mirror of
https://github.com/ruvnet/ruflo.git
synced 2026-09-14 14:01:28 +08:00
rUv
ae9bba523e
* fix(mcp-bridge): streamable-HTTP session cleanup — DELETE + Mcp-Session-Id + 202 (#2425) Closes #2425. Cherry-picked from djimit's commitb905c09f4with extensions: applied symmetrically to BOTH bridges (per ADR-166's "both bridges move together" rule), and added regression coverage in the ADR-166 security lock so this can't silently regress. Streamable-HTTP clients (Codex/RMCP) send `DELETE /mcp` with an `Mcp-Session-Id` header at shutdown. Previously the bridge returned `Cannot DELETE /mcp` which surfaced as noisy client-side cleanup failures after otherwise-successful MCP calls. Changes (both bridges — ruflo/src/mcp-bridge/index.js AND ruflo/src/ruvocal/mcp-bridge/index.js): - Stable per-process `MCP_SESSION_ID = randomUUID()` echoed as the `Mcp-Session-Id` header on every `/mcp*` response - `DELETE /mcp` and `DELETE /mcp/:group` → 204 No Content (idempotent) - `DELETE` added to `Access-Control-Allow-Methods`; `Mcp-Session-Id` added to `Access-Control-Allow-Headers` (both compatible with the ADR-166 CORS allowlist wiring — allowlist controls Origin, not Method/Header) - `notifications/initialized` responses now `202 Accepted` with empty body per the MCP streamable-HTTP spec (was returning a jsonrpc envelope which some clients rejected as unexpected) Regression locks (extends ADR-166 §6 acceptance): - test-security-lock.js — 2 new static checks (#7 delete handler, #8 Mcp-Session-Id header) × 2 bridges = 16/16 - test-runtime-security.mjs — 2 new runtime checks per bridge (R6 DELETE→204 + header echoed; R7 notifications/initialized→202 empty) = 20/20 total green Interaction with ADR-166 auth: DELETE flows through `app.use(requireAuth)` same as the rest of `/mcp*`. When `MCP_AUTH_TOKEN` is set, DELETE without Bearer returns 401 (verified in R2 by contrast; DELETE with Bearer returns 204 verified in R6). Local-only deployments with no token remain zero-friction (the middleware is a no-op). Credit: @djimit for the original patch (branch fix/mcp-session-cleanup-forkbase on djimit/claude-code-flow, commitb905c09f42). Co-Authored-By: djimit <djimit@users.noreply.github.com> Co-Authored-By: RuFlo <ruv@ruv.net> * feat(metaharness): learn + gepa integration — metaharness@0.3.0 / darwin@0.8.0 Version pins (optionalDependencies, both @claude-flow/cli + ruflo wrapper): - @metaharness/darwin ~0.3.1 → ~0.8.0 (GEPA engine + cand-6 genome) - metaharness ~0.2.6 → ~0.2.8 - @metaharness/redblue ~0.1.1 → ~0.1.4 (+ versioned cache dir — pin bumps now invalidate stale installs instead of serving 0.1.1 forever) - @metaharness/kernel ~0.1.0 → ~0.1.2 New surfaces (13 → 15 MCP tools, 11 → 13 CLI subcommands): - metaharness_learn / `ruflo metaharness learn` — upstream ADR-235 GEPA learning run; $0 dry-run default, --run to spend, structured checkout-required payload when the metaharness clone is absent - metaharness_gepa / `ruflo metaharness gepa` — darwin GEPA library surface (genome|validate|render|analyze); gepaOptimize stays library-only (in-process evaluator can't cross subprocess boundary) - skills: harness-learn, harness-gepa Fixes found during validation: - gepa.mjs treats ERR_PACKAGE_PATH_NOT_EXPORTED (stale pre-0.8.0 darwin in ancestor node_modules) as recoverable → falls back to versioned cache - smoke.sh contract counts advanced (scripts 13→15, tools 13→15, subcommands 11→13, CLAUDE.md doc counts) Validated: tsc clean; test-mcp-tools 171/171 across all 15 tools; darwin 0.8.0 evolve/bench/security-bench verbs proven through wrappers; redblue 0.1.4 attack preview proven via new cache path. Co-Authored-By: RuFlo <ruv@ruv.net> --------- Co-authored-by: djimit <djimit@users.noreply.github.com>
5.0 KiB
5.0 KiB
ruflo-metaharness
MetaHarness integration plugin for ruflo. Surfaces the upstream metaharness / harness / @metaharness/darwin CLIs through eleven ruflo skills, honoring ADR-150's architectural constraint that MetaHarness must remain a removable augmentation — never a required runtime dependency.
ADR-150 architectural constraint (load-bearing)
Ruflo remains operational if every MetaHarness package is removed. Every code path in this plugin satisfies four rules:
- Removable — no static
import '@metaharness/*'outside the optional-router path inv3/@claude-flow/cli/src/ruvector/neural-router.ts. - Optional in package.json —
metaharnessis inoptionalDependencies, neverdependencies. - Graceful degradation — every script catches
MODULE_NOT_FOUND/network failure and emits{ degraded: true, reason: 'metaharness-not-available' }JSON, exits 0. The graceful path is the default behavior, not a special case. - CI gate —
no-metaharness-smoke.ymlruns the plugin smoke withnpm install --no-optionaland asserts the contract still passes.
Skills
| Skill | Usage | Description |
|---|---|---|
harness-score |
/harness-score [--path .] [--alert-on-fit-below 70] |
5-dim readiness scorecard (harnessFit/compile/coverage/safety/memory + cost) |
harness-genome |
/harness-genome [--path .] [--alert-on-risk-above 0.5] |
7-section categorical report (repo_type/topology/risk/mcp/test/publish) |
harness-mcp-scan |
/harness-mcp-scan [--path .] [--fail-on high] |
Static MCP security findings — pure-read, no dispatch |
harness-threat-model |
/harness-threat-model [--path .] [--fail-on high] |
Enterprise-grade threat model (clean/low/medium/high + findings) |
harness-mint |
/harness-mint --name <id> --template <id> [--confirm] |
Scaffold a custom harness; DRY-RUN by default; refuses project-root writes |
harness-similarity |
/harness-similarity --a a.json --b b.json [--per-dimension] [--alert-below 0.5] |
ADR-152 §3.1 weighted similarity between two harness fingerprints (cosine + categorical + jaccard) |
harness-oia-audit |
/harness-oia-audit [--path .] [--alert-on-worst high] [--dry-run] |
Composite Phase-2 audit (oia-manifest + threat-model + mcp-scan) into metaharness-audit namespace |
harness-drift-from-history |
/harness-drift-from-history [--baseline-since 7d] [--threshold 0.95] |
1-command drift detection — composes audit-list + oia-audit + audit-trend |
harness-bench |
/harness-bench --op create|verify --repo <path> |
Manage @metaharness/darwin bench suites — fixed evaluation corpora for harness-evolve |
harness-evolve |
/harness-evolve --repo <path> [--generations 3] [--sandbox real|mock|agent] |
Run @metaharness/darwin evolve — mutate seven policy surfaces, sandbox-score variants, promote measured wins |
harness-security-bench |
/harness-security-bench [--population 2] [--cycles 1] [--alert-on-fail] |
"Darwin Shield" / ADR-155 — evolve a security-detection harness against a 10-vuln corpus |
harness-learn |
/harness-learn --host <h> --model <m> --slice <manifest> [--repo <checkout>] [--run] |
metaharness@0.3.0 / upstream ADR-235 — GEPA learning run; $0 dry-run default, --run to spend; needs a metaharness repo checkout |
harness-gepa |
/harness-gepa --op genome|validate|render|analyze [--path <genome.json>] |
darwin@0.8.0 GEPA library surface — genome load/validate/render + transcript failure analysis; gepaOptimize stays library-only |
Phase-0 baseline (ruflo itself, 2026-06-16)
{
"harnessFit": 82,
"compileConfidence": 100,
"taskCoverage": 79,
"toolSafety": 100,
"memoryUsefulness": 40,
"estCostPerRunUsd": 0.048,
"recommendedMode": "CLI + MCP",
"archetype": "typescript-sdk-harness",
"template": "vertical:coding",
"scaffoldReady": true,
"risk_score": 0.27,
"publish_readiness": 0.9
}
Architecture
All skills use subprocess invocation through the _harness.mjs shared helper:
skills/X/SKILL.md → scripts/X.mjs → scripts/_harness.mjs → spawnSync('npx', ['metaharness', …])
↘ on MODULE_NOT_FOUND → emit degraded JSON, exit 0
This means:
- No library import overhead on ruflo's boot path
- 60s hard timeout per subprocess (bounded blast radius)
--jsonflag forced for structured parsing- Graceful degradation is a single helper used by every skill
Cross-links
- ADR-150 — decision + architectural constraint
- Issue #2399 — phase rollout tracker
- Research dossier — full graded-evidence sourcing
- Upstream —
metaharnesssource - ADR-148/149 —
@metaharness/routercost-optimal routing (sibling integration)