age

Age — Staff Engineer Code Review

Amplify the human reviewer's attention. Surface evidence, unknowns, and tradeoffs. Do not render verdicts.

Compatibility Check

Before dispatching agents, verify tilth_* MCP tools are available to plugin sub-agents. If unavailable, fail fast:

ERROR: tilth MCP tools are not exposed to sub-agents.
Upgrade to Claude Code >= 2.1.30 / claude-agent-sdk >= 0.2.63.
See: anthropics/claude-code#13605

Do not attempt graceful degradation. The dim agents require tilth_read and tilth_search; without them every agent silently misses evidence.

Always-Fire Rationale (D-32)

All 9 dims fire on every invocation. Gating by file-type heuristics risks silent misses. Each agent self-noops with scope_match: false + empty observations when its rubric does not apply. Empty dims are tallied ("ran 9; N had findings"); only non-empty dims render as report sections.

Cost: ~35-45K Haiku tokens per /age. This is intentional and acceptable (see spec D-32, D-21-final).

Arguments

/age [<ref>] [--scope <path>] [--comprehensive]

• ref — git ref or range. Default: $(git merge-base origin/main HEAD)..HEAD
• --scope <path> — restrict evidence fetch to this path prefix
• --comprehensive — pass-through hint to dim agents; they may widen review window

Phase 0 — Classify (no LLM)

Parse argv: extract ref, scope, comprehensive flag.

REF="${1:-$(git merge-base origin/main HEAD)..HEAD}"
SLUG="$(echo "$REF" | tr '/..:' '----' | head -c 32)"
RUN_DIR="${TMPDIR:-/tmp}/cheese-flow-age-$(date +%Y%m%d-%H%M%S)-${SLUG}"
mkdir -p "$RUN_DIR"

Phase 1 — Pre-fetch Evidence (parallel; D-22)

Run all fetches in parallel. Merge into $RUN_DIR/evidence-pack.yaml.

Parallel tasks:

git diff --unified=3 $REF > $RUN_DIR/diff.patch
python python/tools/git_diagnose.py precedent \
  --symbols <touched-symbols> --paths <touched-paths> \
  > $RUN_DIR/precedent.json
python python/tools/git_diagnose.py concurrent-prs \
  --paths <touched-paths> \
  > $RUN_DIR/concurrent-prs.json

For each file touched by the diff:

• cheez-read in outline mode → file structure without inline source
• tilth_deps → import/export graph
• cheez-search --kind callers for each touched symbol

Optional — code-review-graph impact radius:

Try get_impact_radius_tool(touched_symbols). If the MCP tool is unavailable (plugin not loaded, tool not found), write:

{"impact_radius": null}

and continue. Do NOT fail the run. Log a one-line notice.

Merge into evidence pack:

# $RUN_DIR/evidence-pack.yaml
ref: <REF>
outlines:
  <path>: <tilth outline>
deps:
  <path>: <tilth_deps output>
callers:
  <symbol>: [<call sites>]
precedent: <precedent.json contents>
concurrent_prs: <concurrent-prs.json contents>
impact_radius: <get_impact_radius output or null>

No inline source content in the pack. Agents use cheez-read for follow-up source reads after reviewing the outline.

Phase 2 — Dispatch All 9 Dim Agents (parallel; D-32)

Spawn all agents in parallel. Each agent:

• Reads $RUN_DIR/evidence-pack.yaml as primary evidence
• Reads $RUN_DIR/diff.patch for the change context
• Writes $RUN_DIR/<dim>.json per the per-agent return contract (see skills/age/references/sidecar-schema.md)

Agents to dispatch:

• age-correctness
• age-security
• age-complexity
• age-encapsulation
• age-spec
• age-precedent
• age-deslop
• age-assertions
• age-nih

Pass to each agent: RUN_DIR, REF, SCOPE, COMPREHENSIVE flag.

When an agent's dim does not apply to the diff, it emits:

{"dimension": "<dim>", "scope_match": false, "observations": [], "stake": "<dim-stake>", "summary": "Dim does not apply to this diff."}

Phase 3 — Synthesize (orchestrator, deterministic; no LLM)

Collect all $RUN_DIR/<dim>.json files.

observations = union(all dim.json observations)
callouts = group_by_locus(observations, window=3 lines, min_dims=2)

group_by_locus: observations from 2+ dims whose anchor.start line numbers fall within 3 lines of each other become a cross-dim callout.

Render Markdown report → .cheese/age/<slug>.md:

# Age Report — <slug>

## Orientation
<1-2 sentence factual description of what the diff does>

Ran 9 dims. <N> had findings. <9-N> were empty (scope_match: false or no observations).

## High-Stake Dimensions
(correctness, security, encapsulation, spec — non-empty only)

## Medium-Stake Dimensions
(complexity, deslop, assertions, nih — non-empty only)

## Advisory Dimensions
(precedent — non-empty only)

## Cross-Dimension Callouts
(loci where 2+ dims agree; omit section if empty)

See skills/age/references/report-template.md for full layout and narrative format rules.

Split observations into sidecar JSON files:

.cheese/age/<slug>.fixes.json:

• observations with a fix field (hash-anchored, syntactically narrow, complete content)

.cheese/age/<slug>.suggestions.json:

• observations with consideration only (no fix)

Both match the schema in skills/age/references/sidecar-schema.md.

Phase 4 — Hand-off (no auto-invoke)

Print:

Age report: .cheese/age/<slug>.md
Fixes:       .cheese/age/<slug>.fixes.json   (<N> entries)
Suggestions: .cheese/age/<slug>.suggestions.json   (<M> entries)

Next step:
  /cure <slug>                              — apply fixes, route suggestions, re-age

Do NOT auto-invoke /cure. The amplifier-pure boundary forbids it (spec D-14-final). The report is the deliverable; the user types the hand-off.

Phase 5 — Cleanup

rm -rf "$RUN_DIR"

Rules

• File I/O from dim agents via cheez-read / cheez-search (NFR-1). No host Read / Grep / Edit, no direct tilth_edit.
• Hash anchors use tilth line:hash strings natively (NFR-2, D-24).
• No numeric scores anywhere in user-facing output (D-5).
• Confidence rendered as low | med | high bucket only.
• Narrative before bucket in every observation (Greptile severity-at-end).
• No writes to production source files during review (FR-8, NFR-6).
• Dim stake is fixed per dim; do not vary at runtime.