retro

Post-run retrospective analysis. After /research:run completes, reads .experiments/state/<run-id>/experiments.jsonl, computes statistical significance, detects dead iterations, flags suspicious metric jumps, generates learning summary with next-hypothesis queue.

NOT for: running experiments (use /research:run); designing experiments (use /research:plan); validating methodology (use /research:judge); verifying paper implementation (use /research:verify); comparing runs from different programs/goals — --compare valid only for same-program, same-metric runs. Read-only — never modifies code, commits, or experiment state.

</objective> <workflow>

Agent Resolution

Agent resolution: load and follow the protocol below. Contains: foundry check + fallback table. research:scientist in same plugin — no fallback needed if research plugin installed.

export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
# loads: unsupported-flag-protocol.md
python "${CLAUDE_PLUGIN_ROOT:-plugins/cc_research}/bin/resolve_shared.py" --persist research-shared --emit agent-resolution.md --emit unsupported-flag-protocol.md || exit 1  # timeout: 5000 — cold resolve; every later site reads the research-shared sentinel instead of re-running python

Retro Mode (Steps T1–T7)

Triggered by retro, retro <run-id>, or retro <run-id> --compare <run-id-2>.

Defaults: --threshold 0.001, --alpha 0.05.

Unsupported flag check: follow the protocol loaded in the Agent Resolution block. Supported flags for this skill: --compare, --threshold, --alpha.

Task tracking: create tasks for T1–T7 at start, all in the same response as the first real tool call. Every later TaskUpdate rides with the next real tool call — never a bookkeeping-only turn; the one standalone call is a completed right before a long report.

Step T1: Locate and load run data

Input resolution (priority order):

Explicit <run-id> arg → read .experiments/state/<run-id>/
No arg → scan .experiments/state/, pick latest dir where state.json has status: completed or status: goal-achieved

None found → stop with error:

No completed run found. Run /research:run first, or provide: /research:retro <run-id>

Newer-in-progress check (only when path 2 used — no explicit run-id given): after selecting completed run, scan .experiments/state/ for any dir with status: running and mtime newer than selected dir. If found, surface warning but don't stop — user may intentionally retro prior completed run:

⚠ Newer in-progress run found: <newer-run-id> (status: running, started <ISO timestamp>). Retro will analyse <selected-run-id> instead. Use /research:retro <run-id> to override.

Load files from .experiments/state/<run-id>/:

state.json: extract goal, best_metric, config (incl. metric.direction), iteration count, best_commit. Compute baseline_metric from iteration 0 in experiments.jsonl.
experiments.jsonl: full iteration history — validate each line parses as JSON. If last line truncated, the T1 setup block below warns and drops it from the sanitized copy $RUN_DIR/experiments-clean.jsonl — produced mechanically once $RUN_DIR exists (never via the Write tool, never under .experiments/state/), path persisted to retro-jsonl-path-${CSID}. All downstream steps (T2 retro_analyze.py, T3 dead-iter scan, T5 scientist) must read sanitized copy — never raw file — so every step sees same iteration set. If JSONL untruncated, sanitized copy byte-identical to raw file.
diary.md: if present, read for qualitative context in T5.

If --compare <run-id-2> present: load second run identically from .experiments/state/<run-id-2>/ — including its own sanitized copy $RUN_DIR/experiments-clean-2.jsonl (same truncation rule, produced by the T1 setup block, path in retro-jsonl2-path-${CSID}); the compare step reads that copy, never the raw file. If not found, stop: "Compare target not found: .experiments/state/<run-id-2>/. Check run ID and retry."

T1 setup — one block: assign RUN_ID_ARG from $ARGUMENTS (first positional non-flag token, empty if absent), resolve $RUN_ID (input resolution above), pre-compute and persist $RUN_DIR for T3:

export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
# retro has no boolean flag; --flags must be non-empty, so a placeholder name. --keep-state: no eval; --inlined: retro keeps no compaction contract, so it clears none
python "${CLAUDE_PLUGIN_ROOT:-plugins/cc_research}/bin/parse-skill-flags.py" --flags retro --value-flags compare,threshold,alpha --keep-state retro --inlined "$ARGUMENTS" >/dev/null  # timeout: 5000
IFS= read -r CLEAN_ARGS < "${TMPDIR:-/tmp}/retro-clean-args-${CSID}" 2>/dev/null || CLEAN_ARGS=""
IFS= read -r VALUE_ALPHA < "${TMPDIR:-/tmp}/retro-alpha-${CSID}" 2>/dev/null || VALUE_ALPHA=""
IFS= read -r VALUE_COMPARE < "${TMPDIR:-/tmp}/retro-compare-${CSID}" 2>/dev/null || VALUE_COMPARE=""
RUN_ID_ARG=$(echo "$CLEAN_ARGS" | awk '{for (i=1; i<=NF; i++) if ($i !~ /^--/) { print $i; exit }}')
RUN_ID_ARG="${RUN_ID_ARG:-}"
echo "$RUN_ID_ARG" > "${TMPDIR:-/tmp}/retro-run-id-${CSID}"  # persist for T3 (vars lost between Bash calls)
echo "${VALUE_ALPHA:-0.05}" > "${TMPDIR:-/tmp}/retro-alpha-${CSID}"
RUN_ID="${RUN_ID_ARG:-$(python "${CLAUDE_PLUGIN_ROOT:-plugins/cc_research}/bin/find_run_id.py" .experiments/state 2>/dev/null)}"  # loads: find_run_id.py
# find_run_id.py stderr suppressed; empty case surfaced to avoid double-slash path
[ -z "$RUN_ID" ] && { echo "! Failed to resolve run ID — no completed run found or bin/find_run_id.py unavailable; check research plugin install."; exit 1; }
echo "$RUN_ID" > "${TMPDIR:-/tmp}/retro-run-id-resolved-${CSID}"
RUN_DIR=$(python "${CLAUDE_PLUGIN_ROOT:-plugins/cc_research}/bin/make_run_dir.py" "retro" ".experiments" 2>/dev/null)  # timeout: 5000
[ -n "$RUN_DIR" ] || { echo "! BLOCKED — retro T1: make_run_dir.py returned empty; check research plugin install"; exit 1; }
mkdir -p "$RUN_DIR/scripts"  # timeout: 3000
echo "$RUN_DIR" > "${TMPDIR:-/tmp}/retro-run-dir-${CSID}"  # T3 + fallback path reload from temp file
RAW_JSONL=".experiments/state/$RUN_ID/experiments.jsonl"
[ -f "$RAW_JSONL" ] || { echo "! BLOCKED — retro T1: $RAW_JSONL not found; run-id '$RUN_ID' has no iteration history"; exit 1; }
# sanitized copy after RUN_DIR exists; truncated last line dropped (sed '$d' — `head -n -1` absent on macOS)
if tail -n 1 "$RAW_JSONL" | python -m json.tool >/dev/null 2>&1; then cp "$RAW_JSONL" "$RUN_DIR/experiments-clean.jsonl"; else echo "⚠ experiments.jsonl last line truncated — dropped from sanitized copy"; sed '$d' "$RAW_JSONL" > "$RUN_DIR/experiments-clean.jsonl"; fi
echo "$RUN_DIR/experiments-clean.jsonl" > "${TMPDIR:-/tmp}/retro-jsonl-path-${CSID}"
echo "$VALUE_COMPARE" > "${TMPDIR:-/tmp}/retro-compare-id-${CSID}"  # empty = no --compare; overwrites a prior retro's value
if [ -n "$VALUE_COMPARE" ]; then
  RAW2=".experiments/state/$VALUE_COMPARE/experiments.jsonl"
  [ -f "$RAW2" ] || { echo "Compare target not found: .experiments/state/$VALUE_COMPARE/. Check run ID and retry."; exit 1; }
  if tail -n 1 "$RAW2" | python -m json.tool >/dev/null 2>&1; then cp "$RAW2" "$RUN_DIR/experiments-clean-2.jsonl"; else echo "⚠ compare run: experiments.jsonl last line truncated — dropped from sanitized copy"; sed '$d' "$RAW2" > "$RUN_DIR/experiments-clean-2.jsonl"; fi
  echo "$RUN_DIR/experiments-clean-2.jsonl" > "${TMPDIR:-/tmp}/retro-jsonl2-path-${CSID}"
fi
echo "RUN_ID=$RUN_ID RUN_DIR=$RUN_DIR"

Step T2: Statistical significance analysis

Run the Wilcoxon signed-rank test via the bundled bin/ script — pure Python with scipy.stats:

export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
IFS= read -r RUN_ID < "${TMPDIR:-/tmp}/retro-run-id-resolved-${CSID}" 2>/dev/null || RUN_ID=""  # re-hydrate RUN_ID from T1 (Check 41: fresh shell)
IFS= read -r ALPHA < "${TMPDIR:-/tmp}/retro-alpha-${CSID}" 2>/dev/null || ALPHA="0.05"
METRIC_DIRECTION=$(python "${CLAUDE_PLUGIN_ROOT:-plugins/cc_research}/bin/read_state_field.py" ".experiments/state/$RUN_ID/state.json" "config.metric.direction" --default "" 2>/dev/null || echo "")  # loads: read_state_field.py
# runs before config persistence wrote `config: {}` — say so; a silent `higher` runs the wrong one-sided test on loss/latency
[ -n "$METRIC_DIRECTION" ] || { echo "⚠ state.json records no config.metric.direction (run predates config persistence) — assuming higher; wrong for a lower-is-better metric"; METRIC_DIRECTION=higher; }
IFS= read -r _RETRO_RUN_DIR < "${TMPDIR:-/tmp}/retro-run-dir-${CSID}" 2>/dev/null || _RETRO_RUN_DIR=""
IFS= read -r RETRO_JSONL < "${TMPDIR:-/tmp}/retro-jsonl-path-${CSID}" 2>/dev/null || RETRO_JSONL="$_RETRO_RUN_DIR/experiments-clean.jsonl"  # re-hydrate sanitized path from T1 (Check 41: fresh shell)
# no raw-jsonl fallback: T2/T3/T5 must share one iteration set
[ -f "$RETRO_JSONL" ] || { echo "! BLOCKED — retro T2: sanitized copy '$RETRO_JSONL' missing — run the T1 setup block first"; exit 1; }
RETRO_RESULT=$(python "${CLAUDE_PLUGIN_ROOT:-plugins/cc_research}/bin/retro_analyze.py" --jsonl "$RETRO_JSONL" --baseline "baseline" --alpha "$ALPHA" --direction "$METRIC_DIRECTION")  # timeout: 30000
RETRO_EXIT=$?
echo "$RETRO_RESULT" > "${TMPDIR:-/tmp}/retro-result-${CSID}"
[ "$RETRO_EXIT" -eq 2 ] && { echo "retro: Input error (exit 2) — run-id '$RUN_ID' missing, malformed, no baseline record, or invalid --alpha value (input: $RETRO_JSONL); no baseline record → that run never finished R4 baseline"; exit 1; }
# effect size: script does not return rank-biserial r directly
EFFECT_R=$(echo "$RETRO_RESULT" | python "${CLAUDE_PLUGIN_ROOT:-plugins/cc_research}/bin/compute_effect_size.py")  # timeout: 5000
echo "retro_result: $RETRO_RESULT"
echo "effect_r: $EFFECT_R"

Contract — script reads JSONL, extracts metric values for ALL iterations with status == "kept", runs one-sided one-sample Wilcoxon signed-rank test of "kept iterations vs single baseline metric" (status == "baseline"). Not a paired test — run records one baseline metric, no per-iteration matched baseline; baseline scalar compared against each kept value. Prints single line of JSON to stdout:

{"significant": bool, "p_value": float, "statistic": float, "n": int} on success
{"significant": false, "p_value": null, "statistic": null, "n": <N>, "reason": "<msg>"} when N < 6 or scipy missing
{"error": "<msg>"} on input error (exit 2 — missing file, malformed JSON, no baseline record)

Exit codes: 0 = significant · 1 = not significant (or insufficient data) · 2 = input error.

Direction handling — script branches on --direction:

higher → alternative = "greater" (improvement = candidate > baseline)
lower → alternative = "less" (improvement = candidate < baseline — for loss, latency, error)

Read direction from state.json config (or infer from goal text), pass via $METRIC_DIRECTION.

Effect size — script does not return rank-biserial r directly; the T2 block computes it via the bundled compute_effect_size.py and prints both retro_result JSON and effect_r.

If --compare: invoke the script a second time on the second run's sanitized copy experiments-clean-2.jsonl (T1 builds it with the same truncation rule as the first run's copy — never the raw file, so both runs are tested on complete iteration sets); downstream report renders second row. Skips itself when no --compare was given:

export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
IFS= read -r CMP_ID < "${TMPDIR:-/tmp}/retro-compare-id-${CSID}" 2>/dev/null || CMP_ID=""
[ -n "$CMP_ID" ] || { echo "retro: no --compare — second run skipped"; exit 0; }
IFS= read -r ALPHA < "${TMPDIR:-/tmp}/retro-alpha-${CSID}" 2>/dev/null || ALPHA="0.05"
IFS= read -r RETRO_JSONL2 < "${TMPDIR:-/tmp}/retro-jsonl2-path-${CSID}" 2>/dev/null || RETRO_JSONL2=""
[ -f "$RETRO_JSONL2" ] || { echo "! BLOCKED — retro T2: sanitized compare copy '$RETRO_JSONL2' missing — run the T1 setup block first"; exit 1; }
IFS= read -r RUN_ID < "${TMPDIR:-/tmp}/retro-run-id-resolved-${CSID}" 2>/dev/null || RUN_ID=""
[ -n "$RUN_ID" ] || { echo "! BLOCKED — retro T2: resolved run-id sentinel missing — run the T1 setup block first"; exit 1; }
METRIC1=$(python "${CLAUDE_PLUGIN_ROOT:-plugins/cc_research}/bin/read_state_field.py" ".experiments/state/$RUN_ID/state.json" "config.metric_cmd" --default "" 2>/dev/null)  # timeout: 5000
METRIC2=$(python "${CLAUDE_PLUGIN_ROOT:-plugins/cc_research}/bin/read_state_field.py" ".experiments/state/$CMP_ID/state.json" "config.metric_cmd" --default "" 2>/dev/null)  # timeout: 5000
[ "$METRIC1" = "$METRIC2" ] || { echo "! BLOCKED — Cannot compare runs with different metrics: $METRIC1 vs $METRIC2"; exit 1; }
[ -n "$METRIC1" ] || echo "⚠ neither run records config.metric_cmd — metric equality unverified"
DIRECTION2=$(python "${CLAUDE_PLUGIN_ROOT:-plugins/cc_research}/bin/read_state_field.py" ".experiments/state/$CMP_ID/state.json" "config.metric.direction" --default "" 2>/dev/null || echo "")  # timeout: 5000
[ -n "$DIRECTION2" ] || { echo "⚠ compare run records no config.metric.direction (predates config persistence) — assuming higher"; DIRECTION2=higher; }
RETRO_RESULT2=$(python "${CLAUDE_PLUGIN_ROOT:-plugins/cc_research}/bin/retro_analyze.py" --jsonl "$RETRO_JSONL2" --baseline "baseline" --alpha "$ALPHA" --direction "$DIRECTION2")  # timeout: 30000
[ "$?" -eq 2 ] && { echo "retro: Input error (exit 2) — compare run '$CMP_ID' malformed or no baseline record (input: $RETRO_JSONL2)"; exit 1; }
EFFECT_R2=$(echo "$RETRO_RESULT2" | python "${CLAUDE_PLUGIN_ROOT:-plugins/cc_research}/bin/compute_effect_size.py")  # timeout: 5000
echo "retro_result_compare: $RETRO_RESULT2"
echo "effect_r_compare: $EFFECT_R2"

Write the combined results (parsed JSON plus computed r) to $RUN_DIR/stats-results.json via Write tool.

Step T3: Dead iteration detection

Definition: dead iteration window = 3+ consecutive iterations (any status) where abs(metric_delta) < threshold (default --threshold 0.001).

Scale check (after loading baseline_metric in T1): if baseline_metric > 100 * threshold, print:

! Threshold advisory: baseline_metric=[value] is >100x the default threshold (0.001).
  For this metric scale, consider: --threshold [baseline_metric * 0.0001:.4f]
  Proceeding with --threshold [threshold] — override with: /research:retro <run-id> --threshold <value>

Apply advisory threshold automatically only when --threshold not explicitly provided by user.

Timeout detection: when scanning reverted iterations, check status field. If status == "timeout": classify as unreliable-timeout (see Notes). Else: flag any reverted iteration where delta is in correct improvement direction (metric moved toward goal) as "possible timeout — verify commit [sha]"; don't count delta as valid.

Scan experiments.jsonl sequentially, skipping iteration 0 (baseline). For each window of 3+ consecutive iterations where abs(delta) < threshold:

Record: start_iter, end_iter, count
Classify type: dead-plateau if all iterations in window have status: kept; dead-churn if mixed kept/reverted/other
Compute wasted_iters = total iterations in all dead windows

Re-hydrate cross-Bash state at the start of every separate Bash invocation in T3 (each Bash call is a fresh shell — $RUN_DIR / $RUN_ID_ARG lost across calls):

export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
IFS= read -r RUN_DIR < "${TMPDIR:-/tmp}/retro-run-dir-${CSID}" 2>/dev/null || RUN_DIR=""
IFS= read -r RUN_ID_ARG < "${TMPDIR:-/tmp}/retro-run-id-${CSID}" 2>/dev/null || RUN_ID_ARG=""
IFS= read -r RUN_ID < "${TMPDIR:-/tmp}/retro-run-id-resolved-${CSID}" 2>/dev/null || RUN_ID=""
IFS= read -r RETRO_JSONL < "${TMPDIR:-/tmp}/retro-jsonl-path-${CSID}" 2>/dev/null || RETRO_JSONL="$RUN_DIR/experiments-clean.jsonl"
# one call reports both missing values; `|| … &&` precedence subtle, trailing `[ -z ] && {…}` leaves exit 1 even when value present
# `exit 1` ends Bash subprocess only — orchestrator must treat non-zero as hard stop, not proceed to T4
python "${CLAUDE_PLUGIN_ROOT:-plugins/cc_research}/bin/require-vars.py" "$RUN_DIR" "retro T3: RUN_DIR missing — T1 must run first" "$RUN_ID" "retro T3: RUN_ID missing — T1 must run first" || exit 1

Write summary to $RUN_DIR/dead-iters.json via Write tool. Format:

{
  "windows": [{"start": 5, "end": 8, "count": 4, "type": "dead-churn"}],
  "total_dead": 4,
  "total_iterations": 20,
  "dead_pct": 20.0
}

Write dead-iteration scan script to $RUN_DIR/scripts/dead-iter-scan.py via Write tool, then execute in a separate Bash call. Never inline Python in the Bash command. (Different from T2: T3 writes a fresh dynamic script per invocation; T2 invokes a static bin/ script.)

Step T4: Suspicious jump detection

Compute per-iteration absolute metric deltas for kept iterations only. Build sliding window of 5 kept iterations to compute running mean and std of deltas.

Flag any single-step improvement where abs(delta) > running_mean + 2 * running_std:

Severity	Condition
HIGH	`abs(delta) > running_mean + 3 * running_std`
MEDIUM	`abs(delta) > running_mean + 2 * running_std` (and not HIGH)

For each flagged jump, record:

iteration, delta, sigma (how many std above mean), commit SHA, files changed (from experiments.jsonl files field)
Label: "suspicious — investigate" — NEVER auto-label "data leakage" or imply causation
Include corresponding diary.md entry for that iteration if present

Minimum data: require ≥6 kept iterations before flagging (need 5 for window + 1 to test). Fewer → skip suspicious-jump detection entirely, write "⚠ Insufficient data for trend analysis (need ≥6 data points, have <N>)" in Suspicious Metric Jumps section of report.

Write to $RUN_DIR/suspicious-jumps.json via Write tool.

Step T5: Scientist learning summary

Pre-compute all file paths before spawning; substitute actual computed values for every <RUN_DIR>/<path> placeholder in prompt below before constructing Agent() call — unsubstituted placeholder reaches agent as literal text: output lands in wrongly-named directory, post-call check reports false timeout. Verify $RUN_DIR/stats-results.json, $RUN_DIR/dead-iters.json, $RUN_DIR/suspicious-jumps.json exist (T2–T4 must complete first).

Agent budget — every spawn prompt must require an envelope even on exhaustion — partial: true plus what was finished.

Spawn research:scientist via Agent(subagent_type="research:scientist", prompt="..."):

Act as research retrospective analyst.

Read:
- experiments-clean.jsonl at <RETRO_JSONL path — the sanitized copy written by T1; fall back to experiments.jsonl if clean copy absent> (full iteration history)
- diary.md at <path> (if exists — for qualitative context)
- stats results at <RUN_DIR>/stats-results.json
- dead iteration summary at <RUN_DIR>/dead-iters.json
- suspicious jumps at <RUN_DIR>/suspicious-jumps.json

Produce a retrospective analysis covering:

1. **Strategy effectiveness**: which agent types (perf/code/ml/arch) had highest kept-rate and average delta? Rank them. Include per-agent iteration count, kept count, mean delta.
2. **Failure pattern analysis**: what approaches were repeatedly tried and reverted? Common failure modes? Group by pattern, not individual iteration.
3. **Diminishing returns**: at which iteration did improvement rate drop below 0.5% per iteration? Was the stopping point appropriate?
4. **Next hypotheses**: based on what worked and failed, generate 3–5 concrete next hypotheses. Write as hypotheses.jsonl-compatible file to <RUN_DIR>/hypotheses.jsonl — one JSON object per line, fields: hypothesis (str), rationale (str), confidence (float 0–1), expected_delta (str like "+2%"), priority (int 1=highest), source: "retro". Do NOT include feasible/blocker/codebase_mapping — feasibility annotation optional here; /research:run treats absent feasibility fields as feasible:true.
5. **Cross-run insights** (only if compare data present in stats-results.json): which run's strategy was more effective and why?

Write full retrospective to <RUN_DIR>/retrospective.md via Write tool.
Include ## Confidence block per quality-gates rules.
Return ONLY: {"status":"done","hypotheses":N,"file":"<RUN_DIR>/retrospective.md","confidence":0.N}

Health monitoring note (CLAUDE.md §6): research:scientist agent runs in background — spawn, then end turn; no filler call, no "waiting" line, no sleep. If completion notification arrives with no output file, or nothing appears for >15 min, treat as timed out.

Post-call timeout check: after Agent() returns, verify:

File $RUN_DIR/retrospective.md exists and has content → success
File missing or empty → set scientist_status = "timed_out", continue to T6; surface with ⏱ in report

Parse returned JSON envelope. Record hypotheses count and confidence for T6.

Step T6: Write retro report

mkdir -p .reports/research  # timeout: 3000
BRANCH=$(git branch --show-current 2>/dev/null | tr '/' '-' || echo 'main')  # timeout: 3000
OUT=$(python "${CLAUDE_PLUGIN_ROOT:-plugins/cc_research}/bin/resolve-anti-overwrite-path.py" .reports/research "retro-$BRANCH-$(date +%Y-%m-%d)")  # timeout: 5000 — counter suffix (-2, -3, …) when the file exists
echo "OUT=$OUT"

Write full report to the printed OUT via Write tool.

---
Title:         Retro — [goal]
Date:          [YYYY-MM-DD]
Scope:         [run-id] / [total] iterations
Focus:         retrospective analysis of ML optimization run
Agents:        research:scientist (T5)
Outcome:       IMPROVED | STALLED | PLATEAU | DIVERGED
Significance:  p=[value] ([significant|not significant] at alpha=[alpha]; kept-only — selection-biased, see caveat below)
Hypotheses:    [N] next steps generated
Confidence:    [score] — [key gaps]
Next steps:    /research:run … --hypothesis | /research:fortify
Path:          → .reports/research/retro-<branch>-<date>.md
---

## Retrospective: <goal>

**Run**: <run-id>
**Date**: <date>
**Iterations**: <total> (<kept> kept, <reverted> reverted, <other> other)
**Baseline**: <metric_key> = <baseline>
**Best**: <metric_key> = <best> (<delta>% improvement)

### Statistical Significance

| Test | N | Statistic | p-value | Significant? | Effect size |
| --- | --- | --- | --- | --- | --- |
| Wilcoxon vs baseline (run-1) | N | ... | ... | YES/NO (alpha=<alpha>) | r=... (<small/medium/large>) |
| Wilcoxon vs baseline (run-2) | N | ... | ... | YES/NO | r=... |

(Row labels drop the `(run-1)`/`(run-2)` suffix — plain "Wilcoxon vs baseline" — when `--compare` is not used, since there is only one run. Second row only if `--compare` used; when it is, add a note below the table: "Each row tests that run's kept iterations against its own baseline independently — not a direct statistical comparison between run-1 and run-2's improvement magnitudes." If N < 6: replace table with descriptive stats table — mean, median, min, max, std — and note "Insufficient data for significance testing (N=<N>)".)

**Effect size interpretation**: |r| < 0.3 = small, 0.3–0.5 = medium, > 0.5 = large.

> **Independence caveat** — Wilcoxon assumes independent samples. Sequential optimization iterations are typically autocorrelated; p-value is indicative only, not formally valid. If `dead_pct > 30%` from the Dead Iterations section, escalate caveat to HIGH: "p-value unreliable — high autocorrelation from dead-plateau windows."

> **Selection-bias caveat** — this test compares only KEPT iterations against the baseline. `/research:run` keeps an iteration only when its metric beats the running best (which starts at the baseline), so every kept value is beyond the baseline in the improvement direction *by construction*: all signed ranks share one sign and the one-sided p-value is a function of N alone (N=6 → p≈0.016). A `YES` in this table therefore records that ≥6 iterations were kept, not that improvement was independently validated. Genuine validation needs held-out data or independent repeat measurements not used for the keep/revert decision.

### Dead Iterations

| Start | End | Count | Type | Severity | Notes | Action |
| --- | --- | --- | --- | --- | --- | --- |
| ... | ... | ... | dead-plateau / dead-churn | MEDIUM | ... | <dead-plateau: raise keep noise floor or retire strategy · dead-churn: stop repeating approach, switch agent/files> |

(Every dead window is MEDIUM regardless of `dead_pct`; the `dead_pct > 30%` escalation applies to the Independence caveat only, not to window severity.)

Total dead: <N> of <total> (<pct>% of compute)

(If no dead windows: "No dead iteration windows detected (threshold=<threshold>)")

### Suspicious Metric Jumps

| Iteration | Delta | Sigma | Severity | Commit | Files Changed |
| --- | --- | --- | --- | --- | --- |
| ... | ... | ... | HIGH/MEDIUM | <sha> | <files> |

(If none: "No suspicious jumps detected")
(If insufficient data: "Insufficient data for jump detection (N=<N>)")

### Anomalies

| Pattern | Iteration(s) | Severity | Evidence | Action |
| --- | --- | --- | --- | --- |
| kept-regression / reverted-improvement / unreliable-timeout / config-repetition / truncated-log / threshold-advisory | ... | HIGH/MEDIUM/LOW | <agent, files, status, delta, commit> | <one concrete next action> |

(One row per detected instance. Pattern name, severity and default action come from `<notes>` Named anomaly patterns. If none: "No named anomalies detected".)

### Strategy Effectiveness

| Strategy | Kept | Tried | Keep-rate | Avg Delta | Best Delta |
| --- | --- | --- | --- | --- | --- |
| ... | ... | ... | ...% | ... | ... |

(From scientist retrospective. If scientist timed out: "Scientist agent timed out — strategy analysis unavailable")

### Failure Patterns
<From scientist retrospective — grouped failure modes>

### Diminishing Returns
<Iteration where improvement rate dropped below 0.5% per iteration, or "not applicable">

### Suggested Next Hypotheses

| # | Hypothesis | Rationale | Expected Delta | Confidence |
| --- | --- | --- | --- | --- |
| 1 | ... | ... | ... | 0.N |

Full retrospective: <RUN_DIR>/retrospective.md
Next hypotheses queue: <RUN_DIR>/hypotheses.jsonl

## Confidence
**Score**: 0.NN — [high|moderate|low]
**Finding confidence** (dead windows, suspicious jumps, classification errors, pattern detection): [high|moderate|low] — independent of statistical test availability
**Statistical test** (Wilcoxon p-value): [available: p=X | unavailable: scipy not installed — descriptive stats only]
**Gaps**:
- (-0.NN) [degraded Finding confidence above, only when moderate/low — named with its cause; statistical-test unavailability, unrun scripts or absent scipy list as (-0.00) unless a specific finding depends on them]
- (-0.NN) [other specific limitations]

Step T7: Terminal summary and follow-up gate

Print compact summary to terminal only — do NOT repeat full report:

---
Retro — <goal>
Run:           <run-id> (<total> iterations, <kept> kept)
Significance:  p=<value> (<significant|not significant> at alpha=<alpha>; kept-only, selection-biased — not independent validation)  [or: N=<N> insufficient]
Effect size:   r=<value> (<small|medium|large>)  [or: n/a]
Dead iters:    <N>/<total> (<pct>%)  [or: none]
Suspicious:    <N> jumps (<severity> — investigate: <sha1>, <sha2>)  [or: none]
Anomalies:     <N> (<n> HIGH — <pattern>@iter <i>, …)  [or: none]
Hypotheses:    <N> next steps generated
-> saved to .reports/research/retro-<branch>-<date>.md
---
Next: /research:run <program.md> --hypothesis <RUN_DIR>/hypotheses.jsonl    [only if scientist_status != "timed_out" AND <RUN_DIR>/hypotheses.jsonl exists]
     /research:fortify <run-id>    ← stress-test top hypothesis before full re-run

If scientist_status == "timed_out" or <RUN_DIR>/hypotheses.jsonl does not exist on disk, omit the --hypothesis Next line entirely and replace with: Next: /research:fortify <run-id> ← scientist analysis unavailable; no hypotheses queue generated.

</workflow> <notes>

Retro read-only — never modifies code, commits, or writes to .experiments/state/<run-id>/
.experiments/retro-<timestamp>/ stores analysis scripts, intermediate JSON, scientist output, hypotheses.jsonl
Retro run dirs don't write result.jsonl — exempt from automated 30-day TTL cleanup (per .claude/rules/foundry-artifact-lifecycle.md, installed via /foundry:setup — requires foundry plugin: no result.jsonl = cleanup skipped); remove manually when done (rm -rf .experiments/retro-*/)
hypotheses.jsonl uses source: "retro" — compatible with --hypothesis flag of /research:run; "retro" extends oracle schema (see protocol.md); feasibility fields omitted, treated as feasible:true by run
--compare requires both runs use same metric; if metric names differ, stop: "Cannot compare runs with different metrics: <metric-1> vs <metric-2>"
Dead iteration threshold (--threshold) should match metric's noise floor — default 0.001 for normalized metrics; adjust for raw values (e.g. --threshold 0.1 for loss in hundreds)
Statistical tests assume metric values are independent samples — if iterations highly correlated (e.g. cumulative optimization), note limitation in report
Requires scipy in active Python environment (pip install scipy) — retro_analyze.py runs Wilcoxon signed-rank test via scipy.stats. Without scipy, test skipped and retro_analyze.py returns {"significant": false, "p_value": null, "reason": "scipy not installed"}; report includes descriptive stats only (mean/median/min/max/std). Install: pip install scipy or uv add scipy.
Named anomaly patterns (use consistently across reports; each instance is one row in the T6 Anomalies table with the severity and default action below):
- kept-regression (HIGH): kept iteration where metric moved in wrong direction (negative delta for higher-is-better, positive delta for lower-is-better). Action: investigate why the keep gate passed; re-test reverting <sha> on top of the current best
- reverted-improvement (MEDIUM): reverted iteration where metric moved in correct direction — reverted for non-metric reasons (performance, OOM, instability); flag as "improvement-when-reverted — consider revisiting with adjusted constraints". Action: check revert cause in diary/logs; revisit with adjusted constraints
- unreliable-timeout (MEDIUM): reverted iteration with status: "timeout" — metric value unreliable; never count delta as valid improvement. Action: count as wasted; re-run with a raised timeout before trusting the delta
- config-repetition (MEDIUM): same agent + same file(s) attempted 3+ times without crossing threshold — flag as "repeated-failure pattern". Action: retire that agent + file combination from the hypothesis queue
- Run-integrity rows in the same table: truncated-log (HIGH) — T1 dropped a truncated last JSONL line; Action: recover or re-measure the dropped iteration before trusting best_metric. threshold-advisory (LOW) — T3 scale check fired; Action: if the default was automatically adjusted, report the applied threshold with no re-run; if the user supplied --threshold, retain it and suggest a re-run with the advisory value.
Findings vs observations: a finding is a row in Dead Iterations, Suspicious Metric Jumps, or Anomalies. Any other remark (state/ledger iteration-count mismatch, keep-gate commentary, diminishing-returns notes) goes in one Observations: line under the Anomalies table, with no severity. Never add an ad-hoc findings table.

</notes>

retro

SKILL.mdplugins/cc_research/skills/retro/SKILL.md