--- name: evolve description: "Start or monitor an evolutionary development loop" --- # ooo evolve - Evolutionary Loop ## Description Start, monitor, or rewind an evolutionary development loop. The loop iteratively refines the ontology and acceptance criteria across generations until convergence. ## Flow ``` Gen 1: Seed(O₁) → Execute(all nodes) → Judge → Gate Gen 2: Wonder(failed nodes) → Reflect(active nodes only) → Execute(active) → Gate(all) Gen 3: Repeat with a smaller active set; frozen PASS nodes are reverified, not regenerated ...until the outcome gate passes, the working set stagnates, ontology converges, or max 30 generations is reached ``` ## Usage ### Start a new evolutionary loop ``` ooo evolve "build a task management CLI" ``` ### Fast mode (ontology-only, no execution) ``` ooo evolve "build a task management CLI" --no-execute ``` ### Check lineage status ``` ooo evolve --status ``` ### Record an isolated full-graph benchmark control Call `ouroboros_evolve_step` for an existing Gen 2+ lineage with `benchmark_control: true`, `execute: true`, and an explicit clean Git `project_dir`. Use a distinct clean worktree at the treatment commit. Normal evolve calls keep `benchmark_control` false and never launch a control arm. ### Rewind to a previous generation ``` ooo evolve --rewind ``` ## Instructions ### Load MCP Tools (Required before Path A/B decision) The Ouroboros MCP tools are often registered as **deferred tools** that must be explicitly loaded before use. **You MUST perform this step before deciding between Path A and Path B.** 1. Use the active runtime's tool-discovery capability to find and load the evolve MCP tools: ``` tool discovery query: "+ouroboros evolve" ``` 2. The tools will typically be named with prefix `mcp__plugin_ouroboros_ouroboros__` (e.g., `ouroboros_evolve_step`, `ouroboros_interview`, `ouroboros_generate_seed`). After runtime tool discovery returns, the tools become callable. 3. If the tools are callable — already exposed, or loaded by discovery — proceed to **Path A**. An empty discovery result for already-exposed tools is expected, not a failure. Proceed to **Path B** only if they are genuinely absent (no Ouroboros MCP server). **IMPORTANT**: Do NOT skip this step. Do NOT assume MCP tools are unavailable just because they don't appear in your immediate tool list. They are almost always available as deferred tools that need to be loaded first. **CRITICAL — deferred-schema guard (prevents "Invalid tool parameters"):** This skill makes `ouroboros_*` MCP calls across multiple turns, and each turn runs in a fresh tool context. A deferred tool's schema loaded on one turn is NOT guaranteed to still be loaded on the next. If you call any `ouroboros_*` MCP tool while its schema is not loaded in the **current** turn, the runtime rejects the call with **"Invalid tool parameters"** before it ever reaches the server. Therefore: **immediately before EVERY `ouroboros_*` MCP call in this skill, re-run the tool-discovery load query for the specific MCP tool or documented tool family you are about to call**. Use `"+ouroboros evolve"` for `ouroboros_evolve_step`, `ouroboros_lineage_status`, and the evolve flow's documented tool family; use `"+ouroboros interview"` before `ouroboros_interview`, `"+ouroboros seed"` before `ouroboros_generate_seed`, and `"+ouroboros lateral"` before `ouroboros_lateral_think`. If a load returns no matching tool (and the tool is not already callable — an empty load for an already-exposed tool is an expected no-op, not absence), switch to the documented fallback / Path B instead of retrying the failing call. ### Path A: MCP Available (loaded via runtime tool discovery above) **Starting a new evolutionary loop:** 1. Parse the user's input as `initial_context` 2. Run the interview: call `ouroboros_interview` with `initial_context` 3. Complete the interview (3+ rounds until ambiguity ≤ 0.2) 4. Generate seed: call `ouroboros_generate_seed` with the `session_id` 5. Call `ouroboros_evolve_step` with: - `lineage_id`: new unique ID (e.g., `lin_`) - `seed_content`: the generated seed YAML - `execute`: `true` (default) for full Execute→Evaluate pipeline, `false` for fast ontology-only evolution (no seed execution) - `benchmark_control`: `false` (default). Set `true` only for a deliberate Gen 2+ full-graph control in an explicit clean Git project/worktree. 6. Check the `action` in the response: - `continue` → Inspect `active_ac_indices`, then call `ouroboros_evolve_step` again with just `lineage_id`. Only active failed/reopened nodes evolve; frozen PASS nodes stay immutable and are boundary-reverified. - `ontology_stable` → This is not success. Call `ouroboros_evolve_step` again for the same `lineage_id` with `execute: true` so the stable Seed goes through Execute→Evaluate. Do not call standalone evaluate or report convergence before that step returns `converged`. - `converged` → Evolution complete! Display final ontology - `stagnated` → Ontology unchanged for 3+ gens. Consider `ouroboros_lateral_think` - `exhausted` → Max 30 generations reached. Display best result - `failed` → Check error, possibly retry. If the error reports an expired lineage owner, first confirm that the prior owner process is dead, then make one explicit recovery call with `recover_expired_claim: true`. Never set this flag for a merely slow or still-running owner. 7. **Repeat step 6** while action is `continue`. Treat `ontology_stable` as the explicit transition above, then process the `execute: true` response using step 6. Stop only on `converged`, `stagnated`, `exhausted`, or `failed`. 8. When the loop terminates, display a result summary with next step: - `converged`: `◆ Current state → next: Ontology converged! Run ooo evaluate for formal verification` - `ontology_stable`: `◆ Current state → next: Run the same lineage with execute=true for Execute→Evaluate; this is not verified convergence yet` - `stagnated`: `◆ Current state → next: ooo unstuck to break through, then ooo evolve --status to resume` - `exhausted`: `◆ Current state → next: ooo evaluate to check best result — or ooo unstuck to try a new approach` - `failed`: `◆ Current state → next: Check the error above. ooo status to inspect session, or ooo unstuck if blocked` **Checking status:** 1. Call `ouroboros_lineage_status` with the `lineage_id` 2. Display: generation count, ontology evolution, convergence progress **Rewinding:** 1. Call `ouroboros_evolve_step` with: - `lineage_id`: the lineage to continue from a rewind point - `seed_content`: the seed YAML from the target generation (Future: dedicated `ouroboros_evolve_rewind` tool) ### Path B: Plugin-only (no MCP tools available) If MCP tools are not available, explain the evolutionary loop concept and suggest installing the Ouroboros MCP server. See [Getting Started](docs/getting-started.md) for install options, then run: ``` ouroboros mcp serve --runtime claude-cli ``` Then add to your runtime's MCP configuration (e.g., `~/.claude/mcp.json` for Claude Code). ## Key Concepts - **Wonder**: "What do we still not know?" - examines evaluation results to identify ontological gaps and hidden assumptions - **Reflect**: "How should the ontology evolve?" - proposes specific mutations to fields, acceptance criteria, and constraints - **Convergence**: Verified success requires the independently evaluated outcome to pass. In ontology-only mode, similarity ≥ 0.95 returns the non-success `ontology_stable` handoff and the same lineage must run once with `execute: true`. Judge-score plateau and the 30-generation cap are also non-success stops. - **Focused evolution**: Gen 1 establishes the baseline. Gen 2+ derives an active node set from failed/regressed verifier results. Only those nodes are open to Wonder, Reflect, and execution; PASS nodes are frozen. The active output should shrink toward zero rather than feeding the full graph back. - **Frugality proof**: A smaller active-node count is a working-set observation, not a savings claim. Gen 2+ records measured total-generation runtime tokens (Wonder, all Reflect attempts, validator/evaluator providers, executor AC attempts, dependency analysis, decomposition policy/attestation/repair, coordinator review, and shadow replay when enabled), wall time, calls, retries, and quality evidence. A PASS requires a paired full-graph control and focused treatment from distinct clean worktrees at the same Git commit, at least 10% fewer total-generation runtime tokens, and no final-gate, evaluation score/stage, drift, reward-hacking, per-AC verdict/score, lineage regression, or TraceGuard degradation. The comparison key covers the complete semantic Seed and previous evaluation. Every expected primary attempt needs exactly one runtime-token receipt and TraceGuard verdict bound to the exact session, primary dispatch, and root identity; every active root must be dispatched; every other generation provider call needs runtime usage; normalized backend/model/tier/mode/effort/permission and provider request configurations must be complete and match the same root AC or auxiliary phase role and preserve per-unit call multiplicity across the paired arms. Shared-unit sequences must be exactly equal; only whole control-only units may be removed. Blank/unknown phase roles are incomplete evidence. Full-graph controls must execute every Seed root; treatment active/frozen sets must exactly partition the Seed. Auxiliary tools, system-prompt identity, request kwargs, and fresh/scoped session mode are configuration-bound. Completion profiles are resolved once into a sealed dispatch config; only registered exact-key, single-attempt, secret-safe adapter attestations whose in-memory endpoint/credential authority reaches the actual call boundary unchanged are proof-eligible. Global or post-attestation mutable routing is ineligible. Malformed attestations invalidate the proof, not evolution. True resumed contexts without semantic identity and unknown/conflicting effective models or unreconciled usage counters are incomplete. Every current Seed AC needs exactly one final verdict against the same final semantic Seed contract. Missing, duplicate, malformed, partial, or opaque evidence is `insufficient_data`. Evidence reads and provider-call capture are capped; oversized or non-finite individual/subtotal/combined usage, cap overflow, and Wonder-only early stops also produce non-PASS durable receipts. Evolve never launches the control automatically because doing so would consume the production savings being measured. - **Judge/Gate split**: Evaluation records score and evidence; deterministic convergence logic decides accept/continue/stagnate. A passing Gen 1 may end immediately—minimum-generation churn is not required after the gate passes. - **No-drift stop**: If rejected evaluation scores move by less than 0.01 for the 3-generation window, stop as `stagnated` and hand off to `ooo unstuck` instead of spending the 30-generation cap. - **Rewind**: Each generation is a snapshot. You can rewind to any generation and branch evolution from there - **evolve_step**: Runs exactly ONE generation per call. Designed for Ralph integration — state is fully reconstructed from events between calls - **execute flag**: `true` (default) runs full Execute→Evaluate each generation. `false` skips execution for fast ontology exploration. Previous generation's execution output is fed into Wonder/Reflect for informed evolution. An `ontology_stable` response from `execute: false` must be followed by the same lineage with `execute: true`; ontology stability alone is never convergence. - **QA verdict**: Each generation's response includes a QA Verdict section (when `execute=true` and `skip_qa` is not set). Use the QA score to track quality progression across generations. Pass `skip_qa: true` to disable ## RFC #1392 State Breadcrumb Footer Your final response MUST end with exactly one breadcrumb footer line: ``` ◆ → next: ``` Derive `` from live session state via `ouroboros_session_status` when that MCP projection is available; otherwise derive it from this skill's actual outcome. Never use a linear `Step N of M` footer because Ouroboros is an evolutionary loop. When the next action is genuinely a choice, list 2-3 honest options in the `next:` clause. The breadcrumb line must be the last line of the response.