multi-agent-paper-orchestrator

name	multi-agent-paper-orchestrator
description	Use when coordinating multiple specialized CFD-AI/SciML paper agents: reviewer, evidence auditor, experiment planner, figure editor, prose editor, and final gatekeeper.

Multi-Agent Paper Orchestrator

Trigger

Use when the user wants a CFD-AI/SciML manuscript improved by multiple roles instead of one monolithic writer:

• reviewer criticizes and finds rejection risks;
• evidence auditor checks claim support and missing facts;
• experiment planner proposes rescue validation;
• figure editor turns claims into figure/table evidence;
• prose editor rewrites the section;
• gatekeeper re-audits the edited draft;
• orchestrator repeats prose_editor -> gatekeeper in a bounded loop when residual blockers are fixable by wording or structure.

This skill can be executed in three modes:

1. Simulated multi-agent mode: one model runs the roles sequentially with strict role separation.
2. Codex native subagent mode: one Codex CLI session spawns project-scoped custom agents from .codex/agents/ and manages them with /agent.
3. External worker mode: the orchestrator launches separate Hermes/Codex/Claude processes and verifies/integrates their outputs.

Use simulated mode for short abstracts. Use Codex native subagent mode when the user wants one Codex CLI session with internal parallel agents. Use external worker mode for durable long runs, separate worktrees, or non-Codex orchestration.

Progressive disclosure

Read in this order:

1. references/multi-agent-paper-roles.md
2. skills/paper-revision-loop-manager/SKILL.md
3. skills/reviewer-audit-toolkit/SKILL.md
4. skills/scientific-journal-writing/SKILL.md
5. skills/sciml-experiment-auditor/SKILL.md
6. skills/figure-and-result-storytelling/SKILL.md
7. rubrics/multi-agent-paper-workflow-rubric.md
8. templates/multi-agent-paper-workflow-report.md

Add topic-specific gold-paper notes only after the manuscript domain is known. Do not give every worker every file by default; source scope is part of the experiment.

Agent roles

Role	Job	Must not do
Orchestrator	define source scope, assign roles, merge outputs, verify final answer	write final prose before receiving audits
Reviewer	identify Fatal/Major risks and likely objections	polish prose
Evidence Auditor	map claims to supplied evidence and TODOs	invent missing solver/data/metric details
Experiment Planner	propose minimum validation, baselines, diagnostics, ablations	claim experiments were run
Figure/Table Editor	design evidence-bearing figures/tables	accept decorative contours as proof
Prose Editor	rewrite into field-native manuscript text	hide missing evidence in smooth language
Gatekeeper	re-audit final draft and list residual blockers	rubber-stamp the editor

Workflow

1. Source-scope contract

Before spawning or simulating roles, write:

• supplied text and evidence;
• files each role may read;
• facts that must remain TODO;
• output expected from each role.

2. Parallel critique pass

Independent roles inspect the same draft from different angles:

• Reviewer: rejection risks;
• Evidence Auditor: claim-evidence map;
• Experiment Planner: rescue experiments and baselines;
• Figure/Table Editor: evidence-bearing visual/table plan.

3. Integration pass

Orchestrator merges disagreements into a single priority ledger:

1. Fatal blockers;
2. Major blockers;
3. wording edits that reduce risk;
4. experiments/tables/figures needed to support stronger claims.

4. Editorial pass

Prose Editor rewrites only after the integration ledger exists. The rewrite must:

• preserve supplied evidence;
• downgrade unsupported claims;
• add TODO placeholders for missing facts;
• follow physical problem -> gap -> method -> evidence -> limitation;
• avoid generic AI academic language.

5. Bounded edit-review loop

After the first editorial pass, Gatekeeper reviews the revised text. If residual blockers are fixable by wording, structure, claim calibration, or TODO placement, the orchestrator runs another edit-review cycle:

prose_editor(cycle N revision + residual-risk table)
-> gatekeeper(original + ledger + previous revision + current revision)

Default limits:

• abstracts: maximum 2 edit-review cycles;
• longer sections/full drafts: maximum 3 edit-review cycles;
• never loop indefinitely.

Stop early when:

• no Fatal/Major blockers remain;
• no new overclaim was introduced by the editor;
• remaining blockers require new experiments/data/figures/tables rather than prose.

If the blocker requires evidence, the editor must downgrade the claim or insert a TODO. Do not polish around missing evidence.

6. Final gate

Gatekeeper compares old vs new text:

• risks removed;
• risks downgraded;
• risks remaining;
• new overclaims introduced by the editor;
• next actions required before submission.

Codex native subagent mode

For users who want one Codex CLI session to spawn several internal agents, configure project-scoped agents under .codex/agents/ and start Codex from the package root. This package includes agent definitions for:

• managing_editor
• cfd_reviewer
• evidence_auditor
• experiment_planner
• figure_table_editor
• prose_editor
• gatekeeper

Prompt pattern:

Use native Codex subagents. Spawn cfd_reviewer, evidence_auditor, experiment_planner, and figure_table_editor in parallel on this draft. Wait for all. Merge their outputs. Then run prose_editor -> gatekeeper -> prose_editor -> gatekeeper as a bounded edit-review loop. Stop when no Fatal/Major blockers remain, remaining blockers need new evidence, or the max cycle count is reached.

Use /agent inside Codex CLI to inspect or switch between active agent threads.

Actual multi-agent launch guidance

When using external Hermes/Codex/Claude workers rather than native Codex subagents, the orchestrator should:

• assign each worker a self-contained prompt with source scope and role boundaries;
• keep workers read-only unless they are explicitly writing a named artifact;
• require structured outputs, not free-form chat;
• verify worker claims by reading generated files and checking source scope;
• integrate outputs centrally instead of letting workers edit the same manuscript concurrently.

Suitable worker prompts:

• Reviewer: “Return Fatal/Major/Minor risks only; no rewrite.”
• Evidence Auditor: “Map each claim to supplied evidence or TODO; do not infer hidden details.”
• Experiment Planner: “Return minimum validation matrix and baselines that rescue claims.”
• Figure Editor: “Return figure/table sequence where each item supports one claim.”
• Prose Editor: “Rewrite only after reading the merged ledger; do not invent evidence.”
• Gatekeeper: “Audit the revised text against the original ledger and list residual blockers.”

Output schema

Use templates/multi-agent-paper-workflow-report.md for full workflows.

Compressed form:

Agent	Output	Key blocker/fix
Reviewer
Evidence Auditor
Experiment Planner
Figure/Table Editor
Prose Editor
Gatekeeper

Then provide:

1. merged blocker ledger;
2. revised text;
3. edit-review cycle log;
4. residual-risk table;
5. stop reason;
6. next experiments/tables/figures.

Anti-patterns

• Calling it multi-agent when one role simply writes everything.
• Letting the prose editor see no reviewer/evidence ledger.
• Allowing workers to invent missing experiments or paper facts.
• Giving every worker the answer key when the benchmark is supposed to test inference.
• Merging worker outputs without resolving contradictions.
• Treating worker self-reports as verified evidence.
• Parallel workers editing the same file without an orchestrator merge step.

Verification

• Source scope is stated for every role.
• Each role returns a distinct artifact.
• The final rewrite traces to reviewer/evidence findings.
• Gatekeeper identifies residual risk rather than approving by default.
• Edit-review loops are bounded and include an explicit stop reason.
• The editor does not run another prose cycle when the remaining blocker requires new evidence rather than wording.
• No hidden solver, metric, citation, data, or runtime detail is invented.
• The workflow can be scored by rubrics/multi-agent-paper-workflow-rubric.md.