Cdo

CDO — Adaptive Multi-Agent Deliberation

Deterministic runtime preference inside digital/leviathan

Inside the Leviathan repo, prefer the plugin-backed CDO runtime defined in plugins/cdo/config.yaml.

Runtime contract (deterministic)

1. cdo run resolves to exec --flow plugins/cdo/flows/cdo-adaptive-deliberation.flow.yaml.
2. The flow loads the selected cdo profile (plugins/cdo/profiles/*.yaml) and validates bounds + schema invariants.
3. The profile loads the selected method recipe (plugins/cdo/recipes/*.yaml).
4. The recipe binds
   • thinking routines: plugins/cdo/catalogs/thinking-routines.yaml
   • strategy states: plugins/cdo/catalogs/igt-strategy-states.yaml
5. runtime_mode=manual exits through manual_fallback_receipt and the configured fallback lane instead of bypassing CDO receipts.
6. Turn synthesis writes claim_verdicts, provenance_refs, and level_tags gates before synthesize_final.

Runtime ownership split: where to change what

• deterministic-code — plugins/cdo/*.yaml, plugins/cdo/flows/*.flow.yaml, plugins/cdo/recipes/*.yaml, plugins/cdo/catalogs/*.yaml, plugins/cdo/schemas/*.yaml, and .lev/pm decision/runtime artifacts that declare contract boundaries.
• dna/flowmind/recipe contracts — plugins/cdo/config.yaml, flowmind_graph nodes, and profile/recipe binding fields that are loaded by plugins/cdo/config.yaml and the flow graph.
• cdo SKILL protocol — /Users/jean-patricksmith/.agents/skills/cdo/SKILL.md itself: seat roles, loop discipline, fallback rules, and update sequence.

When to run plugin-backed CDO

• Default for /cdo use in this repo, including bounded think/deep/full/debug flows.
• Any run that requires method_recipe stage output or scheduler replay.
• Any run where receipts, traceability, and external validation are part of success.

Inside Leviathan, plugin-backed CDO is the preferred path because it is bound by flow contracts and receipts in one deterministic chain (plugins/cdo/config.yaml → plugins/cdo/profiles/*.yaml → plugins/cdo/recipes/*.yaml → flow nodes). Use manual fallback only when the plugin surface cannot be resolved in this repo or the user explicitly requests the legacy multi-agent path.

Manual fallback behavior (preserved)

• Keep manual multi-agent fallback in this SKILL and this protocol as the backup path.
• Use fallback when the plugin surface cannot be resolved in the active runtime or when a user explicitly requests the old multi-agent mode.
• In Leviathan, select it as runtime_mode=manual; the flow records manual_fallback through plugins/cdo/config.yaml#cdo.fallback_policy.manual_mode and .lev/runtime/cdo/manual-fallback.
• Preserve existing seat naming when fallback is used.

Safe skill update pattern

When updating this SKILL:

1. Edit only the skill-level protocol text and cross-reference updates first.
2. Keep deterministic semantics in YAML/flow/plugin files (.flow.yaml, profiles, recipes, schemas).
3. If seat contracts change (seat names, outputs, validator fields), update:
   • this SKILL (clarity of role/sequence)
   • plugins/cdo/recipes/*.yaml that declares those seats
   • matching profile/schema docs.

Code-vs-LLM responsibility boundary (explicit)

• Code must enforce: command resolution, profile/recipe loading, schema validation, scheduler directives, artifact fanout naming, claims gates, receipt append, proof binding, cross-branch closure, and done gates in plugins/cdo/flows/cdo-adaptive-deliberation.flow.yaml.
• LLM must own: seat reasoning, claim interpretation, tension construction, dissent generation, synthesis narrative, and recommendation confidence language.
• Boundary rule: no seat or router code changes should be implied in SKILL text without a matching deterministic contract edit and a verification check that the gate outputs are still producible.

You are a ROUTER. You dispatch agents, collect artifacts, and route to synthesis. You never think, analyze, or synthesize yourself. All reasoning happens in agents.

The one rule you cannot break: Turn N+1's shape comes from Turn N's synthesis directive. You do not pre-plan turns. You do not override the directive. You read the YAML block from synthesis and execute exactly what it says.

steps:

  - id: parse_args
    action: "Parse invocation and resolve preset + modifiers + problem"
    instruction: |
      Parse the input: `/cdo <args,...> <problem>`

      Args are composable, comma-separated. Split tokens and classify:
        - Base preset: quick | think | deep | full | debug
        - Modifiers: hitl, bd, team, adaptive, autoresearch, adaptive-runtime, lev-exec, exec
        - Domain: token after "exec" (dev, arch, or <tag>)
        - Problem: everything remaining

      If no preset specified, GAUGE the problem:
        - Single question, known domain → quick
        - Trade-off or design question → think
        - Multi-stakeholder or high-uncertainty → deep
        - Strategic or high-stakes → full
        - Bug, failure, unexpected behavior → debug

      If modifiers present but no preset → default to deep.
      If "debug" → ignore all modifiers (fixed protocol).
    validation: "preset variable is set to one of: quick, think, deep, full, debug"
    on_failure: "Ask the user to clarify what they want analyzed"

  - id: select_preset
    action: "Load preset config and resolve execution parameters"
    instruction: |
      Apply the preset table:

      | Preset | Width | Max Turns | BD  | Team Mode  | Dashboard | Convergence  |
      |--------|-------|-----------|-----|------------|-----------|--------------|
      | quick  | 1-2   | 1         | No  | Subagents  | No        | N/A          |
      | think  | 2-4   | 2-3       | No  | Subagents  | No        | Perspective  |
      | deep   | 3-8   | 3-5       | Yes | TeamCreate | Yes       | Confidence   |
      | full   | 5-20  | 5-10      | Yes | TeamCreate | Yes       | Resonance    |
      | debug  | 1-3   | 7 fixed   | No  | Subagents  | No        | Turn count   |

      Then apply modifiers — they override specific settings:
        - hitl: user checkpoint between every turn
        - bd: force beads tracking (`bd create epic "CDO: {problem}"`)
        - team: force TeamCreate even for quick/think
        - adaptive: width varies per turn based on synthesis
        - autoresearch | adaptive-runtime: use codex-autoresearch as the long-run scheduler/runtime for CDO. CDO still owns reasoning; autoresearch owns run state, counters, health checks, pause/resume, and exit-gate enforcement.
        - lev-exec: route roles to different models via codex/openrouter
        - exec <domain>: inject domain team shape for T1

      For deep+ or hitl: show planning dashboard before T1 — proposed DAG
      with turns, agents, roles, skills. Two seconds of preview saves minutes.

      Load sub-files only as needed (see references/architecture.md for table).
    validation: "width, max_turns, team_mode, and convergence_type are all set"
    on_failure: "Default to think preset if configuration is ambiguous"

  - id: resolve_autoresearch_scheduler_mode
    action: "If requested, convert CDO into a scheduler-managed autoresearch run"
    instruction: |
      If `autoresearch` or `adaptive-runtime` is present, OR if the problem declares hard scheduling KPIs (`min_turns`, `min_total_agents`, `skills_per_agent`), enter CDO Autoresearch Scheduler Mode.

      This mode is for long-running deliberation only. It is not a normal CDO preset.

      Load the `codex-autoresearch` skill as the runtime contract:
        - `references/core-principles.md`
        - `references/runtime-hard-invariants.md`
        - `references/loop-workflow.md`
        - `references/pivot-protocol.md`
        - `references/health-check-protocol.md`
        - `references/parallel-experiments-protocol.md`

      Then initialize a scheduler state object before Turn 1:

      ```yaml
      cdo_scheduler:
        mode: autoresearch
        run_tag: "cdo-{slug}"
        min_turns: 10              # default for deep unknowns runs unless user specifies otherwise
        min_total_agents: 50       # default for deep unknowns runs unless user specifies otherwise
        skills_per_agent: 2        # default when user requests skill rotation
        adaptive_turn_width: true
        turn_count: 0
        total_agents: 0
        unique_skills: []
        open_tensions: []
        exit_eligible: false
      ```

      The scheduler state is the run contract. The DAG is not.

      Hard rules in this mode:
        - Do not plan all turns upfront.
        - Do not pre-generate a 50-agent roster as the execution plan.
        - Plan only the next scheduling quantum.
        - Every turn chooses strategy from current evidence, open tensions, remaining metrics, and health checks.
        - Final synthesis is blocked until `turn_count >= min_turns` AND `total_agents >= min_total_agents`, unless the user explicitly interrupts with "ship it", "just do it", or another clear stop signal.
        - Each turn must emit a scheduler update.

      Scheduler update format:

      ```yaml
      scheduler_update:
        turn_count: <int>
        agents_this_turn: <int>
        total_agents: <int>
        skills_used_this_turn: [skill-id]
        unique_skills: [skill-id]
        exit_eligible: <bool>
        remaining_turns_min: <int>
        remaining_agents_min: <int>
        next_turn_strategy: research | fanout | debate | negotiate | synthesize | devil_advocate | reduce | checkpoint
      ```

      The mental model is CPU scheduling:
        - a turn is a scheduling quantum
        - agents are runnable tasks
        - skill pairs are execution contexts
        - synthesis is scheduler feedback
        - hard metrics are exit gates
    validation: "If autoresearch/adaptive-runtime or hard scheduling KPIs are present, cdo_scheduler exists and final synthesis is blocked until hard metrics are met"
    on_failure: "Do not continue with normal CDO. Rebuild scheduler state and continue from the next turn."

  - id: execute_turns
    action: "Run the adaptive turn loop"
    instruction: |
      For each turn, execute this sequence:

      COMPOSE: Read previous synthesis directive (or problem statement for T1).
        - Decide width from directive (or preset default for T1)
        - Decide roles from directive (or preset/domain default for T1)
        - For deep+: discover 2-3 skills per agent via skill-discovery
        - Generate agent briefs with role, context, constraints, output format
        - In Autoresearch Scheduler Mode: first read `cdo_scheduler`, compute unmet metrics, then decide this turn's width and strategy. Do not follow a preplanned roster if current evidence says to change strategy.

      DISPATCH: Send agents in parallel.
        - Subagent mode: parallel Agent calls in single message
        - Team mode: SendMessage to teammates, spawn new if needed
        - Each agent writes to: tmp/cdo-{session}/t{N}-{role}.md
        - Agents cannot see each other's work during a turn

      SYNTHESIZE: Dispatch a dedicated synthesis agent (never yourself).
        - Reads ALL turn N artifacts from disk
        - Produces: common ground, tensions, gaps, surprises
        - Anti-groupthink: if >70% agreement, auto-add devil's advocate next turn
        - Emits YAML directive block:
            confidence: <float>
            convergence_met: <bool>
            gaps: [list]
            tensions: [list]
            recommended_next_turn:
              width: <int>
              agents: [{role, skills, focus}]
            scheduler_update: {turn_count, agents_this_turn, total_agents, unique_skills, exit_eligible, remaining_turns_min, remaining_agents_min, next_turn_strategy}

      ADAPT: Check exit criteria.
        - confidence >= threshold AND convergence_met → go to synthesize_final
        - Max turns reached → go to synthesize_final (forced)
        - All tensions resolved, no new gaps in 2 consecutive turns → synthesize_final
        - hitl active → present updated dashboard, user decides
        - Autoresearch Scheduler Mode: even if confidence is high, final synthesis is blocked until hard scheduler metrics are met. If progress stalls for 3 consecutive turns, use codex-autoresearch pivot/refine escalation instead of brute-force repeating the same fanout.
        - Otherwise → next turn using the directive
    validation: "Each turn has artifact files on disk AND a synthesis with YAML directive block"
    on_failure: "If synthesis missing, re-dispatch synthesis agent. If agents produced no output, check briefs and re-dispatch with clearer constraints."

  - id: synthesize_final
    action: "Produce FINAL.md — the only user-facing deliverable"
    instruction: |
      Dispatch the final synthesis agent. It reads ALL artifacts across ALL turns.

      FINAL.md contains:
        - Decision/Answer: the actual output
        - Confidence: numeric + qualitative
        - Key Tensions: what was debated, what won, why
        - Minority Reports: dissenting views preserved, not buried
        - Action Items: concrete next steps if applicable
        - Layer Tags: every structural claim tagged with its hierarchy layer
          (see "Layer Discipline" section). Cross-layer claims MUST cite
          admission evidence at every intermediate layer, or be marked
          CANDIDATE not EQUIVALENT.
        - Language Discipline: use exclusion language, not construction.
          "admissible under probe X," "not-yet-killed," "coupled with,"
          "co-varies under" — NOT "is," "equals," "creates," "drives."

      If bd tracking active: close the epic.
      Turn artifacts are audit trail only — FINAL.md is the deliverable.

      EXTERNAL VALIDATOR GATE (mandatory for external-facing output):
      Before FINAL.md is considered complete, run the external-validator pass
      (see "External Validator Before Broadcast" section). Any claim that
      fails recognition against the ground-truth surface is either retracted
      or downgraded to CANDIDATE.
    validation: "FINAL.md exists at tmp/cdo-{session}/FINAL.md with all six sections AND external-validator pass logged"
    on_failure: "Re-run final synthesis with explicit section checklist + external-validator brief"

Debug Preset — 7-Turn RCA Protocol

When preset is "debug", ignore the adaptive loop above and run this fixed sequence:

debug_steps:

  - id: debug_reproduce
    action: "T1 REPRODUCE — define exact failure condition"
    instruction: |
      Dispatch reproduction specialist. Output: exact steps, expected result,
      actual result, environment. No theorizing, no fixes. Just reproduce.
      Load modes/debug.md for full protocol.
    validation: "01-reproduce.md exists with reproduction steps and confidence level"
    on_failure: "Cannot proceed without reproduction. Ask user for more context."

  - id: debug_isolate
    action: "T2 ISOLATE — find minimal failing case"
    instruction: |
      Dispatch isolation specialist. Strip away everything unnecessary.
      Find the smallest case that still fails.
    validation: "02-isolate.md exists with minimal reproduction and isolation boundary"

  - id: debug_trace
    action: "T3 TRACE — parallel call path + working code comparison"
    instruction: |
      Two agents in parallel:
        Agent A: Trace exact execution path, find divergence point
        Agent B: Find nearest working code path, compare structural differences
    validation: "Both 03a-call-path.md and 03b-working-code.md exist"

  - id: debug_hypothesize
    action: "T4 HYPOTHESIZE — form 2-3 evidence-backed theories"
    instruction: |
      Dispatch hypothesis agent. Exactly 2-3 theories, each citing evidence
      from T3 traces. Ranked by likelihood. No fixes proposed yet.
    validation: "04-hypotheses.md exists with 2-3 hypotheses, each with evidence citations"

  - id: debug_verify
    action: "T5 VERIFY — test each hypothesis"
    instruction: |
      Test each hypothesis. Check predictions, attempt temporary modifications.
      Mark each as CONFIRMED, ELIMINATED, or INCONCLUSIVE.
      Exactly one should be CONFIRMED. If zero → return to T4.
    validation: "05-verified.md exists with exactly one CONFIRMED root cause"
    on_failure: "Return to debug_hypothesize with new evidence from verification"

  - id: debug_fix
    action: "T6 FIX — apply minimal fix"
    instruction: |
      Fix ONLY the confirmed root cause. No refactoring. No improvements.
      No touching unrelated files. If fix exceeds ~20 lines, justify.

      THE 3-FIX ESCALATION RULE:
      Count how many fix attempts have been made for this bug.
      If this is fix attempt 3 or higher → STOP. Do not attempt another fix.
      3+ failed fixes means the architecture is wrong, not the fix.
      Report to user: "3 fixes failed. This is an architectural problem, not
      a bug. Here's what each attempt revealed about the underlying design."
    validation: "06-fix.md exists with files changed, rationale, and scope check"
    on_failure: "If 3+ fixes failed, skip to FINAL with architectural assessment instead of fix"

  - id: debug_validate
    action: "T7 VALIDATE — adversarial validation"
    instruction: |
      Dispatch adversarial validator. Try to break the fix:
        1. Original reproduction case passes?
        2. Edge cases that could still trigger the bug?
        3. Regression — did fix break anything else?
        4. Root cause addressed, not just symptom?
      Verdict: PASS or FAIL. No partial pass. No "looks good enough."
    validation: "FINAL-validation.md exists with PASS or FAIL verdict"
    on_failure: "If FAIL, return to the T-step indicated by the validator. Do not restart from T1."

User Signal Awareness

When the user says any of these, STOP your current approach immediately:

  signals:
    - trigger: "Stop guessing"
      meaning: "You are proposing actions without evidence"
      response: "Drop current approach. Return to evidence gathering. Read code, run commands, trace data flow."

    - trigger: "Ultrathink this"
      meaning: "You are treating symptoms, not causes"
      response: "Widen scope. Question the framing. Is the problem statement itself wrong? Are you solving the right problem?"

    - trigger: "We're stuck?"
      meaning: "Your approach is failing and you haven't acknowledged it"
      response: "Admit the approach isn't working. Enumerate what you've tried and what each attempt revealed. Propose a fundamentally different angle."

    - trigger: "Just do it" / "Ship it"
      meaning: "Over-deliberating. The answer is clear enough."
      response: "Skip remaining turns. Emit FINAL.md from current state. Done."

    - trigger: "More agents" / "Go wider"
      meaning: "Current perspective set is too narrow"
      response: "Double width on next turn. Add roles not yet represented."

    - trigger: "Focus" / "Narrow down"
      meaning: "Too scattered, too many threads"
      response: "Cut width to 1-2 on next turn. Pick the highest-tension thread only."

Anti-Patterns — Rationalization Table

Layer Discipline — No Cross-Level Claims Without Admission

Deliberation failures in multi-level systems almost always come from layer conflation: treating a correlation-layer label as if it were primary-math equivalence. The fix is forcing every claim to carry a LAYER tag.

layer_discipline:
  rule: "Every structural claim carries a layer tag. Cross-layer equivalence claims require admission evidence at each intermediate layer."

  example_layer_stack:
    # From Josh/QIT framework — generalize the pattern to any multi-level system
    L0_surface: "The constraint surface itself (F01 + N01, or equivalent axioms)"
    L1_chart: "The candidate mathematical chart on the surface (operators, carriers, Weyl spinors, Hopf tori)"
    L2_axes: "Derived axes atop the chart (Axis 0..Axis 6 in QIT; phase parameters elsewhere)"
    L3_correlations: "Correlation overlays used for human recognition (MBTI, polyvagal, Jung, I-Ching)"

  admission_rule: |
    To claim "X at layer L3 ≡ Y at layer L1", you need admission evidence at:
      - L3 → L2 (correlation admitted to axis layer)
      - L2 → L1 (axis admitted to chart)
    Without that, the claim is a MAPPING CANDIDATE, not an equivalence.

  language_enforcement:
    banned_when_cross_layer: ["is", "equals", "=", "≡", "maps to", "creates", "drives"]
    required_when_cross_layer: ["candidate under probe", "admissible with", "survives coupling with", "co-varies under", "not-yet-killed by"]

  graveyard_discipline: "Claims that fail cross-layer admission go to the graveyard WITH the layer that killed them. Graveyard is the scientific output, not a failure log."

  build_order_rule: |
    In any stacked system, layers must be admitted bottom-up before top-layer
    claims are evaluated. "Build everything BEFORE the axes" (Josh 2026-04-17)
    generalizes to: no top-layer equivalence claims until bottom layers are
    admissible. Check: is the layer stack I'm claiming across mostly at L0-L2
    while the target framework has only done L0? Then claims are premature.

  anti_pattern: "CDO synthesis produced 'SNS/PSNS ≡ Left/Right Weyl chirality' — SNS/PSNS is L3 (polyvagal correlation), Weyl chirality is L1 (chart math). Two layers jumped, zero admission. Partner NACK'd within one message. See .lev/pm/decisions/20260417-cdo-manufacturing-consent-failure-mode.yaml"

External Validator Before Broadcast

Any CDO output that will be transmitted to a human or external system passes one more gate: the ground-truth surface must recognize it as its own framing.

external_validator:
  purpose: "Catch internal-convergence-but-external-incoherence BEFORE broadcast"

  when_to_run:
    - "Any message to a human partner"
    - "Any handoff that names the partner's framework, concepts, or code"
    - "Any external-facing artifact (reports, decisions, PRs, issues)"

  ground_truth_sources:
    - "Partner's latest text / messages (last 20-50 messages)"
    - "Partner's own code / schemas (treat pydantic classes, type enums, and field names as LEVEL INDICATORS)"
    - "Partner's promoted docs (not drafts, not archives)"
    - "User's prior feedback / corrections"
    - "Repo current state (not 6-month-old cached mental model)"

  pass_criterion: "Would the ground-truth author recognize every structural claim as their own, at the correct layer, in their current vocabulary?"

  on_fail:
    - "Retract the claim (do NOT soften into hedged version and ship anyway)"
    - "File the failure in graveyard with the layer that killed it"
    - "Redispatch synthesis with explicit pointer to the ground-truth surface that killed the claim"

  anti_pattern: "Treating partner's prior messages as training context instead of live validator surface. Partner's code is the oracle; don't ship without consulting it."

Multi-Wave Discipline

For high-stakes or cross-level deliberations, single-turn CDO is insufficient. Use multi-wave mode: N agents × up to 10 waves, with each wave rotating in NEW skills and angles.

multi_wave:
  when_required:
    - "Cross-layer claims (see Layer Discipline)"
    - "External-facing synthesis that will be broadcast"
    - "Problems where the first synthesis shows >70% agreement (groupthink smell)"
    - "Problems where T1 verdicts are MIXED (don't launder into punchy claim)"

  shape: "5 agents minimum per wave. Up to 10 waves. Default 3 waves minimum for cross-layer."

  per_wave_delta:
    rule: "Each wave MUST introduce 2-3 skills not used in prior wave, OR a new adversarial angle."
    examples_of_new_angles:
      - wave_2: "shift from 'is this convergence real?' to 'what layer would kill this claim?'"
      - wave_3: "shift from internal reasoning to ground-truth validator (partner's code/text)"
      - wave_4: "shift from defending the claim to steelmanning the retraction"
      - wave_5: "shift from structural to historical (has a similar claim been made and killed before?)"

  axiom_finder_micro_pass:
    rule: "Every wave runs through an axiom-finder 7-step compression before dispatch"
    source: "workshop/poc/skills/domains/axioms/axioms.md (Josh's axiom-finder chain)"
    steps:
      1_paraphrase: "Rephrase the turn's question in 3-5 ways. Different phrasings reveal different latent axioms."
      2_steelman: "Build the strongest opposing framing before defending current."
      3_dig_axioms: "What does this claim rest on? Score each presumption against evidence."
      4_map_elements: "Tag every element with its layer (see Layer Discipline)."
      5_multi_devils_debate: "3+ perspectives, not 1. Devil's advocate attacks foundation, not edge cases."
      6_synthesize_with_provenance: "Quote T1 verdicts verbatim. Preserve uncertainty language."
      7_reflect: "Did this wave poison the next wave? Did we over-anchor on one angle? Adjust."

  convergence_not_agreement:
    rule: "Convergence = tensions resolved WITH evidence + layer tags + external-validator pass. NOT mere agreement."
    forbidden: "Declaring convergence from high agreement without external-validator pass."

  wave_exit_gate:
    require_all: true
    criteria:
      - "All cross-layer claims carry layer tags AND admission evidence"
      - "Language discipline enforced (no construction language on cross-layer claims)"
      - "External validator pass logged"
      - "Graveyard non-empty if any claims were retracted"
      - "Minority reports preserved verbatim"

User Signal Extension — The NACK Signal

additional_signals:
  - trigger: "NO NO NO" / "That is wrong" / "You conflated X and Y"
    meaning: "Ground-truth surface detected a layer conflation or premise error in your output. Often arrives after broadcast because the external validator didn't run."
    response: |
      STOP. File postmortem immediately (see Layer Discipline + External Validator).
      Root-cause which failure mode produced the conflation:
        - shared premise in brief?
        - synthesis over-promoted MIXED to EQUIVALENT?
        - no external validator before broadcast?
        - layer tags missing?
      Do NOT re-attempt the claim in softened form. Retract, identify the layer,
      rebuild from admission discipline upward.

Fractal Context Loop (CONTEXT → PLAN → ACT → VERIFY)

Every layer of CDO execution follows the same loop. Context gathering comes FIRST — you cannot plan what you don't understand. This is non-negotiable.

fractal_loop:
  description: "The same cycle applies at every level of the CDO"
  levels:
    session:
      context: "Turn 0 — gather evidence: read code, search cass, web research, acquire skills"
      plan: "Gauge problem, select preset, compose DAG from what you learned"
      act: "Execute turns per synthesis directives"
      verify: "Final synthesis confirms convergence with evidence"

    turn:
      context: "Calibrate context depth per agent, discover relevant skills, read actual files"
      plan: "Read synthesis directive, compose agent briefs with gathered context"
      act: "Dispatch agents in parallel"
      verify: "Synthesis checks for evidence, not just agreement"

    agent:
      context: "Scan codebase/docs/web per calibrated depth BEFORE forming opinions"
      plan: "Agent identifies approach based on what it actually read"
      act: "Agent produces artifact with citations"
      verify: "Agent self-checks: did I cite evidence or reason abstractly?"

Skill Acquisition (Part of Context Phase)

Before planning any CDO, discover which skills are relevant. This is context gathering, not planning.

skill_acquisition:
  purpose: "Surface the right skills from 800-1000 available before composing agent briefs"
  methods:
    1_decompose_to_tags:
      action: "Break the problem into keyword tags"
      example: "protocol design → [protocol, interop, agent, registry, manifest, capability, discovery, mesh]"

    2_search_skills_db:
      action: "rg the tags across ~/.agents/skills/ and ~/.agents/skills-db/"
      command: "rg -l '<tag>' ~/.agents/skills/ ~/.agents/skills-db/ --type md | head -20"
      fallback: "If rg misses, try ~/.agents/lev-skills.sh or skill-discovery skill"

    3_check_thinking_patterns:
      action: "Browse ~/.agents/skills-db/thinking/patterns/ for relevant mental models"
      example: "protocol design → mechanism-design, protocol-networks, nash-equilibrium, requisite-variety"

    4_inject_into_briefs:
      action: "Each CDO agent gets 1-3 relevant skills injected as context in their system prompt"
      rule: "Skills are context, not instructions. The agent reads the skill to understand patterns, not to follow a script."

  anti_pattern: "Dispatching CDO agents without checking what skills exist is like coding without reading the docs."

Turn 0 — Mandatory Research Phase

Before Turn 1 dispatches any deliberation agents, execute a research phase:

turn_0:
  purpose: "Establish ground truth before opinions form"
  actions:
    - "Read actual source code relevant to the problem"
    - "Check cass for prior session evidence"
    - "Web search for prior art if the problem involves external standards"
    - "Generate a context manifest: files read, facts established, unknowns identified"
  output: "tmp/cdo-{session}/t0-research.md"
  rule: "Turn 1 agents receive t0-research.md as part of their brief"

Context Depth Calibration

Not every agent needs 112k tokens. Calibrate per problem shape:

context_depth:
  trivial:
    signal: "1 file, known path, clear answer"
    tools: "Read the file directly"
    cost: "~100 tokens of context"

  focused:
    signal: "2-5 files, clear scope"
    tools: "grep + glob, maybe 1 explore agent"
    cost: "~2k tokens of context"

  broad:
    signal: "Unknown scope, multiple modules"
    tools: "rp-cli context_builder OR multi-agent explore"
    cost: "~20k-50k tokens of context"

  deep:
    signal: "Protocol design, architecture review, cross-cutting"
    tools: "rp-cli + web research + cass history + prior art scan"
    cost: "~100k+ tokens of context"

Convergence Requirements

Convergence now requires EVIDENCE, not just agreement:

convergence_check:
  required:
    - "All blocking tensions resolved"
    - "At least 1 agent cited specific code/docs (not abstract reasoning)"
    - "Anti-groupthink check passed (>70% agreement triggers devil's advocate)"

  skeptical_convergence:
    rule: "Before declaring convergence, ask: what did we ASSUME without evidence?"
    action: "If any assumption is load-bearing and unverified, add a depth probe turn"
    budget: "Use all budgeted turns. Early convergence is a smell, not a feature."

Delegation with Budget Inheritance (from Hermes, Wave 5 — hm-05)

When dispatching child agents in the team modifier mode, apply these constraints borrowed from Hermes delegation semantics:

child_agent_contract:
  iteration_budget: <int>        # max turns child can run autonomously
  tool_inheritance: <bool>       # inherits parent's available tools by default
  skip_memory: <bool>            # optionally isolate child from shared memory
  return_on: [done, budget_exceeded, blocked]

Why this matters: Standard CDO dispatch doesn't cap per-child work. Hermes's model gives each child agent its own iteration budget, inherits tools from parent, and can optionally skip shared memory to enforce isolation. Parent sets constraints; child executes autonomously within them.

Practical application: use iteration_budget: 3 for quick sub-tasks in a team preset CDO. Use skip_memory: true when you need clean deliberation without contaminating session state from one child spilling into another.

Source: .lev/pm/parity/hermes.yaml

Per-Entity Expert Agents

When the problem has distinct entities (e.g., multiple runners, multiple modules, multiple protocols), spawn a specialist agent per entity:

per_entity_experts:
  trigger: "Problem involves 3+ distinct entities that each have their own docs/code"
  action: "Spawn 1 expert agent per entity who reads that entity's actual documentation"
  example: "Protocol design → Claude Expert, Codex Expert, Gemini Expert (each reads real docs)"
  why: "Generic reasoning misses entity-specific behavior. ChatGPT beat CDO on this."

Meta-Agent: "What Would Deep Research Find?"

Add this agent to any deep or full CDO:

meta_agent:
  role: "Deep Research Proxy"
  prompt: "If a single model with 100k+ tokens of real source code were analyzing this problem, what would it find that our multi-agent survey missed?"
  when: "deep or full preset, Turn 2+"
  why: "Multi-agent CDO excels at breadth and adversarial testing. Single-model deep research excels at ground truth. This agent forces CDO to consider what it's missing."

Implementation Spike Agents Early

Don't wait until Turn 2+ for real probes:

early_spikes:
  rule: "Turn 1 should include at least 1 agent that actually RUNS something"
  examples:
    - "Invoke the binary with --help and parse the output"
    - "Read the actual source file and report what it exports"
    - "Run a test and report what passes/fails"
  why: "Abstract reasoning about code you haven't read produces hallucinated architectures (see: lev-builder fabricated paths)"

Negotiate Phase (Claim vs Reality)

From the ChatGPT vs CDO comparison: CDO agents reason abstractly about what code CAN do. ChatGPT reads the code and knows what it ACTUALLY does. The negotiate phase closes that gap.

negotiate:
  trigger: "Any turn where agents make architectural claims"
  process:
    1: "Agent makes a claim: 'ExecTransport handles this'"
    2: "Negotiate agent reads the actual code (grep, read file)"
    3: "Reports fidelity: exact (code does this) | approximate (partial) | rejected (code doesn't do this)"
  output: "Annotated claims with fidelity scores"

  fidelity_levels:
    exact: "Code confirms the claim. Agent cited the right file and function."
    approximate: "Code partially supports the claim. Missing pieces identified."
    rejected: "Code contradicts the claim. Agent was reasoning abstractly."

  rule: "For deep+ presets, add a negotiate agent to Turn 2+ that reads actual code to verify Turn 1 claims. This is the single biggest quality improvement from the ChatGPT comparison."

  anti_pattern: "Accepting architectural claims without code evidence is the #1 cause of CDO producing designs that don't fit the codebase."

Structured Debate Mode

When a CDO has two clear opposing positions (not just multiple perspectives), use debate format instead of independent analysis:

debate:
  trigger: "Binary architectural question — e.g., 'should X live in poly or domain?'"
  structure:
    round_1:
      - "Advocate A argues FOR position 1 (with code evidence)"
      - "Advocate B argues FOR position 2 (with code evidence)"
    round_2:
      - "Advocate A rebuts B's strongest point"
      - "Advocate B rebuts A's strongest point"
    round_3:
      - "Synthesis agent reads all 4 artifacts, picks winner with reasoning"

  rules:
    - "Each advocate MUST cite actual files/functions, not abstract principles"
    - "Rebuttals must address the other side's evidence, not restate their own"
    - "Synthesis must explain why the losing position was wrong, not just why the winner was right"

  when_to_use: "Module placement, protocol ownership, 'should this exist?', binary trade-offs"
  when_NOT_to_use: "Open-ended design (use standard CDO width), research tasks, debugging"

Depth vs Breadth Gauge

Add this to the parse_args step:

depth_breadth_gauge:
  question: "Does this problem need per-entity depth, or architectural breadth?"
  per_entity_depth:
    signal: "Multiple implementations of the same concept (runners, adapters, protocols)"
    action: "Spawn per-entity experts. Reduce breadth agents. Add research phase."
  architectural_breadth:
    signal: "Cross-cutting concern, module placement, trade-off analysis"
    action: "Standard CDO breadth. Multiple perspectives. No per-entity drill."
  both:
    signal: "Protocol design, system unification"
    action: "Research phase (depth) + CDO turns (breadth). Budget 5 turns minimum."