strategic-review

Strategic Review

Coordinate modular multi-agent dialectical review under james as Supervisor across three distinct review components and synthesize findings into one reconciled verdict (PASS, REVISE, or FAIL).

Inputs

• Artifact: File path, PKB note ID, inline text, design proposal, specification, or PR reference (owner/repo#N).
• Flags (optional): comment (post findings to review surface), fix (apply minor fixes directly). Without flags, review is advisory.

Review Process

1. Supervisor Calibration: Proportionality, Altitude & Governing Standard (james)

Before dispatching reviewers, the Supervisor (james) loads the artifact, its originating request, the format/quality standard governing the artifact's genre (specs/meta/doc-taxonomy.md -- e.g., specs/meta/writing-specs.md for specifications, craft for agent instructions, disciplinary standards for manuscripts/articles, or task acceptance criteria for task handbacks), and whatever specifications, requirements, or acceptance criteria are available, determines destination repository visibility (public or private), and explicitly calibrates:

• Proportionality (depth, rounds, model selection, and primary-material scope): Weigh the importance, risk, blast radius, and reversibility of the task or artifact at the start:
  • High-stakes / high-risk / architectural work: Allocate deeper refinement, multiple rounds of dialectical cross-examination, smarter, higher-capability models (Model: "pro"), and instruct marsha to audit primary materials.
  • Low-stakes / routine / reversible checks: Keep the review to a fast, single-pass sanity check with lighter models (Model: "flash" or "flash_lite") and without primary-material expeditions to confirm the change is sound without wasting time and money.
• Altitude alignment & artifact-matched standard: Define the target abstraction focus for the review and ensure reviewers apply the standard matching the artifact's genre (never applying task-handback sections like ## Ask to non-task artifacts).

2. Deploy the Three Review Components in Parallel

Dispatch all three specialist reviewers concurrently in a single turn using the calibrated Model selection. Every reviewer's brief is drawn from the shared review briefing (quoting the originating request, governing genre standard, and available specifications/requirements/criteria before the artifact or diff) and returns a verdict in the unified PASS / REVISE / FAIL vocabulary:

• Strategic Direction & Congruence (TypeName pauli): Evaluates the core premise, strategic direction, and congruence with the "seamless web" of existing material in the PKB -- questioning the premise, naming the class of problem, probing for flawed reasoning, untested load-bearing assumptions, negative space, and missing edge cases, and situating the artifact against parent objectives, PKB graph topology, prior art, existing specifications, topic notes, active sibling tasks, parsimony, and lifecycle coherence (PASS, REVISE, or FAIL). Provide Pauli's brief inline in the dispatch payload (parent objective, claimed target nodes, problem definition, and summary of the proposed change) so Pauli operates strictly on the PKB graph without reading repository files or filesystem templates.
• Compliance (TypeName rbg): Evaluates compliance and scope of authority against universal axioms, project-local rules, ultra vires boundaries, data boundaries, and the enforcement pyramid (specs/enforcement/enforcement.md). Confined strictly to artifact text/diff and governing rule/spec sources, returning PASS, REVISE, or FAIL.
• Quality & Requirements Satisfaction (TypeName marsha): Evaluates whether the work actually delivers on its requirements and governing genre standard to an ambitious quality standard ("amazingly good", not merely working) across actual rendered multimodal outputs (never code alone or simulated renders) and runtime execution (PASS, REVISE, or FAIL), and audits primary materials on high-risk requests -- and if no governing standard, specifications, requirements, or acceptance criteria are available, treats that absence itself as an immediate critical failure signal (FAIL).

Every reviewer also judges proportion within its lens: whether the artifact builds more machinery than the request needs, and whether each requested change is worth its cost given the failure's likelihood and consequence. Reviewers immediately self-fix defects that can be resolved without controversy (language updates, data scrubbing under data-boundaries, simple deterministic repairs), commit them with fix(review):, and document them under ### Reviewer Self-Fixes Applied. Uncontroversial self-fixes do not trigger REVISE; issues lacking a reasonably deterministic solution stay itemized findings.

3. Dialectical Deliberation & Critical Synthesis (Proportional to Stakes)

• Fast path (low-stakes sanity check): When proportionality is set to single-pass and reviewers surface no fatal blockers or cross-component contradictions, proceed directly to reconciliation.
• Dialectical rounds (high-stakes or conflicting findings): Read reviewer outputs critically rather than rubber-stamping them. Verify reviewer warrants: a documentation note (README.md, knowledge note) stating a behaviour proves only what the document says, not current working state (require a completed task record with verification evidence or live state, or else record it as an open question); a cited PR or code reference is authoritative only if it is the most recent merged change, reflects current working behaviour, and is backed by an authoritative specification or decision record. When reviewers surface competing tensions (e.g. compliance constraint vs strategic parsimony, or existing spec congruence vs implementation approach), james cross-examines the relevant reviewers with each other's critiques (send_message), demanding explicit defense, concession, or synthesis while holding the tension honestly. Intervene immediately if any reviewer drifts off the calibrated altitude.

4. Reconcile Findings

Synthesize component outputs into a unified findings table and compile itemized self-fixes:

Reviewer	Issue	Feedback	Verdict / Disposition

• Include a calibration header noting the chosen proportionality (model selection and rounds) and target altitude.
• Collapse concordant findings across reviewers into a single row that keeps each reviewer's own verdict/disposition; where they differ, state the disposition adopted and why.
• Give the strategic congruence review (pauli) its own row (PASS, REVISE, or FAIL), even when concordant with other rows.
• Itemize all self-fixes applied across reviewers under ### Reviewer Self-Fixes Applied. If all identified defects were uncontroversial and self-fixed cleanly, do not issue REVISE solely for those items.
• Overall verdict: PASS, REVISE, or FAIL. Any expert reviewer's (pauli, rbg, or marsha) FAIL (including the absence of any specifications, requirements, or acceptance criteria) makes the combined verdict FAIL. The consolidating agent must never downgrade, overrule, or soften an expert reviewer's FAIL (e.g. to REVISE or PASS). If an expert reviewer is judged too harsh, that is the user's responsibility to fix by reviewing that reviewer agent, not the consolidating agent's job to soften. Any REVISE without a FAIL forces REVISE unless all pass (PASS).

5. Action and Reporting

• Reviewer self-fixes: Reviewers apply uncontroversial fixes directly to the working tree / branch, commit immediately (fix(review): <description>), and report them under ### Reviewer Self-Fixes Applied. Issues lacking a reasonably deterministic solution remain as findings.
• Default: Return the calibration summary, reviewer self-fixes applied, synthesis table, and overall verdict (PASS, REVISE, or FAIL) to the caller.
• comment: Run the data-boundaries check against destination visibility before posting; then post reconciled feedback to the PR review or task note.
• fix: Run the data-boundaries check against destination visibility before pushing; apply straightforward mechanical remedies immediately.
• Write the final verdict and evidence to the artifact's task record.