[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M79R12WSWT5SYBDJKSBVCF5V":3,"skill-same:01M79R12WSWT5SYBDJKSBVCF5V":57},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":6,"ownerHtmlUrl":29,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":30,"install":31,"resources":44},"01M79R12WSWT5SYBDJKSBVCF5V","eval-run",null,"Execute a project's eval and produce a results file - locate the golden set, scaffold a runner for the project's stack when none exists, grade every case by its own rule, and write per-case results plus a headline computed from that file. Reports verdict-first against the spec's target and never edits a golden case to raise a score. Use when the user asks to run the eval, score the golden set, or produce eval results.","AaravChadha\u002Facstack","AaravChadha","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F224610398?v=4","01M79QZCAGEMRNH32DKH6VC5RJ","01M79QZCB37TR94RNR32TM7YTV",2,"MIT","https:\u002F\u002Fgithub.com\u002FAaravChadha\u002Facstack\u002Fblob\u002Fmain\u002Fskills\u002Feval-run\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: eval-run\ndescription: \"Execute a project's eval and produce a results file - locate the golden set, scaffold a runner for the project's stack when none exists, grade every case by its own rule, and write per-case results plus a headline computed from that file. Reports verdict-first against the spec's target and never edits a golden case to raise a score. Use when the user asks to run the eval, score the golden set, or produce eval results.\"\nargument-hint: \"[eval-dir | notes]\"\n---\n\n# \u002Feval-run — execute the eval, honestly\n\n`\u002Feval-spec` writes the target before the code exists; `\u002Faudit eval`\nreviews results after; `\u002Fship`'s gate 3 compares a headline to a target.\nAll three assumed a runner nobody produced. This skill is that runner —\nthe step that turns \"the eval is the spec\" from a claim into a loop.\n\nRunning an eval **costs money and time** when the system under test\ncalls a model API. State the case count and that cost before executing,\nthen run — the user asked for a run, so this is not a permission\nrequest (conduct rule 5), it is a disclosure.\n\n`Adjacent skills:` \u002Feval-spec (writes the spec and golden set before any\ncode; \u002Feval-run executes it) · \u002Faudit eval (audits the results this\nproduces, including recomputing the headline) · \u002Fship (gate 3 compares\nthis headline against the spec's target).\n\n\u003C!-- acstack:runtime -->\nRun before the skill's steps — per invocation, not per session (4.36); failures degrade to markdown:\n```bash\nlink=\"$(readlink \"$HOME\u002F.claude\u002Fskills\u002Fhealth\" 2>\u002Fdev\u002Fnull || true)\"   # empty = not symlinked\npack=\"$(dirname \"$(dirname \"$link\")\")\"   # NEVER trust this unless $link was non-empty\nif [ \"${link#\u002F}\" != \"$link\" ] && [ -x \"$pack\u002Fbin\u002Facstack-config\" ] && ! \"$pack\u002Fbin\u002Facstack-config\" runtime | grep -q '=off'; then\n  \"$pack\u002Fbin\u002Facstack-config\" || true          # resolved keys, with sources\n  \"$pack\u002Fbin\u002Facstack-update-check\" || true    # ≤1 fetch\u002Fday; silent ONLY if already checked today\n  \"$pack\u002Fbin\u002Facstack-recall\" || true          # LEARNINGS.md + bug-class names, capped 3KB\nelse\n  echo \"runtime off — proceeding without recall\u002Fupdate-check\"\nfi\n```\n\u003C!-- \u002Facstack:runtime -->\n\n\u003C!-- acstack:principles -->\n## Operating principles\n\n- Be direct. Push back in writing when the plan or the user is wrong. No sycophancy.\n- Never delete a decision. Supersede it: `~~old~~ → **Verdict (YYYY-MM-DD):** new call — reason.`\n- Never fix, tune, or delete a test or eval case to raise a score. Log the miss honestly and leave the case unchanged.\n- Name exact things: regex patterns, function signatures, model names, before → after numbers. Never \"fixed bugs\".\n- Attribution: follow the project's `attribution` setting (default `none`) — no AI-tool mentions in generated docs, no attribution trailers in commits or PRs. Commit with explicit `-m`\u002F`-F` messages only.\n- Config: read `.claude\u002Facstack.md` at the project root (fall back to `~\u002F.claude\u002Facstack.md`) before acting. `## Settings` keys override pack defaults; a `## \u003Cskill-name>` section overrides both. Unknown keys and sections are ignored.\n- Docs: BRIEF.md (frozen seed) \u002F PLAN.md (living plan) \u002F JOURNAL.md (rolling journal). If the repo uses legacy names (PLANNING_PROMPT.md \u002F PLANNING.md \u002F STATUS.md), use those instead — never create both.\n- Recall: if `LEARNINGS.md` exists at the project root, read it before starting.\n- Conduct: follow the `acstack-conduct` block in this repo's AGENTS.md — the word is the mode; the user sets the pace.\n- Hackathon lane: if the project's AGENTS.md carries the `acstack:hackathon-lane` block, only `\u002Fdo` changes the repository during the event. Any other skill that would write a tracked file, commit or push says what it would have done and stops; a change that is not a task goes through the lane's operator route.\n\u003C!-- \u002Facstack:principles -->\n\n**One document set.** Resolve exactly ONE BRIEF\u002FPLAN\u002FJOURNAL set and name\nits path in the report's scope line. If more than one candidate set exists\n— a monorepo, nested products, an `apps\u002F*` tree each with its own docs —\nlist the candidates and STOP. Never pick one silently: a confident answer\nabout the wrong product is worse than no answer (conduct rule 8).\n\n## The sequence\n\n1. **Locate the eval.** `eval\u002Fspec.md` and `eval\u002Fgolden.jsonl` (or the\n   directory named in the argument). Either missing → say WHICH is\n   missing, point at `\u002Feval-spec`, and stop. Never scaffold a golden\n   set here; writing the target is a deliberate act that belongs to\n   \u002Feval-spec, and inventing cases at run time is the never-inflate\n   rule's exact failure mode.\n2. **Prefer the runner that exists.** If `spec.md`'s run command names a\n   runner present in the repo, run THAT and skip step 3 — an existing\n   eval artifact is never regenerated (\u002Feval-spec's step-0 rule).\n3. **Scaffold only when there is none.** Detect the stack:\n   `package.json` → `eval\u002Frun.mjs`; `pyproject.toml` or\n   `requirements.txt` → `eval\u002Frun.py`; neither, or both → **stop and\n   ask**. Never guess a stack. Build from\n   `references\u002Frunner-template.md`.\n\n   **A run command in the spec settles the stack — it is not a guess.** If\n   `spec.md` names `eval\u002Frun.py`, the language is already decided by the\n   committed spec, and stopping to ask which stack to use is a deadlock over\n   a question already answered. Take the extension from the spec's run\n   command and say that is where it came from. Only a spec that names no\n   runner leaves the question genuinely open.\n\n   The runner needs one fact the spec often lacks: **how to invoke the\n   system under test** — a CLI command, an HTTP endpoint, or a function\n   import. If `spec.md` does not say, ask, then record the answer in\n   spec.md's run section as a dated addition (that file is living; the\n   golden set is not).\n\n   **Unattended, the same settled-not-guessed rule applies:** when\n   nobody can answer, a committed document that states the exact\n   invocation — a PLAN exit criterion, an acceptance line — settles the\n   question; record it as the dated addition with the source named.\n   Only when no committed document states the invocation does an\n   unattended run stop and say what is missing. Never derive it from\n   reading the subject's code: the code cannot distinguish its intended\n   interface from an accident, so that is a guess wearing evidence.\n\n   **The invocation carries isolation flags and a pinned subject model, or\n   it is not measuring the subject.** Without isolation the operator's\n   user-level skills, hooks, memory and CLAUDE.md discovery, and output\n   styles all ride into the run — including into the BASELINE arm of an\n   A\u002FB, where they can make a candidate look better or worse than it is.\n   Without a pin the run silently uses whatever the operator or the CLI\n   release defaults to, so two runs are not comparable. Flags per stack are\n   in `references\u002Frunner-template.md`; state what each drops AND what still\n   leaks. An unpinned model STOPS the run — see the template's\n   `SUBJECT_MODEL` guard.\n\n   **Then the non-regression gate.** `references\u002Fregression-gate.py`\n   compares the new results file against the LAST COMMITTED one, per\n   category, and blocks when any category falls. The headline cannot see\n   this: a change that lifts the overall percentage while breaking every\n   refusal case passes both \u002Fship's target gate and the spec's category\n   minimums, since minimums constrain the golden set's composition, not one\n   run against another. No committed baseline PASSES and says so — never\n   silently, because an absent baseline read as a clean pass is the false\n   confidence the gate exists to remove.\n4. **Run it, then read the file.** The runner writes\n   `eval\u002Fresults\u002F\u003CUTC-timestamp>.jsonl`. The headline is computed from\n   that file — overall %, per-category %, refusal % — and never\n   assembled by hand from what the run appeared to do. If the two ever\n   disagree, the file wins and the discrepancy is the finding.\n5. **Report** (shape below). Propose PLAN edits; apply none.\n\n## Grading rules\n\nPer case, from its `grade_rule`, aligned with the canonical grader rules\nin `..\u002Feval-spec\u002Freferences\u002Fgrader-rules.md` and with `\u002Faudit eval`:\n\n- `exact` — normalized equality: trim, collapse whitespace, NFKC, and\n  fold case — unless the case carries `case_sensitive: true`, set when\n  the output's shape is itself part of the contract; then case is kept\n  and a right answer in the wrong case fails on shape. Same answer, not\n  same keystrokes.\n- `concept` — the expected concept is present. The scaffold implements\n  this as normalized substring containment, which is the FLOOR, not the\n  ideal: it is literal enough to produce brittleness. When a case fails\n  here and the answer was right, widen the grader — never touch the\n  case (\u002Faudit eval's \"grader brittleness\" bucket).\n- `numeric-tolerance:\u003Cx>` — parse both sides, compare within ±x\n  absolute, or ±x% relative to the expected value when the suffix is\n  `%` (both forms are documented by \u002Feval-spec's template).\n- `rubric:\u003Cname>` — not machine-gradeable. The runner records the\n  answer, marks the case `needs-rubric-review`, and **excludes it from\n  the headline while naming it**. A human or a judge model scores the\n  rubric's named dimensions; a rubric with unnamed dimensions is a spec\n  defect, not a pass.\n\nDenominator discipline, because a headline is only as honest as what it\ncounts: cases marked `\"status\": \"needs-data\"` are **skipped and reported\nas skipped**, never counted as passes; `\"status\": \"superseded\"` cases\nare excluded entirely; `rubric:` cases are excluded from the headline\nbecause no machine graded them, and are reported for review;\n`acceptable_failure` applies only when the case carries a written\n`reason` (in either shape it is written — a bool with a sibling\n`reason`, or an object), never to an ungraded case, and every\napplication is listed.\n\n**Every exclusion is named in the report.** A case that leaves the\ndenominator quietly moves no percentage at all, so it cannot be caught\nby looking at the number — which makes silent exclusion the most\ndangerous of the false-pass family.\n\n## Hard rules\n\n- **Never edit `golden.jsonl` to raise a score.** A genuinely wrong case\n  is superseded through \u002Feval-spec, with a new id and a written reason.\n  Editing a case to pass is the one thing this pack refuses outright.\n- **Never hand-transcribe the headline.** It comes from the results\n  file, every time.\n- A run that errors part-way reports what completed and what did not —\n  a partial run is never presented as a full score.\n- **A subject that cannot be invoked yields NO SCORE, not 0%.** When\n  the system under test needs a credential or endpoint the environment\n  lacks, record honest per-case errors in the results file, report the\n  headline as measuring the environment rather than the subject, and\n  name what is missing in Scope. Never set, invent, or mock the\n  credential to make the run go — a mocked subject scores the mock.\n- The results file is committed with the run: an eval score whose\n  evidence is not in the repo is an assertion.\n\n## Report shape\n\nVerdict first — `\u003Cheadline>% vs target \u003Cn>%` plus `MEETS TARGET` or\n`BELOW TARGET by \u003Cx>`; when nothing was graded, the verdict is instead\n`NO SCORE — \u003Cwhat is missing>` (per the hard rule above — a zero-graded\nrun has no headline to compare) — then:\n\n- per-category table (category | cases | passed | % | target)\n- failures, grouped by the \u002Faudit eval buckets (prompt issue, grader\n  brittleness, provider flake, data issue, parser issue, genuinely\n  ambiguous) — classification proposed, since confirming it is \u002Faudit's\n  job\n- `acceptable_failure` applications, each with its reason\n- skipped cases and why\n- the results file path, and the exact command that produced it\n- **Scope:** what was NOT run (needs-data cases, categories with no\n  cases, an unreachable system under test)\n\n## Positive control\n\nA golden set containing one deliberately failing case MUST produce a\nsub-100% headline. A runner that reports 100% by construction — because\nit counts skips as passes, swallows errors, or compares nothing — is\nthe false-pass class this pack exists to catch, and it is indetectable\nfrom a green report. `fixtures\u002Feval-run\u002F` seeds exactly that case.\n","# \u002Feval-run — execute the eval, honestly\n\n`\u002Feval-spec` writes the target before the code exists; `\u002Faudit eval`\nreviews results after; `\u002Fship`'s gate 3 compares a headline to a target.\nAll three assumed a runner nobody produced. This skill is that runner —\nthe step that turns \"the eval is the spec\" from a claim into a loop.\n\nRunning an eval **costs money and time** when the system under test\ncalls a model API. State the case count and that cost before executing,\nthen run — the user asked for a run, so this is not a permission\nrequest (conduct rule 5), it is a disclosure.\n\n`Adjacent skills:` \u002Feval-spec (writes the spec and golden set before any\ncode; \u002Feval-run executes it) · \u002Faudit eval (audits the results this\nproduces, including recomputing the headline) · \u002Fship (gate 3 compares\nthis headline against the spec's target).\n\n\u003C!-- acstack:runtime -->\nRun before the skill's steps — per invocation, not per session (4.36); failures degrade to markdown:\n```bash\nlink=\"$(readlink \"$HOME\u002F.claude\u002Fskills\u002Fhealth\" 2>\u002Fdev\u002Fnull || true)\"   # empty = not symlinked\npack=\"$(dirname \"$(dirname \"$link\")\")\"   # NEVER trust this unless $link was non-empty\nif [ \"${link#\u002F}\" != \"$link\" ] && [ -x \"$pack\u002Fbin\u002Facstack-config\" ] && ! \"$pack\u002Fbin\u002Facstack-config\" runtime | grep -q '=off'; then\n  \"$pack\u002Fbin\u002Facstack-config\" || true          # resolved keys, with sources\n  \"$pack\u002Fbin\u002Facstack-update-check\" || true    # ≤1 fetch\u002Fday; silent ONLY if already checked today\n  \"$pack\u002Fbin\u002Facstack-recall\" || true          # LEARNINGS.md + bug-class names, capped 3KB\nelse\n  echo \"runtime off — proceeding without recall\u002Fupdate-check\"\nfi\n```\n\u003C!-- \u002Facstack:runtime -->\n\n\u003C!-- acstack:principles -->\n## Operating principles\n\n- Be direct. Push back in writing when the plan or the user is wrong. No sycophancy.\n- Never delete a decision. Supersede it: `~~old~~ → **Verdict (YYYY-MM-DD):** new call — reason.`\n- Never fix, tune, or delete a test or eval case to raise a score. Log the miss honestly and leave the case unchanged.\n- Name exact things: regex patterns, function signatures, model names, before → after numbers. Never \"fixed bugs\".\n- Attribution: follow the project's `attribution` setting (default `none`) — no AI-tool mentions in generated docs, no attribution trailers in commits or PRs. Commit with explicit `-m`\u002F`-F` messages only.\n- Config: read `.claude\u002Facstack.md` at the project root (fall back to `~\u002F.claude\u002Facstack.md`) before acting. `## Settings` keys override pack defaults; a `## \u003Cskill-name>` section overrides both. Unknown keys and sections are ignored.\n- Docs: BRIEF.md (frozen seed) \u002F PLAN.md (living plan) \u002F JOURNAL.md (rolling journal). If the repo uses legacy names (PLANNING_PROMPT.md \u002F PLANNING.md \u002F STATUS.md), use those instead — never create both.\n- Recall: if `LEARNINGS.md` exists at the project root, read it before starting.\n- Conduct: follow the `acstack-conduct` block in this repo's AGENTS.md — the word is the mode; the user sets the pace.\n- Hackathon lane: if the project's AGENTS.md carries the `acstack:hackathon-lane` block, only `\u002Fdo` changes the repository during the event. Any other skill that would write a tracked file, commit or push says what it would have done and stops; a change that is not a task goes through the lane's operator route.\n\u003C!-- \u002Facstack:principles -->\n\n**One document set.** Resolve exactly ONE BRIEF\u002FPLAN\u002FJOURNAL set and name\nits path in the report's scope line. If more than one candidate set exists\n— a monorepo, nested products, an `apps\u002F*` tree each with its own docs —\nlist the candidates and STOP. Never pick one silently: a confident answer\nabout the wrong product is worse than no answer (conduct rule 8).\n\n## The sequence\n\n1. **Locate the eval.** `eval\u002Fspec.md` and `eval\u002Fgolden.jsonl` (or the\n   directory named in the argument). Either missing → say WHICH is\n   missing, point at `\u002Feval-spec`, and stop. Never scaffold a golden\n   set here; writing the target is a deliberate act that belongs to\n   \u002Feval-spec, and inventing cases at run time is the never-inflate\n   rule's exact failure mode.\n2. **Prefer the runner that exists.** If `spec.md`'s run command names a\n   runner present in the repo, run THAT and skip step 3 — an existing\n   eval artifact is never regenerated (\u002Feval-spec's step-0 rule).\n3. **Scaffold only when there is none.** Detect the stack:\n   `package.json` → `eval\u002Frun.mjs`; `pyproject.toml` or\n   `requirements.txt` → `eval\u002Frun.py`; neither, or both → **stop and\n   ask**. Never guess a stack. Build from\n   `references\u002Frunner-template.md`.\n\n   **A run command in the spec settles the stack — it is not a guess.** If\n   `spec.md` names `eval\u002Frun.py`, the language is already decided by the\n   committed spec, and stopping to ask which stack to use is a deadlock over\n   a question already answered. Take the extension from the spec's run\n   command and say that is where it came from. Only a spec that names no\n   runner leaves the question genuinely open.\n\n   The runner needs one fact the spec often lacks: **how to invoke the\n   system under test** — a CLI command, an HTTP endpoint, or a function\n   import. If `spec.md` does not say, ask, then record the answer in\n   spec.md's run section as a dated addition (that file is living; the\n   golden set is not).\n\n   **Unattended, the same settled-not-guessed rule applies:** when\n   nobody can answer, a committed document that states the exact\n   invocation — a PLAN exit criterion, an acceptance line — settles the\n   question; record it as the dated addition with the source named.\n   Only when no committed document states the invocation does an\n   unattended run stop and say what is missing. Never derive it from\n   reading the subject's code: the code cannot distinguish its intended\n   interface from an accident, so that is a guess wearing evidence.\n\n   **The invocation carries isolation flags and a pinned subject model, or\n   it is not measuring the subject.** Without isolation the operator's\n   user-level skills, hooks, memory and CLAUDE.md discovery, and output\n   styles all ride into the run — including into the BASELINE arm of an\n   A\u002FB, where they can make a candidate look better or worse than it is.\n   Without a pin the run silently uses whatever the operator or the CLI\n   release defaults to, so two runs are not comparable. Flags per stack are\n   in `references\u002Frunner-template.md`; state what each drops AND what still\n   leaks. An unpinned model STOPS the run — see the template's\n   `SUBJECT_MODEL` guard.\n\n   **Then the non-regression gate.** `references\u002Fregression-gate.py`\n   compares the new results file against the LAST COMMITTED one, per\n   category, and blocks when any category falls. The headline cannot see\n   this: a change that lifts the overall percentage while breaking every\n   refusal case passes both \u002Fship's target gate and the spec's category\n   minimums, since minimums constrain the golden set's composition, not one\n   run against another. No committed baseline PASSES and says so — never\n   silently, because an absent baseline read as a clean pass is the false\n   confidence the gate exists to remove.\n4. **Run it, then read the file.** The runner writes\n   `eval\u002Fresults\u002F\u003CUTC-timestamp>.jsonl`. The headline is computed from\n   that file — overall %, per-category %, refusal % — and never\n   assembled by hand from what the run appeared to do. If the two ever\n   disagree, the file wins and the discrepancy is the finding.\n5. **Report** (shape below). Propose PLAN edits; apply none.\n\n## Grading rules\n\nPer case, from its `grade_rule`, aligned with the canonical grader rules\nin `..\u002Feval-spec\u002Freferences\u002Fgrader-rules.md` and with `\u002Faudit eval`:\n\n- `exact` — normalized equality: trim, collapse whitespace, NFKC, and\n  fold case — unless the case carries `case_sensitive: true`, set when\n  the output's shape is itself part of the contract; then case is kept\n  and a right answer in the wrong case fails on shape. Same answer, not\n  same keystrokes.\n- `concept` — the expected concept is present. The scaffold implements\n  this as normalized substring containment, which is the FLOOR, not the\n  ideal: it is literal enough to produce brittleness. When a case fails\n  here and the answer was right, widen the grader — never touch the\n  case (\u002Faudit eval's \"grader brittleness\" bucket).\n- `numeric-tolerance:\u003Cx>` — parse both sides, compare within ±x\n  absolute, or ±x% relative to the expected value when the suffix is\n  `%` (both forms are documented by \u002Feval-spec's template).\n- `rubric:\u003Cname>` — not machine-gradeable. The runner records the\n  answer, marks the case `needs-rubric-review`, and **excludes it from\n  the headline while naming it**. A human or a judge model scores the\n  rubric's named dimensions; a rubric with unnamed dimensions is a spec\n  defect, not a pass.\n\nDenominator discipline, because a headline is only as honest as what it\ncounts: cases marked `\"status\": \"needs-data\"` are **skipped and reported\nas skipped**, never counted as passes; `\"status\": \"superseded\"` cases\nare excluded entirely; `rubric:` cases are excluded from the headline\nbecause no machine graded them, and are reported for review;\n`acceptable_failure` applies only when the case carries a written\n`reason` (in either shape it is written — a bool with a sibling\n`reason`, or an object), never to an ungraded case, and every\napplication is listed.\n\n**Every exclusion is named in the report.** A case that leaves the\ndenominator quietly moves no percentage at all, so it cannot be caught\nby looking at the number — which makes silent exclusion the most\ndangerous of the false-pass family.\n\n## Hard rules\n\n- **Never edit `golden.jsonl` to raise a score.** A genuinely wrong case\n  is superseded through \u002Feval-spec, with a new id and a written reason.\n  Editing a case to pass is the one thing this pack refuses outright.\n- **Never hand-transcribe the headline.** It comes from the results\n  file, every time.\n- A run that errors part-way reports what completed and what did not —\n  a partial run is never presented as a full score.\n- **A subject that cannot be invoked yields NO SCORE, not 0%.** When\n  the system under test needs a credential or endpoint the environment\n  lacks, record honest per-case errors in the results file, report the\n  headline as measuring the environment rather than the subject, and\n  name what is missing in Scope. Never set, invent, or mock the\n  credential to make the run go — a mocked subject scores the mock.\n- The results file is committed with the run: an eval score whose\n  evidence is not in the repo is an assertion.\n\n## Report shape\n\nVerdict first — `\u003Cheadline>% vs target \u003Cn>%` plus `MEETS TARGET` or\n`BELOW TARGET by \u003Cx>`; when nothing was graded, the verdict is instead\n`NO SCORE — \u003Cwhat is missing>` (per the hard rule above — a zero-graded\nrun has no headline to compare) — then:\n\n- per-category table (category | cases | passed | % | target)\n- failures, grouped by the \u002Faudit eval buckets (prompt issue, grader\n  brittleness, provider flake, data issue, parser issue, genuinely\n  ambiguous) — classification proposed, since confirming it is \u002Faudit's\n  job\n- `acceptable_failure` applications, each with its reason\n- skipped cases and why\n- the results file path, and the exact command that produced it\n- **Scope:** what was NOT run (needs-data cases, categories with no\n  cases, an unreachable system under test)\n\n## Positive control\n\nA golden set containing one deliberately failing case MUST produce a\nsub-100% headline. A runner that reports 100% by construction — because\nit counts skips as passes, swallows errors, or compares nothing — is\nthe false-pass class this pack exists to catch, and it is indetectable\nfrom a green report. `fixtures\u002Feval-run\u002F` seeds exactly that case.\n",{"name":5,"description":7,"argument-hint":23},"[eval-dir | notes]","main","6b5be39937ce3823cf8f58c68aadcfdfda839fa0","skills\u002Feval-run",0,"https:\u002F\u002Fgithub.com\u002FAaravChadha\u002Facstack","https:\u002F\u002Fgithub.com\u002FAaravChadha","2026-09-28 13:07:41",{"source":8,"skillName":5,"commands":32,"prompt":42,"zipAvailable":43},[33,36,39],{"tool":34,"command":35},"npx","npx skills add AaravChadha\u002Facstack --skill eval-run",{"tool":37,"command":38},"bunx","bunx skills add AaravChadha\u002Facstack --skill eval-run",{"tool":40,"command":41},"pnpm","pnpm dlx skills add AaravChadha\u002Facstack --skill eval-run","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FAaravChadha\u002Facstack\u002Fblob\u002Fmain\u002Fskills\u002Feval-run\u002FSKILL.md",true,[45,51],{"path":46,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":49,"sourceHtmlUrl":50},"references\u002Fregression-gate.py","SCRIPT","脚本",8186,"https:\u002F\u002Fgithub.com\u002FAaravChadha\u002Facstack\u002Fblob\u002Fmain\u002Fskills\u002Feval-run\u002Freferences\u002Fregression-gate.py",{"path":52,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":55,"sourceHtmlUrl":56},"references\u002Frunner-template.md","REFERENCE","参考文档",22084,"https:\u002F\u002Fgithub.com\u002FAaravChadha\u002Facstack\u002Fblob\u002Fmain\u002Fskills\u002Feval-run\u002Freferences\u002Frunner-template.md",[58,65,72,79,86,93],{"skillNo":59,"name":60,"nameZh":6,"description":61,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":62,"categories":63,"keywords":64,"publishedAt":18},"01M79R3MBE63G66XVKHV652YAD","why","Decision archaeology - answers why the code, config, or plan is the way it is by searching the record in a fixed order: BRIEF constraints, dated PLAN decisions, JOURNAL entries, then git history. Stops at the first real answer, cites it with a date, and says no recorded rationale rather than inventing one. Use when someone asks why something is this way, why a choice was made, or where a decision is written down.","https:\u002F\u002Fgithub.com\u002FAaravChadha\u002Facstack\u002Fblob\u002Fmain\u002Fskills\u002Fwhy\u002FSKILL.md",[],[],{"skillNo":66,"name":67,"nameZh":6,"description":68,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":69,"categories":70,"keywords":71,"publishedAt":18},"01M79R3G7EA313SYFS30H1K51A","verify","Audit a completion claim made by someone else - another session, agent, or teammate - against the running system rather than the diff. Re-derives what the claim's acceptance demands, runs it, and returns CONFIRMED, OVERSTATED, FALSE or UNVERIFIABLE with the command and its output. Never fixes, never edits the claim or the acceptance. Use when handed a \"done\", \"all green\", \"merged\" or \"it works\" you did not produce yourself.","https:\u002F\u002Fgithub.com\u002FAaravChadha\u002Facstack\u002Fblob\u002Fmain\u002Fskills\u002Fverify\u002FSKILL.md",[],[],{"skillNo":73,"name":74,"nameZh":6,"description":75,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":76,"categories":77,"keywords":78,"publishedAt":18},"01M79R3CPD2F05EKE9N8XCMM90","triage","Backlog hygiene sweep - stale items, duplicate pairs, missing acceptance criteria, unblocked-but-unassigned work, and milestone burn, delivered as a report first with only user-approved actions applied. Works on GitHub Issues in tickets mode and on PLAN.md checkboxes in document mode. Use when the user asks to triage, groom, or clean up the backlog or the plan's open tasks.","https:\u002F\u002Fgithub.com\u002FAaravChadha\u002Facstack\u002Fblob\u002Fmain\u002Fskills\u002Ftriage\u002FSKILL.md",[],[],{"skillNo":80,"name":81,"nameZh":6,"description":82,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":83,"categories":84,"keywords":85,"publishedAt":18},"01M79R395S4TS1XENQWWCSBT56","ticket","Capture a brain-dump as a well-formed work item - verb-first title, acceptance criteria, file paths when known, an out-of-scope line. Files a GitHub issue in tickets mode or appends a numbered PLAN.md task in document mode; unknowns are marked TBD, never invented. Use when the user brain-dumps an idea, bug, or task to capture for later, or asks to ticket, file, or note something as work.","https:\u002F\u002Fgithub.com\u002FAaravChadha\u002Facstack\u002Fblob\u002Fmain\u002Fskills\u002Fticket\u002FSKILL.md",[],[],{"skillNo":87,"name":88,"nameZh":6,"description":89,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":90,"categories":91,"keywords":92,"publishedAt":18},"01M79R358JKQ1EB3TQKT2MPSB6","ship","Branch-level release with five gates before the act - clean-state, tests, eval-vs-target, docs drift, and attribution sweep; any failing gate stops the release with its output. Then push, and under push: branch-pr open a report-shaped PR wiring the issue-closing reference in tickets mode or PLAN task IDs in document mode. Use when the user asks to ship, release, cut, or open the PR for a feature or branch.","https:\u002F\u002Fgithub.com\u002FAaravChadha\u002Facstack\u002Fblob\u002Fmain\u002Fskills\u002Fship\u002FSKILL.md",[],[],{"skillNo":94,"name":95,"nameZh":6,"description":96,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":97,"categories":98,"keywords":99,"publishedAt":18},"01M79R2ZXGVB7W1J8ZH8TRGWDN","secure","Security review with a confidence gate - a finding exists only with a concrete exploit scenario and a high\u002Fmedium\u002Flow confidence rating; everything else goes to a worth-hardening list, never inflated. Sweeps five surfaces: auth gates, secrets hygiene, injection and unsafe sinks, unsafe deserialization\u002Fcrypto\u002Ftransport, and LLM tool-use trust boundaries. Reports only, never fixes. Use when the user asks for a security review or to check vulnerabilities, secrets, or auth.","https:\u002F\u002Fgithub.com\u002FAaravChadha\u002Facstack\u002Fblob\u002Fmain\u002Fskills\u002Fsecure\u002FSKILL.md",[],[]]