Temper Core

Temper Core

Stack detection → Quality gates (SUGGEST/WARN/BLOCK) → Confidence scoring (0.0-1.0) → Review memory → Adaptive learning → Metrics.

Stack Detection

1. .claude/temper.config → stack field
2. .claude/presets/*.yaml → stack section
3. Auto-detect: pom.xml→Spring Boot, package.json→Node, pyproject.toml→Python, go.mod→Go, Cargo.toml→Rust
4. Load .claude/packs/stacks/{stack}.md

Pack Resolution (v4.3.0)

Three-tier: project-local > global > built-in. Cached in .temper/pack-manifest.json and consumed by all stage commands (build, review, check, plan, design) for phase-filtered loading.

• .claude/packs/{name}/rules.md (project)
• ~/.claude/packs/{name}/rules.md (global)
• $CLAUDE_PLUGIN_ROOT/.claude/packs/{name}/rules.md (built-in) Packs support link: plugin://name | skill://name and phases: [build, review, ...].

Quality Gates

• SUGGEST: Non-blocking
• WARN: Highlighted, developer decides
• BLOCK: Must fix (security/architecture only)

Confidence & Memory

• Threshold: 0.7 (configurable)
• Review memory: .temper/review-memory.json — auto-suppress after 5 dismissals
• Metrics: .temper/metrics.json

Adaptive Learning (v4.6.0)

Post-review intelligence layer that makes reviews smarter over time. Runs as Step 8.5 in /temper:review and displays in /temper:status.

Three capabilities:

Capability	Trigger	Action
Pattern Detection	Post-review hook (Step 8.5)	Cluster recurring findings by category + file pattern + keywords
Rule Suggestions	Pattern accepted 3+ times at 70%+ rate	Generate pack rule template, queue for promotion
Noise Reduction	Pattern dismissed 5+ times	Auto-suppress, reduce false positives

Key files:

• .temper/learning.json — Learning state (detected patterns, suppressions, suggestions, learning curve)
• .temper/learning/suggestions/ — Generated rule suggestion templates
• .claude/packs/adaptive-learning/rules.md — Promoted rules (user-accepted suggestions)

Graceful degradation: If learning.json is absent, all commands work unchanged. No errors, no warnings.

Full docs: $CLAUDE_PLUGIN_ROOT/.claude-plugin/reference/learning.md

Capabilities (v5.0.0)

Four independently optional capabilities, all enabled by default. Controlled via capabilities section in temper.config.

Capability	Stage	Purpose
Architecture Depth	Review	Module-depth analysis: seams, adapters, locality, leverage, deletion test
Grill Me	Plan, Design	Socratic challenge mode — stress-test plans before building
Teach Me	Plan, Design, Build, Check, Eval	Comprehension companion — teach + quiz the human to mastery at each teaching gate (Review excluded — taught at Build)
Config Suggestions	Check	Suggest CLAUDE.md/AGENTS.md updates based on what was built
HTML Review	Plan	Interactive browser-based plan review with inline comments
Evals	Eval, Check	Behavioral verification via LM-judge + trajectory + plan-time evalsets

Graceful degradation: Each capability checks its config flag. Missing config = all enabled (default-on).

Full Docs

$CLAUDE_PLUGIN_ROOT/.claude-plugin/reference/{command}.md