Run

Sustained metric-improvement loop — reads program.md, iterates specialist ideation agents, commits atomically, auto-rolls back on regression. For long-running automated improvement campaigns.

NOT for: methodology validation before run (use /research:judge); hypothesis generation (use research:scientist agent); one-off feature work (use /develop:feature).

Campaign mode only:

MAX_ITERATIONS:             50 (hard cap); DEFAULT 20 when max_iterations unset in program.md; program.md may raise up to 50; values above 50 clamped to 50 with a warning
MAX_CODEX_RUNS:             10 (cost ceiling for --codex Phase 2c — disable Codex once exceeded)
STUCK_THRESHOLD:            5 consecutive discards → escalation
GUARD_REWORK_MAX:           2 attempts before revert
VERIFY_TIMEOUT_SEC:         120 (local), 300 (--colab)
COLAB_KNOWN_HW:             H100, L4, T4, A100
SUMMARY_INTERVAL:           10 iterations
DIMINISHING_RETURNS_WINDOW: 5 iterations < 0.5% each → warn user and suggest stopping
STATE_DIR:                  .experiments/state/<run-id>/  (timestamped dir per run — see .claude/rules/foundry-artifact-lifecycle.md)
SENTINEL_SLUG_FORMULA: |
  eval "$(bash "${CLAUDE_PLUGIN_ROOT:-plugins/cc_research}/bin/git_slugs.sh")"
  # Sentinel path: ${TMPDIR:-/tmp}/claude-commit-auth-${REPO_SLUG}-${BRANCH_SLUG}  # tmpdir-exempt: user-shell-boundary
  # Bash state is lost between tool calls — re-source git_slugs.sh at each use site; it is the only authorized slug form.

Agent strategy mapping (agent_strategy in config → ideation agent to spawn):

Auto-inference keyword heuristics (when agent_strategy: auto or omitted; checked against ## Goal text AND metric command):

Precedence order (first match wins; ML keywords beat test-framework keywords). ML-specific compound terms (not bare tokens) required — prevents over-triggering on eval/train/val as common words:

• contains accuracy, loss (paired with train_loss/val_loss/eval_loss), f1_score, auc_roc, auroc, train_step, val_acc, eval_loss, epoch, gradient, tensor, overfit, generaliz, regulariz, validation, dropout, weight_decay, lr_schedule, cross_val, precision, recall, OR explicit --scientist flag → ml → research:scientist
• contains time, latency, bench, throughput, memory → perf → foundry:perf-optimizer
• contains pytest, coverage, complexity → code → foundry:sw-engineer
• no keyword match → perf (default fallback) — WARN: print ⚠ No keyword match — defaulting to 'perf' strategy. If this is an ML task, set agent_strategy: ml in program.md. Log resolved agent + reason in state.json strategy_resolution.

Bare tokens eval, train, val (without compound suffix) do NOT trigger ml routing — too common in non-ML contexts (test eval scripts, training-environment configs, validator command names).

Stuck escalation sequence (at STUCK_THRESHOLD consecutive discards):

1. Switch agent type. Rotation by current strategy:

   Current strategy	Next strategy	Escalation agent
   code	ml	research:scientist
   ml	perf	foundry:perf-optimizer
   perf	code	foundry:sw-engineer
   arch	code	foundry:sw-engineer (fallback foundry:solution-architect if sw-engineer unavailable)
   auto	infer from resolved strategy	follow rotation row for whichever concrete strategy auto heuristics resolved to at Step R3 (e.g. auto → resolved ml → next perf → foundry:perf-optimizer)

2. Spawn 2 agents parallel, competing strategies; each writes full analysis to .experiments/state/<run-id>/stuck-escalation-<i>-<agent-type>.md, returns ONLY compact JSON envelope. Use this spawn prompt verbatim (substitute <run-id>, <i>, and strategy):

   Stuck-escalation handoff — iteration <i> after STUCK_THRESHOLD consecutive discards.
   Read `.experiments/state/<run-id>/state.json` for goal, best_metric, baseline, config.
   Read `.experiments/state/<run-id>/experiments.jsonl` for full iteration history.
   Read `.experiments/state/<run-id>/diary.md` for qualitative context (what was tried, why reverted).
   Read `.experiments/state/<run-id>/context-<i>.md` for current iteration's context block.
   Continue from the last completed iteration (do NOT restart from iteration 0).
   Write your full analysis and proposed change to `.experiments/state/<run-id>/stuck-escalation-<i>-<your-strategy>.md`.
   Write a resume point to `.experiments/state/<run-id>/resume.json`: {iteration: <i>, strategy: "<your-strategy>", proposed_change: "<one-line description>"}.
   Return ONLY: {"strategy":"<your-strategy>","description":"...","files_modified":[...],"confidence":0.N,"file":".experiments/state/<run-id>/stuck-escalation-<i>-<your-strategy>.md"}
   

   Consolidation: pick whichever returns delta ≥ 0.1% AND guard pass; if both qualify, pick higher delta.

3. Stop, report progress, surface to user — no blind looping

• Key boundary: end of each Phase 8 in R5 iteration loop — JSONL record appended and state.json updated. Overwrite each iteration; contract always reflects latest in-progress state. Long metric-improvement loops are the primary auto-compact risk.
• Preserve at each boundary: RUN_ID (TMPDIR key), STATE_DIR path, program.md path, current iteration#, best metric, best-commit SHA, experiments.jsonl path.
• Clear at R1 start (stale prior run) and after R6/R7 campaign completion.

Agent Resolution

Agent resolution: load and follow the protocol below. Contains: foundry check + fallback table. If foundry not installed: use table to substitute each foundry:X with general-purpose. Agents this skill uses: foundry:sw-engineer, foundry:linting-expert, foundry:perf-optimizer, foundry:solution-architect, research:scientist.

# loads: compaction-contract.md
export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
_RESEARCH_SHARED=$(python "${CLAUDE_PLUGIN_ROOT:-plugins/cc_research}/bin/resolve_shared.py" 2>/dev/null)  # timeout: 5000
[ -z "$_RESEARCH_SHARED" ] && { echo "! Plugin path resolution failed — ensure research plugin installed and CLAUDE_PLUGIN_ROOT set, or invoke from project root."; exit 1; }
echo "$_RESEARCH_SHARED" > "${TMPDIR:-/tmp}/research-shared-${CSID}"  # cold resolve — every later site reads this sentinel instead of re-running python
cat "$_RESEARCH_SHARED/agent-resolution.md"

CLAUDE_SKILL_DIR resolution — constants block provides default plugins/cc_research/skills/run (source-tree path). Resolve to installed path before use:

export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
CLAUDE_SKILL_DIR=$(ls -td ~/.claude/plugins/cache/borda-ai-rig/research/*/skills/run 2>/dev/null | head -1)
[ -z "$CLAUDE_SKILL_DIR" ] && CLAUDE_SKILL_DIR="$(git rev-parse --show-toplevel 2>/dev/null)/plugins/cc_research/skills/run"
echo "$CLAUDE_SKILL_DIR" > "${TMPDIR:-/tmp}/research-run-skill-dir-${CSID}"

Default Mode (Steps R1–R7)

Triggered by run <goal|file.md>.

Task tracking: create tasks R0–R7 at start. If no --researcher/--architect, mark R0 skipped. If --codex active, create task R5b: Codex co-pilot (iter ?/max) status pending.

Step R0: Hypothesis pre-phase (--researcher / --architect)

If no --researcher/--architect, skip to R1.

Flag combination note: every oracle self-annotates feasibility (feasible/blocker/codebase_mapping are part of the oracle schema — no separate annotation spawn). --researcher alone, --architect alone, and both together are all valid; both together adds architectural hypotheses alongside the research ones.

Follow modes/hypothesis-pipeline.md:

export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
IFS= read -r CLAUDE_SKILL_DIR < "${TMPDIR:-/tmp}/research-run-skill-dir-${CSID}" 2>/dev/null || CLAUDE_SKILL_DIR=""
cat "$CLAUDE_SKILL_DIR/modes/hypothesis-pipeline.md"  # timeout: 5000

Per-iteration hypothesis selection (when --researcher/--architect set, inside R5 loop): pop next from RESEARCH_QUEUE. Append to Phase 2 prompt: "Focus this iteration on testing this hypothesis: <hypothesis text>."

Per-iteration journal hook (inside R5, after Phase 7): if --journal active, append entry to <RUN_DIR>/journal.md after EVERY iteration — regardless of outcome. Entry format: protocol.md (companion file, same skill dir). # loads: protocol.md Journals record kept and reverted iterations so ideation agent learns failed approaches.

Per-iteration checkpoint write (after Phase 7): if --researcher/--architect active, append one line to <RUN_DIR>/checkpoint.json per schema in protocol.md (companion file, same skill dir): {iteration, hypothesis_id, metric_before, metric_after, status: "passed"|"rolled_back"}.

Step R1: Load / build config

--resume flag detection: if --resume in args, extract optional program.md path. Jump to ## Resume Mode. Rest of R1 and R2–R7 skipped.

--hypothesis <path> parsing: if --hypothesis in args, extract path token following it. Verify file exists: [ -f "$HYPOTHESIS_PATH" ]. If not found: print ! --hypothesis <path>: file not found and stop. If found: set hypothesis_override = true. In R5 Phase 2 (Propose change), replace oracle-generated hypothesis with loaded file content — prepend to ideation agent prompt: "Use this pre-specified hypothesis as your starting hypothesis for iteration N: . Validate, refine, and implement it. Do not generate a new hypothesis from scratch."

Auto-detect: first non-flag arg ends in .md → parse as program file. Otherwise → text goal.

Clarification prompt (.md file only): after extracting .md path, inspect next token (before -- flags):

• Absent or starts with -- → clarification_prompt = null
• Quoted string (starts/ends with ") → extract as clarification_prompt, strip quotes
• Bare unquoted token (no --, no ") → accept as clarification_prompt; print: ℹ clarification set to "<token>" (tip: quote multi-word hints — e.g. "/research:run program.md \"focus on sort\" --codex")

After clarification extraction, remaining non-flag tokens (not starting --) are unrecognized. For each, print:

⚠ Unrecognized argument "<token>" — ignored.
  Known positional args: <program.md path> [clarification]
  Known flags: --resume <program.md>, --team, --compute=local|colab|docker, --colab[=HW], --codex, --researcher, --architect, --journal, --hypothesis <path>, --scientist, --codemap, --no-codemap, --keep "<items>"
  If you meant to override the algo, edit the ## Config block in your program.md (algo: sort) and update ## Metric to match.
  If you meant to set a clarification hint, pass it as a quoted string: "/research:run program.md \"sort improvements\" --codex"

export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
# runs under bash — zsh never populates ${BASH_REMATCH[1]}, so --keep "..." was silently resolving empty
python "${CLAUDE_PLUGIN_ROOT:-plugins/cc_research}/bin/extract-keep-flag.py" research-run "$ARGUMENTS"  # timeout: 5000 — parses --keep, clears a stale contract, persists for Phase 8

Unsupported flag check: load and follow the protocol below. Supported flags for this skill: --resume, --team, --compute, --colab, --codex, --researcher, --architect, --journal, --hypothesis, --scientist, --codemap, --no-codemap, --keep.

# loads: unsupported-flag-protocol.md
export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
IFS= read -r _RESEARCH_SHARED < "${TMPDIR:-/tmp}/research-shared-${CSID}" 2>/dev/null || _RESEARCH_SHARED=""  # warm read (Check 41)
cat "$_RESEARCH_SHARED/unsupported-flag-protocol.md"

Codemap auto-detection — structural blast-radius context for modules the experiment edits; on by default when codemap installed + index found. --no-codemap opts out; --codemap is strict (fail if unavailable).

# timeout: 5000
export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
# writes true/false to research-run-codemap-enabled-${CSID}; strict mode exits 1 (already printed ! BLOCKED) if unavailable
CODEMAP_RAW=$(python "${CLAUDE_PLUGIN_ROOT:-plugins/cc_research}/bin/codemap-flag.py" research-run "$ARGUMENTS") || exit 1

loads: codemap-gates.md

When CODEMAP_RAW ≠ off:

export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
IFS= read -r _RESEARCH_SHARED < "${TMPDIR:-/tmp}/research-shared-${CSID}" 2>/dev/null || _RESEARCH_SHARED=""  # warm read (Check 41)
cat "$_RESEARCH_SHARED/codemap-gates.md"

Follow Gate A and Gate B.

If argument is a .md file — read and parse with these rules:

1. Find each ## <Section> heading (case-insensitive).
2. Extract first fenced code block following that heading.
3. Parse block as key: value lines; multi-value = indented - value items. Paths with spaces: wrap in double quotes.
4. Missing required fields (command under ## Metric/## Guard) → stop with error.
5. agent_strategy: auto (or omitted) → apply keyword heuristics from <constants> to ## Goal text and metric command.
6. target under ## Metric: direction: higher → stop when metric ≥ target; direction: lower → stop when metric ≤ target. If target omitted, run until max_iterations.
7. Unrecognized keys/headings → warn once, ignore.
8. ## Notes and # Program: title never parsed — human-only. (# Campaign: accepted as alias.)

If argument is text — auto-detect metric_cmd/guard_cmd from goal string and codebase scan (same as P-P1, non-interactive). config.json not read.

--colab[=HW] parsing: --colab (no =) → compute = "colab", colab_hw = null. --colab=<value> → compute = "colab", colab_hw = <value> (uppercased). Unknown <value> (not in {H100, L4, T4, A100}) → print "⚠ Unknown Colab hardware '<value>' — proceeding with default GPU. Known: H100, L4, T4, A100", set colab_hw = null. --compute=colab (no HW) → compute = "colab", colab_hw = null.

colab_hw in ## Config sets hardware preference (H100, L4, T4, A100); CLI --colab=HW overrides.

Generate run-id = $(date -u +%Y-%m-%dT%H-%M-%SZ). Assign immediately:

export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
RUN_ID=$(date -u +%Y-%m-%dT%H-%M-%SZ)
RUN_DIR=".experiments/${RUN_ID}"  # hypothesis pipeline + journal outputs (per <constants> note)
STATE_DIR=".experiments/state/${RUN_ID}"  # per-iteration artifacts (state.json, experiments.jsonl, diary.md)
mkdir -p "$RUN_DIR" "$STATE_DIR"  # timeout: 5000 — both dirs created before any Write to either
echo "$RUN_ID" > "${TMPDIR:-/tmp}/research-run-id-${CSID}"  # persist for Phase 8 contract write (Check 41: fresh shell)

Note: STATE_DIR (.experiments/state/${RUN_ID}/) is per-iteration artifact dir — distinct from RUN_DIR. Both coexist; see <constants> block.

Create run directory:

.experiments/state/<run-id>/
  state.json         ← iteration count, best metric, status
  experiments.jsonl  ← one line per iteration
  diary.md           ← human-readable research diary (hypothesis → outcome → decision)

Convert program_file to absolute path: realpath "$PROGRAM_FILE" — Resume Mode matches on absolute path.

Write initial state.json (program_file = absolute path to .md or null for text goal):

{
  "run_id": "<run-id>",
  "goal": "<goal>",
  "config": {},
  "program_file": "<absolute path to program.md, or null>",
  "iteration": 0,
  "best_metric": null,
  "best_commit": null,
  "status": "initializing",
  "started_at": "<ISO timestamp>",
  "clarification_prompt": null,
  "colab_hw": null,
  "sandbox_mode": "local"
}

Note: status is "initializing" until all R2 precondition checks pass — resume treats "initializing" as failed-init, not active run. Update to "running" at end of R2 (after all checks pass).

Step R2: Precondition checks

Run all checks before touching code. Fail fast with clear message:

1. Clean git: git status --porcelain → must be empty. If dirty: print dirty files and stop.
2. Not detached HEAD: git rev-parse --abbrev-ref HEAD → must not be HEAD.
3. Metric command numeric: run metric_cmd once; parse stdout for float. If no float: show output and stop.
4. Guard passes: run guard_cmd once; must exit 0. If fails: show output and stop.
5. --colab check: verify mcp__colab-mcp__runtime_execute_code available. If not, print setup instructions (see Colab MCP section) and stop. If --colab=HW (colab_hw non-null): print: Hardware requested: --colab=<colab_hw>. Ensure your Colab notebook running with <colab_hw> GPU.
6. --codex check: distinguish the installed-and-enabled bridge target from absence. claude not on PATH → print ⚠ 'claude' CLI not in PATH — bridge availability cannot be verified. and stop. If claude plugin list lacks bridge@borda-ai-rig, print ⚠ bridge@borda-ai-rig not installed. Install it from the Borda AI Rig marketplace. and stop. If it is disabled, print ⚠ bridge@borda-ai-rig is disabled. Enable it and reload plugins. and stop.
7. compute: docker check: run docker ps via Bash (timeout: 5000). If non-zero: print ⚠ Docker daemon not running. Start Docker Desktop and retry. and stop.
8. Flag conflict: if --colab and --compute=docker both active: print ⚠ --colab and --compute=docker are mutually exclusive. Use one or the other. and stop.
9. --colab + --codex compatibility note (non-blocking): if both flags active, print ℹ --colab + --codex active: Codex Phase 2c will receive colab_hw context so generated code can target the right GPU (H100/T4 bf16 vs fp16). Phase 5 metric verification runs through Colab MCP as usual. and continue. Pass colab_hw to Codex spawn prompt (Phase 2c — see modes/codex-copilot.md).
10. --journal prerequisite: verify --researcher/--architect also set. If neither: print ⚠ --journal requires --researcher or --architect — omit --journal or add a hypothesis pipeline flag. and stop.

--codex-delegation warning (non-blocking): codex-delegation.md ships inside this plugin's own skills/_shared/, so R7 needs no other plugin installed. Verify it resolves:

export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
IFS= read -r _RESEARCH_SHARED < "${TMPDIR:-/tmp}/research-shared-${CSID}" 2>/dev/null || _RESEARCH_SHARED=""  # warm read (Check 41)
[ -f "$_RESEARCH_SHARED/codex-delegation.md" ] || echo "⚠ codex-delegation.md not found under $_RESEARCH_SHARED — R7 Codex delegation will be skipped; reinstall the research plugin."

Set CODEX_DELEGATION_AVAILABLE=true if found, false otherwise. Continue regardless.

Initialize sandbox + timeout variables (after all checks pass — constants YAML block not auto-exported to bash; assign explicitly with ${VAR:-default} to honour environment overrides; ADV-L15 / ADV-M20):

SANDBOX_NETWORK="${SANDBOX_NETWORK:-none}"  # override via program.md Config or environment variable
# Verify timeout — 120s local, 300s Colab per <constants>; bash overrides via VERIFY_TIMEOUT_SEC env var
if [ "${compute:-local}" = "colab" ]; then
    VERIFY_TIMEOUT_SEC="${VERIFY_TIMEOUT_SEC:-300}"
else
    VERIFY_TIMEOUT_SEC="${VERIFY_TIMEOUT_SEC:-120}"
fi
VERIFY_TIMEOUT_MS=$((VERIFY_TIMEOUT_SEC * 1000))
# Ideation Agent() calls are synchronous — no mid-flight poll; after each returns, check its output file and mark timed_out (⏱) if empty.

Initialize sandbox_mode:

• compute: docker (daemon check passed in step 7) → sandbox_mode = "docker". Print: sandbox: Docker daemon reachable — sandbox mode active
• All other cases (compute: local, compute: colab) → sandbox_mode = "local"

Update state.json status to "running" — write only after ALL checks above pass. Resume treats "initializing" as failed-init and skips such runs.

Step R3: Select ideation agent

Apply agent_strategy mapping from <constants>. If auto, apply keyword heuristics to metric_cmd. Log selected agent to state.json.

Step R4: Establish baseline (iteration 0)

Run metric_cmd and guard_cmd. Parse metric value. Append to experiments.jsonl:

{
  "iteration": 0,
  "commit": "<HEAD sha>",
  "metric": 0.0,
  "delta": 0.0,
  "guard": "pass",
  "status": "baseline",
  "description": "baseline",
  "agent": null,
  "confidence": null,
  "timestamp": "<ISO>",
  "files": []
}

Update state.json: best_metric = <baseline>, best_commit = <HEAD sha>.

Print: Baseline: <metric_cmd key> = <value>.

Write initial diary header to .experiments/state/<run-id>/diary.md:

# Research Diary — <goal>

**Run**: <run-id>
**Started**: <ISO timestamp>
**Baseline**: <metric_key> = <baseline value>

---

Then proceed to R5.

Step R5: Iteration loop

# REPO_SLUG / BRANCH_SLUG: source the single authorized slug form (see <constants>)
eval "$(bash "${CLAUDE_PLUGIN_ROOT:-plugins/cc_research}/bin/git_slugs.sh")"  # timeout: 3000
COMMIT_SENTINEL="${TMPDIR:-/tmp}/claude-commit-auth-${REPO_SLUG}-${BRANCH_SLUG}"  # tmpdir-exempt: user-shell-boundary
touch "$COMMIT_SENTINEL"  # timeout: 3000
# trap doesn't survive across Bash calls — commit-guard.js hook (foundry-owned) handles protection instead

Dependency — commit-guard.js (requires foundry plugin): the commit-sentinel dance above (touch at R5, re-touch each phase, rm at cleanup) is enforced by foundry's commit-guard.js PreToolUse hook. That hook ships with the foundry plugin only — research does not bundle it. Standalone install (foundry absent): the sentinel touches become inert and git commit proceeds unguarded. The sentinel logic is still safe to run (touch/rm on a temp file are harmless no-ops without the hook); it simply provides no protection. If you rely on atomic-commit guarding during research:run, install foundry.

Sentinel liveness: touch $COMMIT_SENTINEL after each Phase 8 result write to extend monitoring window — do NOT rely solely on sentinel touched at loop start; slow iterations exceed 15-min TTL. Re-derive slug per SENTINEL_SLUG_FORMULA from <constants> (bash state lost between calls).

--team mode: If --team active, follow modes/team.md and execute Phases A–D in place of standard iteration loop below.

export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
IFS= read -r CLAUDE_SKILL_DIR < "${TMPDIR:-/tmp}/research-run-skill-dir-${CSID}" 2>/dev/null || CLAUDE_SKILL_DIR=""
cat "$CLAUDE_SKILL_DIR/modes/team.md"  # timeout: 5000

--team + --hypothesis combination: combinable. Team mode uses provided hypothesis path and skips oracle/hypothesis-generation phase — hypothesis_override = true applies inside team.md Phase A same as solo mode.

For each iteration i from 1 to max_iterations:

Phase overview (all phases run per iteration):

Command execution rules (apply to ALL phases running external commands):

1. Use Bash tool timeout parameter: Never shell timeout wrapper. Pass timeout: <ms> on Bash tool call itself. (Compound commands are already barred globally — see claude-config.md §Directory Navigation Commands.)
2. No inline multi-line Python: Python logic >3 lines → write to .experiments/state/<run-id>/scripts/script-<i>.py via Write tool, execute with python <path> or uv run python <path>. Two triggers Claude Code always flags: (a) =([0-9.]+) inside -c "..." (false Zsh substitution); (b) multi-line -c "..." with #-prefixed comment lines. Writing to file sidesteps both.
3. No Zsh constructs: Never use =(), <(), >() in Bash commands — even inside quoted strings; Claude Code scans raw command text.
4. Local exploratory scripts writing to real files (scanning config combos, patching JSON, temp overrides): write to .experiments/state/<run-id>/scripts/, run locally with python <path>. Legitimately modify project files — NOT in Docker sandbox.
5. Docker sandbox (when available — see Phase 2a): Phases 4–6 route metric_cmd/guard_cmd through Docker when compute: docker. Phase 2a: read-only hypothesis scripts in sandbox. Scripts writing to project files always run locally.
6. One change per iteration: Never batch-loop over config variants/combos in single Bash/Python call. Each variant = one campaign iteration — loop/measure/compare is campaign framework's job, not ideation agent's.

Phase 0 — Print header

Print iteration header, update R5 task:

[→ Iter N/max_iterations — best so far: <best_metric> (Δ<best_delta_pct>% vs baseline)]

TaskUpdate R5 subject: R5: Iteration N/max_iterations — running

Phase 1 — Build context

Build context for ideation agent, write to file — do NOT accumulate inline in main context:

git log --oneline -10 >.experiments/state/${RUN_ID}/context-${I}.md  # timeout: 3000
tail -10 .experiments/state/${RUN_ID}/experiments.jsonl >>.experiments/state/${RUN_ID}/context-${I}.md  # timeout: 5000
# Fresh repos have <5 commits — fall back to full HEAD diff when shallow
if [ "$(git rev-list HEAD --count 2>/dev/null)" -gt 5 ]; then
    git diff --stat HEAD~5 HEAD >>.experiments/state/${RUN_ID}/context-${I}.md  # timeout: 3000
else
    git diff --stat HEAD >>.experiments/state/${RUN_ID}/context-${I}.md  # timeout: 3000
fi

Codemap structural context (only if CODEMAP_ENABLED=true — re-read from ${TMPDIR:-/tmp}/research-run-codemap-enabled-${CSID}). Cat once, first iteration only — the file is static and stays in context; re-cat only if it is no longer in context (e.g. after a compaction). Re-catting every iteration re-bills ~800 tok × N iterations for identical text:

export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
IFS= read -r _RESEARCH_SHARED < "${TMPDIR:-/tmp}/research-shared-${CSID}" 2>/dev/null || _RESEARCH_SHARED=""  # warm read (Check 41)
cat "$_RESEARCH_SHARED/codemap-context.md"

Execute its block. Leave TARGET_MODULE/TARGET_FN empty for the global central blast-radius baseline, or set TARGET_MODULE to the module the experiment edits (from ## Config) for importer/coverage queries. Append output to context-${I}.md under a ## Structural Context (codemap-py) heading so the Phase 2 ideation agent sees blast-radius before proposing edits.

Codemap output non-empty: also append this codemap-first protocol directly below it in context-${I}.md (own copy — self-contained, no cross-plugin reference), so the Phase 2 spawn prompt's "read context-<i>.md" instruction carries it to the ideation agent: (1) Skill-first — use the Structural Context above before any Grep/Glob/Read aimed at imports, callers, or test coverage for a symbol already listed there. (2) Bounded call budget — symbol not listed → up to 3 additional codemap-py query calls this iteration. (3) Hard stop on query_complete: true (or legacy exhaustive: true) — that result is final for its direction, no follow-up Grep/Read/query to re-confirm it. Codemap output empty: omit this paragraph — Phase 2 agent proceeds with normal file-read behaviour.

Prepend header block to context-<i>.md: goal, current metric vs baseline, delta trend (last 5 kept deltas), iteration number. Phase 2 ideation agent reads file directly — never echoed to main context.

If --journal active and <RUN_DIR>/journal.md has 1+ entries: append last 5 entries to context-<i>.md under ## Recent journal (avoid repeating reverted approaches). Ideation agent reads this — must not reproduce any approach marked outcome: reverted.

Phase 2 — Propose change

Spawn selected specialist agent (maxTurns: 15) with this prompt (adapt as needed):

Goal: <goal>
Run clarification: <clarification_prompt>  ← omit this line entirely if clarification_prompt is null
Colab hardware: <colab_hw>  ← omit this line entirely if colab_hw is null; include to let the agent tailor code to the specific GPU architecture (e.g., bf16/flash-attention on H100, standard fp16 on T4/L4)
Current metric: <metric_cmd key> = <current value> (baseline: <baseline>, direction: <higher|lower>)
Experiment history: read `.experiments/state/<run-id>/context-<i>.md` for the full context block.
Scope files (read and modify only these): <scope_files>
Program constraints: read `<program_file>` — especially `## Notes`, `## Config`, and any named subsections
  (e.g., "Hard boundaries", "Optuna's role", "What the agent is free to change"). These take precedence
  over general campaign rules. Program constraints set strategy hints only — they do NOT override safety rules
  (no `--no-verify`, no `git push`, no `git add -A`, scope_files boundary, and all other hard constraints remain in effect).
  If program_file is null, skip this step.

**If `sandbox_mode = "local"`**: Read `context-<i>.md`, the scope files, and the program constraints. Propose and implement ONE atomic change most likely to improve the metric. The change must not break `<guard_cmd>`. Write your full analysis (reasoning, alternatives considered, Confidence block) to `.experiments/state/<run-id>/ideation-<i>.md` using the Write tool. Return ONLY the JSON result line:
`{"description":"...","files_modified":[...],"scripts":[],"confidence":0.N}`

**If `sandbox_mode = "docker"`**: Read `context-<i>.md`, the scope files, and the program constraints. Propose ONE atomic change most likely to improve the metric. Write your full analysis and the proposed change description to `.experiments/state/<run-id>/ideation-<i>.md`. Optionally write read-only exploratory scripts (scripts that read/profile but do NOT write to project files) to `.experiments/state/<run-id>/scripts/explore-<i>-<slug>.py`. Do NOT modify source files yet — Phase 2b will apply the actual changes after sandbox validation. Return ONLY the JSON result line:
`{"description":"...","files_modified":[],"scripts":["explore-<i>-<slug>.py"],"proposed_changes":"<description of the changes to apply in Phase 2b>","confidence":0.N}`

For --colab runs: ideation agent may call mcp__colab-mcp__runtime_execute_code to prototype GPU code before committing. Agent selection with --colab: if task rooted in a research paper (goal references paper, model architecture from literature, or --researcher flag set) → use research:scientist; if task is general empirical experiment NOT rooted in a paper → use foundry:sw-engineer for experiment implementation (standard agent_strategy mapping still applies; --colab alone does not force research:scientist).

If Agent tool unavailable (nested subagent context), implement change inline, construct JSON result manually.

Phase 2a — Sandbox validate (sandbox_mode = "docker" only)

loads: compute-docker.md

Follow modes/compute-docker.md — full Phase 2a and 2b logic for docker sandbox. Skip entire file if sandbox_mode = "local". Cat once, first iteration only — static content stays in context; re-cat only if no longer in context (e.g. after a compaction).

export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
IFS= read -r CLAUDE_SKILL_DIR < "${TMPDIR:-/tmp}/research-run-skill-dir-${CSID}" 2>/dev/null || CLAUDE_SKILL_DIR=""
cat "$CLAUDE_SKILL_DIR/modes/compute-docker.md"  # timeout: 5000

Phase 2b — Apply change (sandbox_mode = "docker" only)

Skip if sandbox_mode = "local" — handled in compute-docker.md above.

Phase 2c — Codex co-pilot (--codex only)

Follow modes/codex-copilot.md — contains full Phase 2c logic, cost-bounded gate, Codex dispatch prompt, outcome handling, and stuck escalation. Cat once, first --codex iteration only — static content stays in context; re-cat only if no longer in context (e.g. after a compaction).

export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
IFS= read -r CLAUDE_SKILL_DIR < "${TMPDIR:-/tmp}/research-run-skill-dir-${CSID}" 2>/dev/null || CLAUDE_SKILL_DIR=""
cat "$CLAUDE_SKILL_DIR/modes/codex-copilot.md"  # timeout: 5000

Phase 3 — Verify files changed

git diff --stat. If no files changed (no-op): append to JSONL with status: no-op, skip to Phase 8 (log), continue loop.

Phase 4 — Commit change

Refresh commit sentinel before staging — R5 loop can exceed the 15-min sentinel TTL set in R5 setup. Slug computation unavoidably re-run (bash state lost between tool calls); path pattern identical to R5 setup block above:

# refresh sentinel — bash state lost between calls, re-source slug (R5 form)
eval "$(bash "${CLAUDE_PLUGIN_ROOT:-plugins/cc_research}/bin/git_slugs.sh")"  # timeout: 3000
touch "${TMPDIR:-/tmp}/claude-commit-auth-${REPO_SLUG}-${BRANCH_SLUG}"  # timeout: 3000  # tmpdir-exempt: user-shell-boundary

Stage only modified files (never git add -A):

git add <files_modified from agent JSON>  # timeout: 3000
git commit -m "experiment(optimize/i<N>): <description>"  # timeout: 90000

If pre-commit hooks fail:

• Delegate to foundry:linting-expert: provide failing hook output and modified files; ask to fix. Max 2 attempts.
• If still failing after 2 attempts: git restore --staged <files_modified> + git restore <files_modified> to clean up (# <files_modified> = list of files returned by the iteration agent; restricts discard to iteration scope only), append status: hook-blocked, continue loop.

Phase 5 — Verify metric

loads: phase5-metric.md # also loads: codex-copilot.md, colab-setup.md, compute-docker.md, hypothesis-pipeline.md, report.md, resume.md, team.md

Follow modes/phase5-metric.md — metric verification logic for docker, local, and colab sandbox modes. Cat once, first iteration only — static content stays in context; re-cat only if no longer in context (e.g. after a compaction).

export CSID="${CLAUDE_CODE_SESSION_ID:-$PPID}"
IFS= read -r CLAUDE_SKILL_DIR < "${TMPDIR:-/tmp}/research-run-skill-dir-${CSID}" 2>/dev/null || CLAUDE_SKILL_DIR=""
cat "$CLAUDE_SKILL_DIR/modes/phase5-metric.md"  # timeout: 5000

Phase 6 — Run guard

If sandbox_mode = "docker": run guard_cmd in same Docker container as Phase 5 (same flags; no resource limits). Check exit code only.

If sandbox_mode = "local": run guard_cmd directly.

Record pass (exit 0) or fail (non-zero).

Phase 7 — Evaluate outcome

Line count computation (for "gain < 0.1% AND change > 50 lines" row): run before evaluating the condition:

DIFF_SUMMARY=$(git diff --stat HEAD~1..HEAD | tail -1)  # timeout: 3000
INSERTIONS=$(echo "$DIFF_SUMMARY" | grep -oE '[0-9]+ insertion' | grep -oE '[0-9]+' || echo 0)
DELETIONS=$(echo "$DIFF_SUMMARY" | grep -oE '[0-9]+ deletion' | grep -oE '[0-9]+' || echo 0)
CHANGE_LINES=$(( INSERTIONS + DELETIONS ))

git revert HEAD --no-edit — never git reset --hard (preserves history, not in deny list).

Double-revert guard (ADV-H19) — Phase 7 rework→revert can collide with a partial Phase 5 timeout revert or Phase 6 guard-fail revert performed in the same iteration. Always check before issuing the revert:

ALREADY_REVERTED=$(git log --oneline -5 2>/dev/null | grep -c "^[0-9a-f]\+ Revert ") || ALREADY_REVERTED=0  # `|| echo 0` appends a second 0: grep -c prints 0 *and* exits 1 on zero matches
if [ "$ALREADY_REVERTED" -gt 0 ]; then
    echo "Phase 7: prior revert detected (Phase 5/6 already reverted this iteration) — skipping double-revert."
else
    git revert HEAD --no-edit  # timeout: 15000
fi

The guard fires on metric improved AND guard fail (after GUARD_REWORK_MAX attempts exhausted), no improvement, and gain < 0.1% AND change > 50 lines paths — any path that issues a revert after Phase 5 or Phase 6 may have already reverted.

Phase 7a — Write diary

After Phase 7 decision, append one entry to diary.md:

## Iteration N — <ISO timestamp>

**Hypothesis**: <agent's description from Phase 2 JSON — the proposed change and expected improvement>

**Outcome**: <metric_key> = <value> (Δ<delta>% vs baseline) — <kept|reverted|rework|no-op|hook-blocked|timeout>

**Decision**: <one sentence: why the outcome was accepted or rejected — e.g. "Metric improved 1.2% with guard passing" or "Reverted: metric regressed by 0.5%" or "Guard failed after 2 rework attempts">

---

For no-op iterations (no file changes):

## Iteration N — <ISO timestamp>

**Hypothesis**: <descripti

…(truncated)