agent-discussion

Agent Discussion

Use this skill when Codex needs structured external feedback for planning, architecture review, code discussion, or decision support before making a final call.

Current Path

For read-only external feedback, the current path is:

New-AgentDiscussion.ps1
  -> Invoke-ClaudeFeedback.ps1 -Round <1-or-2> -Collect
  -> Invoke-ReasonixFeedback.ps1 -Round <1-or-2> -Collect when Reasonix is a reviewer
  -> Collect-AgentDiscussion.ps1
  -> Codex writes codex-synthesis.md, decision.md, or user-decision-needed.md

If Claude returns a malformed fragment, a file list, or feedback rejected by the format gate, stay in the same discussion folder and use:

Invoke-ClaudeFeedback.ps1 -Round <1-or-2> -RawReadOnlyRecovery -Collect

Do not switch to an unrelated chat or a hand-written direct claude.exe command for this workflow.

Rules

• Codex remains the moderator, evidence evaluator, and final decision maker.
• Discussion mode is read-oriented and file-based.
• Discussion mode does not edit repository files.
• Discussion creation does not launch external agents automatically; use Invoke-ClaudeFeedback.ps1 only when the user wants Claude Code feedback now.
• Discussion mode does not create worktrees, merge branches, apply patches, or accept code changes.
• External agents write Markdown feedback files only.
• Reasonix CLI is the automatic review path. Open-ReasonixDiscussion.ps1 remains manual compatibility only and never counts as completion by itself.
• Codex writes synthesis, user escalation, and final decision files.
• The default limit is two rounds.
• Pick -Mode for the task type so the generated brief.md carries the right output structure. Use article-review for article analysis, strategy-review for strategy/planning, code-review for code review, decision for option selection, and general otherwise.
• Use -Protocol adversarial-discussion when Codex needs debate, synthesis, and a second-round response. Use -Protocol feedback for one-way structured feedback.
• Use -ExpectedSections when the user needs named custom sections. This is preferred over prose format instructions.
• Use -ExpectedOutputFormat only for a complete Markdown template. Prose instructions like Markdown with sections: ... are normalized into a strict template and recorded as custom_normalized.
• Use -ReferencePaths when the external agent needs to inspect article drafts, source files, notes, or bounded folders. Discussion mode copies read-only snapshots into references/files/ and writes references/manifest.md; external agents should read those snapshots rather than original absolute paths.
• Treat ReferencePaths as frozen evidence copied at discussion creation time. Round 2 and -RawReadOnlyRecovery stay on the same frozen snapshots. If the original source files changed and the user wants a re-review of current files, create a fresh discussion with new -ReferencePaths.
• Reference packaging fails on file-count or size truncation, and both CLI reviewers verify manifest inventory, byte counts, hashes, and evidence bundle identity before review.
• The evidence bundle ID is stored in status.json and derived as SHA-256 over ordered snapshot hashes joined by newlines.
• If a major disagreement remains unresolved after two rounds, Codex writes user-decision-needed.md, pauses, and asks the user to decide or redirect.
• Agents are required reviewers. OptionalAgents are attempted and reported but do not block collection. Do not describe a one-reviewer result as a dual audit.
• Use -AuditProfile scientific for consequential code-review and strategy-review work. Round 1 is blind and independent; Round 2 is limited to disputed, blocking, or weak-evidence findings.
• Scientific dual review is complete only when every expected canonical roundN/<agent>.md file exists. Instruction, retry, recovery, and arbitrary Markdown files never count.
• Reviewer-local finding IDs must be unique within each canonical file. Duplicate IDs are rejected rather than silently producing ambiguous calibration records.
• Consensus is not proof. Codex adjudicates evidence and records verified outcomes; it does not majority-vote.