[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7F917VTJWS90304QDH1C3R8":3,"skill-same:01M7F917VTJWS90304QDH1C3R8":46},{"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":25,"commitSha":26,"skillDirectory":27,"forksCount":28,"repositoryHtmlUrl":29,"repositoryHomepageUrl":30,"ownerHtmlUrl":31,"ownerBlogUrl":30,"ownerDescription":6,"updatedAt":32,"install":33,"resources":45},"01M7F917VTJWS90304QDH1C3R8","review",null,"Verify implemented work against its spec with real evidence before any \"done\" claim. Checks the spec's success criteria, runs the surface test command, and runs `steepy-formula graph validate` through the deterministic inline evidence capture.","steepy-it\u002Fsteepy-formula","steepy-it","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F299579966?v=4","01M7F8VT4D9JWJZDQFATW4K3YB","Steepy","01M7F8VT4ZR70KR69DV1RR1Q8W",5,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fsteepy-it\u002Fsteepy-formula\u002Fblob\u002Fmain\u002Fskills\u002Freview\u002FSKILL.md",[],[],"2026-11-16 10:05:34",false,"---\nname: review\ndescription: Verify implemented work against its spec with real evidence before any \"done\" claim. Checks the spec's success criteria, runs the surface test command, and runs `steepy-formula graph validate` through the deterministic inline evidence capture.\nuser-invocable: true\n---\n\u003C!-- steepy:manual-handoff:v1:start -->\n## Manual handoff envelope\n\nManual phase boundaries use this literal shape (the envelope is chat text, never a file):\n\n```yaml\nhandoff: steepy-formula\u002Fv1\nnext: \u003Cphase>\nrequired: \u003Crole\u002Fpath map or none>\nonDemand: \u003Crole\u002Fpath map or none>\n```\n\nAll four keys are mandatory and occur once, in that order. Use explicit `none` for an empty map.\nThe phase input role maps are exact and unique: brainstorm has `required: none`, `onDemand: none`;\nplan has required `spec`, `onDemand: none`; Gear-2 direct implement has required `spec`,\n`onDemand: none`; plan-backed implement has required `plan`, on-demand `source-spec`; either\nimplement alternative additionally authorizes exact `progress-ledger` and `task-results` on demand\nwhen resuming;\nregular Gear-3 `review` has required `criteria`, `task-results`, and `branch-diff`, `onDemand: none`;\nfresh `loop-engineer` has required `goal`, `onDemand: none`; resume may additionally authorize exact\n`loop-ledger` on demand; Gear-4 `review` has required `goal` and `loop-ledger`, `onDemand: none`.\nThus Gear-2 brainstorm → implement maps `spec`; brainstorm → plan maps `spec`; plan → implement\nmaps `plan`, `source-spec`; Gear-3 implement → review maps `criteria`, `task-results`, `branch-diff`;\nloop-engineer → review maps `goal`, `loop-ledger`. The two implement role alternatives and the two\nreview role alternatives are mutually exclusive: never merge or infer them. A role occurs only in\nits mapped phase, alternative, and list.\n\nValidate literally before any work-artifact read. Reject duplicate keys, unknown keys, unknown roles,\nmissing keys, a wrong `next`, and unsafe paths: absolute paths, `..` segments, globs, or anything\noutside the repository. Never infer corrections. An explicit malformed or stale binding fails closed\nwithout discovery. Supporting `onDemand` inputs are never preloaded; read one only for a concrete\nnamed missing fact. Stable routing inputs stay outside the envelope and follow bootstrap routing.\nNative invocation spelling comes only from the active harness. No child inherits general\n`.formula\u002Fwork\u002F**` access; pass only its accepted role paths.\n\nEvery new work artifact starts with verdict\u002Fcontract metadata when that phase requires it, followed\nimmediately by this versioned header before the Markdown title:\n\n```yaml\n\u003C!-- steepy-workflow: v1\nphase: \u003Cbrainstorm|plan|implement|review|loop-engineer|goal-contract>\nstatus: \u003CDRAFT|READY|CONSUMED>\nnext: \u003Cphase|none>\nsource: \u003Csafe exact repo-relative path|none>\nconsumed-by: \u003Csafe exact repo-relative path|none>\n-->\n```\n\nHeader fields are mandatory, unique, and literal. `source` and `consumed-by` must agree with the\nadjacent lifecycle transition and authorize no transitive reads. Headerless artifacts are\nignored by recovery and are available only under explicit user audit scope.\n\n### Exact publication pairs\n\nThese are exact resume capabilities, not paths inferred from provenance. The input retains its\nnormal required role; the output role is added to `onDemand`, as are any existing on-demand companions.\nRegular review retains its required criteria and branch diff. Plan resume adds `onDemand.output-plan`;\nregular Gear-3 review resume adds `onDemand.review-report`. Fresh forms above are unchanged.\nGear-4 controller\u002Fledger handling is excluded from this model-based pair repair.\n\n| Route | Input role | Output role | Companion roles | Input phase | Output phase | Output next |\n| --- | --- | --- | --- | --- | --- | --- |\n| plan | spec | output-plan | none | brainstorm | plan | implement |\n| implement-2-direct | spec | task-results | progress-ledger | brainstorm | implement | none |\n| implement-2-plan | plan | task-results | source-spec,progress-ledger | plan | implement | none |\n| implement-3 | plan | task-results | source-spec,progress-ledger | plan | implement | review |\n| review-3 | task-results | review-report | criteria,branch-diff | implement | review | none |\n\n### Publication interruption prefixes\n\nThis table applies only to the exact authorized pair with matching phase, route and provenance.\nValidate the pair before the fresh-only READY input gate; a resume capability is never inferred.\nInput `next` names the consumer phase until consumed; consumption retains it for plan\u002Fimplement\nand sets it to `none` for review. Output `consumed-by` is `none` throughout this producer's run.\nAn unconsumed input must have `consumed-by: none`; a consumed input must name the exact output.\n\n| Input status | Output status | Action |\n| --- | --- | --- |\n| absent | absent | reject |\n| absent | DRAFT | reject |\n| absent | READY | reject |\n| absent | CONSUMED | reject |\n| DRAFT | absent | reject |\n| DRAFT | DRAFT | reject |\n| DRAFT | READY | reject |\n| DRAFT | CONSUMED | reject |\n| READY | absent | fresh-create |\n| READY | DRAFT | resume-work |\n| READY | READY | finish-consumption |\n| READY | CONSUMED | reject |\n| CONSUMED | absent | reject |\n| CONSUMED | DRAFT | reject |\n| CONSUMED | READY | no-op |\n| CONSUMED | CONSUMED | reject |\n\n`fresh-create` permits a new absent output only in a fresh run, never a missing resume target.\n`resume-work` requires exact output capability and leaves the input READY while work\u002Fgates remain.\nFor `finish-consumption`, revalidate output.source against the exact input path, input.consumed-by\nas `none`, and the recorded approval and verification bound to those input\u002Foutput paths and content\n(excluding lifecycle fields). Recheck current branch\u002FHEAD or uncommitted diff identity where applicable.\nOnly then perform the input-consumption write with input.consumed-by equal to the exact output path.\nFor `no-op`, revalidate the recorded proof with the final reciprocal links instead of the\nunconsumed-input condition; perform no writes or redispatch.\nA missing capability, mismatched path\u002Fphase\u002Fnext\u002Fcontent, absent approval\u002Fproof, or impossible state\nmust fail closed, preserving both files. A READY marker alone is not proof. Do not follow links to\nundeclared evidence; request the exact missing capability or user-delimited scope instead.\nNever select a most-recent output or recover by directory\u002Fbody search.\n\nPublication is ordered: record the phase's applicable approval and verification in the output,\npublish output READY by per-file atomic replacement, then consume the input by per-file atomic\nreplacement. This is not a multi-file atomic operation; READY\u002FREADY is its sole verified intermediate\nprefix. Resume repairs no other state and never repeats completed work or release actions.\nIn manual drive you are the enforcement: no parser, manifest, or validator checks this contract.\n\u003C!-- steepy:manual-handoff:v1:end -->\n\n\n# review\n\nVerify completed work with evidence. Never assert \"done\" without command output to back it.\n\n> **Engine root:** this skill's base directory is `\u003Cengine-root>\u002Fskills\u002Freview\u002F`; engine\n> scripts live two levels up, at `\u003Cengine-root>\u002Fscripts\u002F`. Resolve them relative to the base\n> directory your harness reports for this skill.\n\n## Checklist\n\n- [ ] Declare the hub mode once per run — `steepy-formula hub-mode` (always exit 0) — and stop in `migrate` and `none`, and carry the declared mode into the report (Step 0)\n- [ ] In autopilot, read the supplied phase manifest first; then read the gear and scale ceremony to it (Step 0)\n- [ ] Manifest-backed autopilot: load criteria-only source + task index + diff + standards, regular manual review: load accepted criteria + task-result index + branch diff + graph orientation; Gear 4: load the exact goal contract + loop ledger instead (Step 1)\n- [ ] Run `capture-review-evidence.mjs` once: keep verbatim command output in the current run's\n      `evidence-report.md`, consume only its compact JSON receipt, and judge every criterion in\n      session (Step 2)\n- [ ] For Gear 4, validate goal\u002Fledger provenance and the ledger event commitment, then prove the\n      exact claimed terminal outcome without confusing negative completion with goal success (Steps 1–2)\n- [ ] Sweep unpromoted insights from the active inventory: manifest criteria + task-result index in autopilot; manual insight sweep uses only accepted criteria + task-result index compact status and `signals`; the manual sweep also assembles the Promotion Scenario from the typed candidates file (Step 2)\n- [ ] Confirm the receipt contains a green `coherence-gate` command result (Step 3)\n- [ ] Report: the declared hub-mode, whether the `coherence-gate` is green, the typed candidate list and the quoted typed diff, AND every regular criterion or Gear-4 terminal claim verified\n      from real command output (Step 4)\n- [ ] Manual drive: ask one explicit yes\u002Fchange\u002Fno question on the assembled Promotion Scenario and, on yes, apply it here — `promotion apply \u003Cscenario-id> --yes`, passed only after the human's explicit yes; a no applies nothing (Step 5)\n- [ ] Resolve the version bump (or apply `no-release`) and offer a PR to `main` only for regular\n      review or Gear-4 `GOAL_REACHED`, when the work landed on a separate branch with commits ahead;\n      suppress release\u002FPR for negative terminal outcomes (Step 5)\n\n## Procedure\n\n### Step 0 — Read the gear\n\n**Hub-mode declaration (once per run).** Run `steepy-formula hub-mode` before anything else and\ncarry its declared mode — `graph`, `migrate`, or `none`, each printed with one `evidence:` line —\nthrough the whole run. The command always exits 0: it is a declaration for the report, not a\ngate, and this prose never re-implements its detection — the command is the single authority.\nThe declared mode branches the run:\n\n- **`graph` mode:** proceed — the Step 1 orientation reads, the Promotion Scenario assembly at\n  Step 2, the graph coherence gate wired into the Step 2 capture and checked at Step 3, and the\n  human apply at Step 5.\n- **`migrate` mode:** the repo still carries a pre-graph Markdown hub.\n  Stop and tell the user to run the `graph-migrate` skill (invoke it the way your harness invokes skills) before any other workflow.\n- **Mode `none`:** no graph — the repo is not initialized: tell the user to run the `init` skill\n  first and stop.\n\n**Autopilot preflight:** When the conductor-supplied phase manifest is present, read that\nmanifest before any other task input. Validate its role and scope against the review phase and the\nsupplied correlation identity. Eagerly read every `required` input before acting. Do not preload `onDemand`;\nconsult it only for a concrete named missing fact and record the read in the review\nreport. Phase-manifest authority applies only when `drive: autopilot`; manual-drive discovery remains\nStep 1's behavior.\n\nAutopilot status timestamps must use UTC with three millisecond digits (`YYYY-MM-DDTHH:mm:ss.sssZ`), generated with `new Date().toISOString()`.\n\n**Manifest-backed autopilot:** Take the Step-0 scalar facts from `manifest.contract.verdict`,\n`manifest.contract.gear`, and `manifest.contract.drive`. Require `drive: autopilot` and gear 3, and\nhonor the recorded verdict. Do not read or locate the spec: it is deliberately absent from the review\ninventory, and the criteria-only required artifact supplies the review criteria. Apply the Step 4\u002F5\nAutopilot paragraph below and never ask the user anything. A manifest\u002Foutput failure, success criteria\nunmet, the `graph validate` gate red, or a CONFLICT discovered mid-run → append\n`\u003CISO timestamp> — review — BLOCKED — run-id=\u003Cconductor-supplied> attempt=\u003Cpositive supplied> \u003Creason>`\nto `.formula\u002Fwork\u002Ftasks\u002F\u003Cspec-basename>\u002Fautopilot-status.md` and exit non-zero. A CONFLICT discovered\nmid-run is ALWAYS a halt — doc-wins needs a human, never auto-override.\n\nUse the pinned `manifest.contract.taskResultProtocol` to interpret the task index. In autopilot\nthe task index is the generated `steepy-task-results: v2` JSON block with exact `changedPaths`\narrays; manual drive keeps `- Task` bullets. Never split v2 filenames on commas or expand braces.\nThe conductor replays execution receipts and required task\u002Ffinal approvals before this phase;\nthe index is a projection, and a captured execution alone is never approval. This introduces no\nreviewer permission to open receipt bodies, task reports, the plan, or the progress ledger. Missing\nor drifted evidence fails closed; never infer historical baselines.\nManual drive and Gear 4 retain their existing grammar and capabilities.\n\n**Manual\u002Fno-manifest:** Validate the manual handoff before reading any work artifact. Regular Gear-3\nreview fresh entry accepts exactly `criteria`, `task-results`, and `branch-diff` with `onDemand: none`;\nregular resume adds only `onDemand.review-report`, the exact canonical report paired with that index.\nCheck existence of the canonical report derived from the accepted primary path without opening it.\nAn existing review report without its exact capability requires a stop before body reads or evidence\ncollection; do not overwrite it or select another report. Same-invocation output creation is not\nhistorical resume authority.\nValidate this pair before the fresh-only READY input gate: READY\u002FDRAFT resumes verification,\nREADY\u002FREADY is finish-consumption, and CONSUMED\u002FREADY is no-op after the common proof checks.\nRequire `phase: review`, `next: none`, the exact index as report `source`, and `consumed-by: none`.\nRead the authorized report only for those named progress\u002Fapproval\u002Fproof facts. Verified terminal\npairs retain all supporting-input checks: on every resume, validate the accepted criteria and\nbranch-diff paths against the index metadata and their current content against the report's proof,\nbefore any terminal shortcut. Only then go straight to the human envelope in Step 4 after any sole consumption repair; never rerun\nrelease actions. Gear-4 review accepts exactly `goal` and `loop-ledger` with `onDemand: none` and\ndoes not acquire this regular-review resume route. The task-result index (or the\nGear-4 ledger) is primary: fresh entry requires `READY` \u002F `next: review`, then validate every supporting path and\nits explicit provenance against it. An explicit malformed, stale, unsafe, or mismatched envelope is\nan explicit failure and never falls back to discovery. With no envelope, pathless recovery may\ninspect headers only on candidate task-result indexes (or loop ledgers), never bodies; require user\nselection and then a new exact envelope. Never select by recency.\n- **Artifact present:** honor its gear.\n- **Artifact absent and this is gear-1\u002F2 work:** recompute the coverage verdict inline (COVERED \u002F\n  CONFLICT \u002F GAP with `file:line` evidence → gear). Declare the recomputed\n  verdict, gear, and evidence, then ask the user for one confirmation before the ceremony starts —\n  the user may force a different gear, and the override is recorded in the verdict artifact. Then\n  proceed.\n- **Artifact absent and full gear-3 ceremony is required for this command:** do not proceed\n  blind — tell the user to enter through the upstream workflow chain first, and stop.\n\nScale this skill's ceremony to the gear (see \"Ceremony by gear\" below). This skill dispatches **no\nsubagent reviewer or evidence collector**: `capture-review-evidence.mjs` collects command output\ninline and deterministically, while this session self-verifies every success criterion plus the\n`graph validate` coherence gate, per the \"Review architecture (gear 3)\" governance rule — recorded as a\nRule in the canonical graph. Judgement is never delegated.\n\n**Autopilot correlation.** Before appending any current-protocol status marker, use the `run-id` and\npositive `attempt` supplied by the conductor's `phasePrompt`; that correlated identity is authoritative.\nDo not derive either value from filenames. Do not invent a correlation identity. A child without supplied run\u002Fattempt identity\nis unresolvable: exit non-zero without appending an uncorrelated new-protocol marker. Only versioned, correlated status is supported; uncorrelated markers never complete a phase.\n\n**Session declaration (declare, never block).** Say whether this conversation started clean for this\nphase or has already run an earlier phase of the chain in the same session. A same-session phase\nalready grew the context, and every turn of this phase now pays to carry it — name that cost, but\nnever refuse or gate on it: this is a declaration about the conversation, not a measurement, and\nsound work proceeds either way. If the session's history cannot be determined with certainty, say so\nand proceed. Under `drive: autopilot` this declaration is a no-op: the conductor\n(`scripts\u002Fautopilot.mjs`) already opens a fresh headless session per phase, so the discipline stated\nhere is already enforced by that mechanism, not by this prose.\n\n### Ceremony by gear\n\n- **All gears**: run the surface test command and the `graph validate` coherence gate through\n  the deterministic inline capture in Step 2, and verify each success criterion yourself. There is\n  no subagent evidence reviewer or collector. The complete command bytes are durable on disk; only the bounded receipt\n  enters model context. Verification and judgement are never delegated, at any gear.\n- **Gear 4**: review accepts gear-4 work. Verification is the re-run of the `verifier` from the\n  goal contract, plus the loop ledger as evidence — alongside the existing self-verify above, not\n  replacing it.\n\n### Step 1 — Load bounded review inputs and the hub\n\n**Manual drive, regular Gear 3:** Eagerly read exactly the accepted `criteria`, `task-results`, and\n`branch-diff` artifacts. Never locate or read the spec, plan, progress ledger, task briefs, task\nreports, task reviews, sibling work files, transcripts, or conversations. Treat `task-results` as\nthe primary and validate the criteria and diff paths plus explicit source metadata against it. Then\nread the orientation for the mode declared at Step 0 — `graph`, `migrate`, or `none`; this prose never re-detects it:\n\n- **`graph` mode:** resolve orientation from the canonical graph — run\n  `steepy-formula context resolve --path \u003Cp>...` over the reviewed paths (exit 0 emits\n  `context.json` + `context.md`). The resolved context supplies the owning **surface**, its\n  Rules — the standards-equivalent — its Terms, and the surface test command from its Check\n  item's `ext_v1_command` extension property (fall back to the verification Rule's statement\n  prose when the Check carries no command). The bound specialist agent is\n  `.claude\u002Fagents\u002F\u003Csurface>-agent.md` by naming convention. On exit 10 UNRESOLVED — normal in a freshly graph-first-initialized repo, where no\n  topology bridge links the detected Packages to the curated Components yet — fall back to\n  the deterministic surface targets:\n  `steepy-formula context resolve --target apex:\u003CprojectId>:surface:\u003Cslug(name)>` for each\n  owning surface, every part derived, never guessed — surface names from\n  `.claude\u002Fagents\u002F\u003Cname>-agent.md` (CLAUDE.md's Surfaces section is the fallback listing),\n  the project id from the hub-mode declaration's `hub-mode '\u003CprojectId>'` echo, `slug(name)`\n  lowercase with non-alphanumeric runs collapsed to `-` — and never target the bare project\n  root on its own: it resolves empty. If the resolve fails, diagnose with\n  `steepy-formula graph status` — exit 3,\n  `initialized: no`, or an engine failure means the graph binding is missing or failed: tell the\n  user to run the `init` skill first and stop.\n- **`migrate` mode:** the repo still carries a pre-graph Markdown hub.\n  Stop and tell the user to run the `graph-migrate` skill (invoke it the way your harness invokes skills) before any other workflow.\n- **Mode `none`:** no graph — the repo is not initialized: tell the user to run the `init` skill\n  first and stop.\n\n**Autopilot:** use the validated phase manifest loaded in Step 0 as the authoritative input\ninventory. Its `required` inventory must provide the success-criteria source, task\u002Fresult index,\naggregate diff, relevant standards, and test commands. Exclude per-task transcripts and\nconversations: they are not review inputs. Do not independently preload them. Use `onDemand` only\nunder Step 0's concrete-missing-fact protocol, recording any such read in the report.\n\n**Gear 4 has no spec and no plan.** Eagerly read exactly the accepted goal and loop-ledger paths;\nthe loop ledger is primary. The public capability remains exactly `goal` + `loop-ledger`: do not\nderive or read `events.jsonl`, enumerate the loop directory, or expose any private event path to\nmodel context. Resolve `\u003Cengine-root>` from the skill base and invoke this read-only proof boundary:\n\n```bash\nnode \u003Cengine-root>\u002Fscripts\u002Floop-engineer.mjs --repo-root . --validate-terminal --goal '\u003Cgoal-path>' --ledger '\u003Cloop-ledger-path>'\n```\n\nThe terminal validator accepts only that exact public pair, derives the canonical `events.jsonl`\nidentity internally, replays events as the sole authority, recomputes `events-sha256`, verifies the\ngoal digest plus `phase: goal-contract`, `status: CONSUMED`, and `consumed-by:\n\u003Cloop-ledger-path>` for the goal and the `READY` ledger lifecycle. It binds every receipt fact and\nevent-derived ledger field to replay, recomputes the ledger digest, and accepts runtime-only fields\nonly in their closed canonical form. It also authenticates the event-bound reviewer report digest,\nreviewer diff digest and bytes, final branch-diff projection, and current branch, HEAD, clean tree,\nand snapshot against the retained event-derived branch. It performs zero lifecycle or Git mutation and returns one\nclosed bounded receipt; a nonzero exit or malformed\u002Fmismatched receipt means provenance is\nunverified.\n\nRequire exactly `status: VALIDATED`, the accepted `goal` and `loop-ledger`, and the closed receipt\nfields `run-id`, `outcome`, `goal-succeeded`, `events-sha256`, `mode`, `branch-review`,\n`final-projection`, `attempt-budget`, and `candidate`. `branch-review` supplies exact `verdict`, `issues-found`, review\nand attempt identities, report\u002Fdiff paths, `report-sha256`, and `diff-sha256`; `final-projection`\nsupplies the authenticated diff digest plus current branch, HEAD, clean, and snapshot facts. `attempt-budget` supplies exact\n`reserved`, `budget`, and `exhausted`; `candidate` supplies boolean-green or metric-baseline,\nmetric-best, and strict-improvement facts. Require `outcome` to be exactly `GOAL_REACHED |\nBUDGET_EXHAUSTED | NO_IMPROVEMENT | REVIEW_REJECTED`, and `goal-succeeded: true` only for\n`GOAL_REACHED`. Take these provenance facts from the receipt, never ledger assertions. The goal's\n`verifier` is the active Gear-4 criterion. Read the graph orientation above; never select a loop by recency.\n\n### Step 2 — Verify against the active criteria source\n\nFor each success criterion in the active source, confirm it is met. In manifest-backed autopilot and\nmanual regular review the active source is the required criteria-only artifact.\nRun the surface test command and preserve its **real, verbatim output** on disk; do not copy that\nunbounded transcript into model context.\n\n**Inline deterministic collection.** Resolve the current output path from the accepted primary\nartifact, never by directory discovery:\n\n- regular review: `.formula\u002Fwork\u002Ftasks\u002F\u003Cplan-basename>\u002Fevidence-report.md`;\n- Gear 4: `.formula\u002Fwork\u002Floops\u002F\u003Cslug>\u002Fevidence-report.md`.\n\nThen run this once, passing each command as one safely quoted argument:\n\n```bash\nnode \u003Cengine-root>\u002Fscripts\u002Fcapture-review-evidence.mjs --repo-root . --evidence '\u003Cevidence-path>' --test-command '\u003Cexact surface test command>'\n```\n\nThe coherence gate is the capture's final command: the script's built-in gate runs\n`steepy-formula graph validate` through the packaged CLI (`node \u003Cengine-root>\u002Fscripts\u002Fformula\u002Fcli.mjs`,\nnever a PATH lookup) from the repo root, so the project id resolves from the committed\n`.formula\u002Fconfig.json` binding (or `STEEPY_PROJECT`), labeled `coherence-gate`. Pass no gate flag.\n\nFor Gear 4 append `--verifier-command '\u003Cexact verifier command>'`. The script always runs the\nsurface test, then the optional goal verifier, then the coherence gate; it atomically replaces stale\nevidence at the fixed path and streams combined stdout\u002Fstderr into the new report. It returns one\nbounded JSON receipt on stdout with the artifact path, collection timestamps, aggregate pass state,\nand, for every command, exit\u002Fsignal\u002Ferror state, output byte and line counts, SHA-256, and a bounded\ntail excerpt. A tested command may be red while collection itself exits successfully: use\n`allPassed` and each command's `passed`\u002F`exitCode` fields to determine the review result.\n\nValidate receipt schema version 1, the exact expected artifact path, timestamps from this review\nround, and the ordered command labels (`surface-test`, optional `goal-verifier`, then `coherence-gate`). Do\nnot open or preload the complete `evidence-report.md`: the report is durable verbatim proof, while\nthe receipt is the context boundary. If a named criterion genuinely requires a fact absent from the\nbounded excerpt, read only the smallest exact line or byte range needed from that report, record the\nreason and range in the review report, and never fall back to a whole-file read. A malformed receipt,\nhard capture failure, missing report, hash\u002Fmetadata mismatch, or missing evidence means the affected\ncriterion is not verified; after fixing the cause, rerun the single capture so stale evidence is\nreplaced rather than merged.\n\nThe judgement stays here, in session: whether each success criterion is met, which swept insights\nget promoted, and the final verdict are never delegated. No success criterion is declared met for\nwant of evidence.\n\n**Gear 4:** the criterion is the goal contract and the authenticated terminal outcome, not a spec.\nThe same bounded capture runs the surface test, goal verifier, and the `coherence-gate`\n(`graph validate`) in that order. Require the\ncurrent capture receipt's surface test and that gate to be green. Interpret the fresh\nverifier result against the terminal-validator receipt; a red verifier is expected evidence for\n`BUDGET_EXHAUSTED`, not by itself invalid execution. For metric mode, parse the verifier's final\nstdout line as a finite scalar and compare it with the receipt's metric baseline using the goal's\n`metric-direction`.\n\nApply this closed outcome proof:\n\n- `GOAL_REACHED` requires boolean mode with a green verifier, or metric mode with a strict\n  improvement over the reproduced metric baseline; it also requires an approved branch review via\n  receipt `branch-review.verdict: APPROVED`, `issues-found: false`, `candidate.acceptable: true`, and\n  `goal-succeeded: true`.\n- `BUDGET_EXHAUSTED` requires boolean mode, a red boolean verifier, the full budget spent\n  (`attempt-budget.reserved` equals `budget` and `exhausted: true`), and `goal-succeeded: false`.\n- `NO_IMPROVEMENT` requires metric mode, a verifier value equal to the reproduced metric baseline,\n  no strict improvement in the declared direction, `candidate.metric-strict-improvement: false`,\n  the full budget spent (`attempt-budget.exhausted: true`), and\n  `goal-succeeded: false`.\n- `REVIEW_REJECTED` requires a goal-acceptable candidate: boolean mode must be green, or metric mode\n  must show strict improvement over baseline. It never accepts a red boolean result or a metric\n  equal to baseline. Also require receipt `candidate.acceptable: true`, unresolved review issues\n  (`branch-review.verdict: ISSUES_FOUND` and `issues-found: true`), no remaining mutation budget\n  (`attempt-budget.reserved` equals `budget` and `exhausted: true`), and `goal-succeeded: false`.\n\nThe authenticated terminal receipt is the controller-owned proof of event identity, exact\nbranch-review verdict\u002Fissues, and attempt-budget accounting. This review judges whether its terminal\noutcome is faithfully reproduced; it never reconstructs attempt state from ledger prose. A mismatch\nleaves the ledger READY and produces an issues verdict. For the insight sweep below, the only Gear-4\nsources are the accepted goal's `notes:` and accepted loop-ledger fields.\nBoolean candidate facts describe the retained branch: committed attempts update that state and\ndiscarded review-fix attempts leave the prior retained verifier result unchanged.\n\n**Insight sweep — run it always.** Work discovers knowledge as well as changing it. Use only sources\navailable in the active inventory:\n\n- **Manifest-backed autopilot:** sweep the criteria-only artifact and task-result index. Their criterion\n  IDs, `signals`, and `DONE_WITH_CONCERNS` entries are the declared insight evidence; do not locate the\n  full spec, plan, progress ledger, or per-task transcripts.\n- **Manual\u002Fno-manifest:** sweep only the accepted criteria artifact plus compact status and `signals`\n  in the accepted task-result index, specifically its `DONE_WITH_CONCERNS` entries and `signals`.\n  Never locate richer upstream or sibling artifacts.\n\nA recorded **discovery** occurrence — an implementer that had to discover what to change rather than\nexecute the brief, with DONE_WITH_CONCERNS and `discovery:unplanned` — is a rebuttal to `plan`'s\ntask-cutting criterion (a task that looked pre-decided but wasn't). Report it distinct from other\nconcerns. In both regular-review branches, derive discovery evidence only from the accepted task-result index\nand use the accepted criteria artifact for any mapped criterion; do not invent a criterion absent\nfrom that artifact or read the plan to reconstruct one.\n\n### Discovery signal interpretation\n\n| Status | Signal | Interpretation |\n| --- | --- | --- |\n| DONE_WITH_CONCERNS | discovery:unplanned | unplanned-discovery |\n| DONE_WITH_CONCERNS | other-or-none | ordinary-or-unattributed-concern |\n| DONE | discovery:unplanned | inconsistent-result |\n| DONE | other-or-none | no-discovery-evidence |\n\nNever infer discovery from DONE_WITH_CONCERNS alone. Ordinary or unattributed concerns stay distinct;\nan inconsistent result is reported as a contract defect, not a confirmed occurrence. Signal matching\nis by the exact comma-delimited ID, never substring or a similar-looking spelling. No richer source\nor per-task file read is authorized by a concern. Absence of the signal is no discovery evidence,\nnot proof that no discovery happened.\n\nFor each candidate: already promoted → cite the node in the applied scenario; still valuable →\npropose the promotion to the user as a typed candidate (term, rule, decision, readme); no longer holds → drop it\nwith a one-line reason.\n\n**The sweep assembles the Promotion Scenario.** In manual drive, beyond listing the\ncandidates, the sweep writes the still-valuable ones to the typed candidates file at\n`.formula\u002Fwork\u002Ftasks\u002F\u003Cplan-basename>\u002Fpromotion-candidates.json` (gitignored work area; a gear-4\nreview writes `.formula\u002Fwork\u002Floops\u002F\u003Cslug>\u002Fpromotion-candidates.json` — the loop directory stands in\nfor the plan directory). The file carries the typed schema only — `term` (`term`, `definition`),\n`rule` (`title`, `statement`, `concern`, `strength` of `must|should`, `appliesTo` of an existing\nnode id), `decision` (`title`, `statement`, `context`), `readme` (`note`) — built from the sweep's\nown sources: the spec's deferred proposals, the progress ledger, the per-task files. Then create\nand preview the scenario:\n\n```bash\nsteepy-formula promotion create --project \u003Cid> --candidates \u003Cfile.json> [--proposer \u003Cname>] \u003Cscenario-id>\nsteepy-formula promotion preview \u003Cscenario-id>\n```\n\nThe scenario id convention is `\u003CYYYY-MM-DD>-\u003Ctopic>`. The wrapper is that file's only interpreter\n— terms, rules, and decisions become graph operations; `readme` notes are reported out-of-graph\nand stay on the direct-edit channel — this prose maps nothing itself. A malformed file or a\nviolation exits 2 (fix the file, not the command); an already-used scenario id exits 4\n(`promotion discard` it, or pick a fresh id). Quote the preview's typed diff verbatim into the\nStep 4 report — and apply nothing here: Step 5 owns the apply.\n\n**Autopilot:** a promotion candidate is never a blocker on its own — Step 0's \"never ask; unresolvable → BLOCKED\" governs things the run cannot finish without a human, not documentation opportunities. Carry every candidate into the Step 4 report instead, for the human to act on at the Step 5 version-bump\u002FPR decision. Autopilot review creates NO scenario: a draft assembled now would sit `PROMOTION_STALE` by the time a human reaches Step 5, when the head has moved past this phase. List the typed candidates — same schema, as report entries — and nothing more; the interactive Step 5 session creates, previews, and applies when the human acts.\n\nIf the work changed a durable convention, standard, domain term, architecture decision, user-facing workflow, or README behavior, verify that the durable knowledge was promoted into stable versioned docs.\nSpecs and plans themselves are not sufficient evidence of durable documentation.\nDurable knowledge rides the Promotion Scenario: the typed diff is the proposal\nand the Step 5 human-approved `promotion apply` is the promotion — verify the durable knowledge\nrides the scenario, or its out-of-graph `readme` notes.\n\n### Step 3 — Run the coherence gate (mandatory)\n\nThe Step 2 capture runs the coherence gate as its final command — this is required, not optional.\nTerminal review is the sole Gear-4 owner of this gate; the loop controller does not claim it:\nrequire the current receipt's final command to have `label: coherence-gate`,\n`passed: true`, and `exitCode: 0` — the `graph validate` invocation wired in at Step 2 — with\nits bounded excerpt showing `violations: 0` and a clean revision-chain replay. Plus one sanity\n`steepy-formula context resolve` of a representative path or target from the work just\nreviewed, run in session, confirming the resolver still answers.\n\nThe report path, current-run\ntimestamps, output SHA-256, byte count, and line count bind that receipt to the verbatim transcript\non disk. Do not copy or read the complete transcript merely to satisfy this gate.\n\nThe work is **not done** until the gate is green: any invariant violation or broken replay is incomplete work, and the sanity `context resolve` must answer. Fix the graph before claiming completion.\n\n### Step 4 — Report\n\nReport, with evidence: tests pass\u002Ffail (with output), every active criterion or Gear-4 terminal\ncondition met or not, the gate outcome, and the declared hub-mode. The report also carries\nthe typed candidate list and quotes the Step 2 `promotion preview` typed diff verbatim.\nRegular review is reviewable as complete only when the declared gate is green (`graph validate`) AND your own criterion-by-criterion verification\nfrom Step 2 backs every success criterion with real command output. Gear-4 review is\nreviewable as complete when the surface test and the declared gate are green and every condition\nfor its claimed terminal outcome is proven from the bounded receipt plus accepted provenance.\n\nA faithfully proven negative outcome (`BUDGET_EXHAUSTED`, `NO_IMPROVEMENT`, or `REVIEW_REJECTED`)\nmay receive an approved terminal review. Record it explicitly as terminal completion, not goal\nsuccess, and retain `goal-succeeded: false` in the judgement. Approval confirms faithful execution\nand evidence; it never rewrites the controller outcome to `GOAL_REACHED`.\n\nPersist this review phase's criterion-by-criterion judgement in\n`.formula\u002Fwork\u002Ftasks\u002F\u003Cplan-basename>\u002Freview-report.md`; for Gear 4 use\n`.formula\u002Fwork\u002Floops\u002F\u003Cslug>\u002Freview-report.md`. Cite the evidence artifact plus command hashes and compact\nreceipt fields instead of duplicating command transcripts. Write this terminal review report before\nconsuming the ledger. For Gear 4, include the outcome, goal-success boolean, provenance checks,\nterminal-condition matrix, and exact receipt hashes. For Gear 4 only, after the report is complete and every\napplicable gate is green, consume the READY loop ledger into\n`review-report.md` by setting `status: CONSUMED`, `next: none`, and\n`consumed-by: \u003Creview-report-path>`. A faithfully proven negative outcome is consumed by the same\nterminal-review lifecycle. A Gear-4 interruption or proof mismatch leaves the primary `READY`.\nRegular review instead follows the ordered index\u002Freport publication below.\n\n**Regular Gear 3 publication and resume.** Create a fresh report only at an absent canonical path;\nan existing report requires its exact resume capability, never an overwrite or most-recent fallback.\nStart its workflow header before the title as:\n\n```yaml\n\u003C!-- steepy-workflow: v1\nphase: review\nstatus: DRAFT\nnext: none\nsource: \u003Ctask-result-index-path>\nconsumed-by: none\n-->\n```\n\nRecord the exact criteria\u002Findex\u002Fdiff identities, current branch\u002FHEAD or uncommitted diff identity,\napplicable approval and current receipt-backed verification before publication. After all criteria\nand gates pass, first atomically replace the report as `status: READY`; then atomically replace the\ntask-result index as `status: CONSUMED`, `next: none`, `consumed-by: \u003Creview-report-path>`.\nInterruption before publication leaves READY\u002FDRAFT;\nafter publication READY\u002FREADY is the sole verified prefix. An exact resume revalidates the common\nproof and accepted criteria\u002Findex\u002Fdiff before performing only the input-consumption write. The\nverified CONSUMED\u002FREADY pair is an idempotent no-op; absent proof, a changed input\u002Fdiff, or an\nimpossible pair leaves all artifacts untouched and fails closed. Gear-4 lifecycle remains governed\nby its separate terminal proof above, not this regular index\u002Freport resume table.\n\nPlace the human action immediately beside this terminal envelope; emit no native skill invocation:\n\n```yaml\nhandoff: steepy-formula\u002Fv1\nnext: human\nrequired:\n  evidence: \u003Creview-report-path>\nonDemand: none\n```\n\n**Human action:** inspect the evidence and choose the version-bump\u002FPR action in Step 5.\n\n**Gear-4 negative terminal outcome:** replace that human action with: inspect the evidence; there is\nno release or PR offer for this outcome. Suppress Step 5 completely. A new goal contract, ratified\nby a human for Gear 4, is required for another run; neither this review nor the consumed ledger grants more\nmutation budget.\n\n**Autopilot:** If `drive: autopilot` is present in the verdict contract (Step 0), stop here — this is\nthe last unattended phase of the run, and the version-bump\u002FPR decision below (Step 5) stays human.\nNo scenario was created at Step 2: the typed candidates listed above are what the\ninteractive Step 5 session creates, previews, and applies when the human acts. After the report\nabove, append\n`\u003CISO timestamp> — review — READY_FOR_PR — run-id=\u003Cconductor-supplied> attempt=\u003Cpositive supplied> \u003Cone-line verdict>` to\n`.formula\u002Fwork\u002Ftasks\u002F\u003Cspec-basename>\u002Fautopilot-status.md` and end the session: no version bump, no\nCHANGELOG edit, no PR, no Step 5 question. The run's artifacts and this report stay on disk for a\nhuman to pick up Step 5 later, in an interactive session.\n\n### Step 5 — Offer a pull request\n\nThis step applies to regular review and a successfully proven Gear-4 `GOAL_REACHED` only. A Gear-4\nnegative outcome has already stopped at Step 4 with no release\u002FPR offer and a requirement for a new\ngoal contract before another run.\n\n**The human apply gate.** Before any PR: if Step 2 assembled a Promotion Scenario,\nit is applied here and only here, in manual drive. Ask one explicit yes\u002Fchange\u002Fno question on the\nquoted typed diff; on **yes**, apply it directly: run `steepy-formula promotion apply\n\u003Cscenario-id> --yes [--approver \u003Cname>]`. The human's explicit yes in this interactive session\nIS the approval: pass `--yes` only after the human's explicit yes, only for the scenario id\nwhose preview the human saw, never on silence; `--approver` names the human who said yes\n(default: the OS user). A change re-previews and needs a new yes; the CLI's stale check\n(exit 20) refuses an apply whose base revision moved after the preview. On **no**, apply\nnothing: the scenario stays a draft for a later interactive session; `promotion discard`\ncleans it up if it is abandoned. Never apply inline at Step 2, and never a `--yes` supplied by\na conductor or in autopilot — no human said yes there. In autopilot there is no scenario\nin this review's hands: skip straight to the PR flow below.\n\nCheck the current branch. If it is `main`\u002F`master`, or has no commits ahead of `main`, stop here — there is nothing to propose.\n\n**Applicability check:** this bump step is repo-local, not a plugin feature — `review` runs in any repo the plugin is installed in, but the bump machinery lives in the target repo. Only proceed when the current repo owns it: `scripts\u002Fbump-version.mjs` AND `.github\u002Fworkflows\u002Fversion-gate.yml` both exist at the repo root. If either is missing, skip straight to asking about the PR below.\n\nOtherwise, resolve the version bump before opening the PR:\n1. **Derive the bump type from the accepted criteria, task-result index, and branch diff** (gear 4:\n   from the accepted goal contract and authenticated terminal facts): bugfix → `patch`, feature →\n   `minor`, breaking change → `major`. If that evidence is insufficient, ask the human to resolve\n   the ambiguity; no spec, plan, ledger, or per-task report read is authorized to derive the bump.\n2. **Ask one confirmation question**: state the proposed type and the derivation, e.g. \"This looks like a `minor` bump (new feature) — bump the version before the PR?\"\n3. **On confirm**: run `node scripts\u002Fbump-version.mjs \u003Ctype>` (repo-local tooling, not an engine script — resolve it against the target repo, never `\u003Cengine-root>\u002Fscripts\u002F`, which would bump the engine's own manifests from any repo), add a `## vX.Y.Z (YYYY-MM-DD)` section at the top of `CHANGELOG.md` (below the `# Changelog` title) with short factual bullets summarizing the change, and commit the bump + changelog as one commit on the branch (e.g. `vX.Y.Z`).\n4. **Alternatively, if the work must not ship a release** (docs-only, CI-only): skip the bump and apply the `no-release` GitHub label to the PR instead (`gh pr edit --add-label no-release`, or `--label` at creation).\n\nThe `Version gate` workflow fails an unbumped PR that lacks the `no-release` label; the tag and GitHub Release are created automatically on merge by the `Release` workflow.\n\nThen ask the user: **\"Work is reviewed clean on `\u003Cbranch>`. Create a PR to `main`?\"** On confirmation, push the branch and create the PR following the repo's PR conventions; report the PR URL.\n\nWhatever comes next — a new chain or a fix round — belongs in a new session: this phase's context does not serve it, and every turn there would pay to carry it forward.\n\n## Model Selection\n\nAlways set `model:` explicitly when dispatching a subagent. An omitted model inherits the session model (usually the most capable and most expensive) and silently defeats this policy.\n\nPick the tier by task complexity and risk:\n\n| Signal | Tier |\n|--------|------|\n| Mechanical \u002F transcription (1-2 files, complete spec) | cheap |\n| Integration \u002F multi-file \u002F judgment | standard |\n| Design \u002F architecture \u002F high-risk or subtle change | most-capable |\n\nReviewers floor at **standard** and rise to **most-capable** when the artifact is high-risk or subtle.\n\n**Turn count beats token price.** The cheapest model often takes 2-3× the turns on multi-step work — more wall-clock and context overall. Make a mid-tier model the floor for reviewers and for implementers working from prose; reserve the cheapest tier for transcription tasks whose exact code the plan already contains.\n\nTranslate the tier to a concrete model by judgment at dispatch time using your harness's available models. The three tiers are the whole ladder — `most-capable` is its top rung, not an open-ended \"best available\". A model that sits above that rung is outside the ladder: never reach it from a tier, and record a one-line reason in the ledger when you override to it. The whole-branch review runs at the reviewer tier — `standard`, rising to `most-capable` when the branch is large or subtle — not at whatever model the harness happens to offer. On a harness without per-dispatch model choice, everything runs on the session model — record that in the ledger.\n","\u003C!-- steepy:manual-handoff:v1:start -->\n## Manual handoff envelope\n\nManual phase boundaries use this literal shape (the envelope is chat text, never a file):\n\n```yaml\nhandoff: steepy-formula\u002Fv1\nnext: \u003Cphase>\nrequired: \u003Crole\u002Fpath map or none>\nonDemand: \u003Crole\u002Fpath map or none>\n```\n\nAll four keys are mandatory and occur once, in that order. Use explicit `none` for an empty map.\nThe phase input role maps are exact and unique: brainstorm has `required: none`, `onDemand: none`;\nplan has required `spec`, `onDemand: none`; Gear-2 direct implement has required `spec`,\n`onDemand: none`; plan-backed implement has required `plan`, on-demand `source-spec`; either\nimplement alternative additionally authorizes exact `progress-ledger` and `task-results` on demand\nwhen resuming;\nregular Gear-3 `review` has required `criteria`, `task-results`, and `branch-diff`, `onDemand: none`;\nfresh `loop-engineer` has required `goal`, `onDemand: none`; resume may additionally authorize exact\n`loop-ledger` on demand; Gear-4 `review` has required `goal` and `loop-ledger`, `onDemand: none`.\nThus Gear-2 brainstorm → implement maps `spec`; brainstorm → plan maps `spec`; plan → implement\nmaps `plan`, `source-spec`; Gear-3 implement → review maps `criteria`, `task-results`, `branch-diff`;\nloop-engineer → review maps `goal`, `loop-ledger`. The two implement role alternatives and the two\nreview role alternatives are mutually exclusive: never merge or infer them. A role occurs only in\nits mapped phase, alternative, and list.\n\nValidate literally before any work-artifact read. Reject duplicate keys, unknown keys, unknown roles,\nmissing keys, a wrong `next`, and unsafe paths: absolute paths, `..` segments, globs, or anything\noutside the repository. Never infer corrections. An explicit malformed or stale binding fails closed\nwithout discovery. Supporting `onDemand` inputs are never preloaded; read one only for a concrete\nnamed missing fact. Stable routing inputs stay outside the envelope and follow bootstrap routing.\nNative invocation spelling comes only from the active harness. No child inherits general\n`.formula\u002Fwork\u002F**` access; pass only its accepted role paths.\n\nEvery new work artifact starts with verdict\u002Fcontract metadata when that phase requires it, followed\nimmediately by this versioned header before the Markdown title:\n\n```yaml\n\u003C!-- steepy-workflow: v1\nphase: \u003Cbrainstorm|plan|implement|review|loop-engineer|goal-contract>\nstatus: \u003CDRAFT|READY|CONSUMED>\nnext: \u003Cphase|none>\nsource: \u003Csafe exact repo-relative path|none>\nconsumed-by: \u003Csafe exact repo-relative path|none>\n-->\n```\n\nHeader fields are mandatory, unique, and literal. `source` and `consumed-by` must agree with the\nadjacent lifecycle transition and authorize no transitive reads. Headerless artifacts are\nignored by recovery and are available only under explicit user audit scope.\n\n### Exact publication pairs\n\nThese are exact resume capabilities, not paths inferred from provenance. The input retains its\nnormal required role; the output role is added to `onDemand`, as are any existing on-demand companions.\nRegular review retains its required criteria and branch diff. Plan resume adds `onDemand.output-plan`;\nregular Gear-3 review resume adds `onDemand.review-report`. Fresh forms above are unchanged.\nGear-4 controller\u002Fledger handling is excluded from this model-based pair repair.\n\n| Route | Input role | Output role | Companion roles | Input phase | Output phase | Output next |\n| --- | --- | --- | --- | --- | --- | --- |\n| plan | spec | output-plan | none | brainstorm | plan | implement |\n| implement-2-direct | spec | task-results | progress-ledger | brainstorm | implement | none |\n| implement-2-plan | plan | task-results | source-spec,progress-ledger | plan | implement | none |\n| implement-3 | plan | task-results | source-spec,progress-ledger | plan | implement | review |\n| review-3 | task-results | review-report | criteria,branch-diff | implement | review | none |\n\n### Publication interruption prefixes\n\nThis table applies only to the exact authorized pair with matching phase, route and provenance.\nValidate the pair before the fresh-only READY input gate; a resume capability is never inferred.\nInput `next` names the consumer phase until consumed; consumption retains it for plan\u002Fimplement\nand sets it to `none` for review. Output `consumed-by` is `none` throughout this producer's run.\nAn unconsumed input must have `consumed-by: none`; a consumed input must name the exact output.\n\n| Input status | Output status | Action |\n| --- | --- | --- |\n| absent | absent | reject |\n| absent | DRAFT | reject |\n| absent | READY | reject |\n| absent | CONSUMED | reject |\n| DRAFT | absent | reject |\n| DRAFT | DRAFT | reject |\n| DRAFT | READY | reject |\n| DRAFT | CONSUMED | reject |\n| READY | absent | fresh-create |\n| READY | DRAFT | resume-work |\n| READY | READY | finish-consumption |\n| READY | CONSUMED | reject |\n| CONSUMED | absent | reject |\n| CONSUMED | DRAFT | reject |\n| CONSUMED | READY | no-op |\n| CONSUMED | CONSUMED | reject |\n\n`fresh-create` permits a new absent output only in a fresh run, never a missing resume target.\n`resume-work` requires exact output capability and leaves the input READY while work\u002Fgates remain.\nFor `finish-consumption`, revalidate output.source against the exact input path, input.consumed-by\nas `none`, and the recorded approval and verification bound to those input\u002Foutput paths and content\n(excluding lifecycle fields). Recheck current branch\u002FHEAD or uncommitted diff identity where applicable.\nOnly then perform the input-consumption write with input.consumed-by equal to the exact output path.\nFor `no-op`, revalidate the recorded proof with the final reciprocal links instead of the\nunconsumed-input condition; perform no writes or redispatch.\nA missing capability, mismatched path\u002Fphase\u002Fnext\u002Fcontent, absent approval\u002Fproof, or impossible state\nmust fail closed, preserving both files. A READY marker alone is not proof. Do not follow links to\nundeclared evidence; request the exact missing capability or user-delimited scope instead.\nNever select a most-recent output or recover by directory\u002Fbody search.\n\nPublication is ordered: record the phase's applicable approval and verification in the output,\npublish output READY by per-file atomic replacement, then consume the input by per-file atomic\nreplacement. This is not a multi-file atomic operation; READY\u002FREADY is its sole verified intermediate\nprefix. Resume repairs no other state and never repeats completed work or release actions.\nIn manual drive you are the enforcement: no parser, manifest, or validator checks this contract.\n\u003C!-- steepy:manual-handoff:v1:end -->\n\n\n# review\n\nVerify completed work with evidence. Never assert \"done\" without command output to back it.\n\n> **Engine root:** this skill's base directory is `\u003Cengine-root>\u002Fskills\u002Freview\u002F`; engine\n> scripts live two levels up, at `\u003Cengine-root>\u002Fscripts\u002F`. Resolve them relative to the base\n> directory your harness reports for this skill.\n\n## Checklist\n\n- [ ] Declare the hub mode once per run — `steepy-formula hub-mode` (always exit 0) — and stop in `migrate` and `none`, and carry the declared mode into the report (Step 0)\n- [ ] In autopilot, read the supplied phase manifest first; then read the gear and scale ceremony to it (Step 0)\n- [ ] Manifest-backed autopilot: load criteria-only source + task index + diff + standards, regular manual review: load accepted criteria + task-result index + branch diff + graph orientation; Gear 4: load the exact goal contract + loop ledger instead (Step 1)\n- [ ] Run `capture-review-evidence.mjs` once: keep verbatim command output in the current run's\n      `evidence-report.md`, consume only its compact JSON receipt, and judge every criterion in\n      session (Step 2)\n- [ ] For Gear 4, validate goal\u002Fledger provenance and the ledger event commitment, then prove the\n      exact claimed terminal outcome without confusing negative completion with goal success (Steps 1–2)\n- [ ] Sweep unpromoted insights from the active inventory: manifest criteria + task-result index in autopilot; manual insight sweep uses only accepted criteria + task-result index compact status and `signals`; the manual sweep also assembles the Promotion Scenario from the typed candidates file (Step 2)\n- [ ] Confirm the receipt contains a green `coherence-gate` command result (Step 3)\n- [ ] Report: the declared hub-mode, whether the `coherence-gate` is green, the typed candidate list and the quoted typed diff, AND every regular criterion or Gear-4 terminal claim verified\n      from real command output (Step 4)\n- [ ] Manual drive: ask one explicit yes\u002Fchange\u002Fno question on the assembled Promotion Scenario and, on yes, apply it here — `promotion apply \u003Cscenario-id> --yes`, passed only after the human's explicit yes; a no applies nothing (Step 5)\n- [ ] Resolve the version bump (or apply `no-release`) and offer a PR to `main` only for regular\n      review or Gear-4 `GOAL_REACHED`, when the work landed on a separate branch with commits ahead;\n      suppress release\u002FPR for negative terminal outcomes (Step 5)\n\n## Procedure\n\n### Step 0 — Read the gear\n\n**Hub-mode declaration (once per run).** Run `steepy-formula hub-mode` before anything else and\ncarry its declared mode — `graph`, `migrate`, or `none`, each printed with one `evidence:` line —\nthrough the whole run. The command always exits 0: it is a declaration for the report, not a\ngate, and this prose never re-implements its detection — the command is the single authority.\nThe declared mode branches the run:\n\n- **`graph` mode:** proceed — the Step 1 orientation reads, the Promotion Scenario assembly at\n  Step 2, the graph coherence gate wired into the Step 2 capture and checked at Step 3, and the\n  human apply at Step 5.\n- **`migrate` mode:** the repo still carries a pre-graph Markdown hub.\n  Stop and tell the user to run the `graph-migrate` skill (invoke it the way your harness invokes skills) before any other workflow.\n- **Mode `none`:** no graph — the repo is not initialized: tell the user to run the `init` skill\n  first and stop.\n\n**Autopilot preflight:** When the conductor-supplied phase manifest is present, read that\nmanifest before any other task input. Validate its role and scope against the review phase and the\nsupplied correlation identity. Eagerly read every `required` input before acting. Do not preload `onDemand`;\nconsult it only for a concrete named missing fact and record the read in the review\nreport. Phase-manifest authority applies only when `drive: autopilot`; manual-drive discovery remains\nStep 1's behavior.\n\nAutopilot status timestamps must use UTC with three millisecond digits (`YYYY-MM-DDTHH:mm:ss.sssZ`), generated with `new Date().toISOString()`.\n\n**Manifest-backed autopilot:** Take the Step-0 scalar facts from `manifest.contract.verdict`,\n`manifest.contract.gear`, and `manifest.contract.drive`. Require `drive: autopilot` and gear 3, and\nhonor the recorded verdict. Do not read or locate the spec: it is deliberately absent from the review\ninventory, and the criteria-only required artifact supplies the review criteria. Apply the Step 4\u002F5\nAutopilot paragraph below and never ask the user anything. A manifest\u002Foutput failure, success criteria\nunmet, the `graph validate` gate red, or a CONFLICT discovered mid-run → append\n`\u003CISO timestamp> — review — BLOCKED — run-id=\u003Cconductor-supplied> attempt=\u003Cpositive supplied> \u003Creason>`\nto `.formula\u002Fwork\u002Ftasks\u002F\u003Cspec-basename>\u002Fautopilot-status.md` and exit non-zero. A CONFLICT discovered\nmid-run is ALWAYS a halt — doc-wins needs a human, never auto-override.\n\nUse the pinned `manifest.contract.taskResultProtocol` to interpret the task index. In autopilot\nthe task index is the generated `steepy-task-results: v2` JSON block with exact `changedPaths`\narrays; manual drive keeps `- Task` bullets. Never split v2 filenames on commas or expand braces.\nThe conductor replays execution receipts and required task\u002Ffinal approvals before this phase;\nthe index is a projection, and a captured execution alone is never approval. This introduces no\nreviewer permission to open receipt bodies, task reports, the plan, or the progress ledger. Missing\nor drifted evidence fails closed; never infer historical baselines.\nManual drive and Gear 4 retain their existing grammar and capabilities.\n\n**Manual\u002Fno-manifest:** Validate the manual handoff before reading any work artifact. Regular Gear-3\nreview fresh entry accepts exactly `criteria`, `task-results`, and `branch-diff` with `onDemand: none`;\nregular resume adds only `onDemand.review-report`, the exact canonical report paired with that index.\nCheck existence of the canonical report derived from the accepted primary path without opening it.\nAn existing review report without its exact capability requires a stop before body reads or evidence\ncollection; do not overwrite it or select another report. Same-invocation output creation is not\nhistorical resume authority.\nValidate this pair before the fresh-only READY input gate: READY\u002FDRAFT resumes verification,\nREADY\u002FREADY is finish-consumption, and CONSUMED\u002FREADY is no-op after the common proof checks.\nRequire `phase: review`, `next: none`, the exact index as report `source`, and `consumed-by: none`.\nRead the authorized report only for those named progress\u002Fapproval\u002Fproof facts. Verified terminal\npairs retain all supporting-input checks: on every resume, validate the accepted criteria and\nbranch-diff paths against the index metadata and their current content against the report's proof,\nbefore any terminal shortcut. Only then go straight to the human envelope in Step 4 after any sole consumption repair; never rerun\nrelease actions. Gear-4 review accepts exactly `goal` and `loop-ledger` with `onDemand: none` and\ndoes not acquire this regular-review resume route. The task-result index (or the\nGear-4 ledger) is primary: fresh entry requires `READY` \u002F `next: review`, then validate every supporting path and\nits explicit provenance against it. An explicit malformed, stale, unsafe, or mismatched envelope is\nan explicit failure and never falls back to discovery. With no envelope, pathless recovery may\ninspect headers only on candidate task-result indexes (or loop ledgers), never bodies; require user\nselection and then a new exact envelope. Never select by recency.\n- **Artifact present:** honor its gear.\n- **Artifact absent and this is gear-1\u002F2 work:** recompute the coverage verdict inline (COVERED \u002F\n  CONFLICT \u002F GAP with `file:line` evidence → gear). Declare the recomputed\n  verdict, gear, and evidence, then ask the user for one confirmation before the ceremony starts —\n  the user may force a different gear, and the override is recorded in the verdict artifact. Then\n  proceed.\n- **Artifact absent and full gear-3 ceremony is required for this command:** do not proceed\n  blind — tell the user to enter through the upstream workflow chain first, and stop.\n\nScale this skill's ceremony to the gear (see \"Ceremony by gear\" below). This skill dispatches **no\nsubagent reviewer or evidence collector**: `capture-review-evidence.mjs` collects command output\ninline and deterministically, while this session self-verifies every success criterion plus the\n`graph validate` coherence gate, per the \"Review architecture (gear 3)\" governance rule — recorded as a\nRule in the canonical graph. Judgement is never delegated.\n\n**Autopilot correlation.** Before appending any current-protocol status marker, use the `run-id` and\npositive `attempt` supplied by the conductor's `phasePrompt`; that correlated identity is authoritative.\nDo not derive either value from filenames. Do not invent a correlation identity. A child without supplied run\u002Fattempt identity\nis unresolvable: exit non-zero without appending an uncorrelated new-protocol marker. Only versioned, correlated status is supported; uncorrelated markers never complete a phase.\n\n**Session declaration (declare, never block).** Say whether this conversation started clean for this\nphase or has already run an earlier phase of the chain in the same session. A same-session phase\nalready grew the context, and every turn of this phase now pays to carry it — name that cost, but\nnever refuse or gate on it: this is a declaration about the conversation, not a measurement, and\nsound work proceeds either way. If the session's history cannot be determined with certainty, say so\nand proceed. Under `drive: autopilot` this declaration is a no-op: the conductor\n(`scripts\u002Fautopilot.mjs`) already opens a fresh headless session per phase, so the discipline stated\nhere is already enforced by that mechanism, not by this prose.\n\n### Ceremony by gear\n\n- **All gears**: run the surface test command and the `graph validate` coherence gate through\n  the deterministic inline capture in Step 2, and verify each success criterion yourself. There is\n  no subagent evidence reviewer or collector. The complete command bytes are durable on disk; only the bounded receipt\n  enters model context. Verification and judgement are never delegated, at any gear.\n- **Gear 4**: review accepts gear-4 work. Verification is the re-run of the `verifier` from the\n  goal contract, plus the loop ledger as evidence — alongside the existing self-verify above, not\n  replacing it.\n\n### Step 1 — Load bounded review inputs and the hub\n\n**Manual drive, regular Gear 3:** Eagerly read exactly the accepted `criteria`, `task-results`, and\n`branch-diff` artifacts. Never locate or read the spec, plan, progress ledger, task briefs, task\nreports, task reviews, sibling work files, transcripts, or conversations. Treat `task-results` as\nthe primary and validate the criteria and diff paths plus explicit source metadata against it. Then\nread the orientation for the mode declared at Step 0 — `graph`, `migrate`, or `none`; this prose never re-detects it:\n\n- **`graph` mode:** resolve orientation from the canonical graph — run\n  `steepy-formula context resolve --path \u003Cp>...` over the reviewed paths (exit 0 emits\n  `context.json` + `context.md`). The resolved context supplies the owning **surface**, its\n  Rules — the standards-equivalent — its Terms, and the surface test command from its Check\n  item's `ext_v1_command` extension property (fall back to the verification Rule's statement\n  prose when the Check carries no command). The bound specialist agent is\n  `.claude\u002Fagents\u002F\u003Csurface>-agent.md` by naming convention. On exit 10 UNRESOLVED — normal in a freshly graph-first-initialized repo, where no\n  topology bridge links the detected Packages to the curated Components yet — fall back to\n  the deterministic surface targets:\n  `steepy-formula context resolve --target apex:\u003CprojectId>:surface:\u003Cslug(name)>` for each\n  owning surface, every part derived, never guessed — surface names from\n  `.claude\u002Fagents\u002F\u003Cname>-agent.md` (CLAUDE.md's Surfaces section is the fallback listing),\n  the project id from the hub-mode declaration's `hub-mode '\u003CprojectId>'` echo, `slug(name)`\n  lowercase with non-alphanumeric runs collapsed to `-` — and never target the bare project\n  root on its own: it resolves empty. If the resolve fails, diagnose with\n  `steepy-formula graph status` — exit 3,\n  `initialized: no`, or an engine failure means the graph binding is missing or failed: tell the\n  user to run the `init` skill first and stop.\n- **`migrate` mode:** the repo still carries a pre-graph Markdown hub.\n  Stop and tell the user to run the `graph-migrate` skill (invoke it the way your harness invokes skills) before any other workflow.\n- **Mode `none`:** no graph — the repo is not initialized: tell the user to run the `init` skill\n  first and stop.\n\n**Autopilot:** use the validated phase manifest loaded in Step 0 as the authoritative input\ninventory. Its `required` inventory must provide the success-criteria source, task\u002Fresult index,\naggregate diff, relevant standards, and test commands. Exclude per-task transcripts and\nconversations: they are not review inputs. Do not independently preload them. Use `onDemand` only\nunder Step 0's concrete-missing-fact protocol, recording any such read in the report.\n\n**Gear 4 has no spec and no plan.** Eagerly read exactly the accepted goal and loop-ledger paths;\nthe loop ledger is primary. The public capability remains exactly `goal` + `loop-ledger`: do not\nderive or read `events.jsonl`, enumerate the loop directory, or expose any private event path to\nmodel context. Resolve `\u003Cengine-root>` from the skill base and invoke this read-only proof boundary:\n\n```bash\nnode \u003Cengine-root>\u002Fscripts\u002Floop-engineer.mjs --repo-root . --validate-terminal --goal '\u003Cgoal-path>' --ledger '\u003Cloop-ledger-path>'\n```\n\nThe terminal validator accepts only that exact public pair, derives the canonical `events.jsonl`\nidentity internally, replays events as the sole authority, recomputes `events-sha256`, verifies the\ngoal digest plus `phase: goal-contract`, `status: CONSUMED`, and `consumed-by:\n\u003Cloop-ledger-path>` for the goal and the `READY` ledger lifecycle. It binds every receipt fact and\nevent-derived ledger field to replay, recomputes the ledger digest, and accepts runtime-only fields\nonly in their closed canonical form. It also authenticates the event-bound reviewer report digest,\nreviewer diff digest and bytes, final branch-diff projection, and current branch, HEAD, clean tree,\nand snapshot against the retained event-derived branch. It performs zero lifecycle or Git mutation and returns one\nclosed bounded receipt; a nonzero exit or malformed\u002Fmismatched receipt means provenance is\nunverified.\n\nRequire exactly `status: VALIDATED`, the accepted `goal` and `loop-ledger`, and the closed receipt\nfields `run-id`, `outcome`, `goal-succeeded`, `events-sha256`, `mode`, `branch-review`,\n`final-projection`, `attempt-budget`, and `candidate`. `branch-review` supplies exact `verdict`, `issues-found`, review\nand attempt identities, report\u002Fdiff paths, `report-sha256`, and `diff-sha256`; `final-projection`\nsupplies the authenticated diff digest plus current branch, HEAD, clean, and snapshot facts. `attempt-budget` supplies exact\n`reserved`, `budget`, and `exhausted`; `candidate` supplies boolean-green or metric-baseline,\nmetric-best, and strict-improvement facts. Require `outcome` to be exactly `GOAL_REACHED |\nBUDGET_EXHAUSTED | NO_IMPROVEMENT | REVIEW_REJECTED`, and `goal-succeeded: true` only for\n`GOAL_REACHED`. Take these provenance facts from the receipt, never ledger assertions. The goal's\n`verifier` is the active Gear-4 criterion. Read the graph orientation above; never select a loop by recency.\n\n### Step 2 — Verify against the active criteria source\n\nFor each success criterion in the active source, confirm it is met. In manifest-backed autopilot and\nmanual regular review the active source is the required criteria-only artifact.\nRun the surface test command and preserve its **real, verbatim output** on disk; do not copy that\nunbounded transcript into model context.\n\n**Inline deterministic collection.** Resolve the current output path from the accepted primary\nartifact, never by directory discovery:\n\n- regular review: `.formula\u002Fwork\u002Ftasks\u002F\u003Cplan-basename>\u002Fevidence-report.md`;\n- Gear 4: `.formula\u002Fwork\u002Floops\u002F\u003Cslug>\u002Fevidence-report.md`.\n\nThen run this once, passing each command as one safely quoted argument:\n\n```bash\nnode \u003Cengine-root>\u002Fscripts\u002Fcapture-review-evidence.mjs --repo-root . --evidence '\u003Cevidence-path>' --test-command '\u003Cexact surface test command>'\n```\n\nThe coherence gate is the capture's final command: the script's built-in gate runs\n`steepy-formula graph validate` through the packaged CLI (`node \u003Cengine-root>\u002Fscripts\u002Fformula\u002Fcli.mjs`,\nnever a PATH lookup) from the repo root, so the project id resolves from the committed\n`.formula\u002Fconfig.json` binding (or `STEEPY_PROJECT`), labeled `coherence-gate`. Pass no gate flag.\n\nFor Gear 4 append `--verifier-command '\u003Cexact verifier command>'`. The script always runs the\nsurface test, then the optional goal verifier, then the coherence gate; it atomically replaces stale\nevidence at the fixed path and streams combined stdout\u002Fstderr into the new report. It returns one\nbounded JSON receipt on stdout with the artifact path, collection timestamps, aggregate pass state,\nand, for every command, exit\u002Fsignal\u002Ferror state, output byte and line counts, SHA-256, and a bounded\ntail excerpt. A tested command may be red while collection itself exits successfully: use\n`allPassed` and each command's `passed`\u002F`exitCode` fields to determine the review result.\n\nValidate receipt schema version 1, the exact expected artifact path, timestamps from this review\nround, and the ordered command labels (`surface-test`, optional `goal-verifier`, then `coherence-gate`). Do\nnot open or preload the complete `evidence-report.md`: the report is durable verbatim proof, while\nthe receipt is the context boundary. If a named criterion genuinely requires a fact absent from the\nbounded excerpt, read only the smallest exact line or byte range needed from that report, record the\nreason and range in the review report, and never fall back to a whole-file read. A malformed receipt,\nhard capture failure, missing report, hash\u002Fmetadata mismatch, or missing evidence means the affected\ncriterion is not verified; after fixing the cause, rerun the single capture so stale evidence is\nreplaced rather than merged.\n\nThe judgement stays here, in session: whether each success criterion is met, which swept insights\nget promoted, and the final verdict are never delegated. No success criterion is declared met for\nwant of evidence.\n\n**Gear 4:** the criterion is the goal contract and the authenticated terminal outcome, not a spec.\nThe same bounded capture runs the surface test, goal verifier, and the `coherence-gate`\n(`graph validate`) in that order. Require the\ncurrent capture receipt's surface test and that gate to be green. Interpret the fresh\nverifier result against the terminal-validator receipt; a red verifier is expected evidence for\n`BUDGET_EXHAUSTED`, not by itself invalid execution. For metric mode, parse the verifier's final\nstdout line as a finite scalar and compare it with the receipt's metric baseline using the goal's\n`metric-direction`.\n\nApply this closed outcome proof:\n\n- `GOAL_REACHED` requires boolean mode with a green verifier, or metric mode with a strict\n  improvement over the reproduced metric baseline; it also requires an approved branch review via\n  receipt `branch-review.verdict: APPROVED`, `issues-found: false`, `candidate.acceptable: true`, and\n  `goal-succeeded: true`.\n- `BUDGET_EXHAUSTED` requires boolean mode, a red boolean verifier, the full budget spent\n  (`attempt-budget.reserved` equals `budget` and `exhausted: true`), and `goal-succeeded: false`.\n- `NO_IMPROVEMENT` requires metric mode, a verifier value equal to the reproduced metric baseline,\n  no strict improvement in the declared direction, `candidate.metric-strict-improvement: false`,\n  the full budget spent (`attempt-budget.exhausted: true`), and\n  `goal-succeeded: false`.\n- `REVIEW_REJECTED` requires a goal-acceptable candidate: boolean mode must be green, or metric mode\n  must show strict improvement over baseline. It never accepts a red boolean result or a metric\n  equal to baseline. Also require receipt `candidate.acceptable: true`, unresolved review issues\n  (`branch-review.verdict: ISSUES_FOUND` and `issues-found: true`), no remaining mutation budget\n  (`attempt-budget.reserved` equals `budget` and `exhausted: true`), and `goal-succeeded: false`.\n\nThe authenticated terminal receipt is the controller-owned proof of event identity, exact\nbranch-review verdict\u002Fissues, and attempt-budget accounting. This review judges whether its terminal\noutcome is faithfully reproduced; it never reconstructs attempt state from ledger prose. A mismatch\nleaves the ledger READY and produces an issues verdict. For the insight sweep below, the only Gear-4\nsources are the accepted goal's `notes:` and accepted loop-ledger fields.\nBoolean candidate facts describe the retained branch: committed attempts update that state and\ndiscarded review-fix attempts leave the prior retained verifier result unchanged.\n\n**Insight sweep — run it always.** Work discovers knowledge as well as changing it. Use only sources\navailable in the active inventory:\n\n- **Manifest-backed autopilot:** sweep the criteria-only artifact and task-result index. Their criterion\n  IDs, `signals`, and `DONE_WITH_CONCERNS` entries are the declared insight evidence; do not locate the\n  full spec, plan, progress ledger, or per-task transcripts.\n- **Manual\u002Fno-manifest:** sweep only the accepted criteria artifact plus compact status and `signals`\n  in the accepted task-result index, specifically its `DONE_WITH_CONCERNS` entries and `signals`.\n  Never locate richer upstream or sibling artifacts.\n\nA recorded **discovery** occurrence — an implementer that had to discover what to change rather than\nexecute the brief, with DONE_WITH_CONCERNS and `discovery:unplanned` — is a rebuttal to `plan`'s\ntask-cutting criterion (a task that looked pre-decided but wasn't). Report it distinct from other\nconcerns. In both regular-review branches, derive discovery evidence only from the accepted task-result index\nand use the accepted criteria artifact for any mapped criterion; do not invent a criterion absent\nfrom that artifact or read the plan to reconstruct one.\n\n### Discovery signal interpretation\n\n| Status | Signal | Interpretation |\n| --- | --- | --- |\n| DONE_WITH_CONCERNS | discovery:unplanned | unplanned-discovery |\n| DONE_WITH_CONCERNS | other-or-none | ordinary-or-unattributed-concern |\n| DONE | discovery:unplanned | inconsistent-result |\n| DONE | other-or-none | no-discovery-evidence |\n\nNever infer discovery from DONE_WITH_CONCERNS alone. Ordinary or unattributed concerns stay distinct;\nan inconsistent result is reported as a contract defect, not a confirmed occurrence. Signal matching\nis by the exact comma-delimited ID, never substring or a similar-looking spelling. No richer source\nor per-task file read is authorized by a concern. Absence of the signal is no discovery evidence,\nnot proof that no discovery happened.\n\nFor each candidate: already promoted → cite the node in the applied scenario; still valuable →\npropose the promotion to the user as a typed candidate (term, rule, decision, readme); no longer holds → drop it\nwith a one-line reason.\n\n**The sweep assembles the Promotion Scenario.** In manual drive, beyond listing the\ncandidates, the sweep writes the still-valuable ones to the typed candidates file at\n`.formula\u002Fwork\u002Ftasks\u002F\u003Cplan-basename>\u002Fpromotion-candidates.json` (gitignored work area; a gear-4\nreview writes `.formula\u002Fwork\u002Floops\u002F\u003Cslug>\u002Fpromotion-candidates.json` — the loop directory stands in\nfor the plan directory). The file carries the typed schema only — `term` (`term`, `definition`),\n`rule` (`title`, `statement`, `concern`, `strength` of `must|should`, `appliesTo` of an existing\nnode id), `decision` (`title`, `statement`, `context`), `readme` (`note`) — built from the sweep's\nown sources: the spec's deferred proposals, the progress ledger, the per-task files. Then create\nand preview the scenario:\n\n```bash\nsteepy-formula promotion create --project \u003Cid> --candidates \u003Cfile.json> [--proposer \u003Cname>] \u003Cscenario-id>\nsteepy-formula promotion preview \u003Cscenario-id>\n```\n\nThe scenario id convention is `\u003CYYYY-MM-DD>-\u003Ctopic>`. The wrapper is that file's only interpreter\n— terms, rules, and decisions become graph operations; `readme` notes are reported out-of-graph\nand stay on the direct-edit channel — this prose maps nothing itself. A malformed file or a\nviolation exits 2 (fix the file, not the command); an already-used scenario id exits 4\n(`promotion discard` it, or pick a fresh id). Quote the preview's typed diff verbatim into the\nStep 4 report — and apply nothing here: Step 5 owns the apply.\n\n**Autopilot:** a promotion candidate is never a blocker on its own — Step 0's \"never ask; unresolvable → BLOCKED\" governs things the run cannot finish without a human, not documentation opportunities. Carry every candidate into the Step 4 report instead, for the human to act on at the Step 5 version-bump\u002FPR decision. Autopilot review creates NO scenario: a draft assembled now would sit `PROMOTION_STALE` by the time a human reaches Step 5, when the head has moved past this phase. List the typed candidates — same schema, as report entries — and nothing more; the interactive Step 5 session creates, previews, and applies when the human acts.\n\nIf the work changed a durable convention, standard, domain term, architecture decision, user-facing workflow, or README behavior, verify that the durable knowledge was promoted into stable versioned docs.\nSpecs and plans themselves are not sufficient evidence of durable documentation.\nDurable knowledge rides the Promotion Scenario: the typed diff is the proposal\nand the Step 5 human-approved `promotion apply` is the promotion — verify the durable knowledge\nrides the scenario, or its out-of-graph `readme` notes.\n\n### Step 3 — Run the coherence gate (mandatory)\n\nThe Step 2 capture runs the coherence gate as its final command — this is required, not optional.\nTerminal review is the sole Gear-4 owner of this gate; the loop controller does not claim it:\nrequire the current receipt's final command to have `label: coherence-gate`,\n`passed: true`, and `exitCode: 0` — the `graph validate` invocation wired in at Step 2 — with\nits bounded excerpt showing `violations: 0` and a clean revision-chain replay. Plus one sanity\n`steepy-formula context resolve` of a representative path or target from the work just\nreviewed, run in session, confirming the resolver still answers.\n\nThe report path, current-run\ntimestamps, output SHA-256, byte count, and line count bind that receipt to the verbatim transcript\non disk. Do not copy or read the complete transcript merely to satisfy this gate.\n\nThe work is **not done** until the gate is green: any invariant violation or broken replay is incomplete work, and the sanity `context resolve` must answer. Fix the graph before claiming completion.\n\n### Step 4 — Report\n\nReport, with evidence: tests pass\u002Ffail (with output), every active criterion or Gear-4 terminal\ncondition met or not, the gate outcome, and the declared hub-mode. The report also carries\nthe typed candidate list and quotes the Step 2 `promotion preview` typed diff verbatim.\nRegular review is reviewable as complete only when the declared gate is green (`graph validate`) AND your own criterion-by-criterion verification\nfrom Step 2 backs every success criterion with real command output. Gear-4 review is\nreviewable as complete when the surface test and the declared gate are green and every condition\nfor its claimed terminal outcome is proven from the bounded receipt plus accepted provenance.\n\nA faithfully proven negative outcome (`BUDGET_EXHAUSTED`, `NO_IMPROVEMENT`, or `REVIEW_REJECTED`)\nmay receive an approved terminal review. Record it explicitly as terminal completion, not goal\nsuccess, and retain `goal-succeeded: false` in the judgement. Approval confirms faithful execution\nand evidence; it never rewrites the controller outcome to `GOAL_REACHED`.\n\nPersist this review phase's criterion-by-criterion judgement in\n`.formula\u002Fwork\u002Ftasks\u002F\u003Cplan-basename>\u002Freview-report.md`; for Gear 4 use\n`.formula\u002Fwork\u002Floops\u002F\u003Cslug>\u002Freview-report.md`. Cite the evidence artifact plus command hashes and compact\nreceipt fields instead of duplicating command transcripts. Write this terminal review report before\nconsuming the ledger. For Gear 4, include the outcome, goal-success boolean, provenance checks,\nterminal-condition matrix, and exact receipt hashes. For Gear 4 only, after the report is complete and every\napplicable gate is green, consume the READY loop ledger into\n`review-report.md` by setting `status: CONSUMED`, `next: none`, and\n`consumed-by: \u003Creview-report-path>`. A faithfully proven negative outcome is consumed by the same\nterminal-review lifecycle. A Gear-4 interruption or proof mismatch leaves the primary `READY`.\nRegular review instead follows the ordered index\u002Freport publication below.\n\n**Regular Gear 3 publication and resume.** Create a fresh report only at an absent canonical path;\nan existing report requires its exact resume capability, never an overwrite or most-recent fallback.\nStart its workflow header before the title as:\n\n```yaml\n\u003C!-- steepy-workflow: v1\nphase: review\nstatus: DRAFT\nnext: none\nsource: \u003Ctask-result-index-path>\nconsumed-by: none\n-->\n```\n\nRecord the exact criteria\u002Findex\u002Fdiff identities, current branch\u002FHEAD or uncommitted diff identity,\napplicable approval and current receipt-backed verification before publication. After all criteria\nand gates pass, first atomically replace the report as `status: READY`; then atomically replace the\ntask-result index as `status: CONSUMED`, `next: none`, `consumed-by: \u003Creview-report-path>`.\nInterruption before publication leaves READY\u002FDRAFT;\nafter publication READY\u002FREADY is the sole verified prefix. An exact resume revalidates the common\nproof and accepted criteria\u002Findex\u002Fdiff before performing only the input-consumption write. The\nverified CONSUMED\u002FREADY pair is an idempotent no-op; absent proof, a changed input\u002Fdiff, or an\nimpossible pair leaves all artifacts untouched and fails closed. Gear-4 lifecycle remains governed\nby its separate terminal proof above, not this regular index\u002Freport resume table.\n\nPlace the human action immediately beside this terminal envelope; emit no native skill invocation:\n\n```yaml\nhandoff: steepy-formula\u002Fv1\nnext: human\nrequired:\n  evidence: \u003Creview-report-path>\nonDemand: none\n```\n\n**Human action:** inspect the evidence and choose the version-bump\u002FPR action in Step 5.\n\n**Gear-4 negative terminal outcome:** replace that human action with: inspect the evidence; there is\nno release or PR offer for this outcome. Suppress Step 5 completely. A new goal contract, ratified\nby a human for Gear 4, is required for another run; neither this review nor the consumed ledger grants more\nmutation budget.\n\n**Autopilot:** If `drive: autopilot` is present in the verdict contract (Step 0), stop here — this is\nthe last unattended phase of the run, and the version-bump\u002FPR decision below (Step 5) stays human.\nNo scenario was created at Step 2: the typed candidates listed above are what the\ninteractive Step 5 session creates, previews, and applies when the human acts. After the report\nabove, append\n`\u003CISO timestamp> — review — READY_FOR_PR — run-id=\u003Cconductor-supplied> attempt=\u003Cpositive supplied> \u003Cone-line verdict>` to\n`.formula\u002Fwork\u002Ftasks\u002F\u003Cspec-basename>\u002Fautopilot-status.md` and end the session: no version bump, no\nCHANGELOG edit, no PR, no Step 5 question. The run's artifacts and this report stay on disk for a\nhuman to pick up Step 5 later, in an interactive session.\n\n### Step 5 — Offer a pull request\n\nThis step applies to regular review and a successfully proven Gear-4 `GOAL_REACHED` only. A Gear-4\nnegative outcome has already stopped at Step 4 with no release\u002FPR offer and a requirement for a new\ngoal contract before another run.\n\n**The human apply gate.** Before any PR: if Step 2 assembled a Promotion Scenario,\nit is applied here and only here, in manual drive. Ask one explicit yes\u002Fchange\u002Fno question on the\nquoted typed diff; on **yes**, apply it directly: run `steepy-formula promotion apply\n\u003Cscenario-id> --yes [--approver \u003Cname>]`. The human's explicit yes in this interactive session\nIS the approval: pass `--yes` only after the human's explicit yes, only for the scenario id\nwhose preview the human saw, never on silence; `--approver` names the human who said yes\n(default: the OS user). A change re-previews and needs a new yes; the CLI's stale check\n(exit 20) refuses an apply whose base revision moved after the preview. On **no**, apply\nnothing: the scenario stays a draft for a later interactive session; `promotion discard`\ncleans it up if it is abandoned. Never apply inline at Step 2, and never a `--yes` supplied by\na conductor or in autopilot — no human said yes there. In autopilot there is no scenario\nin this review's hands: skip straight to the PR flow below.\n\nCheck the current branch. If it is `main`\u002F`master`, or has no commits ahead of `main`, stop here — there is nothing to propose.\n\n**Applicability check:** this bump step is repo-local, not a plugin feature — `review` runs in any repo the plugin is installed in, but the bump machinery lives in the target repo. Only proceed when the current repo owns it: `scripts\u002Fbump-version.mjs` AND `.github\u002Fworkflows\u002Fversion-gate.yml` both exist at the repo root. If either is missing, skip straight to asking about the PR below.\n\nOtherwise, resolve the version bump before opening the PR:\n1. **Derive the bump type from the accepted criteria, task-result index, and branch diff** (gear 4:\n   from the accepted goal contract and authenticated terminal facts): bugfix → `patch`, feature →\n   `minor`, breaking change → `major`. If that evidence is insufficient, ask the human to resolve\n   the ambiguity; no spec, plan, ledger, or per-task report read is authorized to derive the bump.\n2. **Ask one confirmation question**: state the proposed type and the derivation, e.g. \"This looks like a `minor` bump (new feature) — bump the version before the PR?\"\n3. **On confirm**: run `node scripts\u002Fbump-version.mjs \u003Ctype>` (repo-local tooling, not an engine script — resolve it against the target repo, never `\u003Cengine-root>\u002Fscripts\u002F`, which would bump the engine's own manifests from any repo), add a `## vX.Y.Z (YYYY-MM-DD)` section at the top of `CHANGELOG.md` (below the `# Changelog` title) with short factual bullets summarizing the change, and commit the bump + changelog as one commit on the branch (e.g. `vX.Y.Z`).\n4. **Alternatively, if the work must not ship a release** (docs-only, CI-only): skip the bump and apply the `no-release` GitHub label to the PR instead (`gh pr edit --add-label no-release`, or `--label` at creation).\n\nThe `Version gate` workflow fails an unbumped PR that lacks the `no-release` label; the tag and GitHub Release are created automatically on merge by the `Release` workflow.\n\nThen ask the user: **\"Work is reviewed clean on `\u003Cbranch>`. Create a PR to `main`?\"** On confirmation, push the branch and create the PR following the repo's PR conventions; report the PR URL.\n\nWhatever comes next — a new chain or a fix round — belongs in a new session: this phase's context does not serve it, and every turn there would pay to carry it forward.\n\n## Model Selection\n\nAlways set `model:` explicitly when dispatching a subagent. An omitted model inherits the session model (usually the most capable and most expensive) and silently defeats this policy.\n\nPick the tier by task complexity and risk:\n\n| Signal | Tier |\n|--------|------|\n| Mechanical \u002F transcription (1-2 files, complete spec) | cheap |\n| Integration \u002F multi-file \u002F judgment | standard |\n| Design \u002F architecture \u002F high-risk or subtle change | most-capable |\n\nReviewers floor at **standard** and rise to **most-capable** when the artifact is high-risk or subtle.\n\n**Turn count beats token price.** The cheapest model often takes 2-3× the turns on multi-step work — more wall-clock and context overall. Make a mid-tier model the floor for reviewers and for implementers working from prose; reserve the cheapest tier for transcription tasks whose exact code the plan already contains.\n\nTranslate the tier to a concrete model by judgment at dispatch time using your harness's available models. The three tiers are the whole ladder — `most-capable` is its top rung, not an open-ended \"best available\". A model that sits above that rung is outside the ladder: never reach it from a tier, and record a one-line reason in the ledger when you override to it. The whole-branch review runs at the reviewer tier — `standard`, rising to `most-capable` when the branch is large or subtle — not at whatever model the harness happens to offer. On a harness without per-dispatch model choice, everything runs on the session model — record that in the ledger.\n",{"name":5,"description":7,"user-invocable":24},true,"main","a5e0492607593191c2b0a1c603cf421996b5e22f","skills\u002Freview",0,"https:\u002F\u002Fgithub.com\u002Fsteepy-it\u002Fsteepy-formula","https:\u002F\u002Fwww.steepy.it","https:\u002F\u002Fgithub.com\u002Fsteepy-it","2026-10-27 16:48:08",{"source":8,"skillName":5,"commands":34,"prompt":44,"zipAvailable":24},[35,38,41],{"tool":36,"command":37},"npx","npx skills add steepy-it\u002Fsteepy-formula --skill review",{"tool":39,"command":40},"bunx","bunx skills add steepy-it\u002Fsteepy-formula --skill review",{"tool":42,"command":43},"pnpm","pnpm dlx skills add steepy-it\u002Fsteepy-formula --skill review","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fsteepy-it\u002Fsteepy-formula\u002Fblob\u002Fmain\u002Fskills\u002Freview\u002FSKILL.md",[],[47,54,61,68,75,82],{"skillNo":48,"name":49,"nameZh":6,"description":50,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":51,"categories":52,"keywords":53,"publishedAt":19},"01M7F91BV8VJM3DS4SA9ZMHMXB","web-check","Run the web surface checks.","https:\u002F\u002Fgithub.com\u002Fsteepy-it\u002Fsteepy-formula\u002Fblob\u002Fmain\u002Ftests\u002Ffixtures\u002Fgood-hub\u002F.claude\u002Fskills\u002Fweb-check\u002FSKILL.md",[],[],{"skillNo":55,"name":56,"nameZh":6,"description":57,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":58,"categories":59,"keywords":60,"publishedAt":19},"01M7F910RRMBN04529ZQZZW76Q","plan","Turn a local working spec into a step-by-step implementation plan. Each step names the owning surface and specialist agent and the surface test command; the plan is written into .formula\u002Fwork\u002Fplans\u002F as a gitignored work artifact.","https:\u002F\u002Fgithub.com\u002Fsteepy-it\u002Fsteepy-formula\u002Fblob\u002Fmain\u002Fskills\u002Fplan\u002FSKILL.md",[],[],{"skillNo":62,"name":63,"nameZh":6,"description":64,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":65,"categories":66,"keywords":67,"publishedAt":19},"01M7F90WV19MCWWJFESM2BNHNN","new-surface","Add a surface: render the graph sub-agent and promote Surface + verification Rule + Check into the graph.","https:\u002F\u002Fgithub.com\u002Fsteepy-it\u002Fsteepy-formula\u002Fblob\u002Fmain\u002Fskills\u002Fnew-surface\u002FSKILL.md",[],[],{"skillNo":69,"name":70,"nameZh":6,"description":71,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":72,"categories":73,"keywords":74,"publishedAt":19},"01M7F90QA1DXYPVFD08V5HFYY1","loop-engineer","Validate and authorize a gear-4 goal contract, invoke the packaged deterministic loop controller, report its result, and hand off terminal evidence to review.","https:\u002F\u002Fgithub.com\u002Fsteepy-it\u002Fsteepy-formula\u002Fblob\u002Fmain\u002Fskills\u002Floop-engineer\u002FSKILL.md",[],[],{"skillNo":76,"name":77,"nameZh":6,"description":78,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":79,"categories":80,"keywords":81,"publishedAt":19},"01M7F90EZHJZKYXEWTJ3BWKWRK","init","Graph-first init — gate on a provisioned Stable Knowledge Graph before writing anything, interview the governed structure, scaffold the disk-minimal harness assets, and land the initial graph through a promotion scenario. Never invents Rules from stack detection.","https:\u002F\u002Fgithub.com\u002Fsteepy-it\u002Fsteepy-formula\u002Fblob\u002Fmain\u002Fskills\u002Finit\u002FSKILL.md",[],[],{"skillNo":83,"name":84,"nameZh":6,"description":85,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":86,"categories":87,"keywords":88,"publishedAt":19},"01M7F908EYQDZ71GA3TQZ0613E","inception","Turn an idea and its starting materials into a minimal, verified, runnable application before the graph binding exists — reconnaissance, architecture decisions, verified versions, one human approval, bootstrap of one representative path, and final verification — then hand off to init and discovery.","https:\u002F\u002Fgithub.com\u002Fsteepy-it\u002Fsteepy-formula\u002Fblob\u002Fmain\u002Fskills\u002Finception\u002FSKILL.md",[],[]]