adversarial-code-loop

BUILD → REVIEW → (FIX → VERIFY)^N → ARBITER on isolated git branches. Git-native: each loop runs on its own branch, changes are committed, reviews inspect git diffs.

Adversarial Code Loop v4

BUILD → REVIEW → (FIX → VERIFY)^N → ARBITER. A sequential pipeline where one model writes code, another critiques the git diff, the first fixes, the second validates, and an optional arbiter resolves the last disagreement. Every loop runs on its own git branch; each BUILD/FIX is a commit; reviews inspect real git diffs; the result is squash- merged into the parent branch (or marked [REJECTED]).

Rule: the orchestrator never writes code directly. This skill delegates code to DEV/FIXER agents (codex, claude-tmux, pi). The orchestrator writes the spec, launches the pipeline, and interprets the results. Never use patch/write/bash to edit code inside a task covered by this skill — always go through the DEV role. If no DEV agent is configured explicitly, use pi with the current model.

Based on Multi-Persona adversarial debate (Smit et al., ICML 2024): each role gets a distinct persona, which improves quality even when both roles share the same model.

When to use: code that must be reviewed by another model before delivery (breaking the echo chamber), critical code (security, auth, money), and well-scoped multi-file refactors (up to ~15 files with a structured spec). Not for simple questions, trivial 1-file changes, or open-ended design exploration.

Installation

Requires the adversarial-common sibling repo (shared engine). One-line install:

curl -fsSL https://raw.githubusercontent.com/chpomob/adversarial-code-loop/main/scripts/install.sh | bash

or, from an existing checkout:

bash scripts/install.sh

Both place adversarial-code-loop and adversarial-common side by side under ~/.hermes/skills (override the target with $1 or $HERMES_HOME).

Overview — what's new in v4

v4 is git-native. Where v3 wrote files directly to the worktree and reviewed a stdin concatenation of file contents, v4 isolates every loop on a dedicated branch and reviews real diffs.

Concern	v3	v4
Isolation	none — writes to live worktree	dedicated branch loop/<feature>/<N>
Review input	concatenated file contents (stdin)	git diff <branch-point>..HEAD
BUILD/FIX output	prose/JSON the orchestrator extracts	files committed by the model
Recovery on failure	manual file salvage	git reset/git checkout to restore
Merge	manual git add -A	squash-merge into parent branch
Rejection	exit code only	[REJECTED] marker commit + branch preserved
Resume	not supported	--resume from state.json
JSON robustness	strict json.loads	strip_json_wrapper parses markdown-fenced JSON
Gates	none	optional --build-cmd / --test-cmd
Result contract	final.json + exit code	final.json + exit code (unchanged, enriched)

Workflow

text

PHASE 0 ──→ GIT SETUP   (detect/init repo, stash dirty tree, record branch-point,
                          create loop/<feature>/<N>, bootstrap git identity, gitignore)
PHASE 1 ──→ BUILD        (DEV writes code, orchestrator stages + commits "build: ...")
                          [optional --build-cmd gate]
PHASE 2 ──→ REVIEW       (model on git diff <branch-point>..HEAD → JSON findings)
PHASE 3 ──→ FIX          (DEV addresses findings, orchestrator commits "fix: ... (round N)")
PHASE 4 ──→ VERIFY       (model checks each finding resolved | rejected | disputed)
   loop 3-4 until APPROVED or --max-loops reached
PHASE 5 ──→ ARBITER      (optional; resolves disputes after max-loops)
                          [optional --test-cmd gate]
MERGE     ──→ squash-merge into parent + evidence tag (APPROVED / ARBITRATED)
              or [REJECTED] marker commit, loop branch preserved (REJECT)

PHASE 0–5 and the merge are implemented as thin wrappers in scripts/phases/ (phase_git, phase_build, phase_review, phase_fix, phase_verify, phase_arbiter). The shared engine — subprocess runner, JSON I/O, provider detection, git operations — lives in the adversarial-common sibling skill.

CLI flags

Resolution order per command role: CLI flag > env var > built-in default. The built-in defaults name specific tools/models (see table) but are overridable; set the env vars or flags to point at your own DEV/REVIEW/ARBITER CLIs.

Flag	Env	Default	Description
--spec	—	(required)	Specification file to implement
--workdir	—	.	Working directory (subprocess cwd, base of --out)
--dev-cmd	ACL_DEV_CMD	codex exec --skip-git-repo-check --sandbox workspace-write	DEV (BUILDER/FIXER) command
--review-cmd	ACL_REVIEW_CMD	pi --provider zai --model glm-5.2	REVIEW (CRITIC/VERIFIER) command
--arbiter-cmd	ACL_ARBITER_CMD	— (unset = no arbiter)	ARBITER (JUDGE) command, optional
--max-loops	—	3	Max FIX/VERIFY cycles
--no-arbiter	—	off	Skip arbitration; REJECT instead
--timeout	—	600	Per-subprocess timeout (s)
--build-cmd	—	—	Build gate run after BUILD (e.g. cargo build)
--test-cmd	—	—	Test gate run before merge (e.g. cargo test)
--no-merge	—	off	On approval, leave the loop branch unmerged
--feature	—	spec filename	Feature name used for branch + artifact dir
--out	—	.adversarial-loop	Artifact output directory (under --workdir if relative)
--resume	—	off	Resume from state.json
--provider-config	—	~/.config/adversarial/providers.yaml	External provider config for quota-aware provider selection (see "Quota-aware provider selection" section)
--force	—	off	Bypass all quota checks for all roles, use first configured provider regardless of state
--force-provider	—	—	Repeatable: --force-provider <role>:<alias> bypasses quota for a single role (e.g. --force-provider review:deepseek). Other roles still check quotas normally

Env-var support is limited by design. As of v4.0.0 the orchestrator honors only the three command env vars above (ACL_DEV_CMD, ACL_REVIEW_CMD, ACL_ARBITER_CMD). ACL_WORKDIR, ACL_MAX_LOOPS, ACL_TIMEOUT, and ACL_OUT_DIR are not read by the current code — pass those values via flags. (The names are reserved so future releases can wire them without breaking existing invocations.)

REVIEW/VERIFY commands are passed through privilege reduction (the pipeline strips known dangerous CLI flags like --dangerously-bypass-approvals-and-sandbox and --yolo from review commands), but the pipeline does NOT enforce OS-level containment (no kernel sandbox, no network cutoff, no filesystem jail). The SandboxMode enum in adversarial_common is advisory metadata, not a security boundary. Reviewers SHOULD use the least-privilege sandbox their CLI provides.

Exit codes

Code	Meaning
0	APPROVED — squash-merged into the parent branch
1	Infrastructure failure — phase crash, timeout, git error, interrupt
2	Usage error — bad flag, missing/unreadable --spec, missing/bad --workdir
3	REJECT — findings unresolved after --max-loops, or --build-cmd/--test-cmd gate failed, or empty BUILD diff. Loop branch is preserved.
4	ARBITRATED — arbiter approved; conditions recorded in final.json

Orchestrators consuming the pipeline should read final.json (the machine-readable contract), not the exit code.

Findings JSON schema

REVIEW output (one model call, validated by phase_review._validate; retried once on malformed JSON):

json

{
  "findings": [
    {"id": "A1",
     "severity": "blocker|major|minor|nit",
     "file": "path/to/file.rs",
     "line": 42,
     "summary": "Short title",
     "evidence": "Why it matters, referencing real code in the diff"}
  ],
  "verdict": "REQUEST_CHANGES|APPROVE|REJECT"
}

line must be an integer (numeric strings are tolerated). Findings lacking an id receive a deterministic auto_<hash> id so VERIFY can track them across rounds.

VERIFY output (validates each finding's resolution against the current diff):

json

{
  "results": [
    {"id": "A1", "status": "resolved|rejected|disputed"}
  ],
  "verdict": "APPROVE|REJECT"
}

• resolved — the problematic code is gone or corrected.
• rejected — the verifier disagrees with the original finding (it was wrong).
• disputed — unclear; stays open for the next round or the arbiter.

Approval requires verdict == APPROVE and every finding settled (resolved or rejected). A finding the verifier rejected does not block approval.

Artifacts

Emitted under <--out>/<feature>/ (auto-appended to .gitignore so they never merge):

File	Phase	Contents
state.json	0	Resumability: completed phases, current loop, branch, branch-point SHA, stash id, findings
00_spec.txt	1	Spec verbatim
01_build.json	1	BUILD result + commit SHA
01_build_gate.json	1	--build-cmd gate (if set)
02_review.json	2	Findings + verdict
03_fix_<N>.json	3	FIX round N result (one per loop)
04_verdict_<N>.json	4	VERIFY round N results + verdict (one per loop)
05_arbiter.json	5	Arbiter verdict + conditions (if run)
06_test_gate.json	6	--test-cmd gate (if set)
final.md	end	Human-readable summary (also the evidence-tag annotation)
final.json	end	Machine-readable contract — verdict, reason, loops, branch, merged, conditions, arbitrated, artifacts_dir

Git workflow

Auto-init. If gitops.detect_enclosing_repo(workdir) finds a parent repo, it is used as-is. Otherwise gitops.auto_init initializes one (initial branch pinned to main). The parent branch is the current branch (or main after auto-init).

Dirty working tree. gitops.stash_dirty runs git stash push -u at PHASE 0 and records stash@{0} in state.json. The stash is popped on every exit path (success, reject, interrupt) via _restore. If git stash pop hits a conflict (the parent branch advanced and touched the same lines), the loop aborts with exit 1 and a human must resolve — the stash is preserved, nothing is lost.

Branch naming. loop/<sanitized-feature>/<N>, where N is one more than the highest existing N under that prefix (starts at 1). --feature is sanitized to a branch-safe slug; default is the --spec filename stem.

Commits. BUILD commits build: <feature> — <summary>; each FIX round commits fix: <feature> — address finding(s) (round N). An empty BUILD diff is still committed (empty commit allowed) but triggers an EMPTY_DIFF REJECT at REVIEW. Git identity (user.name/user.email) is bootstrapped on the loop branch if unset.

Reviews on diffs. REVIEW and VERIFY receive git diff <branch-point>..HEAD, so they see the cumulative change since the branch point — every BUILD + all FIX rounds — and each finding must reference code that actually appears in the diff.

Merge (APPROVED / ARBITRATED). gitops.squash_merge checks out the parent branch, runs git merge --squash <loop-branch>, commits squash: <feature> — adversarial approved, and drops the loop branch. A merge conflict aborts with exit 1 and keeps the loop branch. Before merging, tag_with_evidence creates an annotated tag <loop-branch>-approved carrying final.md (best-effort — a missing file never blocks the merge). --no-merge skips the merge and leaves the loop branch for human review.

Reject (REJECT). gitops.reject_marker records an empty [REJECTED] <feature> — <verdict> commit on the loop branch. The branch is not deleted and not merged, so the rejected work is recoverable.

Language discipline

All internal pipeline text is English: spec files, auto-generated commit messages, personas (builder.md, critic.md, …), findings JSON, verdicts, synthesis reports, and code comments. User-facing summaries (what the orchestrator prints to you) stay in your conversation language. Do not language-switch between roles inside the pipeline — it confuses the model, especially in FIX where it receives an English persona + English review + possibly a non-English spec.

Personas

BUILDER / CRITIC / FIXER / VERIFIER / JUDGE live as text files in ~/.hermes/skills/adversarial-common/personas/ (single source of truth, editable without touching Python). All v4 personas are git-aware: BUILD produces committed code, REVIEW inspects a diff, FIX commits a new round, VERIFY checks findings against the diff. Injection is provider-aware: pi is detected and selects builder-pi.md/fixer-pi.md (tool-based writes instead of markdown/JSON code output — mitigates the prose-overwrite failure, pitfall #6).

Examples (validated)

bash

# Basic — Codex DEV + GLM-5.2 REVIEW, default flags.
python3 ~/.hermes/skills/adversarial-code-loop/scripts/adversarial_loop.py \
  --spec /tmp/spec.md --workdir /path/to/project

# Claude-as-DEV via claude-tmux (Fable 5 / Opus). Use ABSOLUTE paths — `~` expands
# relative to --workdir, not $HOME (pitfall #11). Extended thinking runs 8-12 min,
# so push --timeout up and keep the inner --hard-timeout >= the loop timeout.
python3 ~/.hermes/skills/adversarial-code-loop/scripts/adversarial_loop.py \
  --spec /tmp/spec.md --workdir /path/to/project \
  --dev-cmd "python3 /path/to/claude-tmux.py --model best --timeout 900 --hard-timeout 2400 --max-turns 20" \
  --timeout 2400

# GLM-5.2 DEV + DeepSeek REVIEW (thinking high on both). No Claude quota needed.
python3 ~/.hermes/skills/adversarial-code-loop/scripts/adversarial_loop.py \
  --spec /tmp/spec.md --workdir /path/to/project \
  --dev-cmd  "pi -p --provider zai --model glm-5.2 --thinking high" \
  --review-cmd "pi -p --provider deepseek --model deepseek-v4-pro --thinking high" \
  --max-loops 2 --no-arbiter --timeout 1200

# With build + test gates and a named feature (Rust project).
python3 ~/.hermes/skills/adversarial-code-loop/scripts/adversarial_loop.py \
  --spec /tmp/spec.md --workdir /path/to/project --feature peer-auth \
  --build-cmd "cargo build" --test-cmd "cargo test" \
  --max-loops 3 --timeout 1800

# Arbiter on, no merge (human reviews the loop branch first).
python3 ~/.hermes/skills/adversarial-code-loop/scripts/adversarial_loop.py \
  --spec /tmp/spec.md --workdir /path/to/project \
  --arbiter-cmd "pi -p --provider gemini --model gemini-3-pro" --no-merge

# Resume after an interrupt (reads state.json under --out/<feature>/).
python3 ~/.hermes/skills/adversarial-code-loop/scripts/adversarial_loop.py \
  --spec /tmp/spec.md --workdir /path/to/project --resume

Wrapper compatibility: claude-tmux wrapper v1 rejects --yolo with argparse exit code 2. Omit that flag; its permission-bypass behavior is already the default.

Model pairing notes: Codex is a fast first-choice DEV; GLM-5.2 (pi) reviews thoroughly and reliably returns JSON; DeepSeek REVIEW is slower but finds more findings; Claude (via tmux) is the most thorough reviewer but slowest and quota-bound. Codex FIX often cascades beyond spec scope (migrates consumers, fixes adjacent bugs) — check git diff --stat after every loop before assuming REJECT means the code is wrong.

Quota-aware provider selection (available): The pipeline now supports --provider-config, --force, and --force-provider <role>:<alias> flags. When a provider config is loaded, each phase checks real-time quotas and auto-selects the best available command per role, with fallback chains defined externally. The provider config lives in the user's ~/.config/adversarial/providers.yaml by default. See the spec at adversarial-spec's spec.md for the full design.

Pitfalls

1. --plan mode is NOT wired into adversarial_loop_v4.py's argparse. The actual Python code has no --plan argument. The adversarial_loop.py entry point (which re-exports v4) only accepts --spec; phase_plan.py is not imported and has no CLI entry point. Symptom: passing --plan still reports that --spec is required. Fix: run each step as a separate code loop with --spec pointed at a focused spec. Do NOT rely on --plan; it is not implemented as of 2026-07-15.

1b. Codex --sandbox read-only vs --dangerously-bypass-approvals-and-sandbox. When Codex is the REVIEWER and you add --dangerously-bypass-approvals-and-sandbox, it silently overrides --sandbox read-only to --sandbox danger-full-access, giving the reviewer write access — the opposite of what you want. Fix: for read-only review use --sandbox read-only without the bypass flag (interactive approval only); for a writing DEV use --sandbox danger-full-access --dangerously-bypass-approvals-and-sandbox together. In non-interactive mode the approval flag is required or Codex hangs. 2. Bound the loop with --max-loops. The arbiter settles the last disagreement; it does not extend the loop. 3. GLM JSON wrapped in markdown — parsed in v4, not v3. v3 used strict json.loads and choked on ```json-fenced output. v4's jsonio.strip_json_wrapper strips fences and extracts the largest JSON object, so GLM-5.2 / Claude markdown-wrapped JSON is now parsed. Since P6, every phase (including VERIFY) routes through the shared 3-strategy parser adversarial_common.jsonio.parse_json_output, which tries: (1) markdown stripping, (2) extracting {...} via text.find('{')..rfind('}'), (3) extracting [...] for raw arrays. This makes the pipeline model-agnostic — the same code works regardless of whether the model returns raw JSON, markdown-wrapped JSON, text + JSON, or a JSON array. REVIEW/VERIFY still retry once on malformed JSON. If a model returns prose with no JSON object at all, the phase fails (exit 1) — check the captured stdout in the artifact. 4. Claude extended thinking runs 8-12 min (Fable 5). Pass --timeout 900 --hard-timeout 2400 inside the claude-tmux command and keep the loop's --timeout >= 2400. The inner --timeout controls tmux pane inactivity detection — if Claude goes silent for more than this period, the pane is killed. With extended thinking, Claude can be silent for 12+ minutes even on small codebases (validated 2026-07-13 on a ~2580-line plugin: first review timed out at 600s). Never use --timeout 600 or lower for Fable 5 REVIEW — the first pass always has the longest thinking burst as it reads the full diff and project structure. Set --hard-timeout 2400 (40 min) to survive Verifier passes that require multiple file reads. If Claude repeatedly times out, switch --review-cmd to GLM-5.2 (pi -p --provider zai --model glm-5.2 --thinking high), which is faster (no extended thinking) and equally reliable for JSON output. See references/wrapper-failures.md, references/fable5-timeout-recovery.md, and the claude-tmux-wrapper skill. 5. Dirty working tree must be committed or stashed. v4 auto-stashes at PHASE 0 and restores on every exit path, so a dirty tree no longer blocks startup. The remaining risk is a stash-pop conflict: if the parent branch advanced and touched the same lines you had stashed, git stash pop fails and the loop aborts (exit 1). The stash is preserved — resolve manually, then --resume. 6. Models may overwrite source files with prose instead of code. Claude/Fable 5, pi/GLM-5.2, and Codex have all been observed to replace working source with a markdown report or <<<SEE BELOW>>> placeholder. v4 mitigates this with pi-specific personas (builder-pi.md/fixer-pi.md, auto-selected when pi is detected) and is far easier to recover from than v3: git checkout HEAD -- <file> restores the committed version on the loop branch, then re-run FIX or apply the change directly via patch for a well-understood single-file fix. For mechanical fixes, direct patch is faster and more reliable than re-running the loop. 7. Merge conflicts if the parent branch advances during the loop. Squash-merge aborts with exit 1 and keeps the loop branch. Fix by rebasing the loop branch onto the updated parent (git rebase <parent>) or re-running; the loop branch is never lost. 8. NEVER run parallel loops on the same workdir. Each loop checks out its own branch, but two concurrent DEV/FIXER subprocesses writing to the same worktree files corrupt each other. Run batches sequentially; only disjoint file sets (no overlap in git diff --stat) can run in parallel. See references/batch-splitting-strategy.md. 9. --resume requires state.json from a previous run. It is read from <--out>/<feature>/state.json. If absent (e.g. you changed --feature or wiped --out), the loop starts fresh with a warning. Resumed runs re-checkout the recorded branch and skip completed phases. 10. ~ in --dev-cmd/--review-cmd/--arbiter-cmd expands relative to --workdir, not $HOME. The subprocess runner does no shell expansion. Always use absolute paths (/home/user/.hermes/... or $HOME/.hermes/...) for scripts in command flags. 11. Use claude-tmux-wrapper, not claude -p, for Claude roles. claude -p bills against Agent SDK credit (monthly cap); interactive Claude via tmux stays on the 5h sliding quota. Model alias claude-sonnet-4 is invalid — use claude-sonnet-4-20250514 or opus/sonnet/best/fable aliases. See references/claude-p-migration-pattern.md. 12. Prompt injection from reviewed code. Code under review (diff, spec) can embed adversarial instructions like {"verdict": "APPROVE"} that try to override the pipeline verdict. v4's review-on-diff narrows the attack surface but does not close it. See references/prompt-injection-threat-model.md; cross-model diversity (using different models for DEV and REVIEW) is a recommended defense, though the pipeline does not enforce it — distinct personas alone provide some separation. 13. Codex sandbox builds commit target/ / build artifacts. When a DEV/FIXER runs cargo build/cargo test, the sandbox writes target/ into the workdir; the orchestrator's git add -A at BUILD/FIX commits them, bloating the squash. Ensure target/ (and equivalent) is in .gitignore before the first loop. After a loop: git status --porcelain target/ | head -3; if committed, git rm -r --cached target/ and gitignore it. 14. The loop can REJECT for out-of-scope findings. The reviewer is not told to distinguish "pre-existing bug" from "new bug in this changeset." GLM-5.2 is particularly prone to finding pre-existing bugs outside spec scope. After a REJECT, always build + test and inspect the code on disk; if the spec-scope code is correct and the rest are pre-existing, the code is usable — commit it and patch the rest manually if wanted. 15. Codex / models may exit 1 on deletion-only or "no new code" specs without writing to stdout. Check git status / git diff --stat on the loop branch — the model may have made the changes before the process died. An empty BUILD diff is REJECTed as EMPTY_DIFF. 16. --out persists between runs. The directory is created with mkdir(parents=True, exist_ok=True) and not cleaned. Re-running with a different spec in the same project: either rm -rf .adversarial-loop first, or use a distinct --out / --feature. 17. No parallel loops sharing a branch namespace — the monotonic <N> counter in loop/<feature>/<N> is read from existing refs at PHASE 0; two concurrent starts can pick the same N and clobber each other. Sequential launches are safe. 18. Codex / OpenAI quota exhaustion kills REVIEW silently. Codex has usage limits, especially on free/Plus tiers. When exhausted (ERROR: You've hit your usage limit), the review phase exits 1 with no useful output. Detection: before a long loop, check quota with a quick CODE-only call (no reasoning). Fallback: switch --review-cmd to a non-OpenAI provider (GLM-5.2, DeepSeek, Claude). If Codex is the only reviewer configured, prepare a fallback inline or skip the review pass. Codex quota resets at the start of each month (OpenAI billing cycle). See references/ai-quota-apis.md. 19. User preference: never say "I'll check back in X minutes" without actually doing it. When monitoring a long-running loop, use an explicit polling loop (for i in 1..N; do sleep 30; ls artifacts/; done) or rely on notify_on_complete=true. Passive promises without follow-through frustrate the user. Either monitor actively with a polling loop, or say nothing and let the notification fire. See references/monitoring-long-running-loops.md. Validated 2026-07-14: the user called out the agent twice in one session for saying "I'll check back" without doing it. The agent said "je revérifie dans 3 min" and the reply was "tu as encore menti". This is a hard constraint: either launch a real polling loop now, or use notify_on_complete and stay silent. Never end a turn with a future-monitoring promise.

text

**Concrete pattern that was validated:** launch with
`terminal(background=true, notify_on_complete=true)` and do other work. When
mid-run progress checks are needed, use a compact `for` loop with `sleep 30`
that checks for specific artifact files (`02_review.json`, `loop_1_04_verdict.json`,
`final.json`).

20. DeepSeek via pi requires ~/.pi/agent/auth.json. Hermes stores the DeepSeek API key in ~/.hermes/.env but does NOT export it to subprocesses. To use DeepSeek through pi, create ~/.pi/agent/auth.json with: {"deepseek": {"type": "api_key", "key": "<key>"}}. Extract the key from Hermes via grep DEEPSEEK_API_KEY ~/.hermes/.env (the file has the actual key — Hermes masks it in terminal output but the file is readable by Python). Set permissions to 0600. See references/pi-auth-setup.md. 21. terminal(background=true) with notify_on_complete=true is the recommended monitoring pattern. Long loops (5+ minutes per phase) should run in the background. The preferred approach: launch the loop with background=true + notify_on_complete=true, then work on other tasks. The notification fires automatically on completion. If you must monitor mid-run, use a compact polling loop: for i in 1..N; do sleep 30; ls artifacts/; done. Avoid idle waiting — do other work while the loop runs. 22. DeepSeek V4 Pro VERIFY JSON can be malformed. DeepSeek with --thinking high occasionally wraps JSON in additional markdown or text, causing strip_json_wrapper to fail extraction. The retry also fails because the model repeats the same wrapping. Symptoms: REVIEW succeeds (findings parsed), but VERIFY fails with "invalid JSON after retry". Mitigation: switch --review-cmd to a model that reliably outputs raw JSON (GLM-5.2 is more reliable for VERIFY). Or check the code on the loop branch manually — BUILD and FIX commits are correct even when VERIFY fails. Validated 2026-07-06 with GLM+DeepSeek pairing. 23. Review prompt no longer concatenates code — model reads files directly from the loop branch checkout. The review prompt is under 1K tokens. The reviewer runs git diff HEAD~1..HEAD to see changes and reads files with cat/grep for context. See references/review-on-committed-code.md. 24. GLM-5.2 quota is 80 prompts per rolling 5h (Z.AI Lite). HTTP 429 after 2-3 heavy loops. Recovery: switch to DeepSeek V4 Pro (pi -p --provider deepseek --model deepseek-v4-pro --thinking high) for DEV, or Claude Sonnet for REVIEW. If all providers exhausted, wait 5h for GLM reset. 25. User preference — monitor actively or stay silent. Use polling loops or notify_on_complete=true. Never promise to "check back" without following through. 26. User preference — quality over speed. Always use --thinking high. Set generous timeouts (--timeout 2400). Accept 10-15 min BUILD times. 27. Pipeline workdir == Hermes Agent install directory (fork-as-live-install). When --workdir points at the Hermes Agent repo and Hermes is running that checkout, the pipeline's git operations (branch creation, checkout, squash-merge) operate on the live codebase. A squash-merge into the parent branch (typically main) without --no-merge commits the loop output directly into your running Hermes install — which can leave the install in an inconsistent state mid-change. Always use --no-merge so the loop branch stays isolated for human review and manual merge. After review, merge deliberately: git checkout main && git merge --squash <loop-branch>. Also, auto-stash of dirty trees (pitfall #5) is riskier here: a stash-pop conflict during the pipeline aborts with exit 1 and leaves the working tree in a mixed state while Hermes is trying to run from those same files. Pre-commit or stash manually before launching. See references/fork-as-live-install.md.

28. .gitignore auto-modification leaks into upstream PRs. The pipeline's PHASE 0 appends --out patterns (.adversarial-loop/ by default) to .gitignore so artifacts never get tracked. This is correct for local development, but the .gitignore change ends up in every BUILD commit (via git add -A) and propagates into the squash merge. When the loop output is destined for an upstream PR, drop the .gitignore delta before pushing. After squash-merge into the parent branch: check with git diff HEAD~1..HEAD -- .gitignore; if it shows artifact patterns, restore the upstream version with git checkout HEAD -- .gitignore and amend: git commit --amend --no-edit. For --no-merge loops: inspect .gitignore before the manual merge — the upstream .gitignore likely already has target/ etc., so a diff showing only .adversarial-loop/, *.orig, *.rej is the signal. See references/pre-pr-cleanup.md.

29. Keep REVIEW/VERIFY timeout propagation wired end to end. run_review() and run_verify() accept a timeout parameter and pass it to providers.run_cmd(); both call sites in adversarial_loop.py pass timeout=args.timeout. This makes the pipeline's --timeout apply to all five phases. Preserve all three links when changing phase signatures or dispatch. A regression causes Claude Fable 5 REVIEW or VERIFY to fail with exit code 124: TIMEOUT after 600s even when the caller passed --timeout 2400. See references/fable5-timeout-recovery.md for the validated reproduction and implementation details.

30. claude-tmux wrapper must NOT modify the pipeline prompt. The wrapper exists to capture output via tmux instead of claude -p. Its only addition to the pipeline's stdin is the output-capture instruction; never add behavioral modifiers that duplicate or contradict the pipeline prompt. See references/claude-tmux-prompt-hygiene.md for the validated wrapper pattern.

31. Squash commit naming for upstream PRs. The pipeline's merge commit message format is "squash: <feature> — adversarial approved" (individual BUILD commits use "build: <feature> — <summary>", FIX commits use "fix: <feature> — address finding(s) (round N)"). These are pipeline-internal names that don't follow conventional commits, and upstream reviewers will flag them. After squash-merge into the parent branch, rewrite the squash commit: git commit --amend -m "feat(cli): add on_status_bar_render hook to narrow width tier". For several squash commits stacked together, either rebase and reword each, or squash them all into one conventional-format commit before pushing the branch upstream. Always verify the final commit message with git log --oneline -1 before git push. See references/pre-pr-cleanup.md.

32. pi (GLM-5.2) can review the wrong git repo despite correct cwd. Although pi runs inside subprocess.Popen(cwd=workdir), its internal file-access tools may navigate to a different repository. Symptom: the REVIEW finding references a commit hash and file paths that don't exist in --workdir (e.g., from hermes-agent instead of a plugin repo), claiming an empty diff. Diagnosis: check 02_review.json — if the "file" field says "(commit 92ce650...)" instead of a real file path in your project, pi is in the wrong repo. Workaround: merge the BUILD manually (git merge --squash <loop-branch>); the code on the loop branch is correct, only the review was misdirected. This was validated 2026-07-13 on a 320-line keyring-hardening step where GLM reviewed the hermes-agent repo instead of a plugin repo. After manual merge the code compiled and all 149 tests passed.

33. Fable 5 has its own usage limit separate from Claude Pro's 5h sliding quota. The model can be blocked even when regular Claude Pro quota is green. Symptom: claude-tmux starts, bypasses permissions, reads the prompt, then displays "You've reached your Fable 5 limit" and stops. Fix: switch to --model sonnet or --model opus. Sonnet is preferred for plan-challenger and code-loop REVIEW because it has no extended thinking (faster response, no 12-min silence), reliable JSON output, and lower token cost. See references/fable5-usage-limit.md. Validated: 2026-07-15 — Fable 5 hit limit mid-challenge; Sonnet completed in ~2 min.

34. Codex FIX phase hangs on stdin when the spec is small or findings are minor. Codex prints Reading prompt from stdin... and blocks forever when its generated input does not constitute a complete code-generation request. Symptom: BUILD succeeds, REVIEW returns findings, but FIX exits 1 with Reading prompt from stdin... as the only output. Root cause: the FIX phase embeds findings into a prompt Codex expects to be a full coding task; narrow specs with minor findings can leave Codex waiting. Mitigation (validated 2026-07-15): switch --dev-cmd from Codex to GLM-5.2 (pi -p --provider zai --model glm-5.2 --thinking high) for the problematic step. GLM reliably handles FIX prompts without stdin hang. Permanent fix: ensure the FIX prompt always includes a concrete code-generation request with file paths and expected diff pattern.

35. Untracked spec files (step-P.md) can be lost across sequential code loops.* PHASE 0 runs git stash push -u which captures untracked files. If you prepare multiple step specs in the same repo and run sequential loops, each loop stashes the untracked spec files from the previous loop's stash pop. After a squash-merge, the stash is dropped — and the untracked spec files are gone. Symptom: a code loop with --spec step-P3-spec.md that ran fine earlier now fails with "Spec not found" because the file was stashed and dropped by a previous loop. Fix: add step spec patterns to .gitignore (e.g. step-P*-spec.md) so they are never stashed. Or keep step specs outside the repo's workdir and reference them by absolute path in --spec.

36. hermes update breaks after loops run on the live Hermes install (fork-as-live-install fallout). When the pipeline workdir is the Hermes Agent git install (~/.hermes/hermes-agent, origin=fork, upstream=NousResearch) and loops left the repo sitting on a feature branch with local main carrying squash commits, hermes update fails. Hermes update (fork installs) switches to main and runs git pull --ff-only origin main; with main diverged (local adversarial squashes) and the current branch 1000s of commits behind upstream, the pull can't fast-forward and the update dies on the conflict. Validated 2026-07-31: install sat on feat/status-bar-hook-all-widths (4 commits, 4393 behind upstream/main) with main carrying 3 adversarial squash commits → hermes update failed. Recovery (full recipe with exact commands in references/rebase-pr-onto-upstream.md):

    1. Diagnose: git status, git branch -vv, git log --oneline upstream/main..HEAD, gh pr view <branch> --repo <upstream-repo> --json mergeable,mergeStateStatus (CONFLICTING confirms the drift).
    2. Backup: git branch backup/<name> <parent-branch> + git tag backup/<feature>-pre-rebase HEAD.
    3. git fetch upstream main then git rebase upstream/main on the feature branch (commits replay one at a time; each may conflict).
    4. Conflict resolution: keep BOTH sides — upstream's new code AND the feature's code (validated: upstream had added _status_bar_goal_segment/battery_prefix/focus_label in the same status-bar function the feature was extending with _get_status_bar_plugin_values; the correct merge keeps both methods and folds the feature's parts/plugin_values into upstream's new narrow-width branch).
    5. Non-interactive rebase --continue: GIT_EDITOR=true git rebase --continue — a plain git rebase --continue fails with "There was a problem with the editor" when stdin isn't a TTY (agent terminal). The GIT_EDITOR=true trick applies to any non-interactive rebase/commit.
    6. Verify: python3 -m py_compile <touched files>, targeted venv/bin/python -m pytest <test files>, then the repo's canonical runner (scripts/run_tests.sh <test files> — CI-parity, hermetic env).
    7. git push --force-with-lease origin <branch> → PR flips to MERGEABLE (BLOCKED = review required, normal).
    8. Realign main + sync fork: git branch -f main upstream/main && git checkout main, then git push origin main (fast-forward safe — check git merge-base --is-ancestor origin/main upstream/main first).
    9. Confirm the fix: hermes update --check → "Already up to date."
    10. Leave the install on main (healthy for updates); the feature branch stays for PR work. Only main should track upstream; feature/loop work lives on branches.

37. The loop can APPROVE without satisfying textual acceptance criteria. Validated 2026-07-31 twice on the same step (remove challenge-prompt embedding in adversarial-plan): both loops were APPROVED while phase_challenge.py was left untouched — the BUILD only adjusted tests/docs to match the existing code, the reviewer/verifier accepted the diff, and pytest stayed green. The spec's AC was a grep-level invariant the pipeline never checks mechanically. Fix: when an acceptance criterion is a grep/source invariant (e.g. "no plan_text in phase_challenge.py", "no adversarial_loop_v3"), run the grep yourself AFTER the loop reports APPROVED — do not trust the verdict. Prefer ACs that are enforced by a test the BUILD must add (sentinel-absence assertions) over prose ACs, and say so in the spec ("a test must assert X").

38. REVIEW/VERIFY timeout propagation regression (pitfall #34, re-fixed). After the v4.1.0 fix, the timeout=args.timeout argument was dropped again from run_review() (adversarial_loop_v4.py:886) and the normal-path run_verify() (:1074) call sites — only the concurrent verify path kept it. Symptom: REVIEW exited 124: TIMEOUT after 600s even when --timeout 2400 is passed, on large diffs (e.g. 5000+ line re-tracking diffs). Fixed in 6e7bbed by re-adding timeout=args.timeout to both call sites. When adding phase-parameter plumbing, grep all call sites of phase_review.run_review / phase_verify.run_verify — the concurrent path is a separate call site.

39. The DEV can write OUTSIDE the loop workdir — check sibling repos after every loop. Validated 2026-08-01: the P16b loop (workdir = adversarial-spec) was APPROVED and squash-merged, but the DEV had implemented part of the fix (stash on_pushed plumbing) in the SHARED repo adversarial-common via a relative path (../adversarial-common/...), leaving 4 uncommitted files there that the spec-repo loop never committed and the spec squash didn't include. The loop's git isolation only protects the workdir's branch; any path reachable from the workdir is writable. Fix: after each loop that touches shared modules (adversarial-common), run git status --porcelain on ALL sibling repos before declaring the step done — and commit any out-of-workdir changes with a message naming the originating step.

Retrospective logging

Every pipeline failure is automatically logged to _retrospective/ISSUES.md with:

• Phase name, branch, error message, and last 200 chars of stdout
• Date/time of failure
• Feature name from the spec

Review issues before planning v5:

bash

cat ~/.hermes/skills/adversarial-code-loop/_retrospective/ISSUES.md

To manually add a note about a limitation you noticed, add an entry at the top of _retrospective/ISSUES.md following the same format:

markdown

### YYYY-MM-DD — Short title

- **Model combo:** GLM/DeepSeek/Claude/Codex + (role)
- **Symptom:** What went wrong
- **Root cause:** Why it happened
- **Fix/workaround:** How you worked around it
- **Would fix in v5 by:** Concrete design change

Changelog

• v4.1.3 (2026-07-31): Added pitfall #37 (loop can APPROVE without satisfying textual acceptance criteria — verify grep-level invariants yourself) and pitfall #38 (reviewer timeout propagation regression, re-fixed in 6e7bbed).

• v4.1.2 (2026-07-31): Reconciled the implemented single-spec v4 workflow across 17 files: removed unsupported --plan workflows, obsolete v3 entrypoint/migration documentation, and invalid claude-tmux --yolo examples; clarified that --plan is not implemented; and renumbered the retained pitfalls consecutively.

• v4.1.1 (2026-07-31): Added pitfall #42 (now #36; hermes update breaks after loops on the live install — fork-as-live-install fallout) and reference rebase-pr-onto-upstream.md (validated recipe: backup → rebase onto upstream/main → keep-both-sides conflict resolution → GIT_EDITOR=true git rebase --continue → canonical scripts/run_tests.sh → --force-with-lease push → realign main + sync fork).

• v4.1.0 (2026-07-14): Added pitfalls #25b (multi-repo parent repo auto-init), #26 (claude-tmux --cwd required), #34 (REVIEW/VERIFY timeout propagation), #35 (mid-pipeline model fallback), #36 (plan resume after partial completion), and #38 (pi wrong repo review), using the numbering at release. Added references: fable5-timeout-recovery.md, github-secret-scanning-bypass.md, delegate-task-review-timeout.md. Fixed run_review() and run_verify() timeout propagation bug (timeout defaulted to 600s despite pipeline --timeout). Fixed _fail_phase → fail_phase bug in adversarial_review.py (underscore prefix leaked from runner module).