# chow-compound-engineering-v260-review-pipeline-2026-03-31

## Veille

Compound Engineering v2.60, mandatory code review with confidence scoring, hardened plan→work→review pipeline

## Titre Article

Compound Engineering: 3/31/2026

## Date

2026-03-31

## URL

https://x.com/trevin/status/2038893322333507781

## Keywords

Compound Engineering, mandatory code review, confidence scoring, false positives, agentic pipeline, ce:work, ce:plan, ce:brainstorm, ce:review, test discovery, headless mode, interactive deepening, conditional diagrams, mermaid, PR feedback clustering, document-review, track-based learnings, Claude Code skills, SDLC, software lifecycle, agentic SDLC

## Authors

Trevin Chow

## Ton

**Profile**: Enriched changelog/release notes, technical-casual register, detailed level for practitioners.

**Description**: Chow adopts the format of an annotated changelog, each section detailing an updated skill or feature with the "why" context in addition to the "what." The tone is that of a project maintainer speaking directly to his user community: casual ("we know many of you just wanna rip"), self-deprecating ("I may change my mind on this"), and pragmatic. Each design decision is justified with explicit reasoning ("mandate first, then reduce noise, otherwise we'd be mandating noise"). The writing is dense with technical detail but remains accessible through a conversational style. The target audience consists of active Compound Engineering / Claude Code users who want to understand the changes and their rationale.

## Pense-betes

- v2.60.0 of March 31, 2026 — throughline: tighten the pipeline end-to-end (reviews, plans, reduced friction)
- **ce:review — major change**: headless mode for programmatic invocation + code review made MANDATORY across the entire pipeline (ce:work, ce:brainstorm, ce:plan). Full review by default, limited review requires justification
- 6-level confidence rubric (0.00–1.00) with a 0.60 suppression threshold. 6 targeted false-positive categories: pre-existing issues, style nitpicks, intentional patterns, handled elsewhere, code restatement, generic advice
- ~49% reduction in false positives with no loss of real bug detection
- Intent verification: findings are checked against the PR context (title, body, linked issue) — findings that contradict the PR's goal are suppressed
- Multi-persona consensus: if 2+ persona reviewers flag the same issue, confidence boost of +0.10
- Findings in pipe-delimited tables (no more free-form prose) for scannability and consistency
- Philosophy: "mandate first, then reduce noise — otherwise we'd be mandating noise"
- **ce:work**: now accepts raw prompts without a prior plan. Automatic complexity assessment (trivial → skip ceremony, medium → inline tasks, complex → recommends planning)
- Universal test discovery before implementation. Per-task deliberation: "testing addressed" replaces the binary "tests pass"
- 5th testing-reviewer check: detects behavioral changes (new branches, state mutations, API changes) without corresponding tests
- Detection of auto-generated branch names (e.g. "worktree-printing-ruby-raven") + rename suggestion
- **ce:brainstorm**: bug fixed — Phase 1.1 prevented reading technical files, leading to unverified claims ("table X does not exist" without checking the schema). Now: verification of the current state is allowed, unverified claims are labeled as assumptions
- **ce:plan**: interactive deepening mode — accept, reject, or discuss each agent's findings before integration (instead of auto-merge). Support for /ce:plan "deepen" to invoke enrichment directly
- document-review made mandatory after deepening. Plans flag empty test scenarios on functional units
- **Conditional visual aids**: automatic diagram generation (mermaid/ASCII) when complexity exceeds thresholds (5+ non-linear units, 3+ interacting surfaces). Higher threshold for PR descriptions
- **resolve-pr-feedback**: clustering of PR comments by concern category and spatial proximity. After 2 fix-verify cycles, remaining issues become "recurring patterns." Actionability filter (ignores 👍, badges, wrapper text)
- **document-review** simplified: 3 tiers → 2 tiers (auto / present). Next-step suggestion based on document type
- **ce:compound**: track-based schema — bug-track (full diagnostic) vs knowledge-track (lightweight template). Discoverability check: verifies that docs/solutions/ is referenced in AGENTS.md or CLAUDE.md
- Infrastructure: centralized model field normalization (fixed incorrect Qwen→sonnet mapping), MiniMax support, argument-hints cleanup

## RésuméDe400mots

Trevin Chow details the updates to Compound Engineering culminating in v2.60.0 of March 31, 2026. The throughline is the tightening of the end-to-end pipeline: code reviews are less noisy and now mandatory, plans detect more gaps before implementation, and day-to-day usage friction continues to decrease.

The most significant change concerns ce:review. A headless mode enables programmatic invocation by other skills, which unlocked the next step: making code review mandatory across the entire pipeline (ce:work, ce:brainstorm, ce:plan). Full review is the default level, with limited review requiring explicit justification. To avoid "mandating noise," a 6-level confidence rubric (0.00–1.00) with a 0.60 suppression threshold was added simultaneously. Six false-positive categories are targeted: pre-existing issues, style nitpicks, intentional patterns, cases handled elsewhere, code restatement, and generic advice. Result: a 49% reduction in false positives with no loss of real detection. Findings are checked against the PR context, and multi-persona consensus boosts confidence by 0.10.

ce:work now accepts raw prompts without a prior plan, with automatic complexity assessment. Universal test discovery before implementation ensures code/test synchronization. A new check detects behavioral changes without corresponding tests.

ce:brainstorm fixes a subtle bug where Phase 1.1 prevented agents from reading technical files, producing unverified claims such as "this table doesn't exist" without checking the schema. Now, verification of the current state is allowed while implementation decisions remain deferred to planning.

ce:plan gains an interactive deepening mode allowing each agent's findings to be accepted, rejected, or discussed before integration. Document review is made mandatory after enrichment.

A cross-cutting feature automatically generates diagrams (mermaid/ASCII) when complexity exceeds certain thresholds. resolve-pr-feedback now detects "whack-a-mole" by clustering similar comments. document-review moves from 3 to 2 tiers and suggests the next step based on document type. Finally, ce:compound adopts a track-based schema (bug vs knowledge) with a discoverability check.

## GrapheDeConnaissance

- Compound Engineering —publie→ v2.60.0 (EVENEMENT, 0.99)
- revue de code obligatoire —fait_partie_de→ Compound Engineering (METHODOLOGIE, 0.98)
- ce:review —réduit→ faux positifs de 49% (CONCEPT, 0.95)
- ce:review —utilise→ scoring confiance 6 niveaux (METHODOLOGIE, 0.97)
- ce:work —permet→ prompts bruts sans plan (CONCEPT, 0.95)
- ce:work —utilise→ découverte universelle de tests (METHODOLOGIE, 0.93)
- ce:brainstorm —résout→ bug vérification fichiers techniques Phase 1.1 (CONCEPT, 0.92)
- ce:plan —utilise→ mode deepening interactif (METHODOLOGIE, 0.93)
- Trevin Chow —dirige→ Compound Engineering (METHODOLOGIE, 0.95)
- Compound Engineering —permet→ diagrammes mermaid conditionnels (CONCEPT, 0.9)
- resolve-pr-feedback —utilise→ clustering des commentaires PR similaires (METHODOLOGIE, 0.88)
- ce:compound —utilise→ schema track-based bug vs knowledge (METHODOLOGIE, 0.9)

---
Canonical: https://www.thekb.eu/en/fiches/chow-compound-engineering-v260-review-pipeline-2026-03-31/