docs(adr): ADR-155 — nightly self-learning security harness#2417
Open
ruvnet wants to merge 1 commit into
Open
Conversation
Proposes a single nightly composite GitHub Actions workflow that fans
out into five orthogonal security scan dimensions (deps CVE, MCP static,
MCP active pentest, CodeQL, differential drift) and converges into a
learned triage step that ranks, dedupes, and routes findings.
Three learning loops are scoped (each guarded by its own triple-gate
env flag per the ADR-150 pattern):
A. Per-dimension confidence weighting (KRR over `(finding, dimension,
human_outcome)` tuples)
B. Severity calibration (isotonic regression of CVSS → realized impact
in our stack)
C. Auto-fix bid (out of scope here — reserved for ADR-156)
Triggered by today's CWE-78 incident in agentic-flow ≤ 2.0.13
(GHSA-vcv2-r9jh-99m5): we caught the vulnerability via a reporter, not
via our own gates. This ADR closes that class without inventing new
primitives — composes existing AgentDB persistence, KRR learning,
ADR-026 routing for cost-bounding, ADR-097 budget circuit breaker
pattern, ADR-150 oia-audit static surface, ADR-152 drift detection.
Load-bearing invariant: the harness is ADVISORY by default. The
learning layer learns what to surface vs suppress, NOT what to ship.
HIGH+ findings always alert regardless of suppression; auto-fix PRs
always go through human review; suppression entries have TTL.
Phased: Phase 1 lands the workflow with dumb max-severity ranking
(closes the CWE-78-class regression alone); Phase 2 adds loop A behind
a feature flag once 30 days of outcome data accumulate; Phase 3 adds
loop B; Phase 4 (auto-fix) requires its own ADR.
Co-Authored-By: RuFlo <ruv@ruv.net>
This was referenced Jun 18, 2026
meefs
pushed a commit
to meefs/claude-code-flow
that referenced
this pull request
Jun 22, 2026
…a to 0.2.6
Implements ADR-153 (Darwin Mode integration). Adds the WRITE layer that
closes the loop ADR-150's READ layer opens: score/genome describe a
harness; evolve changes one.
Three new surfaces, all honoring the four ADR-150 architectural constraints
(removable / optional / graceful degradation / CI-absent-path coverage):
- harness-evolve mutate 1 of 7 policy surfaces, sandbox-score, promote
- harness-security-bench upstream's "Darwin Shield" (their own ADR-155)
- harness-bench create/verify bench suites for evolve --bench
Wiring:
- plugins/ruflo-metaharness/scripts/_darwin.mjs — shared subprocess helper
- plugins/ruflo-metaharness/scripts/evolve.mjs — main verb (--confirm gate)
- plugins/ruflo-metaharness/scripts/security-bench.mjs — Darwin Shield wrapper
- plugins/ruflo-metaharness/scripts/bench.mjs — supporting verb
- plugins/ruflo-metaharness/skills/harness-{evolve,security-bench,bench}/SKILL.md
- v3/@claude-flow/cli/src/mcp-tools/metaharness-tools.ts — 3 new MCP tools:
metaharness_evolve, metaharness_security_bench, metaharness_bench
- v3/@claude-flow/cli/package.json optionalDependencies:
+ @metaharness/darwin ~0.3.1
bump metaharness ~0.1.11 -> ~0.2.6
- ruflo/package.json optionalDependencies: mirrored (per the ruvnet#2112 lesson —
wrapper does NOT inherit root overrides; required for transitive pin)
Safety posture (matches mint.mjs convention):
- evolve requires --confirm; without it, returns a dry-run plan
- Ruflo caps: --generations 1..50, --children 1..20, --concurrency 1..8,
--population 1..20, --cycles 1..100 (upstream supports more)
- Upstream exit code 99 (safety-disqualified) propagates VERBATIM —
not remapped — so CI can distinguish "evolution failed" from
"evolution surfaced a safety-tripping mutation"
- All three scripts emit {degraded: true, reason: 'metaharness-darwin-not-available'}
and exit 0 when the optional dep is absent
Connection to ruflo's own ADR-155 (nightly self-learning security harness,
PR ruvnet#2417, issue ruvnet#2418): harness-security-bench is the closest reference
implementation for our Loop A reward-signal sanity check. Running it
periodically gives us the empirical floor — if Darwin Shield's champion
reaches TPR=1/FPR=0 on the seeded corpus, our loop A's gradient signal
is sound.
Tests:
- plugins/ruflo-metaharness/scripts/test-graceful-degradation.mjs extended
with the 3 new scripts; all 16 assertions pass (8 skills × 2 contracts:
exit 0 AND emit "degraded": true)
- npx tsc --noEmit clean on v3/@claude-flow/cli
Co-Authored-By: RuFlo <ruv@ruv.net>
meefs
pushed a commit
to meefs/claude-code-flow
that referenced
this pull request
Jun 22, 2026
MINOR per semver — backward-compatible additions: - 3 new MCP tools: metaharness_evolve, metaharness_security_bench, metaharness_bench - 3 new skills: harness-evolve, harness-security-bench, harness-bench - New optional dep: @metaharness/darwin ~0.3.1 - Umbrella bump: metaharness ~0.1.11 → ~0.2.6 Changes: - Root `package.json` 3.12.4 → 3.13.0 - `v3/@claude-flow/cli` 3.12.4 → 3.13.0 - `ruflo/` wrapper 3.12.4 → 3.13.0 Published to npm with all three legacy dist-tags pointing at 3.13.0: @claude-flow/cli@3.13.0 → latest, alpha, v3alpha claude-flow@3.13.0 → latest, alpha, v3alpha ruflo@3.13.0 → latest, alpha, v3alpha Implements ADR-153 (Darwin Mode integration). All 4 ADR-150 architectural constraints honored — removable, optional, graceful, CI-gated. Connects to ADR-155 (ruvnet#2417, ruvnet#2418) by surfacing the upstream Darwin Shield as an empirical floor for Loop A reward-signal soundness. Verified post-publish: every package × every dist-tag = 3.13.0. Co-Authored-By: RuFlo <ruv@ruv.net>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Decision drivers
Test plan
This is a proposal ADR — no code lands in this PR. Phase 1 implementation will be its own PR after this is merged. The ADR itself:
🤖 Generated with RuFlo