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.