seed - SKILL.md Agent Skill

name: seed description: "Generate validated Seed specifications from interview results" mcp_tool: ouroboros_generate_seed mcp_args: session_id: "$1"

/ouroboros:seed

Generate validated Seed specifications from interview results.

Required Skill Capabilities

ask_user — ask human-judgment questions through the active runtime's user-question surface.
inspect_code — read repo-local agent roles and recover exact context from local files before guessing.
call_mcp — use available Ouroboros MCP tools directly, including runtime tool discovery when a deferred MCP surface must be loaded.
run_shell — run bounded local commands for audit-trail writes and setup steps.
refine_answer — confirm free-form user decisions before treating them as accepted seed revisions.
maintain_ledger — keep QA scores, candidate decisions, rejected proposals, and audit trail keys visible.

Usage

ooo seed [session_id]
/ouroboros:seed [session_id]

Trigger keywords: "crystallize", "generate seed"

Instructions

When the user invokes this skill:

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.

Use the active runtime's call_mcp capability to find and load the seed generation MCP tool through runtime tool discovery when needed:
```
tool discovery query: "+ouroboros seed"
```
The tool will typically be named mcp__plugin_ouroboros_ouroboros__ouroboros_generate_seed (with a plugin prefix). After runtime tool discovery returns, the tool becomes callable through the active runtime's call_mcp capability.
If runtime tool discovery finds the tool → proceed to Path A. If not → proceed to Path B.

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 you are about to call (idempotent — a no-op when the schema is already loaded) so the correct schema is guaranteed present for that call. Use "+ouroboros seed" before ouroboros_generate_seed, "+ouroboros qa" before ouroboros_qa, and "+ouroboros lateral" before ouroboros_lateral_think. If a load ever returns no matching tool, switch to the documented fallback / Path B instead of retrying the failing call.

Path A: MCP Mode (Preferred)

If the ouroboros_generate_seed MCP tool is available (loaded via runtime tool discovery above):

Determine the interview session:
- If session_id provided: Use it directly
- If no session_id: Check conversation for a recent ouroboros_interview session ID
- If none found: Ask the user

Call the MCP tool through the active runtime's call_mcp capability:

Tool: ouroboros_generate_seed
Arguments:
  session_id: <interview session ID>

The tool extracts requirements from persisted interview state, calculates ambiguity score, and generates the Seed YAML.

Seed generation response shapes: Branch only after an actual Seed YAML artifact is available.
- If the response has status: "delegated_to_subagent" and dispatch_mode: "plugin", keep the returned session_id, wait for the plugin-managed subagent result, then extract the Seed YAML from that result. Do not enter the QA loop using the delegation envelope as the artifact.
- If the response directly contains Seed YAML, extract that YAML directly.
- If neither shape yields Seed YAML, stop and ask the user to resume generation or provide the missing artifact; do not fabricate a seed just to satisfy the QA loop.
Continue immediately into the required QA Refinement Loop. Do not present the seed as final, ask for acceptance, or proceed to "After Seed Generation" until QA exits with PASS or the user explicitly accepts a below-threshold best attempt at the loop boundary.

Advantages of MCP mode: Automated ambiguity scoring (must be <= 0.2), structured extraction from persisted interview state, reproducible.

Path B: Plugin Fallback (No MCP Server)

If the MCP tool is NOT available, fall back to agent-based generation:

Read src/ouroboros/agents/seed-architect.md and adopt that role.
Recover the interview requirements before drafting; do not invent missing context:
- If session_id was provided, first identify context for that same session: use current-thread interview Q&A only when it clearly belongs to that session_id, and use current-thread corrections only when they explicitly amend that same interview or seed request.
- If same-session conversation context is incomplete, use the active runtime's inspect_code / run_shell capabilities to look for persisted interview artifacts under the Ouroboros data directory (for example ~/.ouroboros/data/), exported session artifacts, or other exact local records for that ID.
- If both same-session conversation context and a persisted artifact are available, merge them conservatively: keep the persisted transcript as evidence, but let explicit same-thread user corrections or clarifications supersede older persisted wording.
- If no session_id was provided, use current-thread interview Q&A only when it is complete enough to identify one coherent interview; otherwise ask which interview or requirements summary should be seeded.
- If no matching artifact is found, or if local artifacts plus matching conversation history still do not provide enough requirements, ask the user for the missing interview transcript / concise requirement summary, or ask them to run or resume ooo interview. Do not generate a seed from an absent or mismatched transcript.
Generate a Seed YAML specification from the recovered requirements.
Continue immediately into the required QA Refinement Loop. Do not present the seed as final, ask for acceptance, or proceed to "After Seed Generation" until QA exits with PASS or the user explicitly accepts a below-threshold best attempt at the loop boundary.

QA Refinement Loop (Required after generation)

After Path A or Path B produces a seed, do not present it as final yet. Run a QA loop until the seed passes a high quality bar.

The first generation (Path A ouroboros_generate_seed or Path B agent role) runs exactly once and establishes the seed's ontology. From there on, all revisions are direct YAML edits by you (main session) — do not call ouroboros_generate_seed again. It does not accept revision hints, and re-running it would discard the established ontology.

Threshold for seed: pass_threshold: 0.90 (stricter than default 0.80 — seeds are structural specs and must be precise).

Max iterations: 5. Track the highest-scoring seed across all iterations (the "best attempt"). If still not PASS after 5, present that best attempt with its QA verdict and ask the user: accept it as-is, make one final manual edit and accept it below threshold, or escalate to ooo interview / ooo unstuck. If the user chooses one final manual edit, apply exactly that user-specified edit, present the complete edited Seed YAML in a fenced yaml block, and ask for explicit below-threshold acceptance; do not start a sixth QA iteration, rerun QA, or claim the result passed unless the user explicitly asks to rerun QA despite the max-iteration cap. If the user accepts any below-threshold attempt, present the complete accepted Seed YAML in a fenced yaml block before proceeding to "After Seed Generation".

The seed sits inside the Define diamond of Double Diamond — where expansion (Wonder) and convergence (Reflect/Refine/Restate) both happen in service of a single sharp specification. Expansion is not the enemy; unchecked expansion that bypasses the user gate is. The four-phase cycle plus User Adoption Gate is the workflow's primary safeguard.

Loop:

Establish the QA evaluator for this run:
- MCP QA mode: Load the QA tool via the active runtime's call_mcp capability using runtime tool discovery query "+ouroboros qa" if not already loaded.
- Fallback QA mode: If MCP is unavailable, read src/ouroboros/agents/qa-judge.md, adopt that evaluator role, and return its exact JSON schema: lowercase verdict (pass/revise/fail), numeric score, dimensions, differences, suggestions, and reasoning. In this mode there is no MCP-owned qa_session_id; track iteration history in the audit block and local loop ledger instead.
Obtain a QA verdict using the available mode:

MCP QA mode — call QA on the generated seed through the active runtime's call_mcp capability:
```
Tool: ouroboros_qa
Arguments:
  artifact: <the seed YAML>
  quality_bar: "Seed must be internally consistent, acceptance_criteria must be measurable and testable, constraints must be concrete (no vague terms), ontology_schema must cover all entities referenced in goal/criteria, and there must be no contradictions between fields. acceptance_criteria must also be parsimonious: 3-7 outcome-level items, each one independently valuable user-visible outcome, NOT an implementation step. Flag over-fragmentation (more than 7 criteria, or any criterion that is a sub-step of a sibling) as a difference with the same severity as a missing piece — merging redundant criteria is as important as adding missing ones."
  artifact_type: "document"
  pass_threshold: 0.90
  seed_content: <the seed YAML>
  qa_session_id: <reuse across iterations>
  iteration_history: <accumulated>
```
Fallback QA mode — skip the tool call and evaluate the current seed text under the QA Judge role from step 1, using the same quality bar and threshold. Treat the locally produced verdict exactly like the MCP verdict for the PASS/REVISE/FAIL branch below.

QA response shapes: Branch only after a usable verdict is available.
- In MCP QA mode, if the response has status: "delegated_to_subagent" and no verdict payload, keep the returned qa_session_id, wait for the plugin-managed subagent result, then parse that result as the QA verdict. Do not treat the delegation envelope itself as PASS/REVISE/FAIL.
- In MCP QA mode, if the response already includes a scored verdict, parse that inline verdict directly.
- In fallback QA mode, parse the exact QA Judge JSON. Normalize verdict to uppercase only for the branch labels below (pass→PASS, revise→REVISE, fail→FAIL). Treat differences as blocking/revision issues and suggestions as proposed fixes; do not add non-schema fields such as loop_action.
- In all modes, append the parsed verdict plus applied/rejected revision decisions to iteration_history before the next QA pass.
Branch on verdict:
- PASS (>= 0.90): Exit loop. Present the final validated Seed YAML to the user, then proceed to "After Seed Generation" below.
- REVISE (0.40–0.89): Run the Wonder → Reflect → Refine → Restate cycle below, then loop back to step 2.
- FAIL (< 0.40): Stop the loop. The seed has fundamental issues that regeneration likely won't fix. Show the full verdict and recommend ooo interview to revisit requirements, or ooo unstuck to challenge assumptions. Do not proceed to celebration.
On iteration N >= 3, briefly tell the user "Refining seed (iteration N/5)..." so they know progress is being made — but do not dump full verdicts each round; only deltas.
After PASS, show a one-line summary of the journey: Seed passed QA at iteration N/5 with score X.XX.
Immediately after that PASS summary, present the complete final validated Seed YAML in a fenced yaml block. This must happen before any "After Seed Generation" celebration, star prompt, setup prompt, or next-step text.

Wonder → Reflect → Refine → Restate (REVISE branch)

This revision loop mirrors the Double Diamond Define cycle: diverge via multiple perspectives first, then converge through debate, user decision, and structural application. Revisions must NEVER be auto-applied by the main session alone — "No candidate is accepted by default." (Symposium User Adoption Gate)

Four explicit phases per iteration:

Wonder — diverge: collect raw proposals from independent sources
Reflect — debate: surface where sources agree and where they conflict
Refine — user gate: human picks which proposals enter the next seed
Restate — apply: edit YAML in place with accepted items only

Phase 1 — Wonder (diverge): collect raw proposals from available sources

Source 1 — QA Judge (structural, external) The suggestions from the QA verdict. These are gaps, contradictions, and quality issues in the YAML itself. QA cannot see the interview.

Source 2 — Socrates (dialectical, user-intent evidence) You are Socrates — the Socratic facilitator lens from skills/interview/SKILL.md and src/ouroboros/agents/socratic-interviewer.md. Review the current seed YAML against verifiable interview evidence, in this order:

If a session_id exists, first use available persisted interview/session state for that session. Path A may run from ooo seed <session_id> in a fresh conversation, so persisted state can be the only reliable dialectic record.
Use conversation memory when it is available in the current thread.
If no persisted state or conversation evidence is available for a point, mark Socrates output as no Socrates-only proposal: dialectic context unavailable for that point. Do not invent user preferences, rejected scope, or interview nuance.

From the available evidence, surface 2–4 items neither QA nor lateral personas can see:

Did the user emphasize a constraint that got softened or dropped?
Did something the user explicitly rejected sneak back in?
Did the seed flatten nuance the user spent multiple turns clarifying?
Are there silent assumptions the user never agreed to?
Does wording contradict stated priorities (e.g., "MVP in a week" but 8 acceptance criteria)?

If QA and Socrates conflict, do not resolve the conflict silently in Wonder. Carry both candidates into Reflect as a divergent signal, cite the available evidence for each side, and let the Refine user gate choose the resolution. Do not assume the Socratic lens is automatically authoritative; QA can be correct when no user-intent evidence contradicts it.

Source 3 — ouroboros_lateral_think (independent perspectives, MCP-only when available) Attempt to load the MCP tool with the active runtime's call_mcp capability using runtime tool discovery query "+ouroboros lateral" if needed. If the tool loads, call it through the active runtime's call_mcp capability to collect 5 independent MCP personas or isolated perspectives:

Tool: ouroboros_lateral_think
Arguments:
  problem_context: |
    Seed is in REVISE state (QA score X.XX, threshold 0.90).
    Current seed YAML:
    <YAML>
    QA suggestions:
    - <suggestion 1>
    - <suggestion 2>
    Original user goal from interview: <recall>
  current_approach: "The seed as currently drafted (above)."
  persona: "all"
  failed_attempts:
    - <previously rejected candidate from earlier iterations>
    - ...

The 5 personas return distinct revision angles:

hacker: unconventional workarounds (e.g., reframe a constraint instead of adding criteria)
researcher: knowledge the seed assumes but doesn't pin down
simplifier: criteria/constraints to remove for sharper convergence
architect: structural reorganization without expansion
contrarian: challenges to assumptions the seed treats as settled

Parsing persona outputs when lateral MCP is available: Each persona returns free-form prose, not a structured list. After the parallel call returns, read each persona's text and extract its concrete proposals into discrete candidates (one revision per candidate, not bundled). If a persona's output is purely abstract advice with no actionable revision, drop it from the candidate list rather than inventing one. Aim for 1–2 candidates per persona — if a persona produced 5, pick the 2 most concrete and discard the rest.

Lateral response shapes: ouroboros_lateral_think does not have one universal synchronous shape. After calling it with all personas, branch on the returned shape before extracting candidates:

Plugin delegation: If the response has status: "delegated_to_subagent", dispatch_mode: "plugin", and an _subagents array, wait for every plugin-managed subagent result. Extract concrete revision candidates from those returned persona texts. Do not attempt to parse candidates from the envelope prompts themselves.
Inline fallback with dispatch block: If the response returns markdown content plus the hidden sentinel , keep the visible markdown as the lateral scaffold. If the active runtime can dispatch isolated subagents, decode the sentinel JSON (dispatch_mode, persona_count, payloads) and send each payload.prompt + payload.context through that isolated subagent surface, then extract candidates from the returned persona texts. If the runtime cannot dispatch subagents, synthesize candidates directly from the visible inline persona sections.
Inline fallback without dispatch block: Treat the returned markdown as the complete lateral output and synthesize candidates directly from the visible persona sections. Do not split solely on --- if doing so would corrupt user-provided content; prefer section headers and visible persona boundaries.

If runtime tool discovery cannot load ouroboros_lateral_think, do not emulate lateral personas or read persona files directly. Record no lateral proposals: MCP lateral tool unavailable as Source 3 output and proceed with QA plus Socrates/available sources. The QA refinement loop remains required, and the User Adoption Gate still applies to any proposed revision.

Phase 2 — Reflect (debate): structure proposals by agreement and conflict

Do not just dedupe. Read all proposals from the available Wonder sources (Sources 1–2, plus Source 3 only when ouroboros_lateral_think loaded successfully) and surface the structure of the debate:

Convergent signals (strong): same revision proposed by ≥2 independent sources. Example: QA says "criterion 3 is unmeasurable" AND simplifier says "drop criterion 3 or sharpen it" → strong signal to act on criterion 3.
Divergent signals (decisions): sources conflict. Example: researcher says "add User entity to ontology" but simplifier says "remove the User reference from goal — single-user implied". This is a decision the user must resolve, not the main session.
Singleton signals (weaker): one source only. Keep but mark as weaker.
Balance signal: count expansion proposals (add) vs convergence proposals (sharpen/remove). Show the ratio above the user gate as information, not warning — e.g., Balance: 4 expand / 2 sharpen / 1 remove. Both directions are legitimate; the user decides what mix to accept.

Output of Reflect: a tagged candidate list with per-item metadata (sources_backing, type=expand|sharpen|remove|resolve_conflict).

Phase 3 — Refine (User Adoption Gate)

Use the active runtime's ask_user capability with executable single-choice questions only. Do not ask one multi-select question or present options that can be selected contradictorily.

Ask sequential single-choice questions in this order:

For each conflict group, ask one question with exactly one option per mutually exclusive resolution plus "Leave unchanged"; handle the runtime's free-form "Other" response if available. Record the chosen option as accepted and mark the other options in that group rejected.
For non-conflicting convergent signals, ask one single-choice batch question: "Apply all strong non-conflicting revisions, review one by one, or skip them?" If the user chooses review, ask each revision as a Yes/No/Other single-choice question.
For singleton signals, ask one single-choice batch question: "Review singleton revisions one by one, skip all singleton revisions, or other?" If the user chooses review, ask each revision as a Yes/No/Other single-choice question.
Always include a skip option at the batch level: "None of the above / keep current seed for now". If selected during a REVISE iteration, skip applying this candidate batch and return to QA or the max-iteration boundary; do not treat it as below-threshold acceptance unless the user separately chooses an explicit "accept current seed below threshold" option at the loop boundary.

Convergent signals still appear first in summaries, conflicts second, singletons last. Conflict questions must be asked before any non-conflicting batch is applied so contradictory revisions cannot both enter the next seed.

Iteration N/5 — QA score X.XX (REVISE)

Which revisions should enter the next seed?
(Nothing accepted by default. Questions are single-choice and may be sequential.)

Strong (multiple sources agree):
A. [QA + Simplifier] Criterion 3 "easy to use" — sharpen to measurable predicate
B. [QA + Socrates] Re-add "single-user only" constraint dropped from iter-0

Conflicts (mutually exclusive — pick at most one per group):
C1. [Researcher] Add User entity to ontology
C2. [Simplifier] Remove User reference from goal (single-user implied)
C3. Neither — leave ontology untouched on this point

Singletons:
D. [Contrarian] Constraint "no external DB" contradicts criterion 7
E. [Architect] Group 3 user-management criteria under one parent
F. [Hacker] Replace "user authentication" with "device-local key file"

Other:
G. None of the above (exit loop with current seed)
H. Other — describe a different change

Portable gate example:

{
  "questions": [{
    "question": "Conflict: how should the seed handle User in the ontology?",
    "header": "Conflict C",
    "options": [
      {"label": "Add User", "description": "Accept C1 and reject C2/C3"},
      {"label": "Remove User", "description": "Accept C2 and reject C1/C3"},
      {"label": "Leave unchanged", "description": "Accept C3 and reject C1/C2"}
    ],
    "multiSelect": false
  }]
}

Balance line shown above the question: Balance: 4 expand / 2 sharpen / 1 remove (informational, not a warning).

Track all rejected candidates across iterations and pass them as failed_attempts to subsequent ouroboros_lateral_think calls when the MCP lateral tool is available, so personas don't re-propose them.

Phase 4 — Restate (apply accepted only)

Edit the previous seed YAML in place. Apply ONLY user-accepted items. Do not start from scratch. Do not lose fields that were already correct. Do not call ouroboros_generate_seed again — that tool runs only at iter-0.

If the user skips all proposed revisions for a REVISE iteration, keep the current seed unchanged for that iteration and continue the loop or max-iteration boundary. Exit below threshold only after an explicit loop-boundary acceptance choice such as "accept current seed below threshold"; before proceeding to "After Seed Generation", present the complete accepted Seed YAML in a fenced yaml block so the accepted artifact is explicit.

Common edit shapes (both expansion and convergence are legitimate when the user accepted them):

Sharpen: replace vague phrase with measurable predicate ("fast" → "p95 latency < 200ms")
Tighten: harden a soft constraint ("some kind of storage" → "SQLite, single file, no server")
Make implicit explicit: surface a silent assumption as a constraint
Remove: drop a contradicting or redundant criterion
Expand (when accepted): add an ontology entity, criterion, or constraint that fills a gap the user confirmed

Audit trail

After each revision, append a brief audit block to ~/.ouroboros/seed-revisions/<revision_key>.md (create the directory if it doesn't exist) capturing: iteration N, QA score, all candidates with source tag, user's accept/reject decisions, and the resulting diff vs. previous iteration. This makes the convergence path inspectable and lets the user replay decisions later.

Choose revision_key deterministically:

If session_id exists, use that exact session_id.
If no session_id exists (common in Path B), derive a stable seed label from the seed goal or project name plus the current UTC timestamp, for example <slugified-goal>-YYYYMMDDTHHMMSSZ. Once derived, reuse the same key for every iteration in the current seed run.
If the filesystem write is unavailable, include the same audit block in the assistant response instead of silently dropping it.

Format:

## Iteration N — score X.XX

### Candidates
- [A] [QA+Simplifier] sharpen criterion 3 — **accepted**
- [B] [Socrates] re-add single-user constraint — **accepted**
- [C1] [Researcher] add User entity — rejected
- [C2] [Simplifier] remove User from goal — **accepted**
- [D] [Contrarian] resolve no-DB / criterion-7 conflict — rejected
- ...

### Diff vs. iteration N-1
- criteria[2]: "easy to use" → "first-time user completes flow in < 3 clicks"
- constraints: + "single-user only"
- goal: "...for users..." → "...for the single operator..."

Seed Components

The seed contains:

GOAL: Clear primary objective
CONSTRAINTS: Hard limitations (e.g., Python >= 3.12, no external DB)
ACCEPTANCE_CRITERIA: Measurable success criteria
ONTOLOGY_SCHEMA: Data structure definition (name, fields, types)
EVALUATION_PRINCIPLES: Quality principles with weights
EXIT_CONDITIONS: When the workflow should terminate
METADATA: Version, timestamp, ambiguity score, interview ID

Example Output

goal: Build a CLI task management tool
constraints:
  - Python >= 3.12
  - No external database
  - SQLite for persistence
acceptance_criteria:
  - Tasks can be created
  - Tasks can be listed
  - Tasks can be marked complete
ontology_schema:
  name: TaskManager
  description: Task management domain model
  fields:
    - name: tasks
      type: array
      description: List of tasks
    - name: title
      type: string
      description: Task title
evaluation_principles:
  - name: completeness
    description: All requirements are implemented
    weight: 1.0
  - name: usability
    description: CLI commands are clear and easy to use
    weight: 0.7
exit_conditions:
  - name: all_criteria_met
    description: All acceptance criteria pass
    criteria: 100% of acceptance criteria are satisfied
  - name: tests_green
    description: The project test suite passes
    criteria: Required automated tests exit successfully
metadata:
  ambiguity_score: 0.15

After Seed Generation

On successful seed generation, first announce:

Your seed has been crystallized!

Then check ~/.ouroboros/prefs.json for star_asked. If star_asked is not set to true, use the active runtime's ask_user capability with this single question:

{
  "questions": [{
    "question": "If Ouroboros helped clarify your thinking, a GitHub star supports continued development. Ready to unlock Full Mode?",
    "header": "Next step",
    "options": [
      {
        "label": "\u2b50 Star & Setup",
        "description": "Star on GitHub + run ooo setup to enable run, evaluate, status"
      },
      {
        "label": "Just Setup",
        "description": "Skip star, go straight to ooo setup for Full Mode"
      }
    ],
    "multiSelect": false
  }]
}

Star & Setup: Run gh api -X PUT /user/starred/Q00/ouroboros, merge {"star_asked": true} into ~/.ouroboros/prefs.json, then read and execute skills/setup/SKILL.md
Just Setup: Merge {"star_asked": true} into ~/.ouroboros/prefs.json, then read and execute skills/setup/SKILL.md
Other (user provides custom text): Merge {"star_asked": true} into ~/.ouroboros/prefs.json, skip setup

Create ~/.ouroboros/ directory if it doesn't exist. Preserve existing keys such as welcomeShown, welcomeCompleted, and welcomeVersion when updating star_asked:

python3 - <<'PY'
import json, os
path = os.path.expanduser('~/.ouroboros/prefs.json')
os.makedirs(os.path.dirname(path), exist_ok=True)
try:
    with open(path, encoding='utf-8') as f:
        prefs = json.load(f)
    if not isinstance(prefs, dict):
        prefs = {}
except Exception:
    prefs = {}
prefs['star_asked'] = True
with open(path, 'w', encoding='utf-8') as f:
    json.dump(prefs, f, indent=2)
    f.write('\n')
PY

If star_asked is already true, skip the question and just announce:

Your seed has been crystallized!
📍 Next: `ooo run` to execute this seed (requires `ooo setup` first)