Battam1111/Myco

v0.6.15 — 2026-04-29 — Agent-First default for Cycle 自起 闭环 (correct v0.6.14 owner-First regression + 5-critic L0-P1-P5 refactor)

Zero R1-R7 surface deltas; zero new manifest verbs; zero new lint dims; zero subsystem changes; schema v2 unchanged. Corrects v0.6.14's owner-First regression. The autopoietic loop's safety model returns to Agent-First as L0 P1 specifies. The sub-agent fanout pattern's structural blind spot — that 3 same-host critics share unconscious priors — is closed by deriving 5 critics from L0 P1-P5 directly (one per principle).

Why this release (owner observation)

我感觉还是给owner太大权限了, Myco 明明是 Agent-First.

v0.6.14 (shipped same day, hours earlier) introduced two coupled canon defaults that collapsed every auto-craft path to owner-merge-gate, inverting L0 P1's "Myco is a cognitive substrate for an LLM agent. The agent is the sole consumer." The owner observation is structurally correct.

Diagnosis: v0.6.14's sub-agent fanout used 3 critics (mycoparasite/saprotroph/mycorrhiza). All 3 shared an unstated "add more gates = safer" prior. None asked the inverse: "does this design pull humans into the substrate's loop, violating L0 P1?" Same-host critic correlation broke L0 P1 conformance.

Governing craft: docs/primordia/v0_6_15_agent_first_default_for_cycle_autostart_loop_craft_2026-04-29.md. Round 1.5 fanout used 4 critics (3 from v0.6.14 + new endophyte 4th critic with L0_VISION.md-only visibility). 22 dedup'd tensions; 10 P0 blockers all resolved. Endophyte's T7 ("derive critics from L0 P1-P5 directly, not from observed-failure patches") drove the 5-critic outcome.

What landed

Group A — doctrine boundaries

• L2_DOCTRINE/cycle.md § "Cycle 自起 闭环": stripped "owner-merge-gate is the only synapse" + "two-step owner sign for L0/L1/L2 + R-surface PRs" wording (both endophyte-detected L0 P1 violations); added "Agent-First default (v0.6.15+)" + "Winnow gate G7" sub-sections; updated auto-loop chain diagram for 5 critics.
• L2_DOCTRINE/boundary.md § "Sixth seam": "Three fungal critic roles" → "Five fungal critic roles (v0.6.15+, derived from L0 P1-P5)". Each critic anchored to one principle.

Group B — canon governance defaults flipped

auto_evolve_force_high_risk: true → false   # match v0.6.0 governance tiering
auto_evolve_pr_window_skip:   true → false   # restore 7-session/7-day window
auto_evolve_critic_count:        3 → 5      # one per L0 P1-P5

No auto_evolve_owner_paranoia_mode field (per Round 1.5 endophyte T3+T5: undeclared 3rd L0 P1 exception).

Group C — risk_classifier extension + winnow G7 (~190 lines substrate-side guard logic, zero LLM dispatch)

• src/myco/core/risk_classifier.py: new classify_craft_via_path_allowlist(craft_path) reads frontmatter path_allowlist:; recursion-cutter forces HIGH for any craft touching risk_classifier.py / governance.auto_evolve_* keys / .github/workflows/auto_*.yml / cycle/winnow.py / classifier tests. Closes mycoparasite T1's perpetual-motion attack.
• src/myco/cycle/winnow.py: G7 gate requires path_allowlist: list[str] in type: craft frontmatter (post-2026-04-29; pre-v0.6.15 grandfathered).

Group D — primordium 5-critic L0-P1-P5 refactor (Agent-layer)

.claude/agents/primordium.md + <repo>/agents/primordium.md mirror — Autonomous mode section refactored from 3 ad-hoc role-prompts → 5 L0-P1-P5-mapped role-prompts:

| Role | L0 Principle | Visibility |
|------|-------------|------------|
| chytrid (壶菌) | P1 — Only For Agent | L0_VISION.md only |
| rhizomorph (根状菌索) | P2 — Eternal Ingestion | ingestion subsystem code + adapters + L0 P2 |
| mycoparasite (寄生) | P3 — Eternal Evolution | draft only |
| saprotroph (腐生) | P4 — Eternal Iteration | L0/L1/L2 + canon + previous crafts |
| mycorrhiza (菌根) | P5 — Universal Interconnection | src/, tests/, .github/, .claude/, scripts/ |

All 5 names are validated fungal-ecology terms. The L0 P1-P5 mapping is the load-bearing diversity guarantee — future critic additions must name an L0 principle (or revise L0).

Group E — tests (~250 lines)

• tests/contract/test_autopoietic_loop_structural.py: flipped force_high_risk assertion + 3 new tests (pr_window_skip default false, critic_count == 5, no paranoia_mode field). boundary.md sixth-seam check now requires all 5 fungal role names.
• tests/unit/core/test_risk_classifier_recursion_cutter.py (NEW, 12 tests).

What v0.6.15 explicitly does NOT introduce

• auto_evolve_owner_paranoia_mode (endophyte T3+T5): undeclared 3rd L0 P1 exception
• "owner = observer" role redefinition (T2): L0 hasn't authorized it
• Sub-agent fanout meta-lesson in cycle.md (T4): meta-teaching is reverse L0 P1 information flow; lives in this changelog only
• .github/workflows/auto_merge.yml (T5, saprotroph T1): R7/R6 concern + scope; deferred to v0.6.16+ alongside canon round-trip helper

Sub-agent fanout meta-lesson (lives only here, not in doctrine)

3 same-host critics spawned by the same parent agent under the same conditions share an "add more gates = safer" prior. They will not catch L0 P1 inversions because L0 P1 is a reverse constraint (reduce human-loop), and the fanout's natural inertia is the forward direction (add guards).

The v0.6.15 mitigation is structural, not retrospective: critic shape is now derived from L0 P1-P5 directly. Each critic is anchored to one principle and cannot be conceived without naming a principle.

The v0.6.14 → v0.6.15 sequence is the substrate's own demonstration: dogfood the autopoietic loop on its own owner-First regression; the 4th endophyte critic (added explicitly to scan for L0 P1 inversion) caught what the 3-critic fanout missed; endophyte's own meta-critique drove the final 5-critic L0-mapped shape.

Mechanical L0 P1 evidence

• Substrate kernel (src/myco/) LoC delta: +~190 (risk_classifier extension + winnow G7), all guard logic. Zero new LLM dispatch.
• Agent layer (.claude/) LoC delta: +~600 (primordium 5-critic refactor; mostly text).
• grep -r 'owner.*observer\|paranoia\|paranoia_mode\|owner-merge-gate is the' src/myco/ docs/architecture/L2_DOCTRINE/ .claude/agents/ post-v0.6.15: zero hits.

Test count

Pytest: 1544 passed, 1 skipped (was 1529; +15 from new tests). Coverage: 85.06% (≥85% gate sustained).

Pre-flight evidence

• ruff check: All checks passed
• ruff format --check: 315 files clean
• mypy src/myco: 150 source files, 0 issues
• pytest: 1544 passed + 1 skipped, coverage 85.06%
• myco immune: exit 0, findings 76 (= baseline; +0 incremental)
• myco winnow on v0.6.15 craft: pass (G1-G7 all green)

Break from v0.6.14

Two flipped canon defaults (auto_evolve_force_high_risk + auto_evolve_pr_window_skip): substrates that explicitly opted into v0.6.14 strict behavior must set these to true in their substrate's _canon.yaml to opt back in. For substrates not opted into the autopoietic loop (auto_propose_enabled: false master switch — the default), nothing changes operationally.

Follow-ups deferred to v0.6.16+

• Canon round-trip helper for last_winnowed_proposals[].vetoed_at / .status writes (deferred from v0.6.14).
• .github/workflows/auto_merge.yml + canon-LANDED-status-driven auto-merge (operational completion of Agent-First).
• First true-dogfood /myco-evolve invocation against an existing distilled note.
• Critic role-prompt review cycle (every MAJOR release re-audits 5 critics against current L0 wording).