/skills/shared-references/review-tracing.md

Review Tracing Protocol

Purpose

Save full prompt/response pairs for every cross-model reviewer call, enabling:

• Reviewer-independence audit: verify the executor only passed file paths, not summaries
• Reproducibility: threadId preservation allows conversation continuation
• Meta-optimize input: richer data for harness improvement analysis

When to Trace

After every mcp__codex__codex or mcp__codex__codex-reply call that serves a reviewer/critique function. This includes review scoring, experiment auditing, claim verification, idea critique, and patch gating.

Do NOT trace: purely informational LLM calls (e.g., codex exec for code generation that is not a review).

Trace Directory

.aris/traces/<skill-name>/<YYYY-MM-DD>_run<NN>/
  ├── run.meta.json                      # Run-level metadata
  ├── 001-<purpose>.request.json         # Request snapshot
  ├── 001-<purpose>.response.md          # Full response text
  ├── 001-<purpose>.meta.json            # Response metadata
  ├── 002-<purpose>.request.json         # Second call (e.g., reply)
  └── ...

• <skill-name>: the ARIS skill that triggered this call (e.g., auto-review-loop)
• <YYYY-MM-DD>_run<NN>: date + sequential run number (start from 01)
• <purpose>: short kebab-case label (e.g., round-1-review, critique, ideation, audit, patch-gate)

How to Trace

After each reviewer MCP call — including every FAILED attempt in a capability-fallback chain (one trace entry per attempt: --status error + --fallback-reason; the successful entry records the RESOLVED pair) — save the trace using save_trace.sh, resolved through the canonical helper chain (see integration-contract.md §2 — failure policy C, "forensic helper"). The full invocation:

# Resolve $TRACE_HELPER (canonical strict-safe chain; see integration-contract.md §2).
cd "$(git rev-parse --show-toplevel 2>/dev/null || pwd)" || exit 1
if [ -z "${ARIS_REPO:-}" ] && [ -f .aris/installed-skills.txt ]; then
    ARIS_REPO=$(awk -F'\t' '$1=="repo_root"{print $2; exit}' .aris/installed-skills.txt 2>/dev/null) || true
fi
TRACE_HELPER=".aris/tools/save_trace.sh"
[ -f "$TRACE_HELPER" ] || TRACE_HELPER="tools/save_trace.sh"
[ -f "$TRACE_HELPER" ] || { [ -n "${ARIS_REPO:-}" ] && TRACE_HELPER="$ARIS_REPO/tools/save_trace.sh"; }
[ -f "$TRACE_HELPER" ] || TRACE_HELPER=""

if [ -n "$TRACE_HELPER" ]; then
  bash "$TRACE_HELPER" \
    --skill "<skill-name>" \
    --purpose "<purpose>" \
    --model "<model that actually ran — the RESOLVED pair, not the target>" \
    --effort "<effort that actually ran>" \
    --fallback-reason "<why the capability chain stepped down; empty when it didn't>" \
    --status "<ok | fallback_used | error>" \
    --thread-id "<threadId from response>" \
    --prompt "<full prompt as sent>" \
    --response "<full response content>"
else
  # Required fallback: the resolver exhausted all three layers and
  # save_trace.sh is unreachable, but trace artifacts are still
  # required (unless `--- trace: off` was explicitly set on this
  # SKILL invocation). Write the four files below directly per the
  # schemas in "File Schemas", into:
  #   .aris/traces/<skill-name>/<YYYY-MM-DD>_run<NN>/
  #     run.meta.json
  #     <NNN>-<purpose>.request.json
  #     <NNN>-<purpose>.response.md
  #     <NNN>-<purpose>.meta.json
  # Do NOT silently skip — trace_path is load-bearing for any
  # mandatory audit emitting `trace_path` in its artifact (see
  # assurance-contract.md §"Required Audit Artifact Schema").
  echo "WARN: save_trace.sh not resolved; writing trace files directly per review-tracing.md schema." >&2
fi

The helper, when present, handles directory creation, run numbering, and file writing. The fallback branch above documents what to do when the helper is unreachable — the trace is forensic evidence, so "helper missing" never means "skip the trace."

File Schemas

run.meta.json

{
  "skill": "auto-review-loop",
  "run_id": "2026-04-15_run01",
  "started_at": "2026-04-15T14:30:00+08:00",
  "executor": "claude-code",
  "project_dir": "/path/to/project"
}

NNN-<purpose>.request.json

{
  "call_number": 1,
  "purpose": "round-1-review",
  "timestamp": "2026-04-15T14:31:00+08:00",
  "tool": "mcp__codex__codex",
  "model": "gpt-5.6-sol",
  "config": {"model_reasoning_effort": "xhigh"},
  "files_referenced": ["paper/sections/3_method.tex", "results/table1.csv"],
  "prompt": "<full prompt text>"
}

NNN-<purpose>.response.md

The reviewer's full response, verbatim. No truncation, no summarization.

NNN-<purpose>.meta.json

{
  "call_number": 1,
  "purpose": "round-1-review",
  "timestamp": "2026-04-15T14:33:00+08:00",
  "thread_id": "019d8fe0-b25d-...",
  "model": "gpt-5.6-sol",
  "duration_ms": 142000,
  "status": "ok"
}

Configuration

Tracing respects three modes, set via inline parameter --- trace: off | meta | full:

• full (default): save full prompt + full response
• meta: save metadata only (no prompt/response text), useful for sensitive projects
• off: disable tracing entirely

Integration with events.jsonl

After writing a trace, append a compact summary event to .aris/meta/events.jsonl:

{"event":"review_trace","skill":"auto-review-loop","purpose":"round-1-review","thread_id":"...","trace_path":".aris/traces/auto-review-loop/2026-04-15_run01/","status":"ok"}

This allows /meta-optimize to discover traces without reading the full trace files.

Debugging With Traces

Traces are not only audit evidence — they are the first place to look when a verdict is surprising: a score regresses round-to-round, two reviewer backends disagree, or /result-to-claim contradicts an earlier claim. Before re-invoking the reviewer for "a better answer", read the raw transcript and find the moment its judgment actually changed:

# Diff the raw response bodies across the two calls in question
skill=auto-review-loop run=2026-04-15_run01
diff ".aris/traces/$skill/$run/002-round-2.response.md" \
     ".aris/traces/$skill/$run/003-round-3.response.md"

# Grep for the sentence where the assessment turned
grep -En 'however|but|concern|missing|cannot' \
     ".aris/traces/$skill/$run/003-round-3.response.md"

The paragraph where the assessment changed is the causal explanation for the divergence — cite it, don't guess. Re-running the reviewer without reading the trace is tuning by vibe: you get a new opinion, not an explanation.

This is the same muscle ARIS already applies to code failures (the "Read the error — parse traceback, stderr, and log files" step in /experiment-bridge's auto-debug sequence, and /codex:rescue reading tracebacks before a retry) — applied to saved AI-judgment transcripts instead of stderr. The trace is written in English and most of it is the reviewer talking to itself; the discipline is identical: read the primary artifact first, then act on the exact divergence point rather than re-rolling the dice.

Practical triggers:

Surprise	Trace move
Score dropped after a "fix" round	diff the two rounds' .response.md; find which criterion flipped
Two backends disagree (codex vs gemini/manual)	grep both responses for the SAME artifact path; compare what each actually read
Reviewer "forgot" an earlier concern	grep prior rounds for the concern keyword; if present-then-absent, cite it in the next prompt instead of restating from memory
Verdict contradicts a deterministic checker	read the request .md — was the checker's output actually in the files the reviewer was pointed at?

Privacy

• .aris/traces/ should be in .gitignore — traces are project-local, never committed
• Traces may contain sensitive research content; treat them as confidential
• Use --- trace: off for projects with strict confidentiality requirements