codex-orchestrator-report

Report

Use this skill only after orchestration work has finished. Do not start agents, continue tasks, or write an interim report here. Return to ${CLAUDE_PLUGIN_ROOT}/skills/orchestrate/SKILL.md or ${CLAUDE_PLUGIN_ROOT}/skills/workflow/SKILL.md when more work is needed.

Preconditions

Confirm that:

• run_closed.validation contains the complete descriptive validation result;
• every execution has a terminal execution result and every task is terminal;
• run_closed is the final journal entry and contains judgment: passed|blocked;
• no further run work is planned.

If these conditions are not true, stop and return to orchestration. Never edit the journal merely to make the report look complete.

Read the complete journal.jsonl, then ground each claim in its appropriate source:

• actual delivery: final repository state relative to the run_started Git baseline;
• Claude's checks: verification observations and referenced evidence;
• assigned scope: exact prompts;
• agent claims: exact handoffs;
• lifecycle and chronology: journal entries;
• ambiguous process activity: raw events or direct observation;
• decisions and final judgment: decision and run_closed entries.

The journal is working memory, not independent evidence. Use raw events only for ambiguity or debugging. Surface conflicts and missing facts; do not silently repair them.

Required Report Structure

Create the final report.md once, after closure, as a complete Claude-authored report using exactly these five top-level sections in this order. Only replace an existing final report when explicitly asked to correct or regenerate it.

# Report

## Summary

## Changes

## Orchestration Graph

## Consensus

## Final Results

Do not add other ## sections.

Summary

State the original run_started.goal, overall result, run_closed.judgment, and the main reason for that outcome. Mention unresolved work plainly and keep this section concise.

Changes

Compare the final repository state with run_started.repo_head and run_started.repo_status, then connect material delivered changes to their tasks or agents. Do not attribute initially dirty paths without supporting evidence. Cite important repository files and evidence when useful. Distinguish complete, blocked, failed, and intentionally unchanged work. Do not reproduce the journal line by line or treat execution_result.files_changed as mechanical attribution.

Orchestration Graph

Create a readable Mermaid flowchart TD from journal chronology, grounded in handoffs, verification evidence, repository changes, and raw events only when needed. Use this top-to-bottom shape:

• A_CLAUDE{{"Claude Code<br/>planner · orchestrator"}};
• separate non-empty Claude-agent and Codex-agent subgraphs;
• material reviews, checks, decisions, and produced deliverables;
• final judgment and state, including meaningful fix/recheck loops.

Use one node per named agent, ordered by first execution. Include its task, recorded model/effort, main result, and terminal status; combine that agent's resumed executions in its node. Keep a linear run minimal and do not create empty subgraphs. Label edges concisely (assign, review, verified, resolution, consensus, claude_decision, produced, fix required, recheck, accepted). Show only evidence that affected a decision or final judgment. Combine routine passing checks and group related decisions, but keep Claude decisions, user actions, and unresolved outcomes distinct. Mark reconstructed facts as inferred; never infer verification results, decision outcomes, judgments, or terminal status.

Consensus

Summarize consequential decision entries with outcome, basis, resolution, and risk; state when no decision was required. Keep consensus, Claude decisions, user actions, accepted risks, and unresolved outcomes distinct. Put missing basis or residual-risk treatment under Risks / Follow-ups.

Final Results

Use these subsections in this order:

### Gate Result

### Risks / Follow-ups

Under Gate Result, report run_closed.judgment, its summary, and the recorded validation issues and warnings. Validation is an omission check, not evidence of correctness or the source of the judgment. This is a human-facing heading, not a separate journal entry. Do not recompute or soften Claude's recorded judgment.

Under Risks / Follow-ups, list unresolved checks, blocked or failed work, user actions, accepted risks, and concrete next steps. Write None recorded. when nothing remains.

At the end of Final Results, include available Claude Code and Codex CLI versions from run_started in one compact Run metadata bullet. Omit unavailable values. Do not add protocol/schema versions or a Reproducibility section.

Final Check

Reread the finished report. Confirm that the five required ## sections appear once and in order, the Mermaid block is readable, every material claim is grounded, and Final Results faithfully reflects run_closed.judgment, validation output, risks, and follow-ups. Reconcile every numeric total against the journal entries or handoffs it summarizes.