nicsuzor/strategic-review

/strategic-review — Unified Multi-Agent Review

Review any artifact — a document, plan, proposal, or pull request — from several expert perspectives and return one reconciled verdict. Owned by james (reconciliation), but you, the invoking agent, own the orchestration: you deploy the reviewers yourself, because a subagent cannot spawn its own subagents. James is called only at the end, to reconcile.

Inputs

• The artifact: a file path, a PKB id, pasted text, or a pull request (an owner/repo#N ref or URL).
• Action flags (optional): comment, fix, or both. With no flag the review is advisory — you return the verdict to the caller and change nothing.

Procedure (default mode)

1. Gather context

Identify and load the artifact. If it is a PR, load the diff, description, and any prior unresolved review comments (gh pr view, gh pr diff). Select the matching review-context descriptor from review-contexts/ and pass it to every reviewer:

• review-contexts/pr-code.md — code PRs.
• review-contexts/pr-framework.md — framework PRs (skills, agents, hooks, specs).

2. Deploy the three reviewers — in parallel

Spawn all three in a single message (concurrent Agent calls). Give each the artifact, the context, and the descriptor. You spawn them — not james. Subagents cannot spawn subagents, so this fan-out must happen here, at the top level.

• rbg — axiom & rule compliance. Always runs.
• pauli — strategic critique: the premise test, then the "is this in the right place, or a workaround for a root cause that belongs elsewhere?" architectural-fit lens (the 10 cognitive moves). Always runs. Discipline and the worked specimen live in [[references/premise-test.md]].
• marsha — runtime / verification QA. Runs on any outputs to check quality meets our standards.

3. Reconcile via @james

When all three return, dispatch @james with the original artifact plus all three reviewer outputs. James does not re-spawn anyone — it reconciles only. It carries the contradictions, resolves them honestly, and returns one verdict with a synthesis table:

| Agent | Issue | Feedback | Severity |

Merge convergent findings. Since rbg, pauli, and marsha review blind to each other, two or more can independently flag the same underlying defect from their own lens — that's agreement, not contradiction. Collapse it into a single synthesis-table row naming every concurring reviewer and their distinct rationale; never emit one row per reviewer restating the same defect. Reserve separate rows for genuinely distinct issues, and hold genuine disagreements in tension as before.

Severity ladder: REJECT (fundamental — close/redesign) · REVISE (substantial rework, in scope) · FIX (clear correct resolution exists) · TRIVIAL (cosmetic) · ADVISORY (non-blocking follow-up). Overall verdict: APPROVE / REVISE / REJECT.

Agent(subagent_type="aops-core:james",
      prompt="Reconcile these three reviews into one verdict. Do NOT spawn subagents — synthesise only. [artifact + rbg/pauli/marsha outputs]")

4. Act on the verdict — only if asked

• No flag (default): return james's verdict and table to the caller. Change nothing.
• comment: post james's synthesis to the artifact's natural review surface — a PR comment for a PR, an inline note or PKB entry for a document. Scrub all personal info (names, private paths).
• fix: apply every FIX- and TRIVIAL-grade finding directly, without returning to the author. If a fix is substantial, re-run the affected reviewer(s) and fold any new findings into the table. REVISE/REJECT findings are reported, not silently reworked.
• Both: comment and fix.

Whatever the flags, never silently exit: if a write-back action fails, report it and print the full verdict to chat.

Epistemic humility — absence of evidence is not a negative result

Missing evidence licenses "not supported by the available evidence" — never "false / failed / did not happen." A negative verdict requires a held falsifier the reviewer actually holds; where ground truth is unobservable (intent, off-record events, room dynamics), downgrade to ADVISORY (needs primary-source confirmation). Silence in a record is not failure. Apply the same discount symmetrically to flattering and unflattering claims alike. (Source incident: #1891.)