Agent Group Discuss

Agent Group Discuss

This imports the OpenClaw multi-agent discussion and research-review templates into Codex and adapts them to Codex agent tools.

When to use

• group discussion
• panel of agents
• multiple agents with different roles
• multi-agent review
• review-only requests that explicitly ask for multiple agents or a panel
• multi-agent research
• named review templates such as Lakatos, Knuth, Pólya, Graph Reconfiguration Specialist, or Lean Formalization Team

If a request is review-only and does not ask for annotation, stay in this skill for the multi-agent path. Do not route that request to annotated-review unless the user also explicitly asks for annotation.

Supporting files

Read these when relevant:

• TEMPLATES.md for the imported research and review templates
• EXECUTION.md for the Codex execution pattern and per-template round topology
• references/external-cli-agents.md for parent-owned external CLI participant probes, capability profiles, artifact layout, and failure handling
• MODEL_TIERS.md for the live Codex model-tier catalog
• MODEL_TIERS.example.md only as a customization template
• README.md for the structured request shape

If the user requests a named template, or if the task clearly matches one, open TEMPLATES.md and EXECUTION.md before spawning any agents.

Long multi-agent jobs and remote notify

For multi-round / multi-hour panels, the parent should notify via remote-bridge when secrets are configured (same policy as ARL --notify auto):

• on panel start and terminal synthesis
• optionally on each round boundary when wall time is high

Prefer structured remote-bridge send --channel zulip|telegram|both over raw shell notify. Do not block the panel if notify fails. Short single-round spawns need not notify.

Clarification policy

If the request is underspecified, ask only for the minimum needed.

Use this compact question when needed:

Before I start: do you want discussion, review, or research? How many rounds? Any hard constraints? I can choose roles and models automatically if you want.

If the user gives no preference, default to:

• mode: infer from task
• rounds: 2
• roles: 3
• interaction: auto
• role/model selection: automatic

For writing-producing discussions, reviews, final reports, or research syntheses, every role prompt and final synthesis must include a writing-style block referencing writing-style-settings.md. Add math-manuscript-style.md when the task involves mathematical manuscripts, TCS, graph theory, formal proofs, or LaTeX prose. Final artifacts must record style_profile_ref, policy_hash, active_overlays, active_requirement_ids, and style_applied. A bare style_applied: true assertion is not sufficient without the policy-load and requirement-selection evidence recorded by the workflow.

Routing to structured workflows

If the user explicitly wants a more structured, reproducible workflow, prefer prose instead.

If the user says:

• structured workflow
• use prose
• compile the workflow
• deterministic

offer prose. Otherwise stay in this skill.

Confirmation gate

Before any actual multi-agent run:

1. produce the plan
2. show the plan to the user
3. get explicit confirmation
4. only then call spawn_agent

Do not infer consent from silence. If the user already gave an explicit affirmative in the same request after seeing the plan, that counts.

Scope boundary

For discussion, review, investigation, diagnosis, audit, or report-only requests, the orchestration ends with the requested discussion summary, review findings, or report. Do not continue into remediation, file edits, artifact patches, retrieval beyond what is necessary for the review, commits, or cleanup unless the user explicitly asks for those actions.

Codex tool mapping

OpenClaw concepts must be translated to Codex tools as follows:

• OpenClaw sessions_spawn -> Codex spawn_agent
• OpenClaw session follow-up -> Codex send_input
• OpenClaw session wait -> Codex wait_agent
• OpenClaw resume session -> Codex resume_agent
• OpenClaw terminate/cleanup -> Codex close_agent
• OpenClaw local execution -> Codex functions.exec_command
• OpenClaw proof or graph computation helpers -> Codex sagemath, graph-verifier, formal-skeleton-helper, or local functions.exec_command

Participant execution model

• The main agent is the orchestrator.
• A template role is a logical responsibility. A participant is the executable actor assigned to that role.
• Initial participant kinds are codex_spawned and external_cli.
• Use spawn_agent for codex_spawned participants.
• Use references/external-cli-agents.md for external_cli participants.
• Use multi_tool_use.parallel when launching independent codex_spawned participants in the same round.
• Keep role prompts concrete and bounded.
• Run independent roles in parallel.
• Use send_input for later rounds when keeping a codex_spawned participant alive is helpful.
• Use wait_agent once per round or per critical batch, not in a tight polling loop.
• Use close_agent after the run finishes or when a codex_spawned participant is no longer useful.
• Prefer local synthesis unless a dedicated synthesizer role materially improves the result.

For any actual run, follow EXECUTION.md as the detailed orchestration reference.

Codex-specific constraints

Codex cannot hard-restrict each spawned agent's tools the way OpenClaw can. Compensate by doing all of the following in the role prompt:

• scope the role narrowly
• forbid nested sub-agent spawning unless explicitly needed
• forbid file edits unless the role owns a specific write target
• ask for the exact output format needed for the current round
• pass only the minimum prior context needed
• for research-related roles, include the research skill-routing block from EXECUTION.md, narrowed to the role's allowed tools and side-effect limits

For reasoning-only roles, prefer agent_type: "default". Use agent_type: "explorer" only for tightly scoped codebase fact-finding roles. Use fork_context: true only when the role genuinely needs the current thread context; otherwise prefer a compact summarized handoff.

High-level goal

Given a task, you must:

1. classify the task
2. decide which roles are needed
3. decide how many participants are useful
4. assign a model and reasoning effort to each role
5. choose how the participants interact
6. run the requested number of rounds
7. synthesize the result
8. maintain durable recovery state on disk

Input format

Accept either free-form text or a structured block such as:

topic: <topic>
mode: discussion | review | research | mixed
rounds: <integer>
max_agents: <integer>
interaction: auto | star | debate | panel_judge
template: <optional template name>
output: <desired final output format>
constraints:
- <constraint>
- <constraint>

If the user specifies rounds, obey it.

Role selection

Prefer 3 roles by default. Use 4 or more only when the task clearly benefits.

Typical roles by mode:

Discussion

• Optimist
• Skeptic
• Pragmatist
• Judge

Review

• Correctness reviewer
• Edge-case reviewer
• Clarity reviewer
• Synthesizer

Research

• Literature scout
• Hypothesis generator
• Critic / falsifier
• Synthesizer

Model assignment

Each codex_spawned participant gets a model and reasoning_effort through spawn_agent. Refer to MODEL_TIERS.md for the runtime freshness check and role-to-model routing policy, and EXECUTION.md for per-template execution defaults. The active host runtime/tool definitions are the source of truth for available models; update stale checked-in defaults before launching a run from this repo. External CLI participants use probed capability profiles and adapter guidance from references/external-cli-agents.md; do not route them through MODEL_TIERS.md.

Reasoning level classification

Profile selection

If the user specifies a profile, obey it. If the user specifies a model for a specific role, override that role only. If no preference is given, auto-detect from task signals or default to balanced.

Research-task model rule

If the task is a research task, proof verification, mathematical correctness review, manuscript correctness review, or other high-stakes reasoning task:

• default to math-heavy
• use the latest available model with the highest available thinking or reasoning level for every participant
• apply the same latest-model/highest-thinking rule to any permitted nested manager-worker child agents

This rule overrides lower-cost defaults for research-mode runs.

Tier assignment by role type

Record the chosen profile, per-role model, and reasoning effort in state.json.

Token management

Do not promise truncation cannot happen. Keep prompts compact and summarize prior rounds before relaying them.

If a participant response is truncated:

1. note the truncation in state.json
2. re-prompt the same role with compressed context
3. if needed, switch to a model with more headroom from the same or a stronger tier
4. never silently discard a truncated response

Interaction design

Choose:

• star
• debate
• panel_judge

If the user requests one explicitly, obey it.

Round control

If rounds is provided, obey it. Otherwise:

• default to 2 rounds for discussion or review
• default to 2 or 3 rounds for research depending on complexity

Never exceed 5 rounds unless the user explicitly asks.

Timeouts

Ordinary discussion or lightweight review: 10 minutes per round by default. Complex correctness review or research verification: 45 minutes per agent by default, with persistent progress checkpoints every 15 minutes.

If a codex_spawned participant does not respond in time:

1. mark the participant as timed out in responses_received and participant state
2. add a note to pending_work
3. continue the round with available responses
4. do not block the whole run on one unresponsive role

Total run timeout: 30 minutes by default for ordinary runs. For complex correctness reviews, extend the total run budget as needed and keep writing progress checkpoints. If exceeded, write the best available synthesis from completed rounds and mark the run incomplete.

Durable state

Create a run folder under:

• ${AAS_RUNS_ROOT:-$HOME/.local/share/ai-agents-skills/runs}/agent_group_discuss/<run_id>/

Before launching any participant:

1. create the run directory
2. create a lock file with the current timestamp
3. write plan.md
4. write state.json
5. show the plan to the user
6. get explicit confirmation before executing

After each round, write:

• round_01.md
• round_02.md
• round_P1_01.md and similar names for chained phases when needed

After synthesis, write:

• final.md

For long correctness reviews, also write:

• progress_15m.md
• progress_30m.md
• later progress files at the same cadence as needed
• final_report.md

On success, failure, or pause, remove or update the lock state accordingly.

Plan output format

The plan file and the user-facing plan summary must include:

## Run Plan: <run_id>

**Topic:** <topic>
**Mode:** <mode>
**Profile:** <profile>
**Interaction:** <pattern>
**Rounds:** <N>
**Estimated total time:** <X-Y minutes>
**Estimated agent calls:** <parallel batches + follow-up batches + synthesis passes>

### Participant assignments

| # | Participant | Kind | Role | Model/Profile | Output contract | Failure policy | Est. time |
|---|-------------|------|------|---------------|-----------------|----------------|-----------|
| 1 | judge-1 | codex_spawned | Judge | <resolved current frontier model>, R4, xhigh | structured finding list | fail_closed | 2-4 min |
| 2 | reviewer-1 | codex_spawned | Correctness Reviewer | gpt-5.3-codex, R3, high | issue list | partial_allowed | 1-2 min |
| 3 | cli-checker-1 | external_cli | Edge-case Reviewer | profile:copilot-json-2026-05-21 | parseable envelope | fail_closed | 1-2 min |

### Execution plan

- Round 1: roles 1-3 run in parallel
- Round 2: rebuttal or synthesis pass

### Risk notes
- <token budget concerns>
- <roles that may need stronger models>

Time estimation:

• estimate per role from the chosen profile in MODEL_TIERS.md
• parallel roles use the slowest role in the batch
• add sequential synthesis time
• multiply by the number of rounds
• add 1-2 minutes for orchestration overhead

State schema

Use a state.json structure like:

{
  "run_id": "string",
  "topic": "string",
  "mode": "discussion | review | research",
  "profile": "math-heavy | premium | balanced | budget",
  "interaction": "star | debate | panel_judge",
  "template": "string or null",
  "roles": ["string"],
  "participants": {
    "participant_id": {
      "kind": "codex_spawned | external_cli",
      "role": "logical role name",
      "status": "planned | ready | running | succeeded | failed | timed_out | invalid | skipped",
      "required": true,
      "rounds": [1],
      "input_policy": {
        "context_scope": "string",
        "allowed_inputs": ["string"],
        "side_effect_policy": "none | parent_resolved_only"
      },
      "output_contract": {
        "required_fields": ["string"],
        "parser": "string",
        "artifact_paths": ["string"]
      },
      "evidence_policy": {
        "required_evidence": "none | citations | source_refs | artifact_refs",
        "blocked_check_reporting": true
      },
      "timeout": {
        "minutes": 45,
        "retry_count": 0,
        "idempotence": "no_side_effects"
      },
      "failure_policy": "fail_closed | partial_allowed | best_effort",
      "validation": {
        "owner": "parent_orchestrator",
        "status": "unvalidated | accepted | accepted_with_limitations | rejected",
        "notes": []
      },
      "artifacts": {
        "prompt": "artifact ref",
        "raw_output": "artifact ref",
        "parsed_output": "artifact ref",
        "stderr_or_log": "artifact ref",
        "validation_report": "artifact ref"
      },
      "runtime_ref": "agent-id or external profile ref"
    }
  },
  "models": {
    "participant_id": {
      "model": "<resolved runtime model for codex_spawned only>",
      "reasoning_level": "R4",
      "reasoning_effort": "xhigh"
    }
  },
  "rounds_requested": 2,
  "current_round": 0,
  "status": "planning | running | paused | completed | failed",
  "responses_received": {
    "participant_id": true
  },
  "pending_work": ["string"],
  "start_time": "ISO 8601 timestamp",
  "agent_timeout_minutes": 45,
  "estimated_duration_minutes": 10,
  "progress_reports_written": ["progress_15m.md"],
  "recovery_needed": false,
  "notes": ["string"]
}

Participant launch policy

The main agent is the orchestrator. Prefer leaf agents spawned directly by the main agent unless the selected template explicitly enables a bounded manager-worker layer.

Use the role prompt template and per-template execution plans from EXECUTION.md instead of improvising prompt structure ad hoc.

Each participant task must include:

• the role
• the topic
• the role objective
• the expected response format
• the current round number
• only the minimum prior context needed

For codex_spawned participants, launch with spawn_agent and record the returned agent ID in runtime_ref.

For external_cli participants, run only after the parent has recorded a current capability profile and artifact policy according to references/external-cli-agents.md. External CLI output is untrusted until the parent parses it and updates the participant validation status.

External CLI artifacts are parent-owned. Keep raw output, probe logs, command shape records, provider observations, validation files, redaction notes, and recovery state in the AGD run directory. When an AGD result feeds a deep research workflow, add an explicit evidence mapping from participant artifact refs into the research evidence ledger; do not let participant prose directly promote a final claim.

Nested delegation is allowed only when the user-facing plan lists the manager, child cap, provider, resolved model, thinking level, and same-model constraint. Child workers must use the manager's same provider, resolved model, and thinking level, and must not spawn further agents.

For opening statements, ask for:

• short position
• strongest argument
• one uncertainty or caveat

For later rounds, ask each role to:

• respond to the strongest counterpoint
• refine or defend its position
• provide one concession or one rebuttal

Recovery behavior

If a result is missing or the run is disrupted:

1. read state.json
2. inspect which round files exist
3. identify missing required participants from responses_received and participant status
4. set recovery_needed: true
5. if old codex_spawned role agents still exist, use resume_agent or send_input
6. otherwise respawn or rerun only the missing participants with compressed context and a fresh artifact record
7. never discard already completed rounds unless the user asks

If an optional participant fails repeatedly, skip it and note the gap explicitly. If a required participant fails, times out, produces unparseable output, or violates its evidence contract, mark it invalid or failed and mark the run incomplete unless the user explicitly approves a weaker synthesis.

If the user pauses the run:

1. mark status: "paused"
2. note the reason in notes
3. do not spawn new agents
4. wait for in-flight agents to finish if practical
5. tell the user what remains

To resume:

1. read state.json
2. verify the run was paused or failed
3. re-open surviving agents if possible
4. continue from the last consistent checkpoint

Research templates

The imported templates live in TEMPLATES.md. Detailed execution plans live in EXECUTION.md.

The user can request a template by name, or the orchestrator can auto-select one. If auto-selecting, briefly state which template was chosen and why before proceeding. The user can override.

Template auto-selection:

If multiple templates match, prefer the more domain-specific one.

Template chaining is allowed when a task spans multiple concerns. When chaining:

1. run one template per phase
2. pass only surviving claims and strongest proof skeletons forward
3. keep per-phase round files
4. state the planned chain before starting

Mandatory rules for template runs

Before any template begins:

1. produce the plan output
2. restate the target claim in exact terms
3. list assumptions explicitly
4. separate what is given, to be proved, and conjectured
5. identify notation and definitions
6. obtain explicit user confirmation before launching any participants

Show this Step 0 restatement to the user before spawning agents for a high-stakes research template.

If a decisive counterexample or fatal gap is found:

1. stop defending the broken claim
2. switch to diagnosis
3. determine the strongest defensible corrected claim
4. do not keep expanding a broken proof across later rounds

External verification in Codex

When a template calls for computational or formal checking, prefer:

• sagemath for heavy graph-theoretic or algebraic computation
• graph-verifier for lightweight graph checks
• formal-skeleton-helper for Lean-style scaffolding
• functions.exec_command for local scripts or test harnesses

Use the role-vs-orchestrator computation split from EXECUTION.md so verification stays independent when the template requires it.

If no external verification is possible, say so explicitly in the final ledger.

Final output

Return a polished result with:

• Topic
• Mode
• Template used
• Roles used
• Models assigned
• Interaction pattern
• Rounds completed
• Main agreements
• Main disagreements
• Best points by role
• Final synthesis
• Recommended next step

Also include a compact run summary:

• run id
• profile used
• number of agents and their models
• reasoning efforts used
• rounds completed
• actual duration vs estimated duration
• whether recovery was needed
• whether any responses were truncated or models were swapped
• which progress checkpoint files were written

Recommended templates

When this skill is involved, consider these workflow templates (install via the workflow-templates artifact profile, or --with-deps to pull backing skills):

• autonomous-research-loop-runbook -- Bounded autonomous research-loop runbook with four stop conditions, single-path solving, mandatory cross-agent verification, fresh-agent backtracking, and five-lane broker-routed heavy-compute offload with per-lane safety gates.
• cross-agent-adversarial-review -- Producer-never-confirmer adversarial review of a paper, proof, or code artifact across agent families with a fresh-agent confirmation gate.
• engineering-delivery-loop-runbook -- Bounded build-and-deliver loop runbook: single-path implementation with seen-to-fail proof, cross-agent diff verification, behavior-preserving cleanup, and five-lane broker-routed heavy-compute offload with per-lane safety gates.