[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7C4VCJJ7455HMVRWTYGJQ5V":3,"skill-same:01M7C4VCJJ7455HMVRWTYGJQ5V":159},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":14,"repositoryHtmlUrl":27,"repositoryHomepageUrl":6,"ownerHtmlUrl":28,"ownerBlogUrl":29,"ownerDescription":6,"updatedAt":30,"install":31,"resources":44},"01M7C4VCJJ7455HMVRWTYGJQ5V","agada-bench",null,"BrainLayer gold bench: agada bench, BL PR, embedder\u002Fenrichment change, weekly drift, gold build. NOT one-off search.","EtanHey\u002Fgolems","EtanHey","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F93338119?v=4","01M7C4SKC9HJC3B9TAX0D9MZ1E","Etan Heyman","01M7C4SKCT59Z49TTN4WWDDEYD",0,"Apache-2.0","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: agada-bench\ndescription: \"BrainLayer gold bench: agada bench, BL PR, embedder\u002Fenrichment change, weekly drift, gold build. NOT one-off search.\"\n---\n\n# \u002Fagada-bench — Standing BrainLayer Quality Benchmark\n\n> Score live BrainLayer recall against the frozen 4-domain gold corpus. Run whenever BrainLayer changes (PR merges, FM fixes ship, enrichment updates) to detect regressions or confirm improvements.\n\n## The Mission\n\n```\nMISSION = a regression\u002Fimprovement scorecard for the current BrainLayer build.\nNot \"ran some brain_search calls.\" Not \"produced numbers.\"\nDone = bench-report.md exists + per-domain breakdown + regression diff vs prior baseline + brain_store TASK_DONE.\n```\n\nAnything short of a complete scorecard with provenance-tagged hits is a partial run — flag it loudly, don't summarize over it.\n\n---\n\n## Two modes\n\n### Mode 1 (default) — `\u002Fagada-bench` runs the bench\n\nScore live BL recall against the standing gold. Use this 95% of the time. Triggers: post-PR-merge audit, weekly drift check, post-enrichment-schema-change verification, post-embedder-swap regression sweep.\n\nSee `workflows\u002Frun-bench.md` for the 6-wave procedure.\n\n### Mode 2 (rare subcommand) — `\u002Fagada-bench build` extends the standing corpus\n\nAdd a new user-domain to the frozen gold when a distinct retrieval topic becomes important enough to benchmark. Spawns 3 judges in cmux panes and runs the full 7-phase build pipeline.\n\nSee `workflows\u002Fbuild-new-domain.md` for the build pipeline.\n\n---\n\n## What is and isn't\n\n| Is | Isn't |\n|---|---|\n| The standing scorecard for BrainLayer health. | A one-off query test. Just call brain_search yourself. |\n| Anti-placebo (every hit is classified true_hit \u002F echo_fm11 \u002F downstream \u002F uncertain \u002F metadata_gap). | A vanilla recall@K calculator. Provenance matters more than the number. |\n| Frozen gold: 4 domains × 67 queries × 198 pairs (53 L2+L3 relevant). | A flexible-rubric eval. Rubric is v1.1 locked. |\n| Regression-diff-aware (compare vs prior bench summary JSON). | A trend monitor. Each run produces its summary; the operator chains them by passing `--baseline`. |\n| Operates with Python helper + Claude\u002FCodex session driving live brain_search via MCP. | A pure CLI tool. Python can't reach the MCP; live calls happen in the operator's session. |\n\n---\n\n## Invocation\n\n### Bench mode (default)\n\n```bash\n# Run-anywhere invocation from a Claude\u002FCodex session with BL MCP\n\u002Fagada-bench\n\n# Score only a subset of domains (e.g., skip architecture's LOW-POWER noise explicitly)\n\u002Fagada-bench --domains techgym,freelance,recruiting\n\n# With baseline for regression diff\n\u002Fagada-bench --baseline \u003Cartifact-dir>\u002Fprevious-bench-summary.json\n\n# Custom output dir + K values\n\u002Fagada-bench --output-dir \u003Cartifact-dir>\u002F --k 1,3,5,10,20,50\n\n# Include LOW-POWER architecture in aggregates (default: excluded)\n\u002Fagada-bench --include-low-power\n```\n\nWhat the skill does under the hood:\n\n1. `scripts\u002Frun-bench.py prepare` loads the 4 frozen gold paths + each domain's `phase-0b-corpus\u002Fcorpus.jsonl` → emits `bench-queries.jsonl` with 67 queries + verbatim brain_search args.\n2. Operator (Claude\u002FCodex session) reads `workflows\u002Frun-bench.md` and executes Waves 1–6: fires live brain_search per query, classifies each returned chunk for provenance (anti-placebo), appends to `bench-results.jsonl`.\n3. `scripts\u002Frun-bench.py score` reads results, computes recall@K \u002F MRR \u002F precision@5 \u002F placebo rate \u002F regression diff → emits `\u003Cdate>-brainlayer-quality-bench-results.md` + `\u003Cdate>-bench-summary.json`.\n4. brain_store TASK_DONE chunk with headline metrics + verdict.\n\n### Build mode (rare subcommand)\n\n```bash\n# Add a new domain to the standing corpus\n\u002Fagada-bench build --session ~\u002F.claude\u002Fprojects\u002F...\u002F\u003Cnew-session>.jsonl --domain health-whoop\n\n# With 4-way panel (Runs 2-4 historical setup)\n\u002Fagada-bench build \\\n  --session \u003Cjsonl> \\\n  --domain family \\\n  --judges claude,codex,cursor,gemini \\\n  --schema v1.1-3p-1s \\\n  --primary-judges claude,codex,cursor \\\n  --shadow-judge gemini\n```\n\nSee `workflows\u002Fbuild-new-domain.md` for the full 7-phase build pipeline (corpus extract → judge dispatch → liveness gate → crossref → κ → gold-lock → pending-RT).\n\n---\n\n## The standing gold corpus (4 frozen domains)\n\n| Domain | Gold rows | L2+L3 pairs | Power | Source |\n|---|---:|---:|---|---|\n| techgym | 58 | 33 | HIGH | `\u003Cgold-root>\u002Ftechgym\u002Fphase-3-gold\u002Fgold.jsonl` |\n| freelance | 53 | 13 | HIGH | `\u003Cgold-root>\u002Ffreelance\u002Fphase-3-gold\u002Fgold.jsonl` |\n| recruiting | 47 | 6 | BORDERLINE | `\u003Cgold-root>\u002Frecruiting\u002Fphase-3-gold\u002Fgold.jsonl` |\n| architecture | 40 | 1 | **LOW-POWER** | `\u003Cgold-root>\u002Farchitecture\u002Fphase-3-gold\u002Fgold.jsonl` |\n| **Total** | **198** | **53** | — | 67 unique queries |\n\nLOW-POWER means recall@K is statistically meaningless on that domain (single positive sample). Excluded from cross-domain aggregates by default.\n\nVerbatim brain_search args live in each domain's `phase-0b-corpus\u002Fcorpus.jsonl` (sibling dir of gold). See `references\u002Fprior-runs.md` for the full provenance.\n\n---\n\n## Locked decisions (bench mode)\n\n| Setting | Default | Why | Override condition |\n|---|---|---|---|\n| Domains | all 4 | Most representative spread. | `--domains` to subset; useful when isolating a single-domain regression. |\n| LOW-POWER inclusion | excluded | Run 4 (architecture) has L2+L3=1 → recall@K noise dominates. | `--include-low-power` only when specifically auditing architecture. |\n| K values | `1,3,5,10,20,50` | Standard IR eval set. | `--k` to override. |\n| Anti-placebo | mandatory | All BL retrievals must be classified (true_hit \u002F echo_fm11 \u002F downstream \u002F uncertain \u002F metadata_gap). | Cannot disable. Skipping placebo classification = corrupt scorecard. |\n| Rubric | v1.1 frozen | Tonight's gold uses v1.1. Don't iterate. | v2 trigger (see `references\u002Froadmap-v2.md`). |\n| brain_store on completion | yes | TASK_DONE chunk with headline metrics. | `--no-store` for dry-run smokes. |\n\n---\n\n## Locked decisions (build mode)\n\nWhen extending the standing corpus with a new domain, the same defaults from v1 build apply:\n\n| Setting | Default | Why |\n|---|---|---|\n| `--judges` | `claude,codex,gemini` (3) | W3.2 κ matrix — codex is most independent voter; cursor most redundant with claude (κ=0.811 on Runs 2-4 avg \u002F 0.915 on Run 1 alone). |\n| `--schema` | `v1.1-3p` (3 primary, no shadow) | Matches the default 3-judge panel; cleaner than the 3-primary-+-1-shadow split. |\n| `--rubric-version` | `v1.1` | FM12 = 0\u002F145 in production. |\n| `--liveness-check` | `strict` (≥95% rows, fail-loud, exit 2) | Closes W3.1's silent-gemini-absent hole from Run 4. |\n| `--pending-rt-cascade` | `opus-4-7` | Functional queue-routing token accepted by the parser; v1 queues rows without dispatching a model. |\n| Tiebreaker | `claudeJudge` | Cleanest calibration (mean 86.6); 0 hallucinations on Runs 2\u002F3\u002F4. |\n\nResolve the actual cascade adjudicator through `claude.judgment` using\n`node scripts\u002Fmodel-roles.mjs claude.judgment --field model` from the golems checkout.\nChoose effort per `\u002Flarge-plan` phase and pass it explicitly at dispatch.\n\n### Evidence: W3.3 cascade calibration\n\nW3.3 recorded 12 of 13 v1.1 pending-RT cases as FM6-PreCompact single-judge outliers.\nIts original recipe said \"cheap Opus resolves\"; current adjudication uses the role above.\n\nSee `references\u002Fjudge-panel.md` for the full κ rationale.\n\n---\n\n## When to use bench mode\n\n- **After ANY BrainLayer change** that could affect retrieval (PR merges, schema updates, embedder swaps, FM fixes shipping).\n- **Periodic drift checks** (weekly \u002F biweekly) even when nothing visible changed — to catch silent regressions.\n- **Pre-flight before a major BL refactor** — establish the baseline summary that future runs compare against.\n\n## When NOT to use bench mode\n\n- One-off \"does brain_search work\" check — just call brain_search directly.\n- Domain-specific deep-dive that doesn't need the full 198-pair corpus — use `brain_search` ad-hoc.\n- Testing a non-BrainLayer retrieval system — agada-bench is BL-specific.\n\n## When to use build mode\n\n- A new user-domain has emerged. Build a fresh ~50-row gold corpus so the bench can score recall on it.\n- A domain's source session is replaceable with a much-better representative session (rare — usually we add a new domain rather than rebuild an old one).\n\n## When NOT to use build mode\n\n- The 4 existing domains cover your audit need — just run bench mode.\n- You're tempted to \"rebuild techgym with a fresher session.\" The standing gold is frozen for a reason — rebuilding breaks the prior-baseline regression chain.\n- You're tempted to iterate the rubric. Don't. That's a v2 task.\n\n---\n\n## v1 known limitations (deferred to v2)\n\n### Build-mode limits (from extract-corpus.py)\n\nAuto-extraction from session JSONL only has access to what BL renders in the brain_search tool_result text:\n\n1. **chunk_id is a BL-rendered prefix** (e.g., `rt-0c2e3cb8-`), not the full chunk_id. Tonight's hand-curated runs used a live BL DB lookup to resolve prefixes; v1 doesn't.\n2. **chunk_full_content is BL's one-line truncation**, not the verbatim chunk body. Judges grade based on this truncation.\n3. **adjacent_chunks is always empty** — surfacing neighbors requires a live DB lookup.\n\nAll gated by the same fix: extract-corpus.py needs a live `brain_recall(chunk_id_prefix)` lookup to resolve prefixes and pull full bodies. That's a v2 task; see `references\u002Froadmap-v2.md` Phase A.5.\n\n### Bench-mode limits (from run-bench.py)\n\n1. **Python can't call MCP** — live brain_search must happen in the operator's Claude\u002FCodex session.\n2. **Anti-placebo classification is operator-driven** — relies on the operator inspecting chunk_created_iso and source_session_id per the checklist in `workflows\u002Frun-bench.md`. Errors here corrupt every downstream metric.\n3. **Run 4 LOW-POWER**: architecture (1 L2+L3 pair) is excluded from aggregates by default. Acceptable for v1; v2 may rebuild architecture with a richer source session.\n\n---\n\n## v2 roadmap (deferred, documented)\n\nPer W1.7's skeptical pass, v1.1 is SHIP. v2 triggers only when the flip conditions in `references\u002Froadmap-v2.md` fire (FM12 > 0 on any run, second judge dies, etc.). v2 buildout sketches:\n\n- Phase A: Platt-scale verbalized confidence + per-domain calibration (W1.1 + W1.3)\n- Phase A.5: chunk_id-prefix-to-full resolution via live `brain_recall` (extract-corpus.py limitation)\n- Phase B: Abstention τ=0.55 + risk-coverage curve (W1.5)\n- Phase C: Reasoning-tree audit module (W1.2 ADOPT-LITE)\n- Phase D: Production-modal hardening (W1.6 applicable subset)\n\n---\n\n## Composability\n\nThis skill is referenced by:\n- `\u002Forc` — when sweeping multi-domain regressions.\n- The post-PR-merge ritual after BrainLayer PRs land.\n\nThis skill references:\n- `\u002Fnever-fabricate` — anti-placebo classification is mandatory; no \"true_hit\" without verified `chunk_created_iso` vs `query_asked_at`.\n- `\u002Fcmux-agents` — only build mode uses cmux pane spawn (for judge dispatch); bench mode runs entirely in the operator's session.\n- the false-green-gate hook — bench-report.md verdict must say `GREEN | YELLOW | RED` based on actual numbers, not approximations.\n\n---\n\n## Quick reference\n\n### Bench mode (default — the daily-use case)\n\n```bash\n# Run inside a Claude\u002FCodex session with BL MCP connected\npython3 \"$GOLEMS_ROOT\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fscripts\u002Frun-bench.py\" prepare \\\n  --gold techgym:\u003Cgold-root>\u002Ftechgym\u002Fphase-3-gold\u002Fgold.jsonl \\\n  --gold freelance:\u003Cgold-root>\u002Ffreelance\u002Fphase-3-gold\u002Fgold.jsonl \\\n  --gold recruiting:\u003Cgold-root>\u002Frecruiting\u002Fphase-3-gold\u002Fgold.jsonl \\\n  --gold architecture:\u003Cgold-root>\u002Farchitecture\u002Fphase-3-gold\u002Fgold.jsonl \\\n  --output \u003Cartifact-dir>\u002F$(date +%Y-%m-%d)-bench-queries.jsonl\n\n# Operator: read workflows\u002Frun-bench.md, do the 6 waves of brain_search, write bench-results.jsonl\n\npython3 \"$GOLEMS_ROOT\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fscripts\u002Frun-bench.py\" score \\\n  --queries \u003Cartifact-dir>\u002F$(date +%Y-%m-%d)-bench-queries.jsonl \\\n  --results \u003Cartifact-dir>\u002F$(date +%Y-%m-%d)-bench-results.jsonl \\\n  --output \u003Cartifact-dir>\u002F$(date +%Y-%m-%d)-brainlayer-quality-bench-results.md \\\n  --json-out \u003Cartifact-dir>\u002F$(date +%Y-%m-%d)-bench-summary.json \\\n  --baseline \u003Cprevious-summary.json>\n```\n\nOr via the top-level driver:\n\n```bash\nbash \"$GOLEMS_ROOT\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fscripts\u002Frun-agada.sh\"\n```\n\n### Build mode (rare subcommand)\n\n```bash\n# Add a new domain (run once per new domain — a few times a year)\nbash \"$GOLEMS_ROOT\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fscripts\u002Frun-agada.sh\" build \\\n  --session \u003Cnew-domain-session.jsonl> \\\n  --domain \u003Cnew-domain-name> \\\n  --output \u003Cartifact-dir>\u002F2026-MM-DD-agada-\u003Cnew-domain>\u002F\n```\n\n### Prior runs index\n\nSee `references\u002Fprior-runs.md` for the 4 standing-corpus domains and the dates they were locked.\n\n*End SKILL.md. Primary workflow: `workflows\u002Frun-bench.md`. Build workflow: `workflows\u002Fbuild-new-domain.md`. Per-judge briefs: `adapters\u002F`. Standing-gold provenance: `references\u002Fprior-runs.md`.*\n","# \u002Fagada-bench — Standing BrainLayer Quality Benchmark\n\n> Score live BrainLayer recall against the frozen 4-domain gold corpus. Run whenever BrainLayer changes (PR merges, FM fixes ship, enrichment updates) to detect regressions or confirm improvements.\n\n## The Mission\n\n```\nMISSION = a regression\u002Fimprovement scorecard for the current BrainLayer build.\nNot \"ran some brain_search calls.\" Not \"produced numbers.\"\nDone = bench-report.md exists + per-domain breakdown + regression diff vs prior baseline + brain_store TASK_DONE.\n```\n\nAnything short of a complete scorecard with provenance-tagged hits is a partial run — flag it loudly, don't summarize over it.\n\n---\n\n## Two modes\n\n### Mode 1 (default) — `\u002Fagada-bench` runs the bench\n\nScore live BL recall against the standing gold. Use this 95% of the time. Triggers: post-PR-merge audit, weekly drift check, post-enrichment-schema-change verification, post-embedder-swap regression sweep.\n\nSee `workflows\u002Frun-bench.md` for the 6-wave procedure.\n\n### Mode 2 (rare subcommand) — `\u002Fagada-bench build` extends the standing corpus\n\nAdd a new user-domain to the frozen gold when a distinct retrieval topic becomes important enough to benchmark. Spawns 3 judges in cmux panes and runs the full 7-phase build pipeline.\n\nSee `workflows\u002Fbuild-new-domain.md` for the build pipeline.\n\n---\n\n## What is and isn't\n\n| Is | Isn't |\n|---|---|\n| The standing scorecard for BrainLayer health. | A one-off query test. Just call brain_search yourself. |\n| Anti-placebo (every hit is classified true_hit \u002F echo_fm11 \u002F downstream \u002F uncertain \u002F metadata_gap). | A vanilla recall@K calculator. Provenance matters more than the number. |\n| Frozen gold: 4 domains × 67 queries × 198 pairs (53 L2+L3 relevant). | A flexible-rubric eval. Rubric is v1.1 locked. |\n| Regression-diff-aware (compare vs prior bench summary JSON). | A trend monitor. Each run produces its summary; the operator chains them by passing `--baseline`. |\n| Operates with Python helper + Claude\u002FCodex session driving live brain_search via MCP. | A pure CLI tool. Python can't reach the MCP; live calls happen in the operator's session. |\n\n---\n\n## Invocation\n\n### Bench mode (default)\n\n```bash\n# Run-anywhere invocation from a Claude\u002FCodex session with BL MCP\n\u002Fagada-bench\n\n# Score only a subset of domains (e.g., skip architecture's LOW-POWER noise explicitly)\n\u002Fagada-bench --domains techgym,freelance,recruiting\n\n# With baseline for regression diff\n\u002Fagada-bench --baseline \u003Cartifact-dir>\u002Fprevious-bench-summary.json\n\n# Custom output dir + K values\n\u002Fagada-bench --output-dir \u003Cartifact-dir>\u002F --k 1,3,5,10,20,50\n\n# Include LOW-POWER architecture in aggregates (default: excluded)\n\u002Fagada-bench --include-low-power\n```\n\nWhat the skill does under the hood:\n\n1. `scripts\u002Frun-bench.py prepare` loads the 4 frozen gold paths + each domain's `phase-0b-corpus\u002Fcorpus.jsonl` → emits `bench-queries.jsonl` with 67 queries + verbatim brain_search args.\n2. Operator (Claude\u002FCodex session) reads `workflows\u002Frun-bench.md` and executes Waves 1–6: fires live brain_search per query, classifies each returned chunk for provenance (anti-placebo), appends to `bench-results.jsonl`.\n3. `scripts\u002Frun-bench.py score` reads results, computes recall@K \u002F MRR \u002F precision@5 \u002F placebo rate \u002F regression diff → emits `\u003Cdate>-brainlayer-quality-bench-results.md` + `\u003Cdate>-bench-summary.json`.\n4. brain_store TASK_DONE chunk with headline metrics + verdict.\n\n### Build mode (rare subcommand)\n\n```bash\n# Add a new domain to the standing corpus\n\u002Fagada-bench build --session ~\u002F.claude\u002Fprojects\u002F...\u002F\u003Cnew-session>.jsonl --domain health-whoop\n\n# With 4-way panel (Runs 2-4 historical setup)\n\u002Fagada-bench build \\\n  --session \u003Cjsonl> \\\n  --domain family \\\n  --judges claude,codex,cursor,gemini \\\n  --schema v1.1-3p-1s \\\n  --primary-judges claude,codex,cursor \\\n  --shadow-judge gemini\n```\n\nSee `workflows\u002Fbuild-new-domain.md` for the full 7-phase build pipeline (corpus extract → judge dispatch → liveness gate → crossref → κ → gold-lock → pending-RT).\n\n---\n\n## The standing gold corpus (4 frozen domains)\n\n| Domain | Gold rows | L2+L3 pairs | Power | Source |\n|---|---:|---:|---|---|\n| techgym | 58 | 33 | HIGH | `\u003Cgold-root>\u002Ftechgym\u002Fphase-3-gold\u002Fgold.jsonl` |\n| freelance | 53 | 13 | HIGH | `\u003Cgold-root>\u002Ffreelance\u002Fphase-3-gold\u002Fgold.jsonl` |\n| recruiting | 47 | 6 | BORDERLINE | `\u003Cgold-root>\u002Frecruiting\u002Fphase-3-gold\u002Fgold.jsonl` |\n| architecture | 40 | 1 | **LOW-POWER** | `\u003Cgold-root>\u002Farchitecture\u002Fphase-3-gold\u002Fgold.jsonl` |\n| **Total** | **198** | **53** | — | 67 unique queries |\n\nLOW-POWER means recall@K is statistically meaningless on that domain (single positive sample). Excluded from cross-domain aggregates by default.\n\nVerbatim brain_search args live in each domain's `phase-0b-corpus\u002Fcorpus.jsonl` (sibling dir of gold). See `references\u002Fprior-runs.md` for the full provenance.\n\n---\n\n## Locked decisions (bench mode)\n\n| Setting | Default | Why | Override condition |\n|---|---|---|---|\n| Domains | all 4 | Most representative spread. | `--domains` to subset; useful when isolating a single-domain regression. |\n| LOW-POWER inclusion | excluded | Run 4 (architecture) has L2+L3=1 → recall@K noise dominates. | `--include-low-power` only when specifically auditing architecture. |\n| K values | `1,3,5,10,20,50` | Standard IR eval set. | `--k` to override. |\n| Anti-placebo | mandatory | All BL retrievals must be classified (true_hit \u002F echo_fm11 \u002F downstream \u002F uncertain \u002F metadata_gap). | Cannot disable. Skipping placebo classification = corrupt scorecard. |\n| Rubric | v1.1 frozen | Tonight's gold uses v1.1. Don't iterate. | v2 trigger (see `references\u002Froadmap-v2.md`). |\n| brain_store on completion | yes | TASK_DONE chunk with headline metrics. | `--no-store` for dry-run smokes. |\n\n---\n\n## Locked decisions (build mode)\n\nWhen extending the standing corpus with a new domain, the same defaults from v1 build apply:\n\n| Setting | Default | Why |\n|---|---|---|\n| `--judges` | `claude,codex,gemini` (3) | W3.2 κ matrix — codex is most independent voter; cursor most redundant with claude (κ=0.811 on Runs 2-4 avg \u002F 0.915 on Run 1 alone). |\n| `--schema` | `v1.1-3p` (3 primary, no shadow) | Matches the default 3-judge panel; cleaner than the 3-primary-+-1-shadow split. |\n| `--rubric-version` | `v1.1` | FM12 = 0\u002F145 in production. |\n| `--liveness-check` | `strict` (≥95% rows, fail-loud, exit 2) | Closes W3.1's silent-gemini-absent hole from Run 4. |\n| `--pending-rt-cascade` | `opus-4-7` | Functional queue-routing token accepted by the parser; v1 queues rows without dispatching a model. |\n| Tiebreaker | `claudeJudge` | Cleanest calibration (mean 86.6); 0 hallucinations on Runs 2\u002F3\u002F4. |\n\nResolve the actual cascade adjudicator through `claude.judgment` using\n`node scripts\u002Fmodel-roles.mjs claude.judgment --field model` from the golems checkout.\nChoose effort per `\u002Flarge-plan` phase and pass it explicitly at dispatch.\n\n### Evidence: W3.3 cascade calibration\n\nW3.3 recorded 12 of 13 v1.1 pending-RT cases as FM6-PreCompact single-judge outliers.\nIts original recipe said \"cheap Opus resolves\"; current adjudication uses the role above.\n\nSee `references\u002Fjudge-panel.md` for the full κ rationale.\n\n---\n\n## When to use bench mode\n\n- **After ANY BrainLayer change** that could affect retrieval (PR merges, schema updates, embedder swaps, FM fixes shipping).\n- **Periodic drift checks** (weekly \u002F biweekly) even when nothing visible changed — to catch silent regressions.\n- **Pre-flight before a major BL refactor** — establish the baseline summary that future runs compare against.\n\n## When NOT to use bench mode\n\n- One-off \"does brain_search work\" check — just call brain_search directly.\n- Domain-specific deep-dive that doesn't need the full 198-pair corpus — use `brain_search` ad-hoc.\n- Testing a non-BrainLayer retrieval system — agada-bench is BL-specific.\n\n## When to use build mode\n\n- A new user-domain has emerged. Build a fresh ~50-row gold corpus so the bench can score recall on it.\n- A domain's source session is replaceable with a much-better representative session (rare — usually we add a new domain rather than rebuild an old one).\n\n## When NOT to use build mode\n\n- The 4 existing domains cover your audit need — just run bench mode.\n- You're tempted to \"rebuild techgym with a fresher session.\" The standing gold is frozen for a reason — rebuilding breaks the prior-baseline regression chain.\n- You're tempted to iterate the rubric. Don't. That's a v2 task.\n\n---\n\n## v1 known limitations (deferred to v2)\n\n### Build-mode limits (from extract-corpus.py)\n\nAuto-extraction from session JSONL only has access to what BL renders in the brain_search tool_result text:\n\n1. **chunk_id is a BL-rendered prefix** (e.g., `rt-0c2e3cb8-`), not the full chunk_id. Tonight's hand-curated runs used a live BL DB lookup to resolve prefixes; v1 doesn't.\n2. **chunk_full_content is BL's one-line truncation**, not the verbatim chunk body. Judges grade based on this truncation.\n3. **adjacent_chunks is always empty** — surfacing neighbors requires a live DB lookup.\n\nAll gated by the same fix: extract-corpus.py needs a live `brain_recall(chunk_id_prefix)` lookup to resolve prefixes and pull full bodies. That's a v2 task; see `references\u002Froadmap-v2.md` Phase A.5.\n\n### Bench-mode limits (from run-bench.py)\n\n1. **Python can't call MCP** — live brain_search must happen in the operator's Claude\u002FCodex session.\n2. **Anti-placebo classification is operator-driven** — relies on the operator inspecting chunk_created_iso and source_session_id per the checklist in `workflows\u002Frun-bench.md`. Errors here corrupt every downstream metric.\n3. **Run 4 LOW-POWER**: architecture (1 L2+L3 pair) is excluded from aggregates by default. Acceptable for v1; v2 may rebuild architecture with a richer source session.\n\n---\n\n## v2 roadmap (deferred, documented)\n\nPer W1.7's skeptical pass, v1.1 is SHIP. v2 triggers only when the flip conditions in `references\u002Froadmap-v2.md` fire (FM12 > 0 on any run, second judge dies, etc.). v2 buildout sketches:\n\n- Phase A: Platt-scale verbalized confidence + per-domain calibration (W1.1 + W1.3)\n- Phase A.5: chunk_id-prefix-to-full resolution via live `brain_recall` (extract-corpus.py limitation)\n- Phase B: Abstention τ=0.55 + risk-coverage curve (W1.5)\n- Phase C: Reasoning-tree audit module (W1.2 ADOPT-LITE)\n- Phase D: Production-modal hardening (W1.6 applicable subset)\n\n---\n\n## Composability\n\nThis skill is referenced by:\n- `\u002Forc` — when sweeping multi-domain regressions.\n- The post-PR-merge ritual after BrainLayer PRs land.\n\nThis skill references:\n- `\u002Fnever-fabricate` — anti-placebo classification is mandatory; no \"true_hit\" without verified `chunk_created_iso` vs `query_asked_at`.\n- `\u002Fcmux-agents` — only build mode uses cmux pane spawn (for judge dispatch); bench mode runs entirely in the operator's session.\n- the false-green-gate hook — bench-report.md verdict must say `GREEN | YELLOW | RED` based on actual numbers, not approximations.\n\n---\n\n## Quick reference\n\n### Bench mode (default — the daily-use case)\n\n```bash\n# Run inside a Claude\u002FCodex session with BL MCP connected\npython3 \"$GOLEMS_ROOT\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fscripts\u002Frun-bench.py\" prepare \\\n  --gold techgym:\u003Cgold-root>\u002Ftechgym\u002Fphase-3-gold\u002Fgold.jsonl \\\n  --gold freelance:\u003Cgold-root>\u002Ffreelance\u002Fphase-3-gold\u002Fgold.jsonl \\\n  --gold recruiting:\u003Cgold-root>\u002Frecruiting\u002Fphase-3-gold\u002Fgold.jsonl \\\n  --gold architecture:\u003Cgold-root>\u002Farchitecture\u002Fphase-3-gold\u002Fgold.jsonl \\\n  --output \u003Cartifact-dir>\u002F$(date +%Y-%m-%d)-bench-queries.jsonl\n\n# Operator: read workflows\u002Frun-bench.md, do the 6 waves of brain_search, write bench-results.jsonl\n\npython3 \"$GOLEMS_ROOT\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fscripts\u002Frun-bench.py\" score \\\n  --queries \u003Cartifact-dir>\u002F$(date +%Y-%m-%d)-bench-queries.jsonl \\\n  --results \u003Cartifact-dir>\u002F$(date +%Y-%m-%d)-bench-results.jsonl \\\n  --output \u003Cartifact-dir>\u002F$(date +%Y-%m-%d)-brainlayer-quality-bench-results.md \\\n  --json-out \u003Cartifact-dir>\u002F$(date +%Y-%m-%d)-bench-summary.json \\\n  --baseline \u003Cprevious-summary.json>\n```\n\nOr via the top-level driver:\n\n```bash\nbash \"$GOLEMS_ROOT\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fscripts\u002Frun-agada.sh\"\n```\n\n### Build mode (rare subcommand)\n\n```bash\n# Add a new domain (run once per new domain — a few times a year)\nbash \"$GOLEMS_ROOT\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fscripts\u002Frun-agada.sh\" build \\\n  --session \u003Cnew-domain-session.jsonl> \\\n  --domain \u003Cnew-domain-name> \\\n  --output \u003Cartifact-dir>\u002F2026-MM-DD-agada-\u003Cnew-domain>\u002F\n```\n\n### Prior runs index\n\nSee `references\u002Fprior-runs.md` for the 4 standing-corpus domains and the dates they were locked.\n\n*End SKILL.md. Primary workflow: `workflows\u002Frun-bench.md`. Build workflow: `workflows\u002Fbuild-new-domain.md`. Per-judge briefs: `adapters\u002F`. Standing-gold provenance: `references\u002Fprior-runs.md`.*\n",{"name":5,"description":7},"master","3400bc971628f00ab7bf1063993477e211a5508a","skills\u002Fgolem-powers\u002Fagada-bench","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems","https:\u002F\u002Fgithub.com\u002FEtanHey","etanheyman.com","2026-11-15 03:05:08",{"source":8,"skillName":5,"commands":32,"prompt":42,"zipAvailable":43},[33,36,39],{"tool":34,"command":35},"npx","npx skills add EtanHey\u002Fgolems --skill agada-bench",{"tool":37,"command":38},"bunx","bunx skills add EtanHey\u002Fgolems --skill agada-bench",{"tool":40,"command":41},"pnpm","pnpm dlx skills add EtanHey\u002Fgolems --skill agada-bench","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002FSKILL.md",true,[45,51,55,59,63,67,73,77,81,85,89,93,99,103,107,111,115,119,123,127,131,135,139,143,147,151,155],{"path":46,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":49,"sourceHtmlUrl":50},"adapters\u002Fcapabilities.yaml","REFERENCE","参考文档",3174,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fadapters\u002Fcapabilities.yaml",{"path":52,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":53,"sourceHtmlUrl":54},"adapters\u002Fclaude.md",5896,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fadapters\u002Fclaude.md",{"path":56,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":57,"sourceHtmlUrl":58},"adapters\u002Fcodex.md",5749,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fadapters\u002Fcodex.md",{"path":60,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":61,"sourceHtmlUrl":62},"adapters\u002Fcursor.md",5394,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fadapters\u002Fcursor.md",{"path":64,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":65,"sourceHtmlUrl":66},"evals\u002Fevals.json",17496,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fevals\u002Fevals.json",{"path":68,"resourceType":69,"resourceTypeDesc":70,"sizeBytes":71,"sourceHtmlUrl":72},"evals\u002Ffixtures\u002Ftiny-session.jsonl","OTHER","其他",8505,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fevals\u002Ffixtures\u002Ftiny-session.jsonl",{"path":74,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":75,"sourceHtmlUrl":76},"references\u002Ffailure-modes.md",9588,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Freferences\u002Ffailure-modes.md",{"path":78,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":79,"sourceHtmlUrl":80},"references\u002Fgrading-rubric.md",10794,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Freferences\u002Fgrading-rubric.md",{"path":82,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":83,"sourceHtmlUrl":84},"references\u002Fjudge-panel.md",7176,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Freferences\u002Fjudge-panel.md",{"path":86,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":87,"sourceHtmlUrl":88},"references\u002Fprior-runs.md",7632,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Freferences\u002Fprior-runs.md",{"path":90,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":91,"sourceHtmlUrl":92},"references\u002Froadmap-v2.md",7387,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Freferences\u002Froadmap-v2.md",{"path":94,"resourceType":95,"resourceTypeDesc":96,"sizeBytes":97,"sourceHtmlUrl":98},"scripts\u002Fbuild-crossref.py","SCRIPT","脚本",11289,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fscripts\u002Fbuild-crossref.py",{"path":100,"resourceType":95,"resourceTypeDesc":96,"sizeBytes":101,"sourceHtmlUrl":102},"scripts\u002Fbuild-gold.py",17817,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fscripts\u002Fbuild-gold.py",{"path":104,"resourceType":95,"resourceTypeDesc":96,"sizeBytes":105,"sourceHtmlUrl":106},"scripts\u002Fdispatch-judges.py",7070,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fscripts\u002Fdispatch-judges.py",{"path":108,"resourceType":95,"resourceTypeDesc":96,"sizeBytes":109,"sourceHtmlUrl":110},"scripts\u002Fextract-corpus.py",11994,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fscripts\u002Fextract-corpus.py",{"path":112,"resourceType":95,"resourceTypeDesc":96,"sizeBytes":113,"sourceHtmlUrl":114},"scripts\u002Fkappa-matrix.py",7809,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fscripts\u002Fkappa-matrix.py",{"path":116,"resourceType":95,"resourceTypeDesc":96,"sizeBytes":117,"sourceHtmlUrl":118},"scripts\u002Fliveness-check.py",6273,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fscripts\u002Fliveness-check.py",{"path":120,"resourceType":95,"resourceTypeDesc":96,"sizeBytes":121,"sourceHtmlUrl":122},"scripts\u002Froute-pending-rt.py",10014,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fscripts\u002Froute-pending-rt.py",{"path":124,"resourceType":95,"resourceTypeDesc":96,"sizeBytes":125,"sourceHtmlUrl":126},"scripts\u002Frun-agada.sh",10407,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fscripts\u002Frun-agada.sh",{"path":128,"resourceType":95,"resourceTypeDesc":96,"sizeBytes":129,"sourceHtmlUrl":130},"scripts\u002Frun-bench.py",17661,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fscripts\u002Frun-bench.py",{"path":132,"resourceType":95,"resourceTypeDesc":96,"sizeBytes":133,"sourceHtmlUrl":134},"tests\u002Ftest_artifact_truth.py",5477,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Ftests\u002Ftest_artifact_truth.py",{"path":136,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":137,"sourceHtmlUrl":138},"workflows\u002Fadd-domain.md",4642,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fworkflows\u002Fadd-domain.md",{"path":140,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":141,"sourceHtmlUrl":142},"workflows\u002Fbuild-new-domain.md",8721,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fworkflows\u002Fbuild-new-domain.md",{"path":144,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":145,"sourceHtmlUrl":146},"workflows\u002Fkappa-validation.md",6030,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fworkflows\u002Fkappa-validation.md",{"path":148,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":149,"sourceHtmlUrl":150},"workflows\u002Fliveness-check.md",5177,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fworkflows\u002Fliveness-check.md",{"path":152,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":153,"sourceHtmlUrl":154},"workflows\u002Fpending-rt-routing.md",7673,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fworkflows\u002Fpending-rt-routing.md",{"path":156,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":157,"sourceHtmlUrl":158},"workflows\u002Frun-bench.md",12334,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fagada-bench\u002Fworkflows\u002Frun-bench.md",[160,167,174,181,188,195],{"skillNo":161,"name":162,"nameZh":6,"description":163,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":164,"categories":165,"keywords":166,"publishedAt":19},"01M7C53JEF6A6SS20Q0D2YJYM8","yash","Push dependency bugs upstream. Triggers: dependency bug, workaround, monkey-patch, fork\u002Ffix. NOT app bugs.","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fyash\u002FSKILL.md",[],[],{"skillNo":168,"name":169,"nameZh":6,"description":170,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":171,"categories":172,"keywords":173,"publishedAt":19},"01M7C53EPMNSZKHK47VY7WQDP0","whats-new","Diff Claude\u002FCodex\u002FCursor\u002FWispr\u002Fagy changelogs vs local setup. Triggers: whats new, changelog, version bump.","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fwhats-new\u002FSKILL.md",[],[],{"skillNo":175,"name":176,"nameZh":6,"description":177,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":178,"categories":179,"keywords":180,"publishedAt":19},"01M7C53C4TVQ9G7AB3X57DACPB","weave","Orchestrator convergence: mine JSONLs, cite findings, route actions. Triggers: weave, run weave.","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fweave\u002FSKILL.md",[],[],{"skillNo":182,"name":183,"nameZh":6,"description":184,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":185,"categories":186,"keywords":187,"publishedAt":19},"01M7C5342EDDN7YH53KKW511V2","unslop","Cut AI tells while keeping every receipt, number, path, and hedge intact. Triggers: unslop, cut the slop, wall of text, too verbose, rambly, make this readable. NOT for delivery cadence (see i-have-adhd), channel choice, or artifact length governance.","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Funslop\u002FSKILL.md",[],[],{"skillNo":189,"name":190,"nameZh":6,"description":191,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":192,"categories":193,"keywords":194,"publishedAt":19},"01M7C52M0CTDW5DDP4452FJV0G","tmp-block","Guard denying durable writes to \u002Ftmp, \u002Fprivate\u002Ftmp, \u002Fvar\u002Ffolders, $TMPDIR; enforces \u003Crepo>\u002F.worktrees\u002F\u003Cname>. Triggers: tmp guard, TMP-BLOCK deny, WORKTREE-CONVENTION deny, bypass ledger. NOT for storage policy.","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002FSKILL.md",[],[],{"skillNo":196,"name":197,"nameZh":6,"description":198,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":199,"categories":200,"keywords":201,"publishedAt":19},"01M7C52HN4Q9E1Z74Z11K23E2H","t3-visuals","Show\u002Fvisualize\u002Fchart\u002Fdiagram\u002Fcompare\u002Fstatus board in T3 with compact interactive HTML. Not standalone dashboards.","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ft3-visuals\u002FSKILL.md",[],[]]