godmodeskill

内容来源：README.md（说明文档） · 原始地址 · 查看安装指南

原始内容

# 🧠 GodModeSkill

> **`/work`** — a single Claude Code command that runs every plan, feature, and bugfix past **3 different model families** before any merge.

Codex + Gemini + OpenCode (Kimi/DeepSeek) review your work in parallel. Quorum must agree, or Claude revises and retries. Zero token-burn while waiting. Auto-handles git. Always asks before merge.

---

## 👉 Looking for the productized version? Use **Chorus**

GodModeSkill was the bash prototype that proved the multi-LLM peer-review idea. We rebuilt it as a polished cross-platform product:

🎶 **[Chorus](https://chorus.codes)** — `npm install -g chorus-codes`

- Polished cross-platform UI (cockpit at `localhost:5050`) for fan-out reviews + run inspection
- Works with **any** AI CLI (Claude Code, Codex, Gemini, OpenCode, Kimi, Cursor, Windsurf) via MCP
- OpenRouter built in, plus per-voice fallback chains, voice auto-disable on quota, and live cost tracking
- Apache-2.0, open source: **<https://github.com/chorus-codes/chorus>**

GodModeSkill stays here as the original `/work` bash skill for Claude Code users who prefer a script. New features land on Chorus.

---

## ✨ Why this exists

🤖 **Single-model AI coding fails silently.** Code looks reasonable, tests pass — but a subtle bug or design drift only shows up later.

👯 **Same-family models share blind spots.** Three Codex sessions reviewing each other is an echo chamber.

✅ **Lineage diversity is the unlock.** A different model family reading the same code from scratch catches a startling amount of what the author missed.

---

## 🚀 Quick start

### Option 1 — paste-and-go (recommended)

Open Claude Code and paste [`INSTALL.md`](https://github.com/99xAgency/GodModeSkill/blob/HEAD/INSTALL.md). Claude asks which CLIs you have, then installs everything.

### Option 2 — manual

```bash
git clone https://github.com/99xAgency/GodModeSkill.git
cd GodModeSkill
# Read skill/ layout, then copy each file to its target (see "Install layout" below)
```

---

## 🎯 Modes

| Mode | When to use | What happens |
|------|-------------|--------------|
| 📐 `plan` | New feature design | Propose → review → converge → write `planning/<feature>.md` |
| 🛠️ `implement` | Build the agreed plan | Design review → code → test → code review → PR → merge gate |
| 🐛 `major-bug` | Real bug fix | Failing test (TDD) → fix-plan review → code → test → code review → PR |
| ✏️ `minor-bug` | Typo / one-liner | Just fix it. No LLM gates. |

```
/work plan an oauth login flow for the admin panel
/work implement the planning/oauth-login.md design
/work fix bug: detail page crashes on null verdict
/work fix typo in nav header
```

Claude picks the mode, picks 3 reviewers, builds the pack, runs the loop, and asks for sign-off at the merge gate.

---

## 📋 Prerequisites

You need at least **3 different model families**:

- 🟣 **Claude Code** — any plan
- 🟠 **Codex CLI** — ChatGPT subscription (Plus or Pro)
- 🔵 **Gemini CLI** — Google AI Pro
- 🟢 **OpenCode CLI** — one of:
  - OpenCode Go subscription (gives Kimi + DeepSeek + others, flat-fee) ← **recommended**
  - Direct API keys (Moonshot, DeepSeek)

System tools: `tmux`, `jq`, `inotify-tools`, `git`, `python3`.

> ⚠️ Only have 2 families? The lineage quorum will fail (you need 1 of each: codex + gemini + opencode). You can edit `work-converge` to relax the rule, but you lose the diversity insight.

---

## 💸 Cost

**~$0 incremental** if you use plan-priced reviewers (ChatGPT subscription + Google AI Pro + OpenCode Go).

If you use API keys for Kimi/DeepSeek instead of OpenCode Go, expect **~cents per review** depending on pack size (~50KB packs typical).

---

## 🧬 How the lineage quorum works

`work-converge` parses each reviewer's XML response, extracts the `<verdict agree="true|false|partial">` block, and groups by lineage:

| Lineage | Sessions |
|---------|----------|
| 🟠 codex | `cdx-*` |
| 🔵 gemini | `gem-*` |
| 🟢 opencode | `kimi`, `deepseek`, … |

✅ **Quorum passes** only when **at least one agent of each lineage** returns `agree`.
🟡 Partial verdicts count as agree only if no critical/high findings.

| Outcome | Exit | Meaning |
|---------|------|---------|
| 🟢 all 3 agree | `0` | merge gate opens |
| 🔴 any disagree | `1` | revise + retry |
| ⚪ lineage missing | `2` | escalate to user |

---

## ⚡ How the wait is zero-token

When `/work` is waiting for reviewers, Claude is suspended on a **single Bash tool call** running `inotifywait`:

| Metric | Value |
|--------|-------|
| 🪙 Claude tokens during wait | **0** |
| 🔋 CPU during wait | **0** (kernel inotify) |
| ⏱️ Resume latency | **<1 s** (filesystem event) |

A **5-second safety wakeup** also re-checks for permission prompts, modal popups, and provider errors — see *Resilience* below.

---

## ✅ Pre-merge checklist

Before any merge, Claude fills out and shows you:

```
1. Coding principles (CLAUDE.md): [followed | drifted: what + why]
2. Architecture guidelines:        [followed | drifted: what + why]
3. Tests:                          [PASS | FAIL: which]
4. Reviewer consensus:             [agree | overridden: reason]

Merge?
1. Yes — merge now
2. No — fix drift / failures first
3. Override and merge anyway (give reason)
```

Catches a lot of "I think it's done" moments where the model has quietly drifted.

---

## 🛡️ Resilience — what happens when a reviewer breaks

`/work` handles each failure class automatically and only escalates to Claude when it can't recover.

| 💥 Failure | 🔍 Detection (≤5 s) | 🔧 Recovery |
|---|---|---|
| **Provider error** — Moonshot/DeepSeek 5xx, rate limit, gateway hiccup, "Error from provider", context overflow | `ERROR_RE` matches in pane | Retry the nudge **once**. If a 2nd error fires within 60 s: opencode lineage gets a **peer-swap** (kimi ⇄ deepseek). Other lineages get a stub findings file ending `## DONE` so the watcher exits fast and Claude takes over. |
| **TUI modal popup** — opencode "Select agent" / "Select model" / "Ask: No results found" (triggered by `/` or `@` at chunk boundaries during paste) | `OPENCODE_POPUP_RE` matches | Send `Escape` ×2 + re-nudge. **Pre-flight `Escape ×2`** also runs before every opencode nudge (defensive). |
| **Permission prompt** — CLI prompts the reviewer for `cat`/`tee`/`bash` execution | `PROMPT_RE` matches | Auto-approve with the right keystrokes (codex `y`, gemini `2`, opencode `Right Enter` for "Allow always"). Logs the command to `approvals.log`. |
| **Destructive command** — `rm -rf`, `sudo`, `git push --force`, `DROP TABLE`, ~25 patterns | `DESTRUCTIVE_RE` matches | 🛑 **NOT auto-approved.** Agent stays stuck, orchestrator prints `tmux attach -t <session>`, logs to `destructive-blocks.log`. |

**Peer-swap works for ALL lineages, not just opencode:**

| Lineage | Swap chain | When |
|---|---|---|
| 🟠 Codex | `cdx-1 → cdx-2 → cdx-3` (or however many you have) | One ChatGPT account hits its 5h/weekly limit |
| 🔵 Gemini | `gem-1 → gem-2 → …` | First Google AI account hits daily quota |
| 🟢 Opencode | `kimi ⇄ deepseek` | OpenCode Go gateway hiccup on one provider |

`try_lineage_swap` finds any alive agent of the same `type` in `agents.json` not already in the run, copies the prompt over, /clears the partner, nudges. Stays on subscription pricing — picks from agents you've already configured.

**See what's been stuck across runs:**

```bash
work status        # snapshot of current run: who's working, errored, popup-stuck, done
work status --watch # refresh every 2s
work permissions   # last 30 auto-approvals + repeating commands + destructive blocks
```

`work status` replaces ad-hoc `tmux capture-pane` debugging. Surfaces a one-line state per agent (WORKING / ERRORED / POPUP-STUCK / PERM-PROMPT / DONE) plus log-line counts.

`work permissions` surfaces bash patterns that prompt repeatedly. Add them to your CLI config (e.g. `~/.config/opencode/opencode.json`'s `permission.bash` block) so they're allowed permanently — no orchestrator round-trip.

---

## 💬 Per-CLI prompt format

Each CLI's TUI handles multi-line input differently, so `/work` builds a **per-CLI prompt**:

| CLI | 📝 Format | 💡 Why |
|---|---|---|
| 🟠 Codex | Multi-paragraph (bracketed-paste safe) | Handles `[Pasted Content N chars]` blocks correctly |
| 🔵 Gemini | **Single line** with `@/abs/path/to/pack.xml` (inline file attach) | Every `\n` is Submit — multi-paragraph paste fragments into N separate queries |
| 🟢 Opencode | **Single line**, plain-text path (no leading `/` or `@`) | tmux splits paste at ~1.5 KB chunks; `/` or `@` at a chunk boundary opens slash-command or agent-picker mid-paste |

---

## 📦 Pack truncation safety

When a review pack exceeds `MAX_CTX_BYTES` (**800 KB** default — sized for 1M-context reviewers like Opus 4.7 / Gemini 3.1 Pro / Kimi K2.6), `work-pack-build` drops:

1. Journals
2. Memory files
3. Diff (last resort, truncated to 50 KB)

When the diff itself gets truncated, the pack emits:

```xml
<diff truncated="true" original-bytes="N" shown-bytes="50000">…</diff>
<reviewer-instruction priority="critical">
  Verify each finding against the full <code_file> blocks
  before reporting — do NOT rely on the truncated diff alone.
</reviewer-instruction>
```

Prevents reviewers from anchoring on the truncated diff and hallucinating findings about the cut-off portion.

## 🔁 Cross-round packs are diff-only

When a review goes to round 2 (after disagreement + revision), the pack **drops `<related-memory>` and `<related-journals>`** — they don't change between rounds and the reviewer already saw them in round 1. Typical saving: **~65% pack size reduction** on round 2+.

Replaced with a stub:

```xml
<stable-context-skipped reason="round-2-no-change-since-round-1"
                        memory-files-omitted="6" journals-omitted="3"/>
```

Architecture docs, code files, diff, and `<prior-rounds>` (which grows each round) are still included — all the things that matter for re-review. Reviewers are reminded in the `<ask>` block that round-2+ context was intentionally trimmed.

## 🎯 Self-consistency check (anti-hallucination)

Findings now require structured evidence:

```xml
<finding severity="critical" category="bug">
  <claim>Unquoted variable in test causes syntax error on empty input.</claim>
  <evidence>
    <file-path>scripts/greet.sh</file-path>
    <line-number>5</line-number>
    <quoted-line><![CDATA[if [ $name == "admin" ]; then]]></quoted-line>
  </evidence>
  <suggestion>Quote it: if [ "$name" = "admin" ]; then</suggestion>
</finding>
```

`work-converge` greps the cited file for the `<quoted-line>` (whitespace-tolerant). Findings whose quote can't be located are flagged `verified: false` — likely hallucinations. The output JSON exposes:

```json
{ "verified_findings": 7, "unverified_findings": 1,
  "unverified_blockers": [ /* critical/high findings that failed verification */ ] }
```

Claude can then weight unverified findings lower or ignore them. Direct counter to the gem-1-style anchor-bias hallucination class.

---

## ⚙️ Install layout (manual install only)

```
skill/work*           → ~/.local/bin/
skill/agent-status    → ~/.local/bin/
skill/commands/work.md → ~/.claude/commands/work.md
skill/agent-sessions/ → ~/.config/agent-sessions/
                        (rename agents.json.template → agents.json, edit)
skill/gemini/settings.snippet.json → merge into ~/.gemini/settings.json
skill/gemini/GEMINI.snippet.md     → append to  ~/.gemini/GEMINI.md
skill/gemini/policies/00-deny-destructive.toml → ~/.gemini/policies/
skill/opencode/opencode.snippet.json → merge into ~/.config/opencode/opencode.json
skill/examples/AGENTS.md → put in each opencode session working dir
```

Make all `skill/work*` files executable: `chmod +x ~/.local/bin/work*`

---

## 🔧 Customization

| What | Where |
|---|---|
| Quorum rule | `quorum_check()` in `work-converge` |
| Pack cap + context discovery | `MAX_CTX_BYTES` (800 KB) + discovery fns in `work-pack-build` |
| Auto-approve patterns | `PROMPT_RE` in `work` |
| Destructive block list | `DESTRUCTIVE_RE` in `work` |
| Provider-error patterns | `ERROR_RE` in `work` |
| Opencode popup patterns | `OPENCODE_POPUP_RE` in `work` |
| Opencode peer-swap map | `OPENCODE_PARTNER` in `work` (default `kimi ⇄ deepseek`) |
| Mode workflows | `~/.claude/commands/work.md` |

---

## ➕ Adding more agents

Edit `~/.config/agent-sessions/agents.json`. New entries are auto-included in the LRU rotation for their lineage.

- 4th Codex on a new ChatGPT account → `cdx-4` with its own `CODEX_HOME`
- 2nd Gemini → `gem-2` (parallel review)
- Claude reviewer in tmux → `claude-1` with `type: "claude"` (extend `work-converge` quorum logic)

---

## 🛟 Troubleshooting

<details>
<summary><strong>🪧 Reviewer hangs forever</strong></summary>

- Did the findings file end with `## DONE`?
- Is the tmux pane stuck on a permission prompt, modal popup, or "Error from provider" toast?
- Check the per-run log dir `/tmp/work-<task>-r<n>-<ts>/`:
  - `error-retries.log` — provider errors + retries + peer-swaps + escalations
  - `popup-dismissals.log` — opencode modal popups dismissed
  - `approvals.log` — bash prompts auto-approved (with command line)
  - `destructive-blocks.log` — destructive commands refused
- Run `work permissions` for the aggregate view.

</details>

<details>
<summary><strong>🚫 Quorum never passes</strong></summary>

- `agents.json` has at least one alive agent per lineage?
- `agent-status` shows nobody rate-limited?
- Look in the per-run log dir for an `<agent>-findings.md` containing `⚠️ AGENT ESCALATED TO CLAUDE` — orchestrator gave up after retries/swaps. Pick from the 3 numbered options in the stub.

</details>

<details>
<summary><strong>👻 Reviewer reports findings about code that isn't in the diff</strong></summary>

Likely the pack hit the diff truncation cap. Check `pack-meta.json` for `diff_bytes` — if it's `50026`, truncation fired. The reviewer should have seen a `<reviewer-instruction priority="critical">` warning. If they hallucinated anyway, the model isn't following the instruction — file an issue.

**Workaround:** bump `MAX_CTX_BYTES` higher in `work-pack-build`.

</details>

<details>
<summary><strong>📁 Wrong file paths</strong></summary>

- `~/.local/bin` in `$PATH`?
- All scripts use `$HOME` / `Path.home()` — should work for any Unix user.

</details>

---

## 🔒 Safety guarantees

- 🛑 Destructive shell ops blocked at **3 layers**: CLI configs, runtime guard in `work`, Gemini policy file
- 👀 Destructive prompts → agent stays **visibly stuck** so a human notices and decides
- 🚫 **Gemini does NOT run in `yolo` mode** — `yolo` has wiped repos for others (write_file overwriting source with empty content). It runs in `auto_edit`: file-write tools auto-approved, shell prompts caught and approved by orchestrator
- 📜 All blocks logged to `destructive-blocks.log` in the run directory

---

## 📜 License

MIT — do whatever you want.

---

## 🤝 Contributing

PRs welcome. Especially:

- Support for more CLI types (Claude Code as a reviewer, Ollama agents, …)
- Better permission-prompt regex patterns for new CLI versions
- Smarter context discovery in `work-pack-build`

---

## 🌱 Credits

Built incrementally over April 2026 by trial and error. The **lineage diversity** insight came from Reddit discussion about Claude blind spots in self-review loops. The **`/clear` between rounds** rule came from watching opencode CLIs go conversational mid-orchestration. The **provider-error escalation** + **kimi ⇄ deepseek peer-swap** came from real OpenCode Go gateway hiccups. The **single-line gemini prompt** came from spotting that Gemini's TUI was treating each `\n` as Submit.

🧠 GodModeSkill

/work — a single Claude Code command that runs every plan, feature, and bugfix past 3 different model families before any merge.

Codex + Gemini + OpenCode (Kimi/DeepSeek) review your work in parallel. Quorum must agree, or Claude revises and retries. Zero token-burn while waiting. Auto-handles git. Always asks before merge.

👉 Looking for the productized version? Use Chorus

GodModeSkill was the bash prototype that proved the multi-LLM peer-review idea. We rebuilt it as a polished cross-platform product:

🎶 Chorus — npm install -g chorus-codes

Polished cross-platform UI (cockpit at localhost:5050) for fan-out reviews + run inspection
Works with any AI CLI (Claude Code, Codex, Gemini, OpenCode, Kimi, Cursor, Windsurf) via MCP
OpenRouter built in, plus per-voice fallback chains, voice auto-disable on quota, and live cost tracking
Apache-2.0, open source: https://github.com/chorus-codes/chorus

GodModeSkill stays here as the original /work bash skill for Claude Code users who prefer a script. New features land on Chorus.

✨ Why this exists

🤖 Single-model AI coding fails silently. Code looks reasonable, tests pass — but a subtle bug or design drift only shows up later.

👯 Same-family models share blind spots. Three Codex sessions reviewing each other is an echo chamber.

✅ Lineage diversity is the unlock. A different model family reading the same code from scratch catches a startling amount of what the author missed.

🚀 Quick start

Option 1 — paste-and-go (recommended)

Open Claude Code and paste INSTALL.md. Claude asks which CLIs you have, then installs everything.

Option 2 — manual

git clone https://github.com/99xAgency/GodModeSkill.git
cd GodModeSkill
# Read skill/ layout, then copy each file to its target (see "Install layout" below)

🎯 Modes

Mode	When to use	What happens
📐 `plan`	New feature design	Propose → review → converge → write `planning/<feature>.md`
🛠️ `implement`	Build the agreed plan	Design review → code → test → code review → PR → merge gate
🐛 `major-bug`	Real bug fix	Failing test (TDD) → fix-plan review → code → test → code review → PR
✏️ `minor-bug`	Typo / one-liner	Just fix it. No LLM gates.

/work plan an oauth login flow for the admin panel
/work implement the planning/oauth-login.md design
/work fix bug: detail page crashes on null verdict
/work fix typo in nav header

Claude picks the mode, picks 3 reviewers, builds the pack, runs the loop, and asks for sign-off at the merge gate.

📋 Prerequisites

You need at least 3 different model families:

🟣 Claude Code — any plan
🟠 Codex CLI — ChatGPT subscription (Plus or Pro)
🔵 Gemini CLI — Google AI Pro
🟢 OpenCode CLI — one of:
- OpenCode Go subscription (gives Kimi + DeepSeek + others, flat-fee) ← recommended
- Direct API keys (Moonshot, DeepSeek)

System tools: tmux, jq, inotify-tools, git, python3.

⚠️ Only have 2 families? The lineage quorum will fail (you need 1 of each: codex + gemini + opencode). You can edit work-converge to relax the rule, but you lose the diversity insight.

💸 Cost

~$0 incremental if you use plan-priced reviewers (ChatGPT subscription + Google AI Pro + OpenCode Go).

If you use API keys for Kimi/DeepSeek instead of OpenCode Go, expect ~cents per review depending on pack size (~50KB packs typical).

🧬 How the lineage quorum works

work-converge parses each reviewer's XML response, extracts the <verdict agree="true|false|partial"> block, and groups by lineage:

Lineage	Sessions
🟠 codex	`cdx-*`
🔵 gemini	`gem-*`
🟢 opencode	`kimi`, `deepseek`, …

✅ Quorum passes only when at least one agent of each lineage returns agree. 🟡 Partial verdicts count as agree only if no critical/high findings.

Outcome	Exit	Meaning
🟢 all 3 agree	`0`	merge gate opens
🔴 any disagree	`1`	revise + retry
⚪ lineage missing	`2`	escalate to user

⚡ How the wait is zero-token

When /work is waiting for reviewers, Claude is suspended on a single Bash tool call running inotifywait:

Metric	Value
🪙 Claude tokens during wait	0
🔋 CPU during wait	0 (kernel inotify)
⏱️ Resume latency	<1 s (filesystem event)

A 5-second safety wakeup also re-checks for permission prompts, modal popups, and provider errors — see Resilience below.

✅ Pre-merge checklist

Before any merge, Claude fills out and shows you:

1. Coding principles (CLAUDE.md): [followed | drifted: what + why]
2. Architecture guidelines:        [followed | drifted: what + why]
3. Tests:                          [PASS | FAIL: which]
4. Reviewer consensus:             [agree | overridden: reason]

Merge?
1. Yes — merge now
2. No — fix drift / failures first
3. Override and merge anyway (give reason)

Catches a lot of "I think it's done" moments where the model has quietly drifted.

🛡️ Resilience — what happens when a reviewer breaks

/work handles each failure class automatically and only escalates to Claude when it can't recover.

💥 Failure	🔍 Detection (≤5 s)	🔧 Recovery
Provider error — Moonshot/DeepSeek 5xx, rate limit, gateway hiccup, "Error from provider", context overflow	`ERROR_RE` matches in pane	Retry the nudge once. If a 2nd error fires within 60 s: opencode lineage gets a peer-swap (kimi ⇄ deepseek). Other lineages get a stub findings file ending `## DONE` so the watcher exits fast and Claude takes over.
TUI modal popup — opencode "Select agent" / "Select model" / "Ask: No results found" (triggered by `/` or `@` at chunk boundaries during paste)	`OPENCODE_POPUP_RE` matches	Send `Escape` ×2 + re-nudge. Pre-flight `Escape ×2` also runs before every opencode nudge (defensive).
Permission prompt — CLI prompts the reviewer for `cat`/`tee`/`bash` execution	`PROMPT_RE` matches	Auto-approve with the right keystrokes (codex `y`, gemini `2`, opencode `Right Enter` for "Allow always"). Logs the command to `approvals.log`.
Destructive command — `rm -rf`, `sudo`, `git push --force`, `DROP TABLE`, ~25 patterns	`DESTRUCTIVE_RE` matches	🛑 NOT auto-approved. Agent stays stuck, orchestrator prints `tmux attach -t <session>`, logs to `destructive-blocks.log`.

Peer-swap works for ALL lineages, not just opencode:

Lineage	Swap chain	When
🟠 Codex	`cdx-1 → cdx-2 → cdx-3` (or however many you have)	One ChatGPT account hits its 5h/weekly limit
🔵 Gemini	`gem-1 → gem-2 → …`	First Google AI account hits daily quota
🟢 Opencode	`kimi ⇄ deepseek`	OpenCode Go gateway hiccup on one provider

try_lineage_swap finds any alive agent of the same type in agents.json not already in the run, copies the prompt over, /clears the partner, nudges. Stays on subscription pricing — picks from agents you've already configured.

See what's been stuck across runs:

work status        # snapshot of current run: who's working, errored, popup-stuck, done
work status --watch # refresh every 2s
work permissions   # last 30 auto-approvals + repeating commands + destructive blocks

work status replaces ad-hoc tmux capture-pane debugging. Surfaces a one-line state per agent (WORKING / ERRORED / POPUP-STUCK / PERM-PROMPT / DONE) plus log-line counts.

work permissions surfaces bash patterns that prompt repeatedly. Add them to your CLI config (e.g. ~/.config/opencode/opencode.json's permission.bash block) so they're allowed permanently — no orchestrator round-trip.

💬 Per-CLI prompt format

Each CLI's TUI handles multi-line input differently, so /work builds a per-CLI prompt:

CLI	📝 Format	💡 Why
🟠 Codex	Multi-paragraph (bracketed-paste safe)	Handles `[Pasted Content N chars]` blocks correctly
🔵 Gemini	Single line with `@/abs/path/to/pack.xml` (inline file attach)	Every `\n` is Submit — multi-paragraph paste fragments into N separate queries
🟢 Opencode	Single line, plain-text path (no leading `/` or `@`)	tmux splits paste at ~1.5 KB chunks; `/` or `@` at a chunk boundary opens slash-command or agent-picker mid-paste

📦 Pack truncation safety

When a review pack exceeds MAX_CTX_BYTES (800 KB default — sized for 1M-context reviewers like Opus 4.7 / Gemini 3.1 Pro / Kimi K2.6), work-pack-build drops:

Journals
Memory files
Diff (last resort, truncated to 50 KB)

When the diff itself gets truncated, the pack emits:

<diff truncated="true" original-bytes="N" shown-bytes="50000">…</diff>
<reviewer-instruction priority="critical">
  Verify each finding against the full <code_file> blocks
  before reporting — do NOT rely on the truncated diff alone.
</reviewer-instruction>

Prevents reviewers from anchoring on the truncated diff and hallucinating findings about the cut-off portion.

🔁 Cross-round packs are diff-only

When a review goes to round 2 (after disagreement + revision), the pack drops <related-memory> and <related-journals> — they don't change between rounds and the reviewer already saw them in round 1. Typical saving: ~65% pack size reduction on round 2+.

Replaced with a stub:

<stable-context-skipped reason="round-2-no-change-since-round-1"
                        memory-files-omitted="6" journals-omitted="3"/>

Architecture docs, code files, diff, and <prior-rounds> (which grows each round) are still included — all the things that matter for re-review. Reviewers are reminded in the <ask> block that round-2+ context was intentionally trimmed.

🎯 Self-consistency check (anti-hallucination)

Findings now require structured evidence:

<finding severity="critical" category="bug">
  <claim>Unquoted variable in test causes syntax error on empty input.</claim>
  <evidence>
    <file-path>scripts/greet.sh</file-path>
    <line-number>5</line-number>
    <quoted-line><![CDATA[if [ $name == "admin" ]; then]]></quoted-line>
  </evidence>
  <suggestion>Quote it: if [ "$name" = "admin" ]; then</suggestion>
</finding>

work-converge greps the cited file for the <quoted-line> (whitespace-tolerant). Findings whose quote can't be located are flagged verified: false — likely hallucinations. The output JSON exposes:

{ "verified_findings": 7, "unverified_findings": 1,
  "unverified_blockers": [ /* critical/high findings that failed verification */ ] }

Claude can then weight unverified findings lower or ignore them. Direct counter to the gem-1-style anchor-bias hallucination class.

⚙️ Install layout (manual install only)

skill/work*           → ~/.local/bin/
skill/agent-status    → ~/.local/bin/
skill/commands/work.md → ~/.claude/commands/work.md
skill/agent-sessions/ → ~/.config/agent-sessions/
                        (rename agents.json.template → agents.json, edit)
skill/gemini/settings.snippet.json → merge into ~/.gemini/settings.json
skill/gemini/GEMINI.snippet.md     → append to  ~/.gemini/GEMINI.md
skill/gemini/policies/00-deny-destructive.toml → ~/.gemini/policies/
skill/opencode/opencode.snippet.json → merge into ~/.config/opencode/opencode.json
skill/examples/AGENTS.md → put in each opencode session working dir

Make all skill/work* files executable: chmod +x ~/.local/bin/work*

🔧 Customization

What	Where
Quorum rule	`quorum_check()` in `work-converge`
Pack cap + context discovery	`MAX_CTX_BYTES` (800 KB) + discovery fns in `work-pack-build`
Auto-approve patterns	`PROMPT_RE` in `work`
Destructive block list	`DESTRUCTIVE_RE` in `work`
Provider-error patterns	`ERROR_RE` in `work`
Opencode popup patterns	`OPENCODE_POPUP_RE` in `work`
Opencode peer-swap map	`OPENCODE_PARTNER` in `work` (default `kimi ⇄ deepseek`)
Mode workflows	`~/.claude/commands/work.md`

➕ Adding more agents

Edit ~/.config/agent-sessions/agents.json. New entries are auto-included in the LRU rotation for their lineage.

4th Codex on a new ChatGPT account → cdx-4 with its own CODEX_HOME
2nd Gemini → gem-2 (parallel review)
Claude reviewer in tmux → claude-1 with type: "claude" (extend work-converge quorum logic)

🛟 Troubleshooting

🪧 Reviewer hangs forever

Did the findings file end with ## DONE?
Is the tmux pane stuck on a permission prompt, modal popup, or "Error from provider" toast?
Check the per-run log dir /tmp/work-<task>-r<n>-<ts>/:
- error-retries.log — provider errors + retries + peer-swaps + escalations
- popup-dismissals.log — opencode modal popups dismissed
- approvals.log — bash prompts auto-approved (with command line)
- destructive-blocks.log — destructive commands refused
Run work permissions for the aggregate view.

🚫 Quorum never passes

agents.json has at least one alive agent per lineage?
agent-status shows nobody rate-limited?
Look in the per-run log dir for an <agent>-findings.md containing ⚠️ AGENT ESCALATED TO CLAUDE — orchestrator gave up after retries/swaps. Pick from the 3 numbered options in the stub.

👻 Reviewer reports findings about code that isn't in the diff

Likely the pack hit the diff truncation cap. Check pack-meta.json for diff_bytes — if it's 50026, truncation fired. The reviewer should have seen a <reviewer-instruction priority="critical"> warning. If they hallucinated anyway, the model isn't following the instruction — file an issue.

Workaround: bump MAX_CTX_BYTES higher in work-pack-build.

📁 Wrong file paths

~/.local/bin in $PATH?
All scripts use $HOME / Path.home() — should work for any Unix user.

🔒 Safety guarantees

🛑 Destructive shell ops blocked at 3 layers: CLI configs, runtime guard in work, Gemini policy file
👀 Destructive prompts → agent stays visibly stuck so a human notices and decides
🚫 Gemini does NOT run in yolo mode — yolo has wiped repos for others (write_file overwriting source with empty content). It runs in auto_edit: file-write tools auto-approved, shell prompts caught and approved by orchestrator
📜 All blocks logged to destructive-blocks.log in the run directory

📜 License

MIT — do whatever you want.

🤝 Contributing

PRs welcome. Especially:

Support for more CLI types (Claude Code as a reviewer, Ollama agents, …)
Better permission-prompt regex patterns for new CLI versions
Smarter context discovery in work-pack-build

🌱 Credits

Built incrementally over April 2026 by trial and error. The lineage diversity insight came from Reddit discussion about Claude blind spots in self-review loops. The /clear between rounds rule came from watching opencode CLIs go conversational mid-orchestration. The provider-error escalation + kimi ⇄ deepseek peer-swap came from real OpenCode Go gateway hiccups. The single-line gemini prompt came from spotting that Gemini's TUI was treating each \n as Submit.