via Zod (`src/shared/validation/providerSchema.ts`)

598- **Pricing data** syncs from LiteLLM via `src/lib/pricingSync.ts`

599- **Memory/Skills** are cross-cutting: affect MCP tools, request pipeline, and A2A skills

600- **⛔ NEVER close a contributor's PR** after using their code — always merge via GitHub so they get credit. See `.agents/workflows/review-prs.md` for full policy.

601

@@ −1 +1 @@

1−# CLAUDE.md

1+# omniroute — Agent Guidelines

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

3+## Project

5−## Quick Start

5+Unified AI proxy/router — route any LLM through one endpoint. Multi-provider support

6+with **290 provider entries** (OpenAI, Anthropic, Gemini, DeepSeek, Groq, xAI, Mistral, Fireworks,

7+Cohere, NVIDIA, Cerebras, Pollinations, Puter, Cloudflare AI, HuggingFace, DeepInfra,

8+SambaNova, Meta Llama API, Moonshot AI, AI21 Labs, Databricks, Snowflake, and many more)

9+with **MCP Server** (104 tools), **A2A v0.3 Protocol**, and **Electron desktop app**.

610

7−```bash

8−npm install # Install deps (auto-generates .env from .env.example)

9−npm run dev # Dev server at http://localhost:20128

10−npm run build # Production build (Next.js 16 standalone)

11−npm run lint # ESLint (0 errors expected; warnings are pre-existing)

12−npm run typecheck:core # TypeScript check (should be clean)

13−npm run typecheck:noimplicit:core # Strict check (no implicit any)

14−npm run test:coverage # Unit tests + coverage gate (60/60/60/60 — statements/lines/functions/branches)

15−npm run check # lint + test combined

16−npm run check:cycles # Detect circular dependencies

17−```

11+> **Live counts (v3.8.49)**: providers 290 · MCP tools 104 · MCP scopes 30 · A2A skills 6 ·

12+> open-sse services 134 · routing strategies 17 · auto-combo scoring factors 12 ·

13+> DB modules 95 · DB migrations 110 · base tables 17 · search providers 11 ·

14+> i18n locales 42. **Refresh with `npm run check:docs-all`.**

1815

19−### Running Tests

16+## Doc Accuracy Discipline (read before writing any doc)

2017

21−```bash

22−# Single test file (Node.js native test runner — most tests)

23−node --import tsx/esm --test tests/unit/your-file.test.ts

18+> **If `grep -rn "name" src/ open-sse/ bin/` returns nothing, the name does not exist. Do not document it.**

2419

25−# Vitest (MCP server, autoCombo, cache)

26−npm run test:vitest

20+The recurring failure mode in AI-generated docs is _plausible-but-unverified specifics_.

21+Every claim in a `.md` file under `docs/` should be verifiable against the source.

2722

28−# All suites

29−npm run test:all

30−```

23+**Rules (enforced by `npm run check:fabricated-docs`):**

3124

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

25+1. **Never state an API name, endpoint, path, CLI command, or env var without grepping for it first.**

26+ ```bash

27+ grep -rn "theName" src/ open-sse/ bin/

28+ # 0 hits → do not document

29+ ```

30+2. **Never write a line count, file size, migration count, provider count, or strategy count from memory.**

31+ ```bash

32+ wc -l <file> # exact line count

33+ ls <dir>/*.ts | wc -l # file count

34+ ```

35+3. **Every code example should be copy-pasted from real usage or actually run** — not synthesized.

36+ Link to a real call site (`path:line`) instead of inventing a signature.

37+4. **Prefer citing real source (`file.ts:line`) over paraphrasing behavior** — verifiable and self-correcting.

38+5. **A shorter doc that is 100% accurate beats a comprehensive one with fabrications.**

39+ Wrong docs cost more than missing docs, because people trust and act on them.

3340

41+The script `scripts/check/check-fabricated-docs.mjs` extracts every route path, env var, hook

42+name, function name, and file reference from `docs/**/*.md` and verifies each one against the

43+codebase. Run it locally before pushing docs; it runs in CI via `npm run check:docs-all`.

44+

45+## Stack

46+

47+- **Runtime**: Next.js 16 (App Router), Node.js `>=22.0.0 <23 || >=24.0.0 <27`, ES Modules (`"type": "module"`)

48+- **Language**: TypeScript 6.0 (`src/`) + JavaScript (`open-sse/`, `electron/`)

49+- **Database**: better-sqlite3 (SQLite) — `DATA_DIR` configurable, default `~/.omniroute/`

50+- **Streaming**: SSE via `open-sse` internal workspace package

51+- **Styling**: Tailwind CSS v4

52+- **i18n**: next-intl with 42 locales (`src/i18n/messages/`) — refresh with `ls src/i18n/messages/*.json | wc -l`

53+- **Desktop**: Electron (cross-platform: Windows, macOS, Linux)

54+- **Schemas**: Zod v4 for all API / MCP input validation

55+

3456 ---

3557

36−## Project at a Glance

58+## Build, Lint, and Test Commands

3759

38−**OmniRoute** — unified AI proxy/router. One endpoint, 290 LLM providers, auto-fallback.

60+| Command | Description |

61+| ----------------------------------- | ------------------------------------------------------------------ |

62+| `npm run dev` | Start Next.js dev server |

63+| `npm run build` | Production build: `next build` → `.build/next/` + assemble `dist/` |

64+| `npm run build:release` | Clean rebuild + HEAD sentinel (`dist/BUILD_SHA`) — use for deploy |

65+| `npm run start` | Run production build |

66+| `npm run build:cli` | Build CLI package |

67+| `npm run lint` | ESLint on all source files |

68+| `npm run typecheck:core` | TypeScript core type checking |

69+| `npm run typecheck:noimplicit:core` | Strict checking (no implicit any) |

70+| `npm run check` | Run lint + test |

71+| `npm run check:cycles` | Check for circular dependencies |

72+| `npm run electron:dev` | Run Electron app in dev mode |

73+| `npm run electron:build` | Build Electron app for current OS |

3974

40−| Layer | Location | Purpose |

41−| ------------- | ----------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------- |

42−| API Routes | `src/app/api/v1/` | Next.js App Router — entry points |

43−| Handlers | `open-sse/handlers/` | Request processing (chat, embeddings, etc) |

44−| Executors | `open-sse/executors/` | Provider-specific HTTP dispatch |

45−| Translators | `open-sse/translator/` | Format conversion (OpenAI↔Claude↔Gemini) |

46−| Transformer | `open-sse/transformer/` | Responses API ↔ Chat Completions |

47−| Services | `open-sse/services/` | Combo routing, rate limits, caching, etc |

48−| Database | `src/lib/db/` | SQLite domain modules (130 migrations) |

49−| Domain/Policy | `src/domain/` | Policy engine, cost rules, fallback logic |

50−| MCP Server | `open-sse/mcp-server/` | 104 tools (42 base + memory/skill/agentSkill/pool/notion/obsidian/gamification/plugin modules), 3 transports (stdio / SSE / Streamable HTTP), 31 scopes |

51−| A2A Server | `src/lib/a2a/` | JSON-RPC 2.0 agent protocol |

52−| Skills | `src/lib/skills/` | Extensible skill framework |

53−| Memory | `src/lib/memory/` | Persistent conversational memory |

75+**Build output layout:**

5476

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

77+| Directory | Purpose | Gitignored |

78+| --------- | -------------------------------------------------- | ---------- |

79+| `src/` | Application source (TypeScript / TSX) | No |

80+| `.build/` | Build intermediates (`distDir = .build/next`) | Yes |

81+| `dist/` | Shippable bundle assembled by `assembleStandalone` | Yes |

5682

57−---

83+The pipeline is a single `next build` pass — intermediates land in `.build/next/`, the

84+assembled bundle in `dist/`. VPS deploys rsync `dist/` into the remote

85+`/usr/lib/node_modules/omniroute/app/` directory (VPS image path is unchanged).

5886

59−## Request Pipeline

87+### Running Tests

6088

61−```

62−Client → /v1/chat/completions (Next.js route)

63− → CORS → Zod validation → auth? → policy check → prompt injection guard

64− → handleChatCore() [open-sse/handlers/chatCore.ts]

65− → cache check → rate limit → combo routing?

66− → resolveComboTargets() → handleSingleModel() per target

67− → translateRequest() → getExecutor() → executor.execute()

68− → fetch() upstream → retry w/ backoff

69− → response translation → SSE stream or JSON

70− → If Responses API: responsesTransformer.ts TransformStream

71−```

89+```bash

90+# All tests (unit + vitest + ecosystem + e2e)

91+npm run test:all

7292

73−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.

93+# Single test file (Node.js native test runner — most tests use this)

94+node --import tsx/esm --test tests/unit/your-file.test.ts

95+node --import tsx/esm --test tests/unit/plan3-p0.test.ts

96+node --import tsx/esm --test tests/unit/fixes-p1.test.ts

97+node --import tsx/esm --test tests/unit/security-fase01.test.ts

7498

75−**Combo routing** (`open-sse/services/combo.ts`): 19 public strategies (priority, weighted, fill-first, round-robin, p2c, random, least-used, cost-optimized, reset-aware, reset-window, headroom, strict-random, auto, lkgp, context-optimized, cache-optimized, context-relay, 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 13-factor Auto-Combo scoring + the full strategy table and `docs/architecture/RESILIENCE_GUIDE.md` for the 3 resilience layers.

99+# Integration tests

100+node --import tsx/esm --test tests/integration/*.test.ts

76101

77−---

102+# Vitest (MCP server, autoCombo)

103+npm run test:vitest

78104

79−## Resilience Runtime State

105+# E2E with Playwright

106+npm run test:e2e

80107

81−OmniRoute has three related but distinct temporary-failure mechanisms. Keep their

82−scope separate when debugging routing behavior. See the

83−[3-layer resilience diagram](./docs/diagrams/exported/resilience-3layers.svg)

84−(source: [docs/diagrams/resilience-3layers.mmd](./docs/diagrams/resilience-3layers.mmd))

85−for an at-a-glance map.

108+# Protocol clients E2E (MCP transports, A2A)

109+npm run test:protocols:e2e

86110

87−### Provider Circuit Breaker

111+# Ecosystem compatibility tests

112+npm run test:ecosystem

88113

89−**Scope**: whole provider, e.g. `glm`, `openai`, `anthropic`.

114+# Coverage (see CONTRIBUTING.md)

115+npm run test:coverage

116+```

90117

91−**Purpose**: stop sending traffic to a provider that is repeatedly failing at the

92−upstream/service level, so one unhealthy provider does not slow down every request.

118+**For authoritative coverage requirements, test execution, and PR gates, see [`CONTRIBUTING.md`](CONTRIBUTING.md#running-tests).**

93119

94−**Implementation**:

120+---

95121

96−- Core class: `src/shared/utils/circuitBreaker.ts`

97−- Chat gate/execution wiring: `src/sse/handlers/chatHelpers.ts`, `src/sse/handlers/chat.ts`

98−- Runtime status API: `src/app/api/monitoring/health/route.ts`

99−- Shared wrappers: `open-sse/services/accountFallback.ts`

100−- Persisted state table: `domain_circuit_breakers`

122+## Code Style Guidelines

101123

102−**States**:

124+### Formatting (Prettier — enforced via lint-staged)

103125

104−- `CLOSED`: normal traffic is allowed.

105−- `OPEN`: provider is temporarily blocked; callers get a provider-circuit-open response

106− or combo routing skips to another target.

107−- `HALF_OPEN`: reset timeout has elapsed; allow a probe request. Success closes the

108− breaker, failure opens it again.

126+2 spaces · semicolons required · double quotes (`"`) · 100 char width · es5 trailing commas.

127+Always run `prettier --write` on changed files.

109128

110−**Defaults** (`open-sse/config/constants.ts`):

129+### TypeScript

111130

112−- OAuth providers: threshold `3`, reset timeout `60s`.

113−- API-key providers: threshold `5`, reset timeout `30s`.

114−- Local providers: threshold `2`, reset timeout `15s`.

131+- **Target**: ES2022 · **Module**: `esnext` · **Resolution**: `bundler`

132+- `strict: false` — prefer explicit types, don't rely on inference

133+- Path aliases: `@/*` → `src/`, `@omniroute/open-sse` → `open-sse/`, `@omniroute/open-sse/*` → `open-sse/*`

115134

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

135+### ESLint Rules

117136

118−```ts

119−(408, 500, 502, 503, 504);

120−```

137+- **Security (error, everywhere)**: `no-eval`, `no-implied-eval`, `no-new-func`

138+- **Relaxed in `open-sse/` and `tests/`**: `@typescript-eslint/no-explicit-any` = warn

139+- React hooks rules and `@next/next/no-assign-module-variable` disabled in `open-sse/` and `tests/`

121140

122−Do not trip the whole-provider breaker for normal account/key/model errors like most

123−`401`, `403`, or `429` cases. Those usually belong to connection cooldown or model

124−lockout. A generic API-key provider `403` should be recoverable unless it is classified

125−as a terminal provider/account error.

141+### Naming

126142

127−The breaker uses lazy recovery, not a background timer. When `OPEN` expires, reads such

128−as `getStatus()`, `canExecute()`, and `getRetryAfterMs()` refresh the state to

129−`HALF_OPEN`, so dashboards and combo candidate builders do not keep excluding an

130−expired provider forever.

143+| Element | Convention | Example |

144+| ------------------- | -------------------------------- | ------------------------------------ |

145+| Files | camelCase / kebab-case | `chatCore.ts`, `tokenHealthCheck.ts` |

146+| React components | PascalCase | `Dashboard.tsx`, `ProviderCard.tsx` |

147+| Functions/variables | camelCase | `getHealth()`, `switchCombo()` |

148+| Constants | UPPER_SNAKE | `MAX_RETRIES`, `DEFAULT_TIMEOUT` |

149+| Interfaces | PascalCase (`I` prefix optional) | `ProviderConfig` |

150+| Enums | PascalCase (members too) | `LogLevel.Error` |

131151

132−### Connection Cooldown

152+### Imports

133153

134−**Scope**: one provider connection/account/key.

154+- **Order**: external → internal (`@/`, `@omniroute/open-sse`) → relative (`./`, `../`)

155+- **No barrel imports** from `localDb.ts` — import from the specific `db/` module instead

135156

136−**Purpose**: temporarily skip one bad key/account while allowing other connections for

137−the same provider to continue serving requests.

157+### Error Handling

138158

139−**Implementation**:

159+- try/catch with specific error types; always log with context (pino logger)

160+- Never silently swallow errors in SSE streams — use abort signals for cleanup

161+- Return proper HTTP status codes (4xx client, 5xx server)

140162

141−- Write/update path: `src/sse/services/auth.ts::markAccountUnavailable()`

142−- Account selection/filtering: `src/sse/services/auth.ts::getProviderCredentials...`

143−- Cooldown calculation: `open-sse/services/accountFallback.ts::checkFallbackError()`

144−- Settings: `src/lib/resilience/settings.ts`

163+### Security

145164

146−Important fields on provider connections:

165+- **NEVER** commit API keys, secrets, or credentials

166+- Validate all user inputs with Zod schemas

167+- Auth middleware required on all API routes

168+- Never log SQLite encryption keys

169+- Sanitize user content (dompurify for HTML)

170+- **Public upstream OAuth identifiers** (Gemini / Antigravity / Windsurf-style client_id/secret + Firebase Web keys extracted from public CLIs): use `resolvePublicCred()` from `open-sse/utils/publicCreds.ts`, **never** as string literals. Full pattern in `docs/security/PUBLIC_CREDS.md`.

171+- **Error responses** (HTTP / SSE / executor / MCP): use `buildErrorBody()` or `sanitizeErrorMessage()` from `open-sse/utils/error.ts`, **never** put raw `err.stack` / `err.message` in a Response body. Full pattern in `docs/security/ERROR_SANITIZATION.md`.

172+- **`exec()` / `spawn()` with runtime values**: pass via the `env` option, **never** string-interpolate paths/values into the script body. Reference: `src/mitm/cert/install.ts::updateNssDatabases`.

173+- Prefer secure-by-default libraries when available — see [tldrsec/awesome-secure-defaults](https://github.com/tldrsec/awesome-secure-defaults) for the curated list (Helmet.js, DOMPurify, ssrf-req-filter, safe-regex, Google Tink, etc.).

147174

148−```ts

149−rateLimitedUntil;

150−testStatus: "unavailable";

151−lastError;

152−lastErrorType;

153−errorCode;

154−backoffLevel;

155−```

175+---

156176

157−During account selection, a connection is skipped while:

177+## Architecture

158178

159−```ts

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

161−```

179+### Data Layer (`src/lib/db/`)

162180

163−Cooldowns are also lazy: when `rateLimitedUntil` is in the past, the connection becomes

164−eligible again. On successful use, `clearAccountError()` clears `testStatus`,

165−`rateLimitedUntil`, error fields, and `backoffLevel`.

181+All persistence uses SQLite through **95 domain-specific modules** in `src/lib/db/`. Top modules:

166182

167−Default connection cooldown behavior:

183+- Core: `core.ts`, `migrationRunner.ts`, `encryption.ts`, `stateReset.ts`

184+- Providers / catalog: `providers.ts`, `models.ts`, `providerLimits.ts`, `compressionAnalytics.ts`

185+- Routing: `combos.ts`, `modelComboMappings.ts`, `domainState.ts`, `commandCodeAuth.ts`

186+- Auth: `apiKeys.ts`, `secrets.ts`, `registeredKeys.ts`, `sessionAccountAffinity.ts`

187+- Usage / billing: `quotaSnapshots.ts`, `creditBalance.ts`, `usage*.ts`, `compressionCacheStats.ts`

188+- Storage: `backup.ts`, `cleanup.ts`, `jsonMigration.ts`, `healthCheck.ts`, `databaseSettings.ts`

189+- Extension modules: `evals.ts`, `webhooks.ts`, `reasoningCache.ts`, `readCache.ts`, `tierConfig.ts`, `compressionCombos.ts`, `compressionScheduler.ts`, `batches.ts`, `files.ts`, `syncTokens.ts`, `proxies.ts`, `oneproxy.ts`, `upstreamProxy.ts`, `versionManager.ts`, `cliToolState.ts`, `prompts.ts`, `detailedLogs.ts`, `contextHandoffs.ts`, `compression.ts`, `stats.ts`

168190

169−- OAuth base cooldown: `5s`.

170−- API-key base cooldown: `3s`.

171−- API-key `429` should prefer upstream retry hints (`Retry-After`, reset headers, or

172− parseable reset text) when available.

173−- Repeated recoverable failures use exponential backoff:

191+Live count: `ls src/lib/db/*.ts | wc -l` (currently 95). Drift detection: `npm run check:docs-counts`.

192+Schema migrations live in `db/migrations/` (**110 files** as of v3.8.43) and run via `migrationRunner.ts`.

193+`src/lib/localDb.ts` is a **re-export layer only** — never add logic there.

174194

175−```ts

176−baseCooldownMs * 2 ** failureIndex;

177−```

195+#### DB Internals

178196

179−The anti-thundering-herd guard prevents concurrent failures on the same connection from

180−repeatedly extending the cooldown or double-incrementing `backoffLevel`.

197+- **`core.ts`**: `getDbInstance()` returns a singleton `better-sqlite3` instance with WAL

198+ journaling. `SCHEMA_SQL` defines **17 base tables** (verify with `grep -c "CREATE TABLE" src/lib/db/core.ts` minus 1 for the bookkeeping `_omniroute_migrations` table). Helpers: `rowToCamel`, `encryptConnectionFields`.

199+- **`migrationRunner.ts`**: Applies versioned SQL files from `db/migrations/` inside transactions.

200+ Tracks applied migrations in `_omniroute_migrations` table.

201+- **Migrations**: 110 files (`001_initial_schema.sql` → `110_*.sql`).

202+ Each migration is idempotent and runs in a transaction. Live count: `ls src/lib/db/migrations/*.sql | wc -l`.

203+- **Domain modules** import `getDbInstance()` from `core.ts` for all CRUD operations.

204+ Each module owns a specific table/set of tables (e.g., `providers.ts` → `provider_connections`,

205+ `combos.ts` → `combos`). Encryption helpers protect sensitive fields at rest.

206+- **`localDb.ts`** re-exports all domain modules — consumers import from here for convenience.

181207

182−Terminal states are not cooldowns. `banned`, `expired`, and `credits_exhausted` are

183−intended to stay unavailable until credentials/settings change or an operator resets

184−them. Do not overwrite terminal states with transient cooldown state.

208+### API Route Layer (`src/app/api/v1/`)

185209

186−### Model Lockout

210+Next.js App Router routes — each follows a consistent pattern:

187211

188−**Scope**: provider + connection + model.

212+```

213+Route → CORS preflight → Body validation (Zod) → Optional auth (extractApiKey/isValidApiKey)

214+ → API key policy enforcement (enforceApiKeyPolicy) → Handler delegation (open-sse)

215+```

189216

190−**Purpose**: avoid disabling a whole connection when only one model is unavailable or

191−quota-limited for that connection.

217+| Route | Handler | Notes |

218+| ------------------------------- | ------------------------- | ------------------------------------------------------------- |

219+| `chat/completions/route.ts` | `handleChat()` | + prompt injection guard (clones request) |

220+| `responses/route.ts` | `handleChat()` (unified) | Responses API format |

221+| `embeddings/route.ts` | `handleEmbedding()` | Model listing + creation |

222+| `images/generations/route.ts` | `handleImageGeneration()` | Model listing + creation |

223+| `audio/transcriptions/route.ts` | audio handler | Multipart form data |

224+| `audio/speech/route.ts` | TTS handler | Binary audio response |

225+| `videos/generations/route.ts` | video handler | ComfyUI/SD WebUI |

226+| `music/generations/route.ts` | music handler | ComfyUI workflows |

227+| `moderations/route.ts` | moderation handler | Content safety |

228+| `rerank/route.ts` | rerank handler | Document relevance |

229+| `search/route.ts` | search handler | Web search (12 providers per `open-sse/handlers/search.ts:6`) |

192230

193−Examples:

231+**No global Next.js middleware file** — interception is route-specific. Auth is optional

232+(controlled by `REQUIRE_API_KEY` env). Prompt injection guard is unique to chat completions.

194233

195−- Per-model quota providers returning `429`.

196−- Local providers returning `404` for one missing model.

197−- Provider-specific mode/model permission failures such as selected Grok modes.

234+### Request Pipeline (`open-sse/`)

198235

199−Model lockout lives in `open-sse/services/accountFallback.ts` and lets the same

200−connection continue serving other models.

236+The `open-sse/` workspace is the core streaming engine. Full request flow:

201237

202−### Debugging Guidance

238+```

239+Client Request

240+ → src/app/api/v1/.../route.ts (Next.js route)

241+ → open-sse/handlers/chatCore.ts::handleChatCore()

242+ → Semantic/signature cache check

243+ → Rate limit check (rateLimitManager)

244+ → Combo routing? → open-sse/services/combo.ts::handleComboChat()

245+ → resolveComboTargets() → ordered ResolvedComboTarget[]

246+ → For each target: handleSingleModel() (wraps chatCore)

247+ → translateRequest() (open-sse/translator/)

248+ → Convert source format (e.g., OpenAI) → target format (e.g., Claude)

249+ → getExecutor() → provider-specific executor instance

250+ → executor.execute() (BaseExecutor → DefaultExecutor or provider-specific)

251+ → buildUrl() + buildHeaders() + transformRequest()

252+ → fetch() to upstream provider

253+ → Retry logic with exponential backoff

254+ → Response translation back to client format

255+ → If Responses API: responsesTransformer.ts TransformStream

256+ → SSE stream or JSON response to client

257+```

203258

204−- If all keys for a provider are skipped, inspect both provider breaker state and each

205− connection's `rateLimitedUntil`/`testStatus`.

206−- If a provider appears permanently excluded after the reset window, check whether code

207− is reading raw `state` instead of using `getStatus()`/`canExecute()`.

208−- If one provider key fails but others should work, prefer connection cooldown over

209− provider breaker.

210−- If only one model fails, prefer model lockout over connection cooldown.

211−- If a state should self-recover, it should have a future timestamp/reset timeout and a

212− read path that refreshes expired state. Permanent statuses require manual credential

213− or config changes.

259+**Handlers** (`open-sse/handlers/`): `chatCore.ts`, `responsesHandler.ts`, `embeddings.ts`,

260+`imageGeneration.ts`, `videoGeneration.ts`, `musicGeneration.ts`, `audioSpeech.ts`,

261+`audioTranscription.ts`, `moderations.ts`, `rerank.ts`, `search.ts`.

214262

215−---

263+**Upstream headers**: merged after default auth; same header name replaces executor value.

264+**T5 intra-family fallback** recomputes headers using only the fallback model id.

265+Forbidden header names: `src/shared/constants/upstreamHeaders.ts` — keep sanitize,

266+Zod schemas, and unit tests aligned when editing.

216267

217−## Key Conventions

268+### Provider Categories

218269

219−### Code Style

270+- **Free** (2): Qoder AI, Kiro AI

271+- **OAuth** (13): Claude Code, Antigravity, Codex, GitHub Copilot, Cursor, Kimi Coding, Kilo Code, Cline, Kiro, Qoder, Gemini, Windsurf (v3.8), GitLab Duo (v3.8)

272+- **API Key** (120+): OpenAI, Anthropic, Gemini, DeepSeek, Groq, xAI, Mistral, Perplexity,

273+ Together, Fireworks, Cerebras, Cohere, NVIDIA, Nebius, SiliconFlow, Hyperbolic,

274+ HuggingFace, OpenRouter, Vertex AI, Cloudflare AI, Scaleway, AI/ML API, Pollinations,

275+ Puter, Longcat, Alibaba, Kimi, Minimax, Blackbox, Synthetic, Kilo Gateway,

276+ Z.AI, GLM, Deepgram, AssemblyAI, ElevenLabs, Cartesia, PlayHT, Inworld,

277+ NanoBanana, SD WebUI, ComfyUI, Ollama Cloud, Perplexity Search, Serper, Brave, Exa,

278+ Tavily, OpenCode Zen/Go, Bailian Coding Plan, DeepInfra, Vercel AI Gateway,

279+ Lambda AI, SambaNova, nScale, OVHcloud AI, Baseten, PublicAI, Moonshot AI,

280+ Meta Llama API, v0 (Vercel), Morph, Featherless AI, FriendliAI, LlamaGate,

281+ Galadriel, Weights & Biases Inference, Volcengine, AI21 Labs, Venice.ai,

282+ Codestral, Upstage, Maritalk, Xiaomi MiMo, Inference.net, NanoGPT, Predibase,

283+ Bytez, Heroku AI, Databricks, Snowflake Cortex, GigaChat (Sber), CrofAI,

284+ AgentRouter, ChatGPT Web, Baidu Qianfan, AWS Polly, RunwayML, GitLab Duo,

285+ Amazon Q, Empower, Poe, and many more.

286+- **Self-Hosted** (8+): LM Studio, vLLM, Lemonade, Llamafile, Triton, Docker Model Runner, Xinference, Oobabooga

287+- **Custom**: OpenAI-compatible (`openai-compatible-*`) and Anthropic-compatible (`anthropic-compatible-*`) prefixes

220288

221−- **2 spaces**, semicolons, double quotes, 100 char width, es5 trailing commas (enforced by lint-staged via Prettier)

222−- **Imports**: external → internal (`@/`, `@omniroute/open-sse`) → relative

223−- **Naming**: files=camelCase/kebab, components=PascalCase, constants=UPPER_SNAKE

224−- **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)

225−- **TypeScript**: `strict: false`, target ES2022, module esnext, resolution bundler. Prefer explicit types.

289+Providers are registered in `src/shared/constants/providers.ts` with Zod validation at module load.

226290

227−### Database

291+### Executors (`open-sse/executors/`)

228292

229−- **Always** go through `src/lib/db/` domain modules — **never** write raw SQL in routes or handlers

230−- **Never** add logic to `src/lib/localDb.ts` (re-export layer only)

231−- **Never** barrel-import from `localDb.ts` — import specific `db/` modules instead

232−- DB singleton: `getDbInstance()` from `src/lib/db/core.ts` (WAL journaling)

233−- Migrations: `src/lib/db/migrations/` — versioned SQL files, idempotent, run in transactions

293+Provider-specific request executors: `base.ts`, `default.ts`, `cursor.ts`, `codex.ts`,

294+`antigravity.ts`, `github.ts`, `kiro.ts`, `qoder.ts`, `vertex.ts`,

295+`cloudflare-ai.ts`, `opencode.ts`, `pollinations.ts`, `puter.ts`.

234296

235−### Error Handling

297+#### Executor Internals

236298

237−- try/catch with specific error types, log with pino context

238−- Never swallow errors in SSE streams — use abort signals for cleanup

239−- Return proper HTTP status codes (4xx/5xx)

299+- **`base.ts`** (`BaseExecutor`): Abstract base with `buildUrl()`, `buildHeaders()`,

300+ `transformRequest()`, retry logic (exponential backoff), and `execute()`. Subclasses

301+ override URL/header/transform methods for provider-specific behavior.

302+- **`default.ts`** (`DefaultExecutor extends BaseExecutor`): Handles most OpenAI-compatible

303+ providers. Reads provider config from `providerRegistry.ts` to resolve base URL, auth

304+ header format, and request transformations.

305+- **`getExecutor()`** (`executors/index.ts`): Factory that returns the correct executor

306+ instance based on provider ID. Provider-specific executors (Cursor, Codex, Vertex, etc.)

307+ override only what differs from the default.

240308

241−### Security

309+### Translator (`open-sse/translator/`)

242310

243−- **Never** use `eval()`, `new Function()`, or implied eval

244−- Validate all inputs with Zod schemas

245−- Encrypt credentials at rest (AES-256-GCM)

246−- Upstream header denylist: `src/shared/constants/upstreamHeaders.ts` — keep sanitize, Zod schemas, and unit tests aligned when editing

247−- **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.

248−- **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`.

249−- **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`.

250−- **Secure-by-default libraries** ([tldrsec/awesome-secure-defaults](https://github.com/tldrsec/awesome-secure-defaults)): prefer Helmet.js, DOMPurify, ssrf-req-filter, safe-regex, Google Tink over custom implementations whenever adding new security-sensitive surfaces.

311+Translates between API formats (OpenAI-format ↔ Anthropic, Gemini, etc.).

312+Includes request/response translators with helpers for image handling.

251313

252−---

314+#### Translator Internals

253315

254−## Common Modification Scenarios

316+- **`translator/index.ts`**: Exports `translateRequest()` and format constants. Called by

317+ `chatCore.ts` before executor dispatch.

318+- **Flow**: `translateRequest(body, sourceFormat, targetFormat)` → detects source format

319+ (OpenAI, Anthropic, Gemini) → applies the matching translator module → returns

320+ transformed body ready for the target provider.

321+- **Response translation** runs in reverse after upstream response, converting back to

322+ the client's expected format.

255323

256−### Adding a New Provider

324+### Transformer (`open-sse/transformer/`)

257325

258−1. Register in `src/shared/constants/providers.ts` (Zod-validated at load)

259−2. Add executor in `open-sse/executors/` if custom logic needed (extend `BaseExecutor`)

260−3. Add translator in `open-sse/translator/` if non-OpenAI format

261−4. 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

262−5. Register models in `open-sse/config/providerRegistry.ts`

263−6. Write tests in `tests/unit/` (include the publicCreds shape assertion if you added a new embedded default)

326+`responsesTransformer.ts` — transforms Responses API format to/from Chat Completions format.

264327

265−### Adding a New API Route

328+#### Transformer Internals

266329

267−1. Create directory under `src/app/api/v1/your-route/`

268−2. Create `route.ts` with `GET`/`POST` handlers

269−3. Follow pattern: CORS → Zod body validation → optional auth → handler delegation

270−4. Handler goes in `open-sse/handlers/` (import from there, not inline)

271−5. 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`.

272−6. Add tests — including at least one assertion that error responses do not leak stack traces (`!body.error.message.includes("at /")`)

330+- **`createResponsesApiTransformStream()`**: Returns a `TransformStream` that converts

331+ Chat Completions SSE chunks (`data: {"choices":[...]}`) into Responses API SSE events

332+ (`response.output_item.added`, `response.output_text.delta`, etc.).

333+- Used when the client sends a Responses API request: the request is internally converted

334+ to Chat Completions format, dispatched normally, and the response is piped through this

335+ transform stream before reaching the client.

273336

274−### Adding a New DB Module

337+### Services (`open-sse/services/`)

275338

276−1. Create `src/lib/db/yourModule.ts` — import `getDbInstance` from `./core.ts`

277−2. Export CRUD functions for your domain table(s)

278−3. Add migration in `src/lib/db/migrations/` if new tables needed

279−4. Re-export from `src/lib/localDb.ts` (add to the re-export list only)

280−5. Write tests

339+134 service modules in `open-sse/services/` (top-level only; more including sub-dirs like `autoCombo/` and `compression/`). Refresh: `ls open-sse/services/*.ts | wc -l`. Key modules:

340+`combo.ts` (routing engine), `usage.ts`, `tokenRefresh.ts`,

341+`rateLimitManager.ts`, `accountFallback.ts`, `sessionManager.ts`, `wildcardRouter.ts`,

342+`autoCombo/`, `intentClassifier.ts`, `taskAwareRouter.ts`, `thinkingBudget.ts`,

343+`contextManager.ts`, `modelDeprecation.ts`, `modelFamilyFallback.ts`,

344+`emergencyFallback.ts`, `workflowFSM.ts`, `backgroundTaskDetector.ts`, `ipFilter.ts`,

345+`signatureCache.ts`, `volumeDetector.ts`, `contextHandoff.ts`, `compression/` (prompt

346+compression pipeline), and more.

281347

282−### Adding a New MCP Tool

348+#### Prompt Compression Pipeline (`compression/`)

283349

284−1. Add tool definition in `open-sse/mcp-server/tools/` with Zod input schema + async handler

285−2. Register in tool set (wired by `createMcpServer()`)

286−3. Assign to appropriate scope(s)

287−4. Write tests (tool invocation logged to `mcp_audit` table)

350+Modular prompt compression that runs proactively before the existing reactive context manager.

288351

289−### Adding a New A2A Skill

352+- **`strategySelector.ts`**: Selects compression mode based on config, compression combo assignments,

353+ combo overrides, auto-trigger thresholds, and defaults. Priority: assigned compression combo >

354+ combo override > auto-trigger > default mode > off.

355+- **`lite.ts`**: 5 lite-mode techniques: `collapseWhitespace`, `dedupSystemPrompt`,

356+ `compressToolResults`, `removeRedundantContent`, `replaceImageUrls`. Target: 10-15% savings at

357+ <1ms latency.

358+- **`caveman.ts` / `cavemanRules.ts`**: Caveman-style semantic condensation backed by built-in

359+ rules plus file-loaded language packs under `compression/rules/`.

360+- **`engines/rtk/`**: Rule-based terminal/tool-output compression inspired by RTK patterns. Detects

361+ command output classes, applies JSON filter packs, deduplicates repeated lines, strips ANSI/code

362+ noise, and preserves errors/actionable context. The RTK JSON DSL supports replace,

363+ match-output short-circuit, strip/keep, per-line truncation, head/tail/max-line truncation,

364+ inline tests, trust-gated project/global custom filters, and optional redacted raw-output

365+ retention for authenticated recovery.

366+- **`engines/registry.ts`**: Registers engines (`caveman`, `rtk`) and powers stacked pipelines.

367+- **`stats.ts`**: Per-request compression stats tracking (original tokens, compressed tokens,

368+ savings %, techniques used, engine breakdown, compression combo id).

369+- **`types.ts`**: `CompressionMode` (off/lite/standard/aggressive/ultra/rtk/stacked),

370+ `CompressionConfig`, `CompressionStats`, `CompressionResult`.

371+- DB settings in `src/lib/db/compression.ts`, compression combos in

372+ `src/lib/db/compressionCombos.ts`, API routes under `src/app/api/settings/compression/`,

373+ `src/app/api/context/*`, and preview/language-pack routes under `src/app/api/compression/*`.

290374

291−1. Create skill in `src/lib/a2a/skills/` (5 already exist: smart-routing, quota-management, provider-discovery, cost-analysis, health-report)

292−2. Skill receives task context (messages, metadata) → returns structured result

293−3. Register in `A2A_SKILL_HANDLERS` in `src/lib/a2a/taskExecution.ts`

294−4. Expose in `src/app/.well-known/agent.json/route.ts` (Agent Card)

295−5. Write tests in `tests/unit/`

296−6. Document in `docs/frameworks/A2A-SERVER.md` skill table

375+#### Combo Routing Engine (`combo.ts`)

297376

298−### Adding a New Cloud Agent

377+- **`handleComboChat()`**: Entry point for combo-routed requests. Receives the combo config

378+ and iterates through targets in order until one succeeds or all fail.

379+- **`resolveComboTargets()`**: Expands a combo configuration into an ordered array of

380+ `ResolvedComboTarget[]`, each specifying provider + model + account + credentials.

381+- **Strategies** (17): priority, weighted, fill-first, round-robin, P2C, random, least-used, reset-aware (v3.8),

382+ reset-window, cost-optimized, strict-random, auto, lkgp, context-optimized, context-relay, headroom, fusion. Source: `ROUTING_STRATEGY_VALUES` in `src/shared/constants/routingStrategies.ts`.

383+- Each target calls **`handleSingleModel()`** which wraps `handleChatCore()` with

384+ per-target error handling and circuit breaker checks.

299385

300−1. Create agent class in `src/lib/cloudAgent/agents/` extending `CloudAgentBase` (3 already exist: codex-cloud, devin, jules)

301−2. Implement `createTask`, `getStatus`, `approvePlan`, `sendMessage`, `listSources`

302−3. Register in `src/lib/cloudAgent/registry.ts`

303−4. Add OAuth/credentials handling if needed (`src/lib/oauth/providers/`)

304−5. Tests + document in `docs/frameworks/CLOUD_AGENT.md`

386+### Domain Layer (`src/domain/`)

305387

306−### Adding a New Embedded Service

388+Policy engine modules: `policyEngine.ts`, `comboResolver.ts`, `costRules.ts`,

389+`degradation.ts`, `fallbackPolicy.ts`, `lockoutPolicy.ts`, `modelAvailability.ts`,

390+`providerExpiration.ts`, `quotaCache.ts`, `responses.ts`, `configAudit.ts`.

307391

308−1. 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).

309−2. Register the service in `src/lib/services/bootstrap.ts` (add to `SERVICES[]` array and extend `buildSpawnArgsFactory()`).

310−3. Add a DB seed row for the new service in `src/lib/db/migrations/` (`version_manager` table, `status='not_installed'`, `auto_start=0`).

311−4. 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`.

312−5. 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).

313−6. Add a UI tab in `src/app/(dashboard)/dashboard/providers/services/tabs/` reusing `ServiceStatusCard`, `ServiceLifecycleButtons`, `ServiceLogsPanel`.

314−7. Document in `docs/frameworks/EMBEDDED-SERVICES.md` (update §1 service table + §4 API reference) and `docs/openapi.yaml`.

315−8. 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.

392+### MCP Server (`open-sse/mcp-server/`)

316393

317−### Adding a New Guardrail / Eval / Skill / Webhook event

394+**104 tools** total (`TOTAL_MCP_TOOL_COUNT`, `open-sse/mcp-server/server.ts`): a 42-entry base registry (`MCP_TOOLS` in `schemas/tools.ts`, bundling the core / cache / compression / 1proxy / advanced tools) **plus** standalone module sets — memory (3), skill (4), agentSkill (3), pool (6), gamification (8), plugin (8), notion (6), obsidian (22). 3 transports (stdio / SSE / Streamable HTTP). Scoped auth (31 scopes — see `OMNIROUTE_MCP_SCOPES`), Zod schemas. See [`docs/frameworks/MCP-SERVER.md`](docs/frameworks/MCP-SERVER.md).

318395

319−- Guardrail: `src/lib/guardrails/` → docs: `docs/security/GUARDRAILS.md`

320−- Eval suite: `src/lib/evals/` → docs: `docs/frameworks/EVALS.md`

321−- Skill (sandbox): `src/lib/skills/` → docs: `docs/frameworks/SKILLS.md`

322−- Webhook event: `src/lib/webhookDispatcher.ts` → docs: `docs/frameworks/WEBHOOKS.md`

396+**Core tools** (20): get_health, list_combos, get_combo_metrics, switch_combo, check_quota,

397+route_request, cost_report, list_models_catalog, web_search, simulate_route, set_budget_guard,

398+set_routing_strategy, set_resilience_profile, test_combo, get_provider_metrics,

399+best_combo_for_task, explain_route, get_session_snapshot, db_health_check, sync_pricing.

323400

324−---

401+**Cache tools** (2): cache_stats, cache_flush.

325402

326−## Reference Documentation

403+**Compression tools** (5): compression_status, compression_configure, set_compression_engine,

404+list_compression_combos, compression_combo_stats.

327405

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

406+**1proxy tools** (3): oneproxy_fetch, oneproxy_rotate, oneproxy_stats.

329407

330−| Area | Doc |

331−| --------------------------------------------- | ------------------------------------------------------- |

332−| Repo navigation | `docs/architecture/REPOSITORY_MAP.md` |

333−| Architecture | `docs/architecture/ARCHITECTURE.md` |

334−| Engineering reference | `docs/architecture/CODEBASE_DOCUMENTATION.md` |

335−| Auto-Combo (13-factor scoring, 19 strategies) | `docs/routing/AUTO-COMBO.md` |

336−| Resilience (3 mechanisms) | `docs/architecture/RESILIENCE_GUIDE.md` |

337−| Reasoning replay | `docs/routing/REASONING_REPLAY.md` |

338−| Skills framework | `docs/frameworks/SKILLS.md` |

339−| Memory system (FTS5 + Qdrant) | `docs/frameworks/MEMORY.md` |

340−| Cloud agents | `docs/frameworks/CLOUD_AGENT.md` |

341−| Guardrails (PII / injection / vision) | `docs/security/GUARDRAILS.md` |

342−| Public upstream credentials (Gemini/etc.) | `docs/security/PUBLIC_CREDS.md` |

343−| Error message sanitization | `docs/security/ERROR_SANITIZATION.md` |

344−| Evals | `docs/frameworks/EVALS.md` |

345−| Compliance / audit | `docs/security/COMPLIANCE.md` |

346−| Webhooks | `docs/frameworks/WEBHOOKS.md` |

347−| Authorization pipeline | `docs/architecture/AUTHZ_GUIDE.md` |

348−| Stealth (TLS / fingerprint) | `docs/security/STEALTH_GUIDE.md` |

349−| Agent protocols (A2A / ACP / Cloud) | `docs/frameworks/AGENT_PROTOCOLS_GUIDE.md` |

350−| MCP server | `docs/frameworks/MCP-SERVER.md` |

351−| A2A server | `docs/frameworks/A2A-SERVER.md` |

352−| API reference + OpenAPI | `docs/reference/API_REFERENCE.md` + `docs/openapi.yaml` |

353−| Provider catalog (auto-generated) | `docs/reference/PROVIDER_REFERENCE.md` |

354−| Release flow | `docs/ops/RELEASE_CHECKLIST.md` |

355−| Embedded services | `docs/frameworks/EMBEDDED-SERVICES.md` |

356−| Quality gates (~48 scripts, allowlist policy) | `docs/architecture/QUALITY_GATES.md` |

408+**Memory tools** (3): memory_search, memory_add, memory_clear.

357409

358−---

410+**Skill tools** (4): skills_list, skills_enable, skills_execute, skills_executions.

359411

360−## Testing

412+**Agent-skill tools** (3): A2A skill discovery / invocation bridges.

361413

362−| What | Command |

363−| ----------------------- | --------------------------------------------------------------------------- |

364−| Unit tests | `npm run test:unit` |

365−| Single file | `node --import tsx/esm --test tests/unit/file.test.ts` |

366−| Vitest (MCP, autoCombo) | `npm run test:vitest` |

367−| E2E (Playwright) | `npm run test:e2e` |

368−| Protocol E2E (MCP+A2A) | `npm run test:protocols:e2e` |

369−| Ecosystem | `npm run test:ecosystem` |

370−| Coverage gate | `npm run test:coverage` (60/60/60/60 — statements/lines/functions/branches) |

371−| Coverage report | `npm run coverage:report` |

414+**Gamification tools** (8): levels, badges, leaderboard, and community-federation queries.

372415

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

416+**Plugin tools** (8): plugin marketplace listing, install/enable/disable, and runtime inspection.

374417

375−**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.

418+**Notion tools** (6) + **Obsidian tools** (22): knowledge-base read/write integrations (the largest tool family — vault search, note CRUD, WebDAV-backed file ops).

376419

377−**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.

420+#### MCP Internals

378421

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

422+- **Tool registration**: Each tool is an object with `{ name, description, inputSchema: ZodSchema,

423+handler: async (args) => {...} }`. Zod validates inputs before the handler fires.

424+- **`createMcpServer()`** and **`startMcpStdio()`** exported from `mcp-server/index.ts`.

425+ `createMcpServer()` wires all tool sets; `startMcpStdio()` launches the stdio transport.

426+- **Transports**: stdio (CLI `omniroute --mcp`), SSE (`/api/mcp/sse`), Streamable HTTP

427+ (`/api/mcp/stream`). All share the same tool/scope engine.

428+- **Scopes** (30): Control which tool categories an API key can access. Enforcement happens

429+ before handler dispatch.

430+- **Audit**: Every tool invocation is logged to SQLite (`mcp_audit` table) with tool name,

431+ args, success/failure, API key attribution, and timestamp.

380432

381−1. **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.

382−2. **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.

383−3. "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.

433+### A2A Server (`src/lib/a2a/`)

384434

385−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).

435+JSON-RPC 2.0, SSE streaming, Task Manager with TTL cleanup.

436+Agent Card at `/.well-known/agent.json`.

437+Skills (6): `smartRouting.ts`, `quotaManagement.ts`, `providerDiscovery.ts`, `costAnalysis.ts`, `healthReport.ts`, `listCapabilities.ts`.

386438

387−**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.

439+#### A2A Internals

388440

389−---

441+- **`taskManager.ts`**: State machine lifecycle for tasks: `submitted → working →

442+completed | failed | canceled`. Tasks have TTL and are cleaned up automatically.

443+- **JSON-RPC methods**: `message/send` (sync), `message/stream` (SSE), `tasks/get`,

444+ `tasks/cancel`. Dispatched via `POST /a2a`.

445+- **Skills**: Registered in a DB-backed registry. Each skill receives task context

446+ (messages, metadata) and returns structured results. `quotaManagement.ts` summarizes

447+ quota; `smartRouting.ts` recommends routing decisions.

448+- **Agent Card**: `/.well-known/agent.json` exposes capabilities, skills, and metadata

449+ for client auto-discovery.

390450

391−## Planning & Research Artifacts (superpowers, deep-research)

451+### ACP Module (`src/lib/acp/`)

392452

393−`_tasks/` is a **separate, isolated git repository** that is gitignored by the main

394−repo (`.gitignore` → `_tasks/`). It is the canonical home for working artifacts —

395−plans, specs/designs, research, hand-offs — so they stay **versioned in their own

396−repo** instead of polluting the main OmniRoute tree.

453+Agent Communication Protocol registry and manager.

397454

398−**Hard rule — never write superpowers / planning / research output under `docs/` or

399−the repo root.** The superpowers skills ship with defaults that point at `docs/…`

400−(`writing-plans` → `docs/superpowers/plans/`, `brainstorming` → `docs/superpowers/specs/`).

401−Those defaults are **overridden here**. Whenever you invoke superpowers (or any

402−plan/spec/research generator) in this project, save to `_tasks/` instead, using the

403−same filename convention:

455+### Memory System (`src/lib/memory/`)

404456

405−| Artifact (skill) | Default (do NOT use) | Save here instead |

406−| ---------------------------------- | ------------------------- | ------------------------------------------------------------- |

407−| Plans (`writing-plans`) | `docs/superpowers/plans/` | `_tasks/superpowers/plans/YYYY-MM-DD-<feature>.md` |

408−| Specs / design (`brainstorming`) | `docs/superpowers/specs/` | `_tasks/superpowers/specs/YYYY-MM-DD-<topic>-design.md` |

409−| Research (`deep-research`, ad-hoc) | `docs/research/` | `_tasks/research/…` |

410−| Hand-offs (`/handoff`) | — | `_tasks/hands-off/<YYYY-MM-DD>_<branch>_v<versão>_sess-<id>/` |

457+Extraction, injection, retrieval, summarization, and store modules for persistent

458+conversational memory across sessions.

411459

412−When a superpowers skill announces a path like "saved to `docs/superpowers/plans/…`",

413−rewrite it to the `_tasks/…` equivalent before writing. Commit those artifacts inside

414−the `_tasks/` repo (`git -C _tasks …`), never in the main repo.

460+### Skills System (`src/lib/skills/`)

415461

416−## Git Workflow

462+Extensible skill framework: registry, executor, sandbox, built-in skills,

463+custom skill support, interception, and injection.

417464

418−```bash

419−# Never commit directly to main

420−git checkout -b feat/your-feature

421−git commit -m "feat: describe your change"

422−git push -u origin feat/your-feature

423−```

465+#### Skills Internals

424466

425−**Branch prefixes**: `feat/`, `fix/`, `refactor/`, `docs/`, `test/`, `chore/`

467+- **`registry.ts`**: DB-backed skill registration and discovery. Skills have metadata

468+ (name, description, version, enabled status) stored in SQLite.

469+- **`executor.ts`**: Execution engine with configurable timeout and retry logic.

470+ Receives skill name + input, looks up the skill, runs it in the sandbox.

471+- **`sandbox.ts`**: Isolation layer for custom (user-provided) skills. Limits resource

472+ access and execution time.

473+- **Built-in skills**: Ship with OmniRoute (e.g., quota management, routing). Located

474+ alongside the registry.

475+- **Interception/Injection**: Skills can intercept requests in the pipeline (pre/post

476+ processing) or inject context into prompts.

426477

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

478+### Compliance (`src/lib/compliance/`)

428479

429−**Husky hooks**:

480+Policy index for compliance enforcement.

430481

431−- **pre-commit**: lint-staged + `check-docs-sync` + `check:any-budget:t11` + `check:tracked-artifacts`

432−- **pre-push**: intentionally light (PATH/npm sanity only). `any-budget` + `tracked-artifacts`

433− already run on pre-commit; re-running them on every push was pure double-pay. CI still

434− enforces both. (Was Fase 6A.12 full pre-push gate; folded into pre-commit in #6716.)

482+### MITM Proxy (`src/mitm/`)

435483

436−### Worktree isolation (MANDATORY for every development task)

484+MITM proxy capability with certificate management, DNS handling, and target routing.

437485

438−Multiple sessions/agents work this repo in parallel. The main checkout is **shared**, so a

439−`git checkout`/branch switch in it silently discards another session's uncommitted work and

440−yanks the branch out from under whatever else is running (incidents: 2026-06-05, 2026-06-13).

486+### Middleware (`src/middleware/`)

441487

442−**Rule: never develop on the shared main checkout. Every task gets its own git worktree on its

443−own dedicated branch, and you MUST confirm the base branch with the operator before creating it.**

488+Request middleware including `promptInjectionGuard.ts`.

444489

445−1. **Ask first — which base branch?** Before creating anything, ask the operator (via

446− `AskUserQuestion`, unless they already told you) from which branch the new worktree/branch

447− should be cut. Do NOT assume `main` or "whatever I'm on" — the answer is usually the active

448− `release/vX.Y.Z`, but it can be another feature/release branch. Get the base explicitly.

449−2. **Create an isolated worktree + branch off that base** (never reuse the main checkout).

450− **🔴 MANDATORY PATH: every worktree lives under `.claude/worktrees/` — and nowhere else.**

451− This is the single canonical location (the same dir the native `EnterWorktree` tool uses). It

452− is gitignored AND in the `tsconfig.json` / `.dockerignore` excludes, so worktrees never leak

453− into the build scope. **Never** use `.worktrees/`, repo-root, or any other path — a worktree

454− outside `.claude/worktrees/` (a) escapes the build-scope excludes and poisons `next build` (the

455− `tsconfig` `include: **/*` globs ~70× the codebase → OOM; incident 2026-06-25) and (b) scatters

456− worktrees across two dirs.

490+### Guardrails (`src/lib/guardrails/`)

457491

458− ```bash

459− BASE_BRANCH="release/vX.Y.Z" # ← the branch the operator confirmed in step 1

460− TASK="feat/your-feature" # feat/ fix/ refactor/ docs/ test/ chore/

461− git fetch origin "$BASE_BRANCH"

462− git worktree add ".claude/worktrees/${TASK##*/}" -b "$TASK" "origin/$BASE_BRANCH"

463− cd ".claude/worktrees/${TASK##*/}"

464− # Reuse the main checkout's node_modules to skip a per-worktree npm install.

465− # HARD LINKS (`cp -al`), never a symlink: ~5s for the whole tree and near-zero extra

466− # disk (the inodes are shared), and unlike a symlink it does not break the dev server.

467− cp -al "$(git -C <main_checkout> rev-parse --show-toplevel)/node_modules" node_modules

468− ```

492+Hot-reloadable guardrails framework (3 built-in: pii-masker, prompt-injection, vision-bridge). Fail-open. The `pii-masker` guardrail is registered and runs on every request, but its data-mutating logic is **opt-in** and OFF by default — it only redacts when `PII_REDACTION_ENABLED` (request) / `PII_RESPONSE_SANITIZATION` (response + streaming) are enabled (both `defaultValue: "false"`); with them off, payloads pass through untouched. A request can additionally opt OUT of any guardrail via header (`x-omniroute-disabled-guardrails`). Never make PII default-on (Hard Rule #20). See [`docs/security/GUARDRAILS.md`](docs/security/GUARDRAILS.md).

469493

470− **Never `ln -s` node_modules.** Turbopack rejects a symlink that resolves outside the

471− project root, so `npm run dev` dies with a FATAL panic (`Symlink [project]/node_modules

472− is invalid, it points out of the filesystem root`) while typecheck, lint and the test

473− runners all keep passing — the error names "filesystem root", not the worktree, so it

474− reads like a Next/build bug and costs real time to trace (incident 2026-07-31, #9043).

494+### Cloud Agents (`src/lib/cloudAgent/`)

475495

476− In Claude Code prefer the native `EnterWorktree` tool (it already creates worktrees under

477− `.claude/worktrees/`): create the worktree with the command above, then call `EnterWorktree`

478− with its `path`.

496+`CloudAgentBase` abstract class + 3 agents (codex-cloud, devin, jules). Tasks persisted in `cloud_agent_tasks`; management auth required. See [`docs/frameworks/CLOUD_AGENT.md`](docs/frameworks/CLOUD_AGENT.md).

479497

480−3. **Work, commit, push, open the PR — all from inside the worktree.** Never `git checkout` a

481− different branch inside a worktree another session might share.

482−4. **Tear down only your own** worktree + branch when done, from the main checkout:

483− `git worktree remove .claude/worktrees/<dir>` then `git branch -D <task>`. Never blanket-delete

484− `fix/*`/`feat/*` — other sessions keep their own; delete only the branches you created, by name.

485−5. **Never touch another session's worktree, branch, or uncommitted changes.** If `git worktree

486−list` shows worktrees you didn't create, leave them alone. End every session with the main

487− checkout back on the branch it started on (the active `release/vX.Y.Z`, never `main`).

498+### Evals (`src/lib/evals/`)

488499

489−---

500+Generic eval framework: `evalRunner.ts`, `runtime.ts`. Targets: combo / model / suite-default. See [`docs/frameworks/EVALS.md`](docs/frameworks/EVALS.md).

490501

491−## Environment

502+### Webhooks (`src/lib/webhookDispatcher.ts`)

492503

493−- **Runtime**: Node.js ≥22.0.0 <23 || ≥24.0.0 <27, ES Modules. This is the **only supported** runtime for the published `omniroute` 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 omniroute`) 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.

494−- **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`).

495−- **TypeScript**: 6.0+, target ES2022, module esnext, resolution bundler

496−- **Path aliases**: `@/*` → `src/`, `@omniroute/open-sse` → `open-sse/`, `@omniroute/open-sse/*` → `open-sse/*`

497−- **Default port**: 20128 (API + dashboard on same port)

498−- **Data directory**: `DATA_DIR` env var, defaults to `~/.omniroute/`

499−- **Key env vars**: `PORT`, `JWT_SECRET`, `API_KEY_SECRET`, `INITIAL_PASSWORD`, `REQUIRE_API_KEY`, `APP_LOG_LEVEL`

500−- Setup: `cp .env.example .env` then generate `JWT_SECRET` (`openssl rand -base64 48`) and `API_KEY_SECRET` (`openssl rand -hex 32`)

504+HMAC-signed delivery, exponential backoff, auto-disable after 10 failures. 7 event types. See [`docs/frameworks/WEBHOOKS.md`](docs/frameworks/WEBHOOKS.md).

501505

502−---

506+### Authorization Pipeline (`src/server/authz/`)

503507

504−## Quality Gates & Ratchets

508+`classify → policies → enforce`. 3 route classes (PUBLIC / CLIENT_API / MANAGEMENT). See [`docs/architecture/AUTHZ_GUIDE.md`](docs/architecture/AUTHZ_GUIDE.md).

505509

506−OmniRoute has **~48 quality-gate scripts** (`scripts/check/` + `scripts/quality/`) wired

507−across **9 gate-running jobs** in `.github/workflows/ci.yml` (`lint`, `quality-gate`,

508−`quality-extended`, `docs-sync-strict`, `i18n-ui-coverage`, `i18n`, `pr-test-policy`,

509−`test-vitest`, `sonarqube`), plus the `quality.yml` fast-gates job (PR→`release/**`) and

510−3 nightly workflows (`nightly-property`, `nightly-resilience`, `nightly-llm-security`;

511−`nightly-mutation` once merged). Full inventory, per-job breakdown, and operational

512−procedures are in [`docs/architecture/QUALITY_GATES.md`](docs/architecture/QUALITY_GATES.md).

510+### Reasoning Replay (`src/lib/db/reasoningCache.ts` + `open-sse/services/reasoningCache.ts`)

513511

514−**Quick reference:**

512+Hybrid in-memory + SQLite cache for `reasoning_content`. Re-injects on multi-turn for strict providers (DeepSeek V4, Kimi K2, Qwen-Thinking, GLM, xiaomi-mimo). See [`docs/routing/REASONING_REPLAY.md`](docs/routing/REASONING_REPLAY.md).

515513

516−- Gates in jobs `lint` + `docs-sync-strict`: pass/fail policy gates —

517− fix the violation or add an allowlist entry with a justification comment + tracking issue.

518−- Gates in job `quality-gate`: ratchet — metrics (ESLint warnings, code coverage, duplication,

519− complexity) must not regress vs `quality-baseline.json`. Update via

520− `npm run quality:ratchet -- --update` when a metric genuinely improves.

521−- Job `test-vitest` runs `npm run test:vitest` (MCP tools, autoCombo, cache) — blocking.

522− `test:vitest:ui` is advisory until UI component tests are triaged.

514+### Tunnels (`src/lib/{cloudflaredTunnel,ngrokTunnel}.ts` + `src/app/api/tunnels/`)

523515

524−**Allowlist policy (short form):** Fix the cause; use the allowlist only for pre-existing

525−violations you cannot fix in the same PR. Add a comment with justification + issue number.

526−Stale allowlist entries (suppressing a violation that no longer exists) will be caught by

527−the stale-enforcement added in Fase 6A.3.

516+Cloudflare Quick/Named, ngrok, Tailscale Funnel. See [`docs/ops/TUNNELS_GUIDE.md`](docs/ops/TUNNELS_GUIDE.md).

528517

518+### Adding a New Provider

519+

520+1. Register in `src/shared/constants/providers.ts`

521+2. Add executor in `open-sse/executors/` (if custom logic needed)

522+3. Add translator in `open-sse/translator/` (if non-OpenAI format)

523+4. Add OAuth config in `src/lib/oauth/constants/oauth.ts` (if OAuth-based)

524+5. Add models in `open-sse/config/providerRegistry.ts`

525+

529526 ---

530527

531−## Hard Rules

528+## Subdirectory AGENTS.md Files

532529

533−1. Never commit secrets or credentials

534−2. Never add logic to `localDb.ts`

535−3. Never use `eval()` / `new Function()` / implied eval

536−4. Never commit directly to `main`

537−5. Never write raw SQL in routes — use `src/lib/db/` modules

538−6. Never silently swallow errors in SSE streams

539−7. Always validate inputs with Zod schemas

540−8. Always include tests when changing production code

541−9. 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`.

542−10. Never bypass Husky hooks (`--no-verify`, `--no-gpg-sign`) without explicit operator approval.

543−11. 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`.

544−12. 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`.

545−13. 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`.

546−14. 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`.

547−15. 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`.

548−16. 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 OmniRoute — 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.

549−17. 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`.

550−18. 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.

551−19. 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".

552−20. PII redaction/sanitization is **opt-in — never on by default**. OmniRoute 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`.

553−21. **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/OmniRoute --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.

554−22. **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):

555− - **(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).

556− - **(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.)

530+- **[`src/lib/db/AGENTS.md`](src/lib/db/AGENTS.md)** — SQLite persistence, domain modules, migrations

531+- **[`open-sse/services/AGENTS.md`](open-sse/services/AGENTS.md)** — Routing engine, combo resolution, strategy selection

557532

533+## Reference Documentation (docs/)

534+

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

536+

537+| Area | Doc |

538+| ------------------------------------------ | --------------------------------------------------------------------------------------------------------------- |

539+| Repo navigation | [`docs/architecture/REPOSITORY_MAP.md`](docs/architecture/REPOSITORY_MAP.md) |

540+| Architecture | [`docs/architecture/ARCHITECTURE.md`](docs/architecture/ARCHITECTURE.md) |

541+| Engineering reference | [`docs/architecture/CODEBASE_DOCUMENTATION.md`](docs/architecture/CODEBASE_DOCUMENTATION.md) |

542+| Auto-Combo (12-factor, 18 strategies) | [`docs/routing/AUTO-COMBO.md`](docs/routing/AUTO-COMBO.md) |

543+| Resilience (3 layers) | [`docs/architecture/RESILIENCE_GUIDE.md`](docs/architecture/RESILIENCE_GUIDE.md) |

544+| Skills | [`docs/frameworks/SKILLS.md`](docs/frameworks/SKILLS.md) |

545+| Memory | [`docs/frameworks/MEMORY.md`](docs/frameworks/MEMORY.md) |

546+| Cloud agents | [`docs/frameworks/CLOUD_AGENT.md`](docs/frameworks/CLOUD_AGENT.md) |

547+| Guardrails | [`docs/security/GUARDRAILS.md`](docs/security/GUARDRAILS.md) |

548+| Evals | [`docs/frameworks/EVALS.md`](docs/frameworks/EVALS.md) |

549+| Compliance | [`docs/security/COMPLIANCE.md`](docs/security/COMPLIANCE.md) |

550+| Webhooks | [`docs/frameworks/WEBHOOKS.md`](docs/frameworks/WEBHOOKS.md) |

551+| Authz | [`docs/architecture/AUTHZ_GUIDE.md`](docs/architecture/AUTHZ_GUIDE.md) |

552+| Stealth | [`docs/security/STEALTH_GUIDE.md`](docs/security/STEALTH_GUIDE.md) |

553+| Reasoning replay | [`docs/routing/REASONING_REPLAY.md`](docs/routing/REASONING_REPLAY.md) |

554+| Agent protocols (A2A / ACP / Cloud) | [`docs/frameworks/AGENT_PROTOCOLS_GUIDE.md`](docs/frameworks/AGENT_PROTOCOLS_GUIDE.md) |

555+| MCP server | [`docs/frameworks/MCP-SERVER.md`](docs/frameworks/MCP-SERVER.md) |

556+| A2A server | [`docs/frameworks/A2A-SERVER.md`](docs/frameworks/A2A-SERVER.md) |

557+| API reference | [`docs/reference/API_REFERENCE.md`](docs/reference/API_REFERENCE.md) + [`docs/openapi.yaml`](docs/openapi.yaml) |

558+| Provider catalog (auto-generated) | [`docs/reference/PROVIDER_REFERENCE.md`](docs/reference/PROVIDER_REFERENCE.md) |

559+| Tunnels | [`docs/ops/TUNNELS_GUIDE.md`](docs/ops/TUNNELS_GUIDE.md) |

560+| Electron desktop | [`docs/guides/ELECTRON_GUIDE.md`](docs/guides/ELECTRON_GUIDE.md) |

561+| Release flow | [`docs/ops/RELEASE_CHECKLIST.md`](docs/ops/RELEASE_CHECKLIST.md) |

562+| Quality gates (35 gates, allowlist policy) | [`docs/architecture/QUALITY_GATES.md`](docs/architecture/QUALITY_GATES.md) |

563+| Cluster opt-in profiles (memory, bifrost) | [`docs/architecture/cluster-decisions.md`](docs/architecture/cluster-decisions.md) |

564+

558565 ---

559566

560−## PII & Stream Sanitization Learnings

567+## Fork / Upstream Workflow

561568

562−### 1. Regex Security (ReDoS)

569+This repository is a fork of `diegosouzapw/OmniRoute`. Keep fork-only operational

570+changes (for example GHCR image publishing, personal deployment workflows, or local

571+automation) out of upstream contribution PRs.

563572

564−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.

573+When preparing a PR for upstream, always start the work branch from the upstream

574+**default branch** — the active `release/vX.Y.Z` line (today `release/v3.8.49`).

575+Never branch from `main`: `main` only receives release squash-merges, so a branch

576+cut there is weeks behind and produces conflict-heavy PRs

577+(see `CONTRIBUTING.md` and `docs/ops/BRANCHING_MODEL.md`):

565578

566−### 2. SSE Snapshot Handling

579+```bash

580+git fetch upstream

581+# the default branch is the active release line, e.g. release/v3.8.49

582+git switch -c <branch-name> upstream/release/vX.Y.Z

583+```

567584

568−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.

585+Only cherry-pick or reapply the changes intended for the upstream PR.

569586

570−### 3. Database Handles in Tests

587+---

571588

572−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.

589+## Review Focus

590+

591+- **DB ops** go through `src/lib/db/` modules, never raw SQL in routes

592+- **Provider requests** flow through `open-sse/handlers/`

593+- **MCP/A2A pages** are tabs inside `/dashboard/endpoint`, not standalone routes

594+- **No memory leaks** in SSE streams (abort signals, cleanup)

595+- **Rate limit headers** must be parsed correctly

596+- All API inputs validated with **Zod schemas**

597+- **Provider constants** validated at module load via Zod (`src/shared/validation/providerSchema.ts`)

598+- **Pricing data** syncs from LiteLLM via `src/lib/pricingSync.ts`

599+- **Memory/Skills** are cross-cutting: affect MCP tools, request pipeline, and A2A skills

600+- **⛔ NEVER close a contributor's PR** after using their code — always merge via GitHub so they get credit. See `.agents/workflows/review-prs.md` for full policy.

573601