Claude-native telemetry: waste & efficiency detectors

Problem

Counting tokens shows where the fire is, not why it started. A retrospective needs actionable, evidence-linked findings: specific patterns that wasted context or signalled inefficient collaboration, each with a remediation.

Capability

A registry of rule-based detectors over normalized turns that emit findings {detector, severity, evidence, estimatedWasteTokens?, dimension, remediation}. Each detector is a small pure function with configurable thresholds.

Initial detectors

Giant tool output; repeated file reads; tool/Bash overuse; thrash; error storm; retry storm; one-shot failure; skill over-firing; model substitution; cost outlier.

Behaviour

• Interface Detector.run(stats, baseline?) => Finding[]; no I/O inside detectors.
• Every finding cites sessionId + uuid (or toolUseId) — never a vague claim.
• Configurable thresholds; conservative defaults; documented as uncalibrated on small samples.

Acceptance criteria

• Each detector has a triggering and a non-triggering fixture (true/false-positive control).
• Every finding carries a resolvable evidence reference.
• Config changes (disable / threshold) change output deterministically.

Dimensions: Collaboration Efficiency, Quality & Maintainability, Delivery Predictability. Tier 3. Depends on #36, #37, #38. PRD: docs/prd/features/04-waste-efficiency-detection.md.

Read docs/prd/features/04-waste-efficiency-detection.md and review the contracts from dependencies #36, #37, and #38 before tracing normalized turns. Implement the Detector.run interface and initial detector registry with configurable thresholds, then use triggering and non-triggering fixtures to verify deterministic output and resolvable evidence references.