[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7800KJRST3T08D64ZP2TT05":3,"skill-same:01M7800KJRST3T08D64ZP2TT05":263},{"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":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":6,"ownerHtmlUrl":29,"ownerBlogUrl":30,"ownerDescription":31,"updatedAt":32,"install":33,"resources":46},"01M7800KJRST3T08D64ZP2TT05","validate-recommendation",null,"Validate (Recommended) options in AskUserQuestion via N configurable panelists (~\u002F.claude\u002Fpanel\u002Fconfig.yml). Triggered by the validate-recommendation hook; do not invoke manually.","ArangoGutierrez\u002Fclaude-toolkit","ArangoGutierrez","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F15933089?v=4","01M77ZX10BNKRH5SRK782MGZQR","Carlos Eduardo Arango Gutierrez","01M77ZX1108AW5SJRFWGFKPKTA",4,"MIT","https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: validate-recommendation\ndescription: Validate (Recommended) options in AskUserQuestion via N configurable panelists (~\u002F.claude\u002Fpanel\u002Fconfig.yml). Triggered by the validate-recommendation hook; do not invoke manually.\n---\n\n# Validate Recommendation\n\nYou were invoked because the `validate-recommendation` hook fired on an\n`AskUserQuestion` call that contained a `(Recommended)` option marker.\nYour job: dispatch an N-panelist review (where N comes from\n`~\u002F.claude\u002Fpanel\u002Fconfig.yml`), aggregate the verdicts into a JSON\ndirective, and act on it.\n\nThe panel composition is **configurable** and ships as three panelists:\nan adversarial DA, which is told which option was recommended, plus a\nblinded PE and QA, which are not. Any of them can be turned off, and new\nones added, in `config.yml`. This skill is config-driven; it does NOT\nassume a fixed panel — always enumerate the enabled set from the config.\n\n## Inputs\n\nThe hook wrote tool input to a session state file:\n\n```\nSTATE_FILE=\"${TMPDIR:-\u002Ftmp}\u002Fclaude-panel-${CLAUDE_SESSION_ID:-$PPID}.json\"\n```\n\nWhen `CLAUDE_SESSION_ID` is unset, both hook and skill resolve `$PPID`\nto the same value (the Claude Code parent process), so the skill reads\nexactly the file the hook wrote. Two concurrent CC sessions get\ndifferent `$PPID`s and therefore different state-file paths — the old\n`-unknown` sentinel collided across sessions.\n\nRead it via the `Read` tool. If the file is missing, emit the\nuser-visible message \"Panel state file missing; asking the original\nquestion.\" and fall back to issuing the original `AskUserQuestion`\nunmodified. Stop here.\n\nThe panel config:\n\n```\nCONFIG=\"${HOME}\u002F.claude\u002Fpanel\u002Fconfig.yml\"\n```\n\nState file keys you'll use:\n- `tool_input.questions` — array of question objects (1-4 per the\n  AskUserQuestion schema)\n- `timeout_seconds` — per-panelist budget (default 90)\n\n## Setup\n\n```bash\nexport PYTHONPATH=\"$HOME\u002F.claude${PYTHONPATH:+:$PYTHONPATH}\"\n```\n\nThis ensures `import tool` (the shared engine) resolves at runtime.\nAll subsequent `python3.12 -m panel …` calls inherit this setting.\n\n### 1. Validate the config\n\nRun via the `Bash` tool:\n\n```bash\ncd \"${HOME}\u002F.claude\u002Fskills\u002Fvalidate-recommendation\" && \\\n    \"${CLAUDE_PANEL_PYTHON:-python3.12}\" -m panel lint-config --config \"$CONFIG\"\n```\n\nIf exit code is non-zero, fall back: print \"Panel disabled: config\ninvalid (see lint-config output).\" and re-issue the original\n`AskUserQuestion` unmodified. Stop.\n\n### 2. Read enabled panelists from config\n\nUse `jq` via Bash to enumerate enabled panelists:\n\n```bash\njq -r '.panelists[] | select(.enabled) | \"\\(.id)|\\(.role)|\\(.backend)|\\(.subagent_type \u002F\u002F \"\")\"' \\\n    \u003C(\"${CLAUDE_PANEL_PYTHON:-python3.12}\" -c \"\nimport yaml, json, sys\nprint(json.dumps(yaml.safe_load(open('$CONFIG'))))\n\")\n```\n\nThe yaml→json conversion runs once at skill startup. Capture the output\ninto a variable; each line is `\u003Cid>|\u003Crole>|\u003Cbackend>|\u003Csubagent_type>`\nfor one enabled panelist.\n\n(If `python3.12 -c` for yaml→json is awkward in your sandbox, the\nequivalent is\n`cd \"${HOME}\u002F.claude\u002Fskills\u002Fvalidate-recommendation\" && python3.12 -m panel lint-config --config \"$CONFIG\"`,\nwhich prints the same data in human-readable form; you can grep its output\nfor the `- \u003Cid>` lines and extract role\u002Fbackend from them. The `cd` is\nrequired here too: a Bash tool call does not inherit the cwd of step 1's,\nand without it `panel` resolves to the `~\u002F.claude\u002Fpanel` config directory.)\n\n### 2a. Resolve the panel profile\n\nThe default panel is all three panelists: an adversarial DA plus a blinded\nPE and QA. There is no team auto-detect — the ensemble is the standard\noperating mode, because at N=1 the hard threshold is ceil(1\u002F2) = 1, which\nmakes SOFT-DISSENT unreachable and turns every single OVERTURN into a hard\ninterrupt.\n\nOne override, for cost control:\n\n- `CLAUDE_PANEL_PROFILE=solo` disables PE and QA for this dispatch, leaving\n  DA only. Read it via Bash: `echo \"${CLAUDE_PANEL_PROFILE:-}\"`. Use it when\n  the extra two Claude subagent spawns are not worth it.\n\nIf solo is requested, write a resolved config with `pe` and `qa` set to\n`enabled: false` and use it as `$CONFIG_RESOLVED` for the REST of this run\n(both the section-2 enumeration AND the §4 aggregator — `panel aggregate`\nre-derives its panelist set from its `--config`, so it would otherwise\nsilently disagree with what was dispatched).\n\n```bash\nCONFIG_RESOLVED=\"${TMPDIR:-\u002Ftmp}\u002Fpanel-config-resolved-${CLAUDE_SESSION_ID:-$PPID}.yml\"\n\"${CLAUDE_PANEL_PYTHON:-python3.12}\" - \"$CONFIG\" \"$CONFIG_RESOLVED\" \u003C\u003C'PY'\nimport sys, yaml\ncfg = yaml.safe_load(open(sys.argv[1]))\nfor p in cfg.get(\"panelists\", []):\n    if p.get(\"id\") in (\"pe\", \"qa\"):\n        p[\"enabled\"] = False\nyaml.safe_dump(cfg, open(sys.argv[2], \"w\"))\nPY\n```\n\nThen re-run the section-2 enumeration against `$CONFIG_RESOLVED` (it now\nlists `da` only) and pass `$CONFIG_RESOLVED` to the §4 aggregator.\n\nOtherwise use the stock `$CONFIG` unchanged and skip to step 3.\n\n### 3. Create the per-session workdir\n\n```bash\nWORKDIR=\"${HOME}\u002F.claude\u002Fpanel\u002Fwork\u002F${CLAUDE_SESSION_ID:-$PPID}\"\nmkdir -p \"$WORKDIR\" && chmod 0700 \"$WORKDIR\"\n```\n\nThis is the directory where per-panelist verdict files land.\n`panel aggregate` will read `${WORKDIR}\u002F\u003Cid>.verdict` for each enabled\npanelist after fan-out completes.\n\n## Per-question dispatch\n\nFor EACH question in `tool_input.questions` that has an option labeled\nwith `(Recommended)` AND NOT `(Recommended; Panel-flagged)`:\n\n### 1. Build the per-panelist prompt files\n\nWrite the question payload to a JSON file, then let the CLI build each\npanelist's prompt. Blinding is applied automatically from each panelist's\n`blind` flag — do NOT hand-write prompt bodies, and do NOT mention the\nrecommended option to a blinded panelist.\n\n```bash\nQPAYLOAD=\"${TMPDIR:-\u002Ftmp}\u002Fpanel-question-${CLAUDE_SESSION_ID:-$PPID}-q\u003CN>.json\"\n# Write JSON: {\"question\": ..., \"options\": [{\"label\":..., \"description\":...}],\n#              \"recommended_label\": ..., \"reasoning\": ...}\n# reasoning comes from the recommended option's description; if there is\n# none, use \"(no reasoning supplied)\". NEVER invent hidden reasoning.\n\n# The cd is load-bearing, exactly as in the lint-config, dispatch and\n# aggregate blocks. $PYTHONPATH is $HOME\u002F.claude, so from any other cwd\n# `python -m panel` resolves `panel` to ~\u002F.claude\u002Fpanel — the CONFIG\n# directory — which has no __init__.py and shadows the real package as a\n# namespace package. The call then dies with \"No module named\n# panel.__main__\" and writes no prompt file at all.\ncd \"${HOME}\u002F.claude\u002Fskills\u002Fvalidate-recommendation\"\n\nfor PID in \u003Ceach enabled panelist id>; do\n  \"${CLAUDE_PANEL_PYTHON:-python3.12}\" -m panel build-prompt \\\n    --panelist \"$PID\" \\\n    --config \"${CONFIG_RESOLVED:-$CONFIG}\" \\\n    --question-file \"$QPAYLOAD\" \\\n    --output \"${TMPDIR:-\u002Ftmp}\u002Fpanel-prompt-${CLAUDE_SESSION_ID:-$PPID}-q\u003CN>-${PID}.txt\"\ndone\n```\n\nWhere `\u003CN>` is the question index (0-based).\n\nEach panelist is then dispatched with ITS OWN prompt file. A blinded\npanelist's file has the recommendation removed, the `(Recommended)` marker\nstripped, and the options deterministically reordered.\n\nThe payload's `reasoning` field is read only for UNBLINDED panelists;\n`build-prompt` substitutes an empty string for a blinded one. That is why\nthe payload may carry the literal `(no reasoning supplied)` safely. Never\ncall the underlying `build_prompt_body` yourself with a blinded panelist and\na non-empty reasoning — including that literal, which is a non-empty string:\nit raises `ValueError` by design. Go through the CLI and the question is\nmoot.\n\n### 2. Fan out N panelists in ONE message\n\n**This is the parallelism point. All panelist dispatches MUST be in a\nsingle message so they run concurrently.**\n\nFor EACH enabled panelist from the config enumeration above:\n\n- **If `backend` starts with `nat-` (e.g., `nat-nim`):** include a\n  `Bash` tool call in your fan-out message. Export `CLAUDE_SESSION_ID`\n  from the state file's `session_id` field in the SAME command —\n  `panel\u002Ftrace.py` reads it from the environment for verdict-line\n  telemetry, and it is otherwise never a real exported variable (only\n  ever referenced with a `${CLAUDE_SESSION_ID:-$PPID}` fallback), which\n  is why trace lines were showing `session=unknown`:\n\n  ```bash\n  export CLAUDE_SESSION_ID=\"$(jq -r '.session_id' \"${TMPDIR:-\u002Ftmp}\u002Fclaude-panel-${CLAUDE_SESSION_ID:-$PPID}.json\")\" && \\\n  cd \"${HOME}\u002F.claude\u002Fskills\u002Fvalidate-recommendation\" && \\\n      \"${CLAUDE_PANEL_PYTHON:-python3.12}\" -m panel dispatch \\\n      --panelist \"\u003Cid>\" \\\n      --config \"${CONFIG_RESOLVED:-$CONFIG}\" \\\n      --persona \"${HOME}\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpersonas\u002F\u003Crole-lowercase>.md\" \\\n      --prompt-file \"${TMPDIR:-\u002Ftmp}\u002Fpanel-prompt-${CLAUDE_SESSION_ID:-$PPID}-q\u003CN>-\u003Cid>.txt\" \\\n      --output \"${WORKDIR}\u002F\u003Cid>.verdict\"\n  ```\n\n  The `--prompt-file` is THIS panelist's file from step 1, not a shared\n  one. Passing another panelist's file hands a blinded seat the\n  recommendation.\n\n  `dispatch.py` writes the verdict file directly (mode 0600). Auth\n  comes from per-backend env vars (`nat-nim`: `$PANEL_DA_API_KEY`,\n  falling back to `$NVIDIA_API_KEY`; `nat-anthropic`:\n  `$ANTHROPIC_API_KEY`; `nat-openai`: `$OPENAI_API_KEY`).\n\n- **If `backend` is `claude-subagent`:** include an `Agent` tool call\n  in your fan-out message:\n  - `subagent_type`: the panelist's `subagent_type` from config (e.g.,\n    `principal-engineer`, `qa-engineer`)\n  - `description`: short, e.g., `\"Panel \u003Crole> review\"`\n  - `prompt`: the concatenation of the persona file's `# System prompt`\n    section, the `# One-shot example` section (if present), and the\n    contents of THIS panelist's prompt file from step 1\n    (`…-q\u003CN>-\u003Cid>.txt`), copied verbatim.\n\n  Read both files via the `Read` tool BEFORE composing the fan-out\n  message: the persona at\n  `${HOME}\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpersonas\u002F\u003Crole-lowercase>.md`,\n  and the panelist's own prompt file. Never re-compose the prompt body by\n  hand — for a blinded panelist that is the leak.\n\nAll `Bash` + `Agent` calls go in ONE message. The framework executes\nthem concurrently.\n\n### 3. Collect verdict files after fan-out returns\n\nWhen the fan-out message's tool results come back:\n\n- **`nat-*` panelists:** `dispatch.py` already wrote the verdict file.\n  Nothing to do.\n- **`claude-subagent` panelists:** the `Agent` tool returned a string.\n  Write the ENTIRE response verbatim to the verdict file via the\n  `Write` tool:\n\n  ```\n  ${WORKDIR}\u002F\u003Cid>.verdict\n  ```\n\n  Do NOT alter, summarize, or extract from the Agent response. The\n  aggregator does the parsing, and which lines it wants depends on the\n  panelist: an UNBLINDED one emits `VERDICT:`\u002F`RATIONALE:`\u002F`ALTERNATIVE:`,\n  a BLINDED one emits `CHOICE:`\u002F`RATIONALE:`. `aggregate()` maps a blinded\n  `CHOICE:` onto the same contract — agreement with the recommended label\n  is HOLD, disagreement is OVERTURN naming that option — comparing\n  marker- and case-insensitively. Surrounding prose is ignored, and a\n  blinded reply with no `CHOICE:` line becomes ERROR for that panelist.\n\n### 4. Run the aggregator\n\nSame rule as dispatch: export `CLAUDE_SESSION_ID` from the state file's\n`session_id` inside this SAME command, so the aggregator's trace lines\ncarry the real session instead of `session=unknown`.\n\n```bash\nDIRECTIVE_JSON=$(export CLAUDE_SESSION_ID=\"$(jq -r '.session_id' \"${TMPDIR:-\u002Ftmp}\u002Fclaude-panel-${CLAUDE_SESSION_ID:-$PPID}.json\")\" && \\\n    cd \"${HOME}\u002F.claude\u002Fskills\u002Fvalidate-recommendation\" && \\\n    \"${CLAUDE_PANEL_PYTHON:-python3.12}\" -m panel aggregate \\\n    --config \"${CONFIG_RESOLVED:-$CONFIG}\" \\\n    --verdicts-dir \"$WORKDIR\" \\\n    --recommended-label \"\u003Crecommended_label from state for THIS question>\")\n```\n\nCapture stdout into `$DIRECTIVE_JSON`. It's a single-line JSON object.\n\n### 5. Parse the directive\n\nExtract fields via `jq`:\n\n```bash\nVERDICT=$(jq -r '.verdict' \u003C\u003C\u003C \"$DIRECTIVE_JSON\")\nSUMMARY=$(jq -r '.summary' \u003C\u003C\u003C \"$DIRECTIVE_JSON\")\nESCALATE=$(jq -r '.escalate_to_user \u002F\u002F false' \u003C\u003C\u003C \"$DIRECTIVE_JSON\")\nGATE=$(jq -r '.rationale_gate_passed' \u003C\u003C\u003C \"$DIRECTIVE_JSON\")\n```\n\nValues:\n- `VERDICT` ∈ `{HOLD, SOFT-DISSENT, HARD-DISSENT, ERROR}`\n- `SUMMARY` — one-line user-facing text, already sanitized\n- `ESCALATE` — `true` for HARD-DISSENT in Phase 3c, `false` otherwise\n- `GATE` — `true`\u002F`false`\u002F`null` per the rationale gate semantics\n\n## Acting on the directive\n\nBehavior matrix:\n\n### `VERDICT == \"HOLD\"`\n\nBehavior depends on `${CLAUDE_PANEL:-on}`:\n\n**Default (`on`)** — auto-take mode:\n\nThe recommendation stands. Take the recommended option without asking\nthe user. Print a brief message:\n\n> Panel validated `\u003Crecommended_label>` for \"\u003Cquestion excerpt>\".\n> \u003CSUMMARY>\n> Proceeding.\n\nThen continue the work as if the user had picked the recommended\noption. The `AskUserQuestion` for this question is NEVER issued.\n\n**`advise`** — advisory mode (preserves user agency):\n\nDo NOT auto-take. Re-issue `AskUserQuestion` for this question with\nthe panel's affirmative annotation appended. Augmented payload:\n\n- **Question text**: original question + two newlines +\n  `**Panel validated:** \u003CSUMMARY>` (positive phrasing).\n- **Options**: identical, EXCEPT the recommended option's label has\n  `(Recommended)` swapped for `(Recommended; Panel-flagged)`. The hook\n  detects this marker and skips the panel on the re-ask.\n\n### `VERDICT == \"SOFT-DISSENT\"`\n\nRe-issue `AskUserQuestion` for this question. Augmented payload:\n\n- **Question text**: original + two newlines + `\u003CSUMMARY>` (the\n  `**Panel review:**` line). Use it verbatim.\n- **Options**: identical, EXCEPT the recommended option's label is\n  swapped to `(Recommended; Panel-flagged)`.\n\n### `VERDICT == \"HARD-DISSENT\"`\n\nIn Phase 3c, HARD-DISSENT also escalates to the user (Phase 5 will add\nre-think cycles). Same payload as SOFT-DISSENT, but include severity\nin the augmented note:\n\n- **Question text**: original + two newlines + `Panel HARD-DISSENT:\n  \u003CSUMMARY>` (severity-clear prefix).\n- **Options**: identical, EXCEPT marker swap to\n  `(Recommended; Panel-flagged)`.\n\nWhen Phase 5 lands, this branch will check `re_brainstorm` in the\ndirective and emit a re-think markdown directive (no `AskUserQuestion`)\nwhen present. Until then, every HARD-DISSENT escalates the question to\nthe user with full panel feedback.\n\n### `VERDICT == \"ERROR\"`\n\nSomething went wrong (panelist backend down, malformed responses,\nmissing files, aggregator parse error). For this question, re-issue\nthe original `AskUserQuestion` unmodified — same options, same\n`(Recommended)` marker — BUT swap the marker to\n`(Recommended; Panel-flagged)` for loop safety.\n\nPrint a brief explanation:\n\n> Panel evaluation failed for \"\u003Cquestion excerpt>\" (see hook trace\n> for detail). Asking the question directly.\n\n## Cleanup\n\nAfter processing ALL questions (HOLD + DISSENT + ERROR paths), remove\nevery file this run created — **one explicit path per `rm -f` argument,\nnever a glob (`*`)**. Under zsh, an unmatched glob aborts the ENTIRE\ncommand line, not just that argument — so a single `rm -f a b* c` where\n`b*` matches nothing removes NEITHER `a` NOR `c` either. This is the\nroot cause of the leak this section exists to close: a bare\n`panel-prompt-*-q*.txt` glob that matched nothing silently left\n`$STATE_FILE` on disk too, and the hook's re-entry guard then bypassed\nthe NEXT, unrelated paneled question it saw.\n\n```bash\nrm -rf \"$WORKDIR\"\nrm -f \"$STATE_FILE\"\n```\n\nThen remove each question payload and each per-panelist prompt file you\nactually wrote, one `rm -f` per exact path — using the same 0-based\nquestion indices you dispatched in the per-question loop above, crossed\nwith the panelist ids you enumerated (you know both sets; do not glob\n`-q*.json` or `-q*-*.txt`):\n\n```bash\nrm -f \"${TMPDIR:-\u002Ftmp}\u002Fpanel-question-${CLAUDE_SESSION_ID:-$PPID}-q0.json\"\nrm -f \"${TMPDIR:-\u002Ftmp}\u002Fpanel-prompt-${CLAUDE_SESSION_ID:-$PPID}-q0-da.txt\"\nrm -f \"${TMPDIR:-\u002Ftmp}\u002Fpanel-prompt-${CLAUDE_SESSION_ID:-$PPID}-q0-pe.txt\"\nrm -f \"${TMPDIR:-\u002Ftmp}\u002Fpanel-prompt-${CLAUDE_SESSION_ID:-$PPID}-q0-qa.txt\"\n# ... then the same four lines for q1, q2, ... — one payload plus one\n# prompt file per enabled panelist, for every question index you actually\n# dispatched. Use the ids from YOUR enumeration, not this example's:\n# solo profile writes only the da file, and a config with other panelists\n# writes theirs. Skip indices that were never a (Recommended) question —\n# there is no file for them, and rm -f on a missing path is a no-op.\n#\n# ${CLAUDE_SESSION_ID:-$PPID} must match the fallback used when the files\n# were built. A bare ${CLAUDE_SESSION_ID} resolves to an empty string in a\n# session where it is unset, and removes a path nothing ever wrote.\n```\n\nIf step 2a resolved a solo profile, `$CONFIG_RESOLVED` holds the exact\nresolved-config path already — remove that path directly (it is a\nsingle concrete path, not a glob):\n\n```bash\nif [ -n \"${CONFIG_RESOLVED:-}\" ]; then\n    rm -f \"$CONFIG_RESOLVED\"\nfi\n```\n\nIf you crash before cleanup, stale files are harmless PROVIDED\n`CLAUDE_SESSION_ID` is unique per session (different session = different\npath). The hook's re-entry guard on `$STATE_FILE` is now stale-aware —\nage >= 600s discards the leftover and replaces it with a fresh trigger\ninstead of bypassing indefinitely — so a leftover `$STATE_FILE` from a\ncrashed run self-heals within 10 minutes even if this cleanup never\nruns. `panel gc` will eventually reap stale workdirs (Phase 6).\n\n## Failure modes you must handle gracefully\n\n| What goes wrong | Behavior |\n|---|---|\n| State file missing | Print fallback message; re-issue original `AskUserQuestion`. |\n| `panel lint-config` fails | Print \"Panel disabled: config invalid\"; re-issue original. |\n| `~\u002F.claude\u002Fpanel\u002Fconfig.yml` missing | Caught by `lint-config`; same fallback. |\n| Persona file missing for an ENABLED panelist | Caught up front by `lint-config`'s persona\u002Fblind cross-check; same fallback as an invalid config. A disabled panelist's persona is never loaded. |\n| `panel dispatch` crashes (caller-bug exit 1) | Verdict file not written. Aggregator coerces to ERROR for that panelist. Severity decides per failure_mode. |\n| `panel dispatch` exits 0 but writes ERROR verdict | Normal path. Severity decides per failure_mode. |\n| `Agent` tool call errors \u002F returns garbled output | Write verbatim to verdict file; aggregator coerces to ERROR for that panelist when its expected line is missing — `VERDICT:` unblinded, `CHOICE:` blinded. |\n| `panel aggregate` crashes (non-zero exit) | Fall back. Re-issue original with marker swap. |\n| `$WORKDIR` unwritable | Fall back. Print \"Panel infrastructure unavailable\". |\n| Missing API keys (e.g., `$PANEL_DA_API_KEY`) | `dispatch.py` writes ERROR verdict; aggregator emits ERROR directive (at N=1) or degrades (at N>=3 with graceful failure_mode). User-facing message ends up as re-ask original. |\n\nThe whole panel is best-effort. The user-visible question ALWAYS\nsurvives.\n\n### Loop safety on fallback\n\nWhen you fall back to re-issuing the original `AskUserQuestion`, the\noriginal payload still has `(Recommended)` — meaning the hook will\nfire AGAIN. The hook has a re-entry guard: if a state file for the\ncurrent session already exists when the hook runs, and it is fresh\n(age \u003C 600s — always true for an immediate fallback re-issue), the\nhook removes it and approves (exit 0) without dispatching the skill\nagain. So your fallback re-issue is mechanically safe.\n\nAs a courtesy to future readers, swap the marker to\n`(Recommended; Panel-flagged)` in your fallback re-issue (it documents\nintent even though the re-entry guard makes it a no-op for safety).\n\n## Multi-question parallelism (optional optimization)\n\nIf the `AskUserQuestion` call had multiple questions each with a\n`(Recommended)` marker, you MAY dispatch panel calls for ALL\nquestions in parallel: fan out the N panelists × M questions in one\nmessage. Aggregate runs per question after both panelists return.\n\nFor v1, sequential per-question dispatch is fine. Optimize only if\nmulti-question recommendations become common (they aren't currently).\n\n## What you must NOT do\n\n- Do NOT echo `$PANEL_DA_API_KEY` (or any other API key) in any\n  user-facing text or trace output.\n- Do NOT write API keys to any file.\n- Do NOT modify or summarize the Agent tool's response before writing\n  it to the verdict file. Write verbatim; let the aggregator parse.\n- Do NOT modify the `AskUserQuestion` call beyond augmenting question\n  text (for dissents) or swapping the marker (for ERROR or fallback).\n- Do NOT re-invoke this skill after auto-taking a HOLD recommendation.\n  The work continues as if the user picked.\n- Do NOT introduce a panelist not in `config.yml`. New panelists ship\n  via `config.yml` + a persona file under `personas\u002F\u003Crole>.md` — never\n  ad-hoc in this file.\n- Do NOT call `python3.12 -m panel ...` for JSON parsing. Use `jq`.\n  Python invocation is reserved for `lint-config`, `build-prompt`,\n  `dispatch`, and `aggregate`.\n- Do NOT bypass the hook re-entry guard by clearing the state file\n  yourself. The guard exists to break loops on fallback paths.\n","# Validate Recommendation\n\nYou were invoked because the `validate-recommendation` hook fired on an\n`AskUserQuestion` call that contained a `(Recommended)` option marker.\nYour job: dispatch an N-panelist review (where N comes from\n`~\u002F.claude\u002Fpanel\u002Fconfig.yml`), aggregate the verdicts into a JSON\ndirective, and act on it.\n\nThe panel composition is **configurable** and ships as three panelists:\nan adversarial DA, which is told which option was recommended, plus a\nblinded PE and QA, which are not. Any of them can be turned off, and new\nones added, in `config.yml`. This skill is config-driven; it does NOT\nassume a fixed panel — always enumerate the enabled set from the config.\n\n## Inputs\n\nThe hook wrote tool input to a session state file:\n\n```\nSTATE_FILE=\"${TMPDIR:-\u002Ftmp}\u002Fclaude-panel-${CLAUDE_SESSION_ID:-$PPID}.json\"\n```\n\nWhen `CLAUDE_SESSION_ID` is unset, both hook and skill resolve `$PPID`\nto the same value (the Claude Code parent process), so the skill reads\nexactly the file the hook wrote. Two concurrent CC sessions get\ndifferent `$PPID`s and therefore different state-file paths — the old\n`-unknown` sentinel collided across sessions.\n\nRead it via the `Read` tool. If the file is missing, emit the\nuser-visible message \"Panel state file missing; asking the original\nquestion.\" and fall back to issuing the original `AskUserQuestion`\nunmodified. Stop here.\n\nThe panel config:\n\n```\nCONFIG=\"${HOME}\u002F.claude\u002Fpanel\u002Fconfig.yml\"\n```\n\nState file keys you'll use:\n- `tool_input.questions` — array of question objects (1-4 per the\n  AskUserQuestion schema)\n- `timeout_seconds` — per-panelist budget (default 90)\n\n## Setup\n\n```bash\nexport PYTHONPATH=\"$HOME\u002F.claude${PYTHONPATH:+:$PYTHONPATH}\"\n```\n\nThis ensures `import tool` (the shared engine) resolves at runtime.\nAll subsequent `python3.12 -m panel …` calls inherit this setting.\n\n### 1. Validate the config\n\nRun via the `Bash` tool:\n\n```bash\ncd \"${HOME}\u002F.claude\u002Fskills\u002Fvalidate-recommendation\" && \\\n    \"${CLAUDE_PANEL_PYTHON:-python3.12}\" -m panel lint-config --config \"$CONFIG\"\n```\n\nIf exit code is non-zero, fall back: print \"Panel disabled: config\ninvalid (see lint-config output).\" and re-issue the original\n`AskUserQuestion` unmodified. Stop.\n\n### 2. Read enabled panelists from config\n\nUse `jq` via Bash to enumerate enabled panelists:\n\n```bash\njq -r '.panelists[] | select(.enabled) | \"\\(.id)|\\(.role)|\\(.backend)|\\(.subagent_type \u002F\u002F \"\")\"' \\\n    \u003C(\"${CLAUDE_PANEL_PYTHON:-python3.12}\" -c \"\nimport yaml, json, sys\nprint(json.dumps(yaml.safe_load(open('$CONFIG'))))\n\")\n```\n\nThe yaml→json conversion runs once at skill startup. Capture the output\ninto a variable; each line is `\u003Cid>|\u003Crole>|\u003Cbackend>|\u003Csubagent_type>`\nfor one enabled panelist.\n\n(If `python3.12 -c` for yaml→json is awkward in your sandbox, the\nequivalent is\n`cd \"${HOME}\u002F.claude\u002Fskills\u002Fvalidate-recommendation\" && python3.12 -m panel lint-config --config \"$CONFIG\"`,\nwhich prints the same data in human-readable form; you can grep its output\nfor the `- \u003Cid>` lines and extract role\u002Fbackend from them. The `cd` is\nrequired here too: a Bash tool call does not inherit the cwd of step 1's,\nand without it `panel` resolves to the `~\u002F.claude\u002Fpanel` config directory.)\n\n### 2a. Resolve the panel profile\n\nThe default panel is all three panelists: an adversarial DA plus a blinded\nPE and QA. There is no team auto-detect — the ensemble is the standard\noperating mode, because at N=1 the hard threshold is ceil(1\u002F2) = 1, which\nmakes SOFT-DISSENT unreachable and turns every single OVERTURN into a hard\ninterrupt.\n\nOne override, for cost control:\n\n- `CLAUDE_PANEL_PROFILE=solo` disables PE and QA for this dispatch, leaving\n  DA only. Read it via Bash: `echo \"${CLAUDE_PANEL_PROFILE:-}\"`. Use it when\n  the extra two Claude subagent spawns are not worth it.\n\nIf solo is requested, write a resolved config with `pe` and `qa` set to\n`enabled: false` and use it as `$CONFIG_RESOLVED` for the REST of this run\n(both the section-2 enumeration AND the §4 aggregator — `panel aggregate`\nre-derives its panelist set from its `--config`, so it would otherwise\nsilently disagree with what was dispatched).\n\n```bash\nCONFIG_RESOLVED=\"${TMPDIR:-\u002Ftmp}\u002Fpanel-config-resolved-${CLAUDE_SESSION_ID:-$PPID}.yml\"\n\"${CLAUDE_PANEL_PYTHON:-python3.12}\" - \"$CONFIG\" \"$CONFIG_RESOLVED\" \u003C\u003C'PY'\nimport sys, yaml\ncfg = yaml.safe_load(open(sys.argv[1]))\nfor p in cfg.get(\"panelists\", []):\n    if p.get(\"id\") in (\"pe\", \"qa\"):\n        p[\"enabled\"] = False\nyaml.safe_dump(cfg, open(sys.argv[2], \"w\"))\nPY\n```\n\nThen re-run the section-2 enumeration against `$CONFIG_RESOLVED` (it now\nlists `da` only) and pass `$CONFIG_RESOLVED` to the §4 aggregator.\n\nOtherwise use the stock `$CONFIG` unchanged and skip to step 3.\n\n### 3. Create the per-session workdir\n\n```bash\nWORKDIR=\"${HOME}\u002F.claude\u002Fpanel\u002Fwork\u002F${CLAUDE_SESSION_ID:-$PPID}\"\nmkdir -p \"$WORKDIR\" && chmod 0700 \"$WORKDIR\"\n```\n\nThis is the directory where per-panelist verdict files land.\n`panel aggregate` will read `${WORKDIR}\u002F\u003Cid>.verdict` for each enabled\npanelist after fan-out completes.\n\n## Per-question dispatch\n\nFor EACH question in `tool_input.questions` that has an option labeled\nwith `(Recommended)` AND NOT `(Recommended; Panel-flagged)`:\n\n### 1. Build the per-panelist prompt files\n\nWrite the question payload to a JSON file, then let the CLI build each\npanelist's prompt. Blinding is applied automatically from each panelist's\n`blind` flag — do NOT hand-write prompt bodies, and do NOT mention the\nrecommended option to a blinded panelist.\n\n```bash\nQPAYLOAD=\"${TMPDIR:-\u002Ftmp}\u002Fpanel-question-${CLAUDE_SESSION_ID:-$PPID}-q\u003CN>.json\"\n# Write JSON: {\"question\": ..., \"options\": [{\"label\":..., \"description\":...}],\n#              \"recommended_label\": ..., \"reasoning\": ...}\n# reasoning comes from the recommended option's description; if there is\n# none, use \"(no reasoning supplied)\". NEVER invent hidden reasoning.\n\n# The cd is load-bearing, exactly as in the lint-config, dispatch and\n# aggregate blocks. $PYTHONPATH is $HOME\u002F.claude, so from any other cwd\n# `python -m panel` resolves `panel` to ~\u002F.claude\u002Fpanel — the CONFIG\n# directory — which has no __init__.py and shadows the real package as a\n# namespace package. The call then dies with \"No module named\n# panel.__main__\" and writes no prompt file at all.\ncd \"${HOME}\u002F.claude\u002Fskills\u002Fvalidate-recommendation\"\n\nfor PID in \u003Ceach enabled panelist id>; do\n  \"${CLAUDE_PANEL_PYTHON:-python3.12}\" -m panel build-prompt \\\n    --panelist \"$PID\" \\\n    --config \"${CONFIG_RESOLVED:-$CONFIG}\" \\\n    --question-file \"$QPAYLOAD\" \\\n    --output \"${TMPDIR:-\u002Ftmp}\u002Fpanel-prompt-${CLAUDE_SESSION_ID:-$PPID}-q\u003CN>-${PID}.txt\"\ndone\n```\n\nWhere `\u003CN>` is the question index (0-based).\n\nEach panelist is then dispatched with ITS OWN prompt file. A blinded\npanelist's file has the recommendation removed, the `(Recommended)` marker\nstripped, and the options deterministically reordered.\n\nThe payload's `reasoning` field is read only for UNBLINDED panelists;\n`build-prompt` substitutes an empty string for a blinded one. That is why\nthe payload may carry the literal `(no reasoning supplied)` safely. Never\ncall the underlying `build_prompt_body` yourself with a blinded panelist and\na non-empty reasoning — including that literal, which is a non-empty string:\nit raises `ValueError` by design. Go through the CLI and the question is\nmoot.\n\n### 2. Fan out N panelists in ONE message\n\n**This is the parallelism point. All panelist dispatches MUST be in a\nsingle message so they run concurrently.**\n\nFor EACH enabled panelist from the config enumeration above:\n\n- **If `backend` starts with `nat-` (e.g., `nat-nim`):** include a\n  `Bash` tool call in your fan-out message. Export `CLAUDE_SESSION_ID`\n  from the state file's `session_id` field in the SAME command —\n  `panel\u002Ftrace.py` reads it from the environment for verdict-line\n  telemetry, and it is otherwise never a real exported variable (only\n  ever referenced with a `${CLAUDE_SESSION_ID:-$PPID}` fallback), which\n  is why trace lines were showing `session=unknown`:\n\n  ```bash\n  export CLAUDE_SESSION_ID=\"$(jq -r '.session_id' \"${TMPDIR:-\u002Ftmp}\u002Fclaude-panel-${CLAUDE_SESSION_ID:-$PPID}.json\")\" && \\\n  cd \"${HOME}\u002F.claude\u002Fskills\u002Fvalidate-recommendation\" && \\\n      \"${CLAUDE_PANEL_PYTHON:-python3.12}\" -m panel dispatch \\\n      --panelist \"\u003Cid>\" \\\n      --config \"${CONFIG_RESOLVED:-$CONFIG}\" \\\n      --persona \"${HOME}\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpersonas\u002F\u003Crole-lowercase>.md\" \\\n      --prompt-file \"${TMPDIR:-\u002Ftmp}\u002Fpanel-prompt-${CLAUDE_SESSION_ID:-$PPID}-q\u003CN>-\u003Cid>.txt\" \\\n      --output \"${WORKDIR}\u002F\u003Cid>.verdict\"\n  ```\n\n  The `--prompt-file` is THIS panelist's file from step 1, not a shared\n  one. Passing another panelist's file hands a blinded seat the\n  recommendation.\n\n  `dispatch.py` writes the verdict file directly (mode 0600). Auth\n  comes from per-backend env vars (`nat-nim`: `$PANEL_DA_API_KEY`,\n  falling back to `$NVIDIA_API_KEY`; `nat-anthropic`:\n  `$ANTHROPIC_API_KEY`; `nat-openai`: `$OPENAI_API_KEY`).\n\n- **If `backend` is `claude-subagent`:** include an `Agent` tool call\n  in your fan-out message:\n  - `subagent_type`: the panelist's `subagent_type` from config (e.g.,\n    `principal-engineer`, `qa-engineer`)\n  - `description`: short, e.g., `\"Panel \u003Crole> review\"`\n  - `prompt`: the concatenation of the persona file's `# System prompt`\n    section, the `# One-shot example` section (if present), and the\n    contents of THIS panelist's prompt file from step 1\n    (`…-q\u003CN>-\u003Cid>.txt`), copied verbatim.\n\n  Read both files via the `Read` tool BEFORE composing the fan-out\n  message: the persona at\n  `${HOME}\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpersonas\u002F\u003Crole-lowercase>.md`,\n  and the panelist's own prompt file. Never re-compose the prompt body by\n  hand — for a blinded panelist that is the leak.\n\nAll `Bash` + `Agent` calls go in ONE message. The framework executes\nthem concurrently.\n\n### 3. Collect verdict files after fan-out returns\n\nWhen the fan-out message's tool results come back:\n\n- **`nat-*` panelists:** `dispatch.py` already wrote the verdict file.\n  Nothing to do.\n- **`claude-subagent` panelists:** the `Agent` tool returned a string.\n  Write the ENTIRE response verbatim to the verdict file via the\n  `Write` tool:\n\n  ```\n  ${WORKDIR}\u002F\u003Cid>.verdict\n  ```\n\n  Do NOT alter, summarize, or extract from the Agent response. The\n  aggregator does the parsing, and which lines it wants depends on the\n  panelist: an UNBLINDED one emits `VERDICT:`\u002F`RATIONALE:`\u002F`ALTERNATIVE:`,\n  a BLINDED one emits `CHOICE:`\u002F`RATIONALE:`. `aggregate()` maps a blinded\n  `CHOICE:` onto the same contract — agreement with the recommended label\n  is HOLD, disagreement is OVERTURN naming that option — comparing\n  marker- and case-insensitively. Surrounding prose is ignored, and a\n  blinded reply with no `CHOICE:` line becomes ERROR for that panelist.\n\n### 4. Run the aggregator\n\nSame rule as dispatch: export `CLAUDE_SESSION_ID` from the state file's\n`session_id` inside this SAME command, so the aggregator's trace lines\ncarry the real session instead of `session=unknown`.\n\n```bash\nDIRECTIVE_JSON=$(export CLAUDE_SESSION_ID=\"$(jq -r '.session_id' \"${TMPDIR:-\u002Ftmp}\u002Fclaude-panel-${CLAUDE_SESSION_ID:-$PPID}.json\")\" && \\\n    cd \"${HOME}\u002F.claude\u002Fskills\u002Fvalidate-recommendation\" && \\\n    \"${CLAUDE_PANEL_PYTHON:-python3.12}\" -m panel aggregate \\\n    --config \"${CONFIG_RESOLVED:-$CONFIG}\" \\\n    --verdicts-dir \"$WORKDIR\" \\\n    --recommended-label \"\u003Crecommended_label from state for THIS question>\")\n```\n\nCapture stdout into `$DIRECTIVE_JSON`. It's a single-line JSON object.\n\n### 5. Parse the directive\n\nExtract fields via `jq`:\n\n```bash\nVERDICT=$(jq -r '.verdict' \u003C\u003C\u003C \"$DIRECTIVE_JSON\")\nSUMMARY=$(jq -r '.summary' \u003C\u003C\u003C \"$DIRECTIVE_JSON\")\nESCALATE=$(jq -r '.escalate_to_user \u002F\u002F false' \u003C\u003C\u003C \"$DIRECTIVE_JSON\")\nGATE=$(jq -r '.rationale_gate_passed' \u003C\u003C\u003C \"$DIRECTIVE_JSON\")\n```\n\nValues:\n- `VERDICT` ∈ `{HOLD, SOFT-DISSENT, HARD-DISSENT, ERROR}`\n- `SUMMARY` — one-line user-facing text, already sanitized\n- `ESCALATE` — `true` for HARD-DISSENT in Phase 3c, `false` otherwise\n- `GATE` — `true`\u002F`false`\u002F`null` per the rationale gate semantics\n\n## Acting on the directive\n\nBehavior matrix:\n\n### `VERDICT == \"HOLD\"`\n\nBehavior depends on `${CLAUDE_PANEL:-on}`:\n\n**Default (`on`)** — auto-take mode:\n\nThe recommendation stands. Take the recommended option without asking\nthe user. Print a brief message:\n\n> Panel validated `\u003Crecommended_label>` for \"\u003Cquestion excerpt>\".\n> \u003CSUMMARY>\n> Proceeding.\n\nThen continue the work as if the user had picked the recommended\noption. The `AskUserQuestion` for this question is NEVER issued.\n\n**`advise`** — advisory mode (preserves user agency):\n\nDo NOT auto-take. Re-issue `AskUserQuestion` for this question with\nthe panel's affirmative annotation appended. Augmented payload:\n\n- **Question text**: original question + two newlines +\n  `**Panel validated:** \u003CSUMMARY>` (positive phrasing).\n- **Options**: identical, EXCEPT the recommended option's label has\n  `(Recommended)` swapped for `(Recommended; Panel-flagged)`. The hook\n  detects this marker and skips the panel on the re-ask.\n\n### `VERDICT == \"SOFT-DISSENT\"`\n\nRe-issue `AskUserQuestion` for this question. Augmented payload:\n\n- **Question text**: original + two newlines + `\u003CSUMMARY>` (the\n  `**Panel review:**` line). Use it verbatim.\n- **Options**: identical, EXCEPT the recommended option's label is\n  swapped to `(Recommended; Panel-flagged)`.\n\n### `VERDICT == \"HARD-DISSENT\"`\n\nIn Phase 3c, HARD-DISSENT also escalates to the user (Phase 5 will add\nre-think cycles). Same payload as SOFT-DISSENT, but include severity\nin the augmented note:\n\n- **Question text**: original + two newlines + `Panel HARD-DISSENT:\n  \u003CSUMMARY>` (severity-clear prefix).\n- **Options**: identical, EXCEPT marker swap to\n  `(Recommended; Panel-flagged)`.\n\nWhen Phase 5 lands, this branch will check `re_brainstorm` in the\ndirective and emit a re-think markdown directive (no `AskUserQuestion`)\nwhen present. Until then, every HARD-DISSENT escalates the question to\nthe user with full panel feedback.\n\n### `VERDICT == \"ERROR\"`\n\nSomething went wrong (panelist backend down, malformed responses,\nmissing files, aggregator parse error). For this question, re-issue\nthe original `AskUserQuestion` unmodified — same options, same\n`(Recommended)` marker — BUT swap the marker to\n`(Recommended; Panel-flagged)` for loop safety.\n\nPrint a brief explanation:\n\n> Panel evaluation failed for \"\u003Cquestion excerpt>\" (see hook trace\n> for detail). Asking the question directly.\n\n## Cleanup\n\nAfter processing ALL questions (HOLD + DISSENT + ERROR paths), remove\nevery file this run created — **one explicit path per `rm -f` argument,\nnever a glob (`*`)**. Under zsh, an unmatched glob aborts the ENTIRE\ncommand line, not just that argument — so a single `rm -f a b* c` where\n`b*` matches nothing removes NEITHER `a` NOR `c` either. This is the\nroot cause of the leak this section exists to close: a bare\n`panel-prompt-*-q*.txt` glob that matched nothing silently left\n`$STATE_FILE` on disk too, and the hook's re-entry guard then bypassed\nthe NEXT, unrelated paneled question it saw.\n\n```bash\nrm -rf \"$WORKDIR\"\nrm -f \"$STATE_FILE\"\n```\n\nThen remove each question payload and each per-panelist prompt file you\nactually wrote, one `rm -f` per exact path — using the same 0-based\nquestion indices you dispatched in the per-question loop above, crossed\nwith the panelist ids you enumerated (you know both sets; do not glob\n`-q*.json` or `-q*-*.txt`):\n\n```bash\nrm -f \"${TMPDIR:-\u002Ftmp}\u002Fpanel-question-${CLAUDE_SESSION_ID:-$PPID}-q0.json\"\nrm -f \"${TMPDIR:-\u002Ftmp}\u002Fpanel-prompt-${CLAUDE_SESSION_ID:-$PPID}-q0-da.txt\"\nrm -f \"${TMPDIR:-\u002Ftmp}\u002Fpanel-prompt-${CLAUDE_SESSION_ID:-$PPID}-q0-pe.txt\"\nrm -f \"${TMPDIR:-\u002Ftmp}\u002Fpanel-prompt-${CLAUDE_SESSION_ID:-$PPID}-q0-qa.txt\"\n# ... then the same four lines for q1, q2, ... — one payload plus one\n# prompt file per enabled panelist, for every question index you actually\n# dispatched. Use the ids from YOUR enumeration, not this example's:\n# solo profile writes only the da file, and a config with other panelists\n# writes theirs. Skip indices that were never a (Recommended) question —\n# there is no file for them, and rm -f on a missing path is a no-op.\n#\n# ${CLAUDE_SESSION_ID:-$PPID} must match the fallback used when the files\n# were built. A bare ${CLAUDE_SESSION_ID} resolves to an empty string in a\n# session where it is unset, and removes a path nothing ever wrote.\n```\n\nIf step 2a resolved a solo profile, `$CONFIG_RESOLVED` holds the exact\nresolved-config path already — remove that path directly (it is a\nsingle concrete path, not a glob):\n\n```bash\nif [ -n \"${CONFIG_RESOLVED:-}\" ]; then\n    rm -f \"$CONFIG_RESOLVED\"\nfi\n```\n\nIf you crash before cleanup, stale files are harmless PROVIDED\n`CLAUDE_SESSION_ID` is unique per session (different session = different\npath). The hook's re-entry guard on `$STATE_FILE` is now stale-aware —\nage >= 600s discards the leftover and replaces it with a fresh trigger\ninstead of bypassing indefinitely — so a leftover `$STATE_FILE` from a\ncrashed run self-heals within 10 minutes even if this cleanup never\nruns. `panel gc` will eventually reap stale workdirs (Phase 6).\n\n## Failure modes you must handle gracefully\n\n| What goes wrong | Behavior |\n|---|---|\n| State file missing | Print fallback message; re-issue original `AskUserQuestion`. |\n| `panel lint-config` fails | Print \"Panel disabled: config invalid\"; re-issue original. |\n| `~\u002F.claude\u002Fpanel\u002Fconfig.yml` missing | Caught by `lint-config`; same fallback. |\n| Persona file missing for an ENABLED panelist | Caught up front by `lint-config`'s persona\u002Fblind cross-check; same fallback as an invalid config. A disabled panelist's persona is never loaded. |\n| `panel dispatch` crashes (caller-bug exit 1) | Verdict file not written. Aggregator coerces to ERROR for that panelist. Severity decides per failure_mode. |\n| `panel dispatch` exits 0 but writes ERROR verdict | Normal path. Severity decides per failure_mode. |\n| `Agent` tool call errors \u002F returns garbled output | Write verbatim to verdict file; aggregator coerces to ERROR for that panelist when its expected line is missing — `VERDICT:` unblinded, `CHOICE:` blinded. |\n| `panel aggregate` crashes (non-zero exit) | Fall back. Re-issue original with marker swap. |\n| `$WORKDIR` unwritable | Fall back. Print \"Panel infrastructure unavailable\". |\n| Missing API keys (e.g., `$PANEL_DA_API_KEY`) | `dispatch.py` writes ERROR verdict; aggregator emits ERROR directive (at N=1) or degrades (at N>=3 with graceful failure_mode). User-facing message ends up as re-ask original. |\n\nThe whole panel is best-effort. The user-visible question ALWAYS\nsurvives.\n\n### Loop safety on fallback\n\nWhen you fall back to re-issuing the original `AskUserQuestion`, the\noriginal payload still has `(Recommended)` — meaning the hook will\nfire AGAIN. The hook has a re-entry guard: if a state file for the\ncurrent session already exists when the hook runs, and it is fresh\n(age \u003C 600s — always true for an immediate fallback re-issue), the\nhook removes it and approves (exit 0) without dispatching the skill\nagain. So your fallback re-issue is mechanically safe.\n\nAs a courtesy to future readers, swap the marker to\n`(Recommended; Panel-flagged)` in your fallback re-issue (it documents\nintent even though the re-entry guard makes it a no-op for safety).\n\n## Multi-question parallelism (optional optimization)\n\nIf the `AskUserQuestion` call had multiple questions each with a\n`(Recommended)` marker, you MAY dispatch panel calls for ALL\nquestions in parallel: fan out the N panelists × M questions in one\nmessage. Aggregate runs per question after both panelists return.\n\nFor v1, sequential per-question dispatch is fine. Optimize only if\nmulti-question recommendations become common (they aren't currently).\n\n## What you must NOT do\n\n- Do NOT echo `$PANEL_DA_API_KEY` (or any other API key) in any\n  user-facing text or trace output.\n- Do NOT write API keys to any file.\n- Do NOT modify or summarize the Agent tool's response before writing\n  it to the verdict file. Write verbatim; let the aggregator parse.\n- Do NOT modify the `AskUserQuestion` call beyond augmenting question\n  text (for dissents) or swapping the marker (for ERROR or fallback).\n- Do NOT re-invoke this skill after auto-taking a HOLD recommendation.\n  The work continues as if the user picked.\n- Do NOT introduce a panelist not in `config.yml`. New panelists ship\n  via `config.yml` + a persona file under `personas\u002F\u003Crole>.md` — never\n  ad-hoc in this file.\n- Do NOT call `python3.12 -m panel ...` for JSON parsing. Use `jq`.\n  Python invocation is reserved for `lint-config`, `build-prompt`,\n  `dispatch`, and `aggregate`.\n- Do NOT bypass the hook re-entry guard by clearing the state file\n  yourself. The guard exists to break loops on fallback paths.\n",{"name":5,"description":7},"main","26b9a4abffc3b1c9c228a44076c9d3f682ea5b53",".claude\u002Fskills\u002Fvalidate-recommendation",1,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit","https:\u002F\u002Fgithub.com\u002FArangoGutierrez","https:\u002F\u002Fwww.linkedin.com\u002Fin\u002Feduardo-arango\u002F","Senior Systems Software Engineer.\r\n\r\nEnvironmental engineer lost to distributed systems, HPC and cloud computing.","2026-10-13 16:16:20",{"source":8,"skillName":5,"commands":34,"prompt":44,"zipAvailable":45},[35,38,41],{"tool":36,"command":37},"npx","npx skills add ArangoGutierrez\u002Fclaude-toolkit --skill validate-recommendation",{"tool":39,"command":40},"bunx","bunx skills add ArangoGutierrez\u002Fclaude-toolkit --skill validate-recommendation",{"tool":42,"command":43},"pnpm","pnpm dlx skills add ArangoGutierrez\u002Fclaude-toolkit --skill validate-recommendation","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002FSKILL.md",true,[47,53,57,63,67,71,75,79,83,87,91,95,99,103,107,111,115,119,123,127,131,135,139,143,147,151,155,159,163,167,171,175,179,183,187,191,195,199,203,207,211,215,219,223,227,231,235,239,243,247,251,255,259],{"path":48,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":51,"sourceHtmlUrl":52},"detect-team-context_test.sh","SCRIPT","脚本",1449,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fdetect-team-context_test.sh",{"path":54,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":55,"sourceHtmlUrl":56},"detect-team-context.sh",1121,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fdetect-team-context.sh",{"path":58,"resourceType":59,"resourceTypeDesc":60,"sizeBytes":61,"sourceHtmlUrl":62},"fixtures\u002Fda_hold.txt","REFERENCE","参考文档",388,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Ffixtures\u002Fda_hold.txt",{"path":64,"resourceType":59,"resourceTypeDesc":60,"sizeBytes":65,"sourceHtmlUrl":66},"fixtures\u002Fda_overturn_b.txt",237,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Ffixtures\u002Fda_overturn_b.txt",{"path":68,"resourceType":59,"resourceTypeDesc":60,"sizeBytes":69,"sourceHtmlUrl":70},"fixtures\u002Fda_response_api_error.json",109,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Ffixtures\u002Fda_response_api_error.json",{"path":72,"resourceType":59,"resourceTypeDesc":60,"sizeBytes":73,"sourceHtmlUrl":74},"fixtures\u002Fda_response_hold.json",748,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Ffixtures\u002Fda_response_hold.json",{"path":76,"resourceType":59,"resourceTypeDesc":60,"sizeBytes":77,"sourceHtmlUrl":78},"fixtures\u002Fda_response_malformed.json",352,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Ffixtures\u002Fda_response_malformed.json",{"path":80,"resourceType":59,"resourceTypeDesc":60,"sizeBytes":81,"sourceHtmlUrl":82},"fixtures\u002Fda_response_null_content.json",334,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Ffixtures\u002Fda_response_null_content.json",{"path":84,"resourceType":59,"resourceTypeDesc":60,"sizeBytes":85,"sourceHtmlUrl":86},"fixtures\u002Fda_response_overturn_missing_alt.json",341,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Ffixtures\u002Fda_response_overturn_missing_alt.json",{"path":88,"resourceType":59,"resourceTypeDesc":60,"sizeBytes":89,"sourceHtmlUrl":90},"fixtures\u002Fda_response_overturn_no_alt.json",399,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Ffixtures\u002Fda_response_overturn_no_alt.json",{"path":92,"resourceType":59,"resourceTypeDesc":60,"sizeBytes":93,"sourceHtmlUrl":94},"fixtures\u002Fda_response_overturn.json",742,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Ffixtures\u002Fda_response_overturn.json",{"path":96,"resourceType":59,"resourceTypeDesc":60,"sizeBytes":97,"sourceHtmlUrl":98},"fixtures\u002Fmalformed.txt",101,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Ffixtures\u002Fmalformed.txt",{"path":100,"resourceType":59,"resourceTypeDesc":60,"sizeBytes":101,"sourceHtmlUrl":102},"fixtures\u002Fpe_hold.txt",223,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Ffixtures\u002Fpe_hold.txt",{"path":104,"resourceType":59,"resourceTypeDesc":60,"sizeBytes":105,"sourceHtmlUrl":106},"fixtures\u002Fpe_overturn_c.txt",232,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Ffixtures\u002Fpe_overturn_c.txt",{"path":108,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":109,"sourceHtmlUrl":110},"panel\u002F__init__.py",86,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002F__init__.py",{"path":112,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":113,"sourceHtmlUrl":114},"panel\u002F__main__.py",87,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002F__main__.py",{"path":116,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":117,"sourceHtmlUrl":118},"panel\u002Faggregate.py",5316,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Faggregate.py",{"path":120,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":121,"sourceHtmlUrl":122},"panel\u002Fcli.py",9206,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Fcli.py",{"path":124,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":125,"sourceHtmlUrl":126},"panel\u002Fconfig.py",7051,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Fconfig.py",{"path":128,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":129,"sourceHtmlUrl":130},"panel\u002Fdispatch.py",5067,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Fdispatch.py",{"path":132,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":133,"sourceHtmlUrl":134},"panel\u002Fpersonas.py",3199,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Fpersonas.py",{"path":136,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":137,"sourceHtmlUrl":138},"panel\u002Fprompt.py",3021,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Fprompt.py",{"path":140,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":141,"sourceHtmlUrl":142},"panel\u002Fsanitize.py",977,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Fsanitize.py",{"path":144,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":145,"sourceHtmlUrl":146},"panel\u002Fseverity.py",10184,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Fseverity.py",{"path":148,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":149,"sourceHtmlUrl":150},"panel\u002Fstats.py",2296,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Fstats.py",{"path":152,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":153,"sourceHtmlUrl":154},"panel\u002Ftelemetry.py",4117,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Ftelemetry.py",{"path":156,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":157,"sourceHtmlUrl":158},"panel\u002Ftests\u002F__init__.py",0,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Ftests\u002F__init__.py",{"path":160,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":161,"sourceHtmlUrl":162},"panel\u002Ftests\u002Fconftest.py",2431,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Ftests\u002Fconftest.py",{"path":164,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":165,"sourceHtmlUrl":166},"panel\u002Ftests\u002Ftest_aggregate.py",16842,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Ftests\u002Ftest_aggregate.py",{"path":168,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":169,"sourceHtmlUrl":170},"panel\u002Ftests\u002Ftest_cli_aggregate.py",2823,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Ftests\u002Ftest_cli_aggregate.py",{"path":172,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":173,"sourceHtmlUrl":174},"panel\u002Ftests\u002Ftest_cli_build_prompt.py",4177,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Ftests\u002Ftest_cli_build_prompt.py",{"path":176,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":177,"sourceHtmlUrl":178},"panel\u002Ftests\u002Ftest_cli_dispatch.py",3321,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Ftests\u002Ftest_cli_dispatch.py",{"path":180,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":181,"sourceHtmlUrl":182},"panel\u002Ftests\u002Ftest_cli_exit_codes.py",3091,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Ftests\u002Ftest_cli_exit_codes.py",{"path":184,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":185,"sourceHtmlUrl":186},"panel\u002Ftests\u002Ftest_cli_lint_config.py",7716,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Ftests\u002Ftest_cli_lint_config.py",{"path":188,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":189,"sourceHtmlUrl":190},"panel\u002Ftests\u002Ftest_cli_stats.py",6816,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Ftests\u002Ftest_cli_stats.py",{"path":192,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":193,"sourceHtmlUrl":194},"panel\u002Ftests\u002Ftest_config.py",11126,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Ftests\u002Ftest_config.py",{"path":196,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":197,"sourceHtmlUrl":198},"panel\u002Ftests\u002Ftest_dispatch.py",13567,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Ftests\u002Ftest_dispatch.py",{"path":200,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":201,"sourceHtmlUrl":202},"panel\u002Ftests\u002Ftest_personas.py",7616,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Ftests\u002Ftest_personas.py",{"path":204,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":205,"sourceHtmlUrl":206},"panel\u002Ftests\u002Ftest_prompt.py",5403,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Ftests\u002Ftest_prompt.py",{"path":208,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":209,"sourceHtmlUrl":210},"panel\u002Ftests\u002Ftest_sanitize.py",1268,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Ftests\u002Ftest_sanitize.py",{"path":212,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":213,"sourceHtmlUrl":214},"panel\u002Ftests\u002Ftest_severity.py",13545,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Ftests\u002Ftest_severity.py",{"path":216,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":217,"sourceHtmlUrl":218},"panel\u002Ftests\u002Ftest_shipped_config_template.py",2312,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Ftests\u002Ftest_shipped_config_template.py",{"path":220,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":221,"sourceHtmlUrl":222},"panel\u002Ftests\u002Ftest_skill_doc_invocations.py",6043,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Ftests\u002Ftest_skill_doc_invocations.py",{"path":224,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":225,"sourceHtmlUrl":226},"panel\u002Ftests\u002Ftest_stats.py",4293,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Ftests\u002Ftest_stats.py",{"path":228,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":229,"sourceHtmlUrl":230},"panel\u002Ftests\u002Ftest_telemetry.py",9799,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Ftests\u002Ftest_telemetry.py",{"path":232,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":233,"sourceHtmlUrl":234},"panel\u002Ftests\u002Ftest_trace.py",2849,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Ftests\u002Ftest_trace.py",{"path":236,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":237,"sourceHtmlUrl":238},"panel\u002Ftests\u002Ftest_verdict.py",2145,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Ftests\u002Ftest_verdict.py",{"path":240,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":241,"sourceHtmlUrl":242},"panel\u002Ftrace.py",2720,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Ftrace.py",{"path":244,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":245,"sourceHtmlUrl":246},"panel\u002Fverdict.py",1808,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpanel\u002Fverdict.py",{"path":248,"resourceType":59,"resourceTypeDesc":60,"sizeBytes":249,"sourceHtmlUrl":250},"personas\u002Fda.md",2899,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpersonas\u002Fda.md",{"path":252,"resourceType":59,"resourceTypeDesc":60,"sizeBytes":253,"sourceHtmlUrl":254},"personas\u002Fpe.md",2923,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpersonas\u002Fpe.md",{"path":256,"resourceType":59,"resourceTypeDesc":60,"sizeBytes":257,"sourceHtmlUrl":258},"personas\u002Fqa.md",3017,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002Fpersonas\u002Fqa.md",{"path":260,"resourceType":59,"resourceTypeDesc":60,"sizeBytes":261,"sourceHtmlUrl":262},"README.md",12249,"https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fvalidate-recommendation\u002FREADME.md",[264,271,278,285,292,299],{"skillNo":265,"name":266,"nameZh":6,"description":267,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":268,"categories":269,"keywords":270,"publishedAt":19},"01M7802B5QY9YG45W2HR7ZM17S","update-cursor-settings","Modify Cursor\u002FVSCode user settings in settings.json. Use when you want to change editor settings, preferences, configuration, themes, font size, tab size, format on save, auto save, keybindings, or any settings.json values.","https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.cursor\u002Fskills-cursor\u002Fupdate-cursor-settings\u002FSKILL.md",[],[],{"skillNo":272,"name":273,"nameZh":6,"description":274,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":275,"categories":276,"keywords":277,"publishedAt":19},"01M780277C2PMVHZQGN6W0VWNB","update-cli-config","View and modify Cursor CLI configuration settings in ~\u002F.cursor\u002Fcli-config.json. Use when the user wants to change CLI settings, configure permissions, switch approval mode, enable vim mode, toggle display options, configure sandbox, or manage any CLI preferences.","https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.cursor\u002Fskills-cursor\u002Fupdate-cli-config\u002FSKILL.md",[],[],{"skillNo":279,"name":280,"nameZh":6,"description":281,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":282,"categories":283,"keywords":284,"publishedAt":19},"01M78023CDY74JXKCZTDJTYVQF","statusline","Configure a custom status line in the CLI. Use when the user mentions status line, statusline, statusLine, CLI status bar, prompt footer customization, or wants to add session context above the prompt.","https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.cursor\u002Fskills-cursor\u002Fstatusline\u002FSKILL.md",[],[],{"skillNo":286,"name":287,"nameZh":6,"description":288,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":289,"categories":290,"keywords":291,"publishedAt":19},"01M780209KQZBVMRX43KMZH5TT","split-to-prs","Split current work into small reviewable PRs. Use when the user asks to split a chat, set of changes, branch, or PR.","https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.cursor\u002Fskills-cursor\u002Fsplit-to-prs\u002FSKILL.md",[],[],{"skillNo":293,"name":294,"nameZh":6,"description":295,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":296,"categories":297,"keywords":298,"publishedAt":19},"01M7801XXCWKR3PACKNX01RXMT","shell","Runs the rest of a \u002Fshell request as a literal shell command. Use only when the user explicitly invokes \u002Fshell and wants the following text executed directly in the terminal.","https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.cursor\u002Fskills-cursor\u002Fshell\u002FSKILL.md",[],[],{"skillNo":300,"name":301,"nameZh":6,"description":302,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":303,"categories":304,"keywords":305,"publishedAt":19},"01M7801NF9TC7NWNTZKXQCZXAW","sdk","Guide users building apps, scripts, CI pipelines, or automations on top of the Cursor SDK - TypeScript (`@cursor\u002Fsdk`) or Python (`cursor-sdk` \u002F `cursor_sdk`). Use when the user mentions integrating, installing, or writing code against the Cursor SDK; says `Agent.create`, `Agent.prompt`, `Agent.resume`, `agent.send`, `run.stream`, `run.messages`, `CursorAgentError`, `@cursor\u002Fsdk`, `cursor-sdk`, or `cursor_sdk`; asks to run Cursor agents programmatically from a script, CI\u002FCD pipeline, GitHub Action, backend service, or other code outside the Cursor IDE; wants to pick between local and cloud runtime, configure MCP servers for an SDK agent, or handle streaming, cancellation, or errors; or is wiring Cursor into an automation, bot, or REST `\u002Fv1\u002Fagents` migration. Use eagerly rather than answering from memory; the SDK surface evolves and this skill is the source of truth for the external packages.","https:\u002F\u002Fgithub.com\u002FArangoGutierrez\u002Fclaude-toolkit\u002Fblob\u002Fmain\u002F.cursor\u002Fskills-cursor\u002Fsdk\u002FSKILL.md",[],[]]