Battam1111/Myco

v0.6.14 — 2026-04-29 — Cycle 自起 fruit—winnow—molt 闭环 + sub-agent fanout craft critique pattern

Zero R1-R7 surface deltas; zero new manifest verbs; zero new lint dims; zero subsystem changes; schema v2 unchanged. Mechanizes the bridge from sporulate (distilled note) to a candidate kernel mutation, with owner-merge-gate as the only remaining synapse. Substrate kernel adds 0 lines of LLM dispatch; the autopoietic loop runs entirely in the Claude Code Agent layer (Agent tool sub-agent fanout — analogous host-mediated mechanism to v0.6.0's existing exception #2 for myco fruit content authoring via MCP sampling).

Why this release (owner observation 2026-04-29)

After the neat-freak ingestion exercise earlier the same day, the owner observed:

通过这次实践，我是明白了一件事情，就是 Myco 吃完之后似乎并不会直接进化自己的内核啊？

The metabolic flow (eat → assimilate → sporulate) and the morphogenetic flow (fruit → winnow → molt) were doctrinally separated and procedurally bridged by 5 manual agent-prose handoffs. Most insights captured in distilled notes never reached the kernel because the bridge was too long. v0.6.14 mechanizes 4 of 5 handoffs.

Governing craft: docs/primordia/v0_6_14_cycle_autostart_fruit_winnow_molt_loop_craft_2026-04-29.md. Round 1.5 was authored via 3 real parallel Agent tool calls (mycoparasite / saprotroph / mycorrhiza fungal critics with disjoint visibility scopes). 28 raw tensions → 17 dedupe + 1 derived → 8 P0 blockers all resolved before any code was written. agentIds preserved in craft frontmatter for audit.

The auto-loop chain

eat → assimilate → sporulate (distilled note)
                       │
                       │  /myco-evolve <slug>  (manual trigger; owner runs)
                       ▼
                  primordium (autonomous mode)
                  Spawns 3 parallel sub-agents via Task tool, each with
                  disjoint visibility: mycoparasite (draft only) / saprotroph
                  (doctrine only) / mycorrhiza (src only). Synthesizes
                  Round 1.5 T-tensions; HIGH veto from any critic forces
                  abort to DRAFT. Round 2 + Round 3.
                       │ if LANDED
                       ▼
                  winnow (shape gate)
                       │
                       ▼
                  hypha (read-only feasibility trace)
                       │
                       ▼
                  anamorph (only if schema delta needed)
                       │
                       ▼
                  stipe --branch-only
                  Branch: fruiting/<slug>-<YYYY-MM-DD> (fungal taxonomy;
                  NOT "auto-craft/"). Implements scope per craft path_allowlist
                  (anti-prompt-injection guard); runs gate quintet; commits;
                  pushes branch (NOT main); opens PR via gh pr create with
                  ≤300-char summary + repo-relative path link (NOT full text;
                  prevents internal-path leakage).
                       │
                       ▼
                  Owner: merge to ship, OR close-without-merge.
                  - merge → kernel evolves
                  - close-without-merge → auto_revert.yml deletes branch +
                    posts vetoed_intent JSON to substrate-wide tracking issue
                  - next senesce reaps tracking-issue comments → queues
                    .myco_state/auto_evolve_vetoed_pending.json

What landed (Group A: doctrine boundaries)

• L2_DOCTRINE/cycle.md gains §"Cycle 自起 fruit—winnow—molt 闭环 (v0.6.14+)" describing the auto-loop chain, sub-agent fanout protocol, governance.auto_evolve_force_high_risk owner-gate, vetoed_intent reaping path, convergence + budget guards, and the mechanical L0 P1 evidence claim.
• L2_DOCTRINE/boundary.md gains §"Sixth seam: GitHub-side critic-fanout + auto-revert (v0.6.14+)" describing /myco-evolve as 6th slash, fungal critic taxonomy, auto_revert.yml, MP1 host-signature extension, plugin-mirror discipline, and 6th surface invariant ("auto-craft branches use fruiting/<slug>-<date> prefix").
• src/myco/providers/ excreted (__init__.py + README.md deleted). Reserved at v0.5.6 as named, path-scoped, canon-gated escape hatch from MP1; through 7 minor releases (v0.5.6 → v0.6.13) it remained empty by design. v0.6.14's auto-loop runs entirely in the Agent layer; the kernel-side escape hatch is unused infrastructure. Future provider coupling, if ever needed, requires its own L0 P1 amendment craft and a fresh contract-bumping molt — NOT a pre-baked escape hatch sitting empty for years. Historical note migrated to boundary.md "Future axis".
• _canon.yaml::system.llm_policy enum reduced 3 → 2 values (forbidden, opt-in). Drops providers-declared per the providers/ excretion above.

What landed (Group B: 11 new canon governance fields, all opt-in defaults)

# v0.6.14 Cycle 自起 fruit—winnow—molt 闭环
governance:
  # ... (existing fields preserved) ...
  auto_propose_enabled: false                          # Master switch
  auto_evolve_min_wall_clock_seconds_between: 600      # Rate limit (10 min)
  auto_evolve_critic_count: 3                          # mycoparasite + saprotroph + mycorrhiza
  auto_evolve_branch_prefix: "fruiting/"               # Fungal taxonomy
  auto_evolve_distilled_hash_cooldown_senesce: 7       # Anti-feedback loop
  auto_evolve_force_high_risk: true                    # Forces owner-gate
  auto_evolve_pr_window_skip: true                     # PR-merge is sole gate
  auto_evolve_min_distilled_severity: medium           # LOW distilled doesn't trigger
  auto_evolve_daily_budget_usd: null                   # Owner-set cap
  auto_evolve_tracking_issue_id: null                  # Seed via seed script
  recognized_authoring_hosts:                          # MP1 host-signature whitelist
    - "claude-code-agent"
    - "cursor-agent"
    - "claude-desktop-agent"
    - "cowork-agent"
    - "human"

What landed (Group C: 5 boundary surfaces, all in Agent layer with plugin-mirror discipline)

• .claude/agents/primordium.md + <repo>/agents/ mirror: gains §"Autonomous mode (v0.6.14+ — Round 1.5 critic fanout)" with quarantine pre-step, 3 fungal critic role-prompt templates (mycoparasite / saprotroph / mycorrhiza, real fungal-ecology terms), severity rubric, veto-vote semantics. Frontmatter tools: adds Task. Recursion-forbidden line explicitly excepts primordium-only autonomous mode; the other 4 subagents (hypha / autolysis / stipe / anamorph) preserve their forbiddance.
• .claude/commands/myco-evolve.md + mirror (new): orchestrator slash command. 7 pre-flight gates (master switch, tracking issue seeded, rate limit, daily budget, distilled-hash cooldown, severity threshold, distilled exists). 8-step orchestration chain. Refusal vs halt semantics documented.
• .claude/agents/stipe.md + mirror: gains §"--branch-only mode (v0.6.14+ — for the auto-evolve loop)" with phase delta vs default 9-phase pipeline, branch-only invocation contract (--target-craft + --path-allowlist + --branch-only flags), PR body construction (privacy-conscious: ≤300-char summary + repo-relative path, never full craft text), refusal modes specific to --branch-only mode (canon master switch off, rate limit hit, budget exceeded, distilled hash cooldown, no tracking issue, path_allowlist touches L0/L1/protocol).
• .github/workflows/auto_revert.yml (new): triggers on pull_request.closed && !merged && head_ref starts-with fruiting/. Deletes branch via gh api DELETE; posts vetoed_intent JSON comment to substrate-wide tracking issue. Idempotent + deny-by-default permissions (contents: read at top-level, contents: write only at job level for branch deletion).
• scripts/seed_auto_evolve_tracking_issue.py (new): one-shot script that creates the substrate-wide tracking issue via gh issue create and writes the issue number to _canon.yaml::system.governance.auto_evolve_tracking_issue_id. Idempotent (subsequent runs no-op).

What landed (Group D: 2 substrate-side mechanical guards — substrate's only changes)

Both are guards, not LLM dispatch. The substrate kernel's L0 P1 stance is unchanged.

• src/myco/homeostasis/dimensions/mechanical/mp1_no_provider_imports.py extended:
  • Part 1 (existing v0.5.6 behavior): scan src/myco/** for LLM provider SDK imports. Provider/ path-skip removed (directory excreted; nonexistence is stronger guard).
  • Part 2 (new v0.6.14): scan docs/primordia/*.md files with type: craft frontmatter for required authored_by: field naming a recognized host from canon.governance.recognized_authoring_hosts. Crafts without recognized signature → HIGH finding → winnow refuses. Non-craft primordia (handoff notes, audits, design-notes, release-notes) are exempt — host-signature is craft-specific.
• src/myco/cycle/senesce.py full mode gains _reap_vetoed_intents step that reads new vetoed_intent comments via gh issue view (shell-out; no LLM call), parses JSON blobs from comment bodies, and queues them to .myco_state/auto_evolve_vetoed_pending.json. Cursor-tracked + idempotent. Canon round-trip (writing vetoed_at: into last_winnowed_proposals[]) is deferred to v0.6.15+ with a dedicated atomic-canon-write helper that preserves comments + ordering. The pending-queue file IS the source of truth for vetoed_at until then.

What landed (Group E: tests + bookkeeping)

• tests/contract/test_autopoietic_loop_structural.py (new, ~360 lines, 18 test functions): structural contract tests covering canon governance fields (11 new with type checks), L0 vocabulary discipline (branch prefix is fruiting/, NOT auto-craft/), MP1 host-signature scope (every type: craft craft has authored_by:), providers/ excretion, llm_policy enum reduction, sixth-seam doctrine sections, myco-evolve plugin-mirror byte-identity, stipe --branch-only declaration, auto_revert.yml shape, seed script existence + AST validity. Does NOT mock the Task tool — behavioral validation is human integration testing post-merge.
• tests/unit/boundary/test_subagent_and_command_surface.py extended: _EXPECTED_COMMANDS 5 → 6 (adds myco-evolve); docstring v0.6.11+ → v0.6.14+; new tests test_only_primordium_has_task_in_tools + test_only_primordium_mentions_autonomous_mode prevent the autonomous-mode exception from accidentally proliferating to other subagents.
• tests/unit/verbs/senesce/test_senesce.py: payload-shape test gains vetoed_intent_reap key.
• tests/unit/homeostasis/dimensions/test_mp1_no_provider_imports.py: test_mp1_ignores_providers_directory replaced with test_mp1_no_longer_ignores_providers_directory asserting the v0.6.14 path-skip removal.
• 20 craft files in docs/primordia/ gain authored_by: human frontmatter (5 from earlier session in scope; 15 legacy crafts bulk-added via one-shot Python script). The v0.6.14 craft itself carries authored_by: claude-code-agent.

Mechanical L0 P1 evidence (the load-bearing claim of v0.6.14)

| Layer | LoC delta | Nature |
|-------|-----------|--------|
| Substrate kernel (src/myco/) | +~70 / -135 | Guards (MP1 ext + senesce reaper) + providers excretion. Zero new LLM dispatch. |
| Agent layer (.claude/, .github/, scripts/) | +~2400 / -0 | Subagent specs, slash command, workflow, seed script — all prose instructions to LLMs running in Claude Code, not in substrate process. |

grep -r 'Task\|sampling\|provider' src/myco/cycle/ returns zero v0.6.14-added hits. The autopoietic loop's intelligence lives entirely in the Agent layer (Claude Code's Task tool sub-agent fanout, analogous to MCP sampling — both are host-mediated mechanisms that do not require substrate-side provider SDK imports).

Test count

Pytest: 1504 passed + 1 skipped (was 1475; +29 from new contract test + extended boundary surface tests + the senesce reaper payload key).

Pre-flight evidence

• ruff check src tests scripts: All checks passed
• ruff format --check src tests scripts: 313 files clean
• mypy src/myco: 150 source files, 0 issues (was 151; -1 from providers/ excretion)
• pytest -q -n auto --dist loadfile: 1504 passed, 1 skipped (~40s)
• myco immune: exit 0, findings 76 (= baseline pre-v0.6.14; +0 incremental — MP1 host-signature check passes on all 20 authored_by:-tagged crafts)

Break from v0.6.13

None for kernel users. The autopoietic loop is opt-in via canon.governance.auto_propose_enabled: false (default). Existing substrates that don't enable the loop see no behavioral change. The MP1 host-signature requirement applies to crafts going forward; v0.6.13 substrates with no authored_by: field on their existing crafts will see HIGH findings on next myco immune run only on type: craft files, and the fix is a single-line frontmatter addition (authored_by: human).

For substrate authors enabling the autopoietic loop on their own substrate:

1. Set canon.governance.auto_propose_enabled: true.
2. Run python scripts/seed_auto_evolve_tracking_issue.py to seed the tracking issue.
3. Set canon.governance.auto_evolve_daily_budget_usd: <number> if desired.
4. Mark distilled notes for auto-evolve consideration with auto_propose: true in their frontmatter (per craft Round 2 R-T2; the cooldown + severity gates still apply).

Follow-ups deferred to v0.6.15+

• Canon round-trip helper for last_winnowed_proposals[].vetoed_at (currently queued in .myco_state/auto_evolve_vetoed_pending.json).
• Two-step owner sign for L0/L1/L2 + R-surface PRs (/sign auto-craft comment requirement).
• myco ramify --agent <name> + --command <name> flag extensions (deferred since v0.6.11; not load-bearing for v0.6.14 either).
• First true-dogfood run of /myco-evolve against the existing neat-freak distilled note (notes/distilled/d_khazix-neat-freak-isomorphism-with-myco-senesce.md) to validate the chain end-to-end.