[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M79DMDSD4Z004V0CB01EYKT4":3,"skill-same:01M79DMDSD4Z004V0CB01EYKT4":46},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":23,"commitSha":24,"skillDirectory":25,"forksCount":26,"repositoryHtmlUrl":27,"repositoryHomepageUrl":28,"ownerHtmlUrl":29,"ownerBlogUrl":30,"ownerDescription":6,"updatedAt":31,"install":32,"resources":45},"01M79DMDSD4Z004V0CB01EYKT4","update",null,"Update this project's INSPIRE runtime to the installed plugin's version: report drift, re-materialize unchanged files, and refuse to clobber local edits. Use after \u002Fplugin update inspire reports a newer base.","Genomcore\u002Finspire","Genomcore","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F67845649?v=4","01M79DGNHB5DVE1Q8TC7F3K4FY","01M79DGNM14QF64G5NWT5X00SY",0,"MIT","https:\u002F\u002Fgithub.com\u002FGenomcore\u002Finspire\u002Fblob\u002Fmain\u002Fplugin\u002Fskills\u002Fupdate\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: update\ndescription: \"Update this project's INSPIRE runtime to the installed plugin's version: report drift, re-materialize unchanged files, and refuse to clobber local edits. Use after \u002Fplugin update inspire reports a newer base.\"\n---\n\n# \u002Finspire:update — re-materialize the runtime\n\nBrings a project's materialized runtime up to the plugin's current base. This is the\n**materialize-only** stage: it does not reconcile lessons. Lesson classification and\nskill rebuilding are specified in\n[`adr-runtime-lifecycle-and-lessons`](https:\u002F\u002Fgithub.com\u002FGenomcore\u002Finspire\u002Fblob\u002Fmain\u002Fdocs\u002Fadr\u002Fadr-runtime-lifecycle-and-lessons.md)\nD6, and building them is future work — nothing here approximates them in the meantime.\n\n## Preconditions\n\n1. **`.inspire.lock` is not required, and a missing one is not evidence of anything.**\n   `--mode plan` identifies the project's version from what is actually on disk,\n   scoring it against every shipped manifest; the lock is only a tie-break hint. A\n   pre-0.3 install may legitimately have no lock at all — the `install.sh`-era\n   installer wrote one only when both a manifest and `jq` were present — and those\n   are exactly the projects this skill exists to migrate. So never read a missing lock\n   as evidence that INSPIRE was never installed here, and never redirect the operator\n   to `\u002Finspire:init` over it: init refuses a pre-0.3 tree and points back here, so that\n   advice is a closed loop with no exit. If the tree genuinely holds no INSPIRE\n   runtime, `--mode plan` says so itself by exiting 1 — let it.\n2. `jq` and `yq` (Mike Farah v4) are on `PATH` — `materialize.sh` hard-requires both.\n3. A clean-enough working tree that the operator can review a diff. If `git status\n   --porcelain` is non-empty, warn and ask whether to continue.\n\n## Step 1 — Plan\n\nRun `--mode plan` (`--mode drift-check` is accepted as a deprecated alias). It is\n**read-only** — it writes nothing to the project, and that is asserted by the test\nsuite, not just claimed here:\n\n```bash\n\"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Fmaterialize.sh\" \\\n  --mode plan \\\n  --plugin-root \"${CLAUDE_PLUGIN_ROOT}\" \\\n  --project-root \"$(git rev-parse --show-toplevel)\"\n```\n\nIt detects the project's version (scored against every shipped manifest — never read\nblindly from the lock, which is only a tie-break hint), verifies the layout, enumerates\nthe hop chain the upgrade would run, classifies every file the project has against what\nthe target version ships, and emits:\n\n- **stdout** — JSON: `source_version`, `target_version`, `score`, `layout`, `chain` (the\n  hop versions that would run), `verdicts` (a count per verdict — `noop`\u002F`replace`\u002F`keep`\u002F\n  `ask`\u002F`create`\u002F`restore`\u002F`delete` — tallied over the *merged* hop-journal + classify\n  stream, the same union the stderr footer counts, never a classify-only tally), and `ask`\n  (the array of paths that actually need a decision, the same union). `--mode update`'s own\n  stdout carries the identical `ask` field: whatever a hop asked and no `--take-base`\u002F\n  `--take-mine` answered is still listed there once the run is done (Step 3).\n- **stderr** — the grouped human report (RUNTIME \u002F KNOWLEDGE BASE \u002F HARNESS \u002F PRODUCT \u002F\n  LEFT ALONE). In `--mode update` (never in `--mode plan`, which writes nothing at all)\n  the same report is also saved to `.inspire\u002Flast-upgrade.log`, overwritten on each run —\n  so what an upgrade did stays auditable after the session ends.\n\n**Relay the stderr report verbatim.** If it exits 1, relay the refusal and stop — do not\nattempt a workaround. Exit 1 means a precondition failed (an unidentifiable project, a\ndowngrade, a layout the tool cannot verify), and every one of those is a decision only the\noperator or a plugin update can resolve — this includes the pre-0.3, `install.sh`-era\nlayout, which is no longer a special case: it is simply the longest hop chain there is,\nand `--mode plan`\u002F`--mode update` migrate it like any other upgrade.\n\nVersion comparison and the never-downgrade guard are enforced in the script itself\n(`version_cmp`), in both `--mode plan` and `--mode update` — do not compare\n`inspire_version` to the plugin's `version` yourself, and do not second-guess the exit\ncode.\n\nA newer release's brand-new KB layer file (a new README, a new starter template) is\ndelivered automatically too: KB seeding runs additively in **both** `init` and `update` —\na path already on disk is left alone, only a path the project lacks is added. That\ndirection stays fully automatic. As of 0.7.0 a versioned hop can also go the other way and\n*retire* a KB seed file the project no longer needs — a proven-derivable one is deleted\nsilently, a diverged one lands in `ask` like any other conflict (Step 2). So \"nothing\nmanual\" no longer covers the whole picture: seeding is still silent, retirement sometimes\nis not.\n\n## Step 2 — Decide\n\nOnly the paths in `ask` need a decision — everything else `--mode plan` already resolved\n(`noop`\u002F`replace`\u002F`keep`\u002F`create`\u002F`restore`\u002F`delete` all apply without asking). If `ask`\nis empty, skip straight to Step 3.\n\nOtherwise, one `AskUserQuestion` covering the whole batch: **keep mine** (every `ask`\npath keeps the operator's file), **keep base** (every `ask` path takes the new version),\nor **go one by one**. **Retirement asks are excluded from both batch shortcuts.** Every\n`ask` path under `inspire_kb\u002F` is one — content classification never runs against\n`inspire_kb\u002F` at all (Rule 2; no shipped manifest lists a KB path), so a KB-rooted `ask`\nhas no source but a hop proposing to retire a seed file it could not prove derivable. A\nbatch \"keep base\" must never delete a KB index as a side effect of accepting ordinary\nskill\u002Fhook updates, and a batch \"keep mine\" must not wave a retirement through unseen\neither — the operator sees what they are declining, not just the safe outcome by accident.\nItemize every retirement ask instead, one by one, below.\n\n**Two more kinds of `ask` are excluded from both shortcuts, for that same reason.** A\n**collision** ask — a path the target version newly ships where the operator already has\na file — carries the file Rule 2 protects, so a batch \"keep base\" must not replace it as a\nside effect of accepting ordinary skill updates. A **split-case** ask — a `SKILL.md` with\n`create` rows beside it under the same skill's `references\u002F` — needs the two-copies and\nnew-home disclosure below, which neither batch answer can give. Itemize both one by one\ntoo; the batch shortcuts are for asks whose two outcomes an operator can weigh from the\npath alone.\n\nFor a content `ask` (anywhere outside `inspire_kb\u002F`), ask per file: **mine** \u002F **base** \u002F\n**merge**.\n- `mine` \u002F `base` resolve exactly like the batch choice, scoped to that file.\n- **`merge` is the one case where your judgement is legitimate.** Read both the\n  operator's file and the matching file under `${CLAUDE_PLUGIN_ROOT}\u002Fbase\u002F...`, write a\n  combined file that preserves the operator's intent, save it at the operator's path, then\n  resolve that path as `mine` in Step 3 (the file now on disk — the merged one — is the\n  one to keep). Nowhere else in this flow should you edit file content: everywhere else\n  the resolution is \"pick one side,\" and the script performs it.\n\n**The split case — an `ask` on a `SKILL.md` with `create` rows under the same skill's\n`references\u002F`.** Read that pairing off the grouped **stderr** report, which lists both\nkinds of row: the plan JSON's `verdicts` is a tally, not a path list, and update-mode\nstdout carries only `ask`, so neither one tells you which paths are the `create` side.\nThe pairing means the monolith's content has a new home: the operator's divergent edit\nmay belong in one of the new reference files rather than in `SKILL.md`. Say so when you\nask, and name the shape of keeping mine — a **kept monolith\nwith the new references beside it**, two live copies of the same guidance — so the choice\nis made with that in view.\n\nTaking the split is a `merge` whose combined content lands at a *different* path than the\nconflicted one, and **the sequencing is load-bearing: place content only after\n`--mode update` has finished.** That run reclassifies from disk before it writes, so a\nfile you create at a `create` path stops being a creation — it becomes an `ask` (\"new in\nthis release, and you already have a different file here\"), the unresolved default keeps\n*your* file, and the reference INSPIRE ships never lands at all. So: resolve the\n`SKILL.md` itself as `--take-base`, let the run install the new references, then write the\noperator's content into the appropriate one and tell them where it went. Never pass\n`--take-base` \u002F `--take-mine` for a `create` path: it matches no `ask` row, and the script\nwarns that the resolution \"matched nothing\" while changing nothing.\n\nFor a retirement `ask` (any path under `inspire_kb\u002F` — see the batch note above for the\ndiscriminator), the choice is **keep** \u002F **retire** only. There is no `merge`: there is no\nbase-side file to merge against either way — the index seeds left `base\u002Fkb\u002F` in this same\nrelease, and the ADR index (`01_adr\u002F_index.md` in the layered layout) never shipped one to begin with. **keep**\nleaves the file on disk untouched (the default if left unresolved); **retire** deletes it.\nHand both back the same way, as `--take-mine` \u002F `--take-base` in Step 3.\n\nLeaving a path unresolved is a valid choice, not an error: an unresolved `ask` defaults to\n**keeping the operator's file** — `materialize.sh` does this itself if you pass nothing\n(see Rule 2).\n\n## Step 3 — Apply\n\nRe-run with `--mode update`, plus one `--take-base \u003Cpath>` or `--take-mine \u003Cpath>` per\nresolved `ask` path — paths come straight back from the plan's `ask` array\n(project-relative, no leading `\u002F`, no `..`; `--skip` is accepted as a deprecated alias for\n`--take-mine`):\n\n```bash\n\"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Fmaterialize.sh\" \\\n  --mode update \\\n  --plugin-root \"${CLAUDE_PLUGIN_ROOT}\" \\\n  --project-root \"$(git rev-parse --show-toplevel)\" \\\n  --source-root \"\u003Cfrom stack.md>\" --prototype-root \"\u003Cfrom stack.md>\" \\\n  [--take-base \u003Cpath>]... [--take-mine \u003Cpath>]...\n```\n\nThe two flags mean different things depending on which kind of `ask` they resolve. For a\ncontent `ask`, `--take-base \u003Cpath>` replaces the file with the new base version and\n`--take-mine \u003Cpath>` keeps the operator's. For a retirement `ask`, there is no base file to\ntake, so `--take-base \u003Cpath>` retires it instead — it deletes the file — and\n`--take-mine \u003Cpath>` keeps it (`hops\u002F0.7.0.sh`'s `_h7_settle`: an explicit `--take-mine`\noutranks everything, an explicit `--take-base` outranks the verdict and calls the hop's own\n`rm`, and anything left unmatched falls through to `hop_ask`). Either kind left unresolved\ndefaults to keeping the operator's file — `_apply_resolutions` for a content `ask`, the\nsame fallthrough for a retirement one — doing nothing is always the safe choice.\n\nRead `source_root` and `prototype_root` from `inspire_kb\u002F00_bootstrap\u002Fstack.md` — update\nnever re-asks; those were settled at init and changing them is a Shape change owned by\n`\u002Finspire-bootstrap stack`.\n\nRun with `--dry-run` first if the operator wants to see the exact file list before\nwriting — the decision logic is identical in both, only the write is skipped, so the\npreview cannot diverge from what happens next.\n\n**On exit 2, relay the failure and tell the operator to re-run — that is now the correct\nrecovery, not a hopeful guess:**\n- a hop that already moved a path is a silent no-op the next time (a missing source is\n  treated as \"already done,\" never an error);\n- a failed layout hop stops the run *before* `.inspire.lock` is rewritten, so the\n  version still reads the old one. That is deliberate and it is what makes the re-run\n  work: a lock claiming a migration that did not happen would make the next run score\n  the project as already migrated and never retry the hop. Tell the operator the\n  version was left as it was — it is not a second failure to explain away;\n- the content merge converges: a file already replaced with the new base version hashes\n  identical to that base version, so a second run reads it as already matching and leaves\n  it alone rather than reapplying or re-flagging it.\n\nThere is nothing to hand-patch and no different flags to pass. Re-running\n`\u002Finspire:update` with the same command is the whole recovery.\n\n## Step 4 — Report\n\nRelay the final report and every entry in the JSON's `warnings` array verbatim — these\nare conditions the run could not fix on the operator's behalf (a `.gitignore` excluding\nthe runtime is the common one, which silently keeps the update out of the commit\nentirely). Skill edits need no session restart: `.claude\u002Fskills\u002F` already existed at\nsession start, so edits to files inside it load live. **The agents class is the\ncaveat.** When this run created `.claude\u002Fagents\u002F` for the first time — every project\ncoming from 0.7 or earlier — say plainly that a directory which appears mid-session may\nnot be discovered until the next one, so an agent the project adds there can need a fresh session\nbefore `\u002Fagents` lists it. Then remind the operator to review the git diff and commit.\n\n**A layered KB.** When a warning says `inspire_kb\u002F` is still on the layered layout\n(`01_adr` … `05_screens`), say plainly that the migration is **mandatory**: from 0.15.0\nthe skills and validators read the specification only from `inspire_kb\u002F01_spec\u002F`, and\nthis update moved nothing in the KB, because it never does. The next step is\n`\u002Finspire:migrate`, after the update is committed. Skip Steps 5, 8 and 9 — they act on\n`01_spec\u002F` — and run them after the migration instead.\n\n**Wikilinks.** When a warning counts wikilinks in `inspire_kb\u002F`, say that from 0.17.0\na KB link is a relative Markdown link to the target file, that this update changed no\nlink, and that `links-resolve` reports each wikilink until it is converted. The next\nstep is `\u002Finspire:migrate`, after the update is committed: on a KB already on the\nconcept layout it converts the links and moves nothing.\n\n## Step 5 — Backfill the UI specification\n\nA release can add parts of the screen layer an older vault never wrote — the token\nfile, the supported viewports, the `layouts\u002F` tier, a screen's `## Render` table.\nMaterializing ships the rules and the validators for them; it never writes product\ncontent. This step closes the gap through the skills that own those files, so the\noperator ends the update with a specification the new release can read, not a list\nof warnings.\n\n1. **Detect.** Run, from the project root:\n\n   ```bash\n   .inspire\u002Fbin\u002Fdesign-system-coherence.sh inspire_kb\u002F01_spec\n   .inspire\u002Fbin\u002Fscreen-presentation.sh inspire_kb\u002F01_spec\n   ```\n\n   No finding from either, or no `inspire_kb\u002F01_spec\u002F_design-system\u002Fdesign-system.md` at\n   all (a project with no screen layer yet, or one still on the layered layout, which\n   `\u002Finspire:migrate` moves first): skip to Step 6. Findings that only count\n   `generated` or `undecided` values are not gaps — they are decisions still open,\n   and this step leaves them to the operator.\n2. **Ask once**, in one `AskUserQuestion` with two questions:\n   - **The reference** the gaps are measured from, per\n     `.claude\u002Fskills\u002F_references\u002Fui-reference.md` § Which reference. Options: **the\n     prototype** (recommended, offered first, when `prototype_root` in\n     `inspire_kb\u002F00_bootstrap\u002Fstack.md` is not `none` and holds a project), **another\n     reference** (the operator types a URL, a path or a description in \"Other\"),\n     **no reference — generate by common sense**, and **skip for now**.\n   - **The narrowest supported width**, only when `design-system.md` carries no\n     `## Supported viewports`: laptop (1024px) and up, tablet (768px) and up, phone\n     (390px) and up. A product decision, never inferred.\n3. **Run the two backfills in order**, passing the answers: first\n   `\u002Finspire-bootstrap design-system backfill` (the token file and the three\n   sections — every later row cites them), then `\u002Finspire-screens backfill` (the\n   layouts, the `**Layout:**` lines, the `## Render` tables, the component entries).\n   Each follows its own reference file; this skill writes none of it.\n4. **Relay both reports.** When either generated a value, say so plainly and\n   first: how many values, in which files, that nobody has decided them, and that\n   `\u002Finspire-bootstrap design-system` confirms them one by one. A generated value\n   is a proposal, and the operator must know which ones they are before they\n   commit.\n\n**Skip for now** leaves the vault exactly as Step 3 left it, and says so: the two\nvalidators keep reporting every gap as a warning at every lifecycle, never a block,\nand running the two backfills later closes them the same way.\n\n## Step 6 — Nudge\n\nFor each kept **skill** edit (an `ask` resolved to `mine`, or a plain `keep` verdict —\nnever a kept `inspire_kb\u002F` retirement ask: that is the operator declining to retire a seed\nfile, not a skill edit, and it earns no lesson-capture suggestion): tell the operator the\ncustomization belongs in `\u002Finspire-lesson note`. Say plainly that\nnothing is remembered by this flow — the same file will be classified as `ask` (or\n`keep`) again on the next upgrade, and that recurrence is deliberate: it is what stands in\nfor the lesson-reconciliation half (D6) that does not exist yet.\n\nFor each kept **validator** edit: there is no capture mechanism for it — a lesson has no\ncategory for bash. Say it belongs upstream: open an issue or a PR against INSPIRE.\n\n## Step 7 — Trust report\n\nIf `.inspire\u002Fbin\u002Ftrust.sh` exists, offer — never run unprompted — `.inspire\u002Fbin\u002Ftrust.sh\nreport`. What its groups mean is explained at\n`.claude\u002Fskills\u002F_references\u002Ftrust-stamps.md`; point there rather than restating it.\n\n## Step 8 — Single-file ADRs\n\nIf `.inspire\u002Fbin\u002Fadr-split.sh` exists, run `.inspire\u002Fbin\u002Fadr-split.sh --list`. It writes\nnothing. When it lists files, say how many ADRs are still single files, and offer —\nnever run unprompted — `\u002Finspire-adr migrate`, which splits each one into a decision and\na rationale. The KB is the operator's (`adr-upgrade-path` D10), so the upgrade itself\nnever splits them.\n\n## Step 9 — Heading keys\n\nIf `.inspire\u002Fbin\u002Fheading-keys.sh` exists, run `.inspire\u002Fbin\u002Fheading-keys.sh --list`. It\nwrites nothing. When it lists files, say how many KB artifacts still name their\nsections by the English name (`## Fields`) and **recommend** rewriting them into the\n`{ID}: {label}` shape (`_references\u002Foutput-language.md`): the KB then has one heading\nshape, the one every skill now writes, and its headings read in `output_language`. Present it as the\nrecommended option, not as one of two equal choices. The KB is the operator's\n(`adr-upgrade-path` D10), so rewrite only on their yes:\n\n1. Run `.inspire\u002Fbin\u002Fheading-keys.sh`. Each rewritten heading carries its section id,\n   and keeps its label or takes the English name as one: `## Fields` becomes\n   `## ENT4: Fields`.\n2. If `output_language` in `inspire_kb\u002F00_bootstrap\u002Fproject.md` is not English, translate\n   every English label the script wrote, in the files it reported. Change only the text\n   after `: `. The id stays verbatim.\n3. Run `.inspire\u002Fbin\u002Freview.sh inspire_kb` and report anything it finds.\n\nThe validators still read the English name as its section, so a project that declines\nkeeps validating.\n\n## What this skill is, and what is missing from it\n\nUpdate has two halves, and only one of them can be a script:\n\n| Half | Nature | Where it lives | Status |\n|---|---|---|---|\n| Version compare, manifest-driven content classification, the hop chain, copying, settings merge, lock rewrite | **deterministic** | `materialize.sh` | **complete** |\n| Re-deriving each skill as `new base + surviving lessons` | **semantic** — a judgement about meaning, not a file operation | this skill | **not built** |\n\nThe second half is `adr-runtime-lifecycle-and-lessons` D6: classify every lesson against the\nnew base (absorbed \u002F untouched \u002F partial \u002F contradicted), then rebuild each affected skill\nfrom `theirs + apply(surviving lessons)`. It cannot be scripted — it is a semantic merge over\nprose, which is exactly why that ADR rejected textual three-way merge for skills.\n\nUntil it exists, this skill's honest behaviour is to **ask rather than approximate**: a\nconflicted file (an `ask` path) is never silently reconciled. Step 2 puts every one of them\nto the operator — mine \u002F base \u002F merge — and `merge` is the only point in this whole flow\nwhere the agent's own judgement about file content is legitimate; everywhere else the\nresolution is \"pick one side,\" and the script performs it. The deterministic half is\ngenuinely useful on its own — it brings validators, hooks and untouched skills forward\nsafely, migrating even a pre-0.3 project in the process — and it is the substrate the\nsemantic half will call.\n\n**Refining that second half is future work**, specified by D6 of the ADR linked at the\ntop of this file and tracked by the operator rather than promised here. Nothing in this\nskill waits on it: everything above is what the current release actually does.\n\n## Rules\n\n1. **All runtime file operations go through `materialize.sh`.** No copying, hashing\n   or settings editing in prose. The one KB write this flow causes is Step 5's, and\n   it goes through the skills that own those files, never through this one. Content classification is `--mode plan` (`--mode drift-check` is\n   accepted as a deprecated alias).\n2. **Never silently overwrite an operator's file.** An `ask` row left unresolved defaults\n   to keeping the operator's version — doing nothing is how work survives. Only\n   `--take-base` moves a conflicted path to the new base version; only a `merge` you write\n   yourself combines the two.\n\n   **The guarantee's exact scope: a path the target version does not ship.**\n   Classification compares disk against the SOURCE version's manifest\n   (`plugin\u002Fmanifests\u002F\u003Cversion>.json`) and the current `base\u002F` — not against any lock;\n   the lock's old `files` map is gone. A file the *project* authored at a path no\n   manifest lists and the new base does not carry — e.g.\n   `.claude\u002Fskills\u002Finspire-code-nestjs\u002Freferences\u002Fmystack.md` — always classifies as `keep`\n   (\"yours — INSPIRE never shipped this\") and is never touched or deleted, even though it\n   sits inside a skill directory INSPIRE owns. This is the fix for the pre-0.3 behaviour\n   (`install.sh` used to `rm -rf` a whole owned entry, destroying exactly these files) —\n   already shipped, not a roadmap item.\n\n   **Where the target version does ship that path, the guarantee is narrower — and a\n   skill's `references\u002F` is exactly such a place.** INSPIRE ships reference files under\n   `inspire-*\u002Freferences\u002F`, and 0.7.0 ships a batch of new ones, so a file the project\n   authored there can **collide** with a newly shipped one. It then classifies `ask` —\n   \"new in this release, and you already have a different file here\" — and an unresolved\n   `ask` keeps the operator's file, exactly as everywhere else. It is still never silently\n   touched or deleted; what changes is that the operator gets a question where before there was\n   nothing to ask. Project content still belongs in the KB, in a `98_lessons` node, or in\n   the project's `CLAUDE.md` — not inside an `inspire-*` directory — but a file left there\n   by mistake survives rather than being silently destroyed.\n\n   `inspire_kb\u002F` is out of scope for *classification* — content classification never runs\n   against it at all, because no shipped manifest lists a KB path in the first place. KB\n   seeding is a separate, purely additive path that runs in both `init` and `update` (see\n   Step 1): a path already on disk is never replaced, only a missing one is added.\n   `--mode update` is **not**, though, flatly non-destructive under `inspire_kb\u002F` as of\n   0.7.0: a versioned hop may retire a KB seed file the project no longer needs to keep —\n   a proven-derivable one is deleted silently, and any divergence from the derivation lands\n   in `ask` rather than a silent delete (Step 2's retirement branch). Two further exceptions\n   are purely additive and never clobber: `seed_design_system` creates\n   `01_spec\u002F_design-system\u002Fdesign-system.md` (never in a KB still on the layered layout), and `seed_claude_md` creates `CLAUDE.md`, **only if\n   absent**.\n3. **No writes happen before Step 2's decisions are gathered.** `--mode plan` itself\n   writes nothing, whatever the operator decides — that is asserted by the test suite, not\n   just documented here.\n4. **Never downgrade.** A plugin older than the project's detected version is an error, not\n   an update — enforced by `version_cmp` in the script, in both `--mode plan` and\n   `--mode update`.\n5. **No lesson reconciliation here.** Classification, archiving and skill rebuilding are\n   D6, and unbuilt. This skill does not read `98_lessons\u002F`. Do not approximate it — telling\n   the operator an edit was kept is truthful; silently merging it is not.\n6. **Do not re-ask the product roots.** Read them from `stack.md`; changing them is a Shape\n   change owned by `\u002Finspire-bootstrap stack`.\n","# \u002Finspire:update — re-materialize the runtime\n\nBrings a project's materialized runtime up to the plugin's current base. This is the\n**materialize-only** stage: it does not reconcile lessons. Lesson classification and\nskill rebuilding are specified in\n[`adr-runtime-lifecycle-and-lessons`](https:\u002F\u002Fgithub.com\u002FGenomcore\u002Finspire\u002Fblob\u002Fmain\u002Fdocs\u002Fadr\u002Fadr-runtime-lifecycle-and-lessons.md)\nD6, and building them is future work — nothing here approximates them in the meantime.\n\n## Preconditions\n\n1. **`.inspire.lock` is not required, and a missing one is not evidence of anything.**\n   `--mode plan` identifies the project's version from what is actually on disk,\n   scoring it against every shipped manifest; the lock is only a tie-break hint. A\n   pre-0.3 install may legitimately have no lock at all — the `install.sh`-era\n   installer wrote one only when both a manifest and `jq` were present — and those\n   are exactly the projects this skill exists to migrate. So never read a missing lock\n   as evidence that INSPIRE was never installed here, and never redirect the operator\n   to `\u002Finspire:init` over it: init refuses a pre-0.3 tree and points back here, so that\n   advice is a closed loop with no exit. If the tree genuinely holds no INSPIRE\n   runtime, `--mode plan` says so itself by exiting 1 — let it.\n2. `jq` and `yq` (Mike Farah v4) are on `PATH` — `materialize.sh` hard-requires both.\n3. A clean-enough working tree that the operator can review a diff. If `git status\n   --porcelain` is non-empty, warn and ask whether to continue.\n\n## Step 1 — Plan\n\nRun `--mode plan` (`--mode drift-check` is accepted as a deprecated alias). It is\n**read-only** — it writes nothing to the project, and that is asserted by the test\nsuite, not just claimed here:\n\n```bash\n\"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Fmaterialize.sh\" \\\n  --mode plan \\\n  --plugin-root \"${CLAUDE_PLUGIN_ROOT}\" \\\n  --project-root \"$(git rev-parse --show-toplevel)\"\n```\n\nIt detects the project's version (scored against every shipped manifest — never read\nblindly from the lock, which is only a tie-break hint), verifies the layout, enumerates\nthe hop chain the upgrade would run, classifies every file the project has against what\nthe target version ships, and emits:\n\n- **stdout** — JSON: `source_version`, `target_version`, `score`, `layout`, `chain` (the\n  hop versions that would run), `verdicts` (a count per verdict — `noop`\u002F`replace`\u002F`keep`\u002F\n  `ask`\u002F`create`\u002F`restore`\u002F`delete` — tallied over the *merged* hop-journal + classify\n  stream, the same union the stderr footer counts, never a classify-only tally), and `ask`\n  (the array of paths that actually need a decision, the same union). `--mode update`'s own\n  stdout carries the identical `ask` field: whatever a hop asked and no `--take-base`\u002F\n  `--take-mine` answered is still listed there once the run is done (Step 3).\n- **stderr** — the grouped human report (RUNTIME \u002F KNOWLEDGE BASE \u002F HARNESS \u002F PRODUCT \u002F\n  LEFT ALONE). In `--mode update` (never in `--mode plan`, which writes nothing at all)\n  the same report is also saved to `.inspire\u002Flast-upgrade.log`, overwritten on each run —\n  so what an upgrade did stays auditable after the session ends.\n\n**Relay the stderr report verbatim.** If it exits 1, relay the refusal and stop — do not\nattempt a workaround. Exit 1 means a precondition failed (an unidentifiable project, a\ndowngrade, a layout the tool cannot verify), and every one of those is a decision only the\noperator or a plugin update can resolve — this includes the pre-0.3, `install.sh`-era\nlayout, which is no longer a special case: it is simply the longest hop chain there is,\nand `--mode plan`\u002F`--mode update` migrate it like any other upgrade.\n\nVersion comparison and the never-downgrade guard are enforced in the script itself\n(`version_cmp`), in both `--mode plan` and `--mode update` — do not compare\n`inspire_version` to the plugin's `version` yourself, and do not second-guess the exit\ncode.\n\nA newer release's brand-new KB layer file (a new README, a new starter template) is\ndelivered automatically too: KB seeding runs additively in **both** `init` and `update` —\na path already on disk is left alone, only a path the project lacks is added. That\ndirection stays fully automatic. As of 0.7.0 a versioned hop can also go the other way and\n*retire* a KB seed file the project no longer needs — a proven-derivable one is deleted\nsilently, a diverged one lands in `ask` like any other conflict (Step 2). So \"nothing\nmanual\" no longer covers the whole picture: seeding is still silent, retirement sometimes\nis not.\n\n## Step 2 — Decide\n\nOnly the paths in `ask` need a decision — everything else `--mode plan` already resolved\n(`noop`\u002F`replace`\u002F`keep`\u002F`create`\u002F`restore`\u002F`delete` all apply without asking). If `ask`\nis empty, skip straight to Step 3.\n\nOtherwise, one `AskUserQuestion` covering the whole batch: **keep mine** (every `ask`\npath keeps the operator's file), **keep base** (every `ask` path takes the new version),\nor **go one by one**. **Retirement asks are excluded from both batch shortcuts.** Every\n`ask` path under `inspire_kb\u002F` is one — content classification never runs against\n`inspire_kb\u002F` at all (Rule 2; no shipped manifest lists a KB path), so a KB-rooted `ask`\nhas no source but a hop proposing to retire a seed file it could not prove derivable. A\nbatch \"keep base\" must never delete a KB index as a side effect of accepting ordinary\nskill\u002Fhook updates, and a batch \"keep mine\" must not wave a retirement through unseen\neither — the operator sees what they are declining, not just the safe outcome by accident.\nItemize every retirement ask instead, one by one, below.\n\n**Two more kinds of `ask` are excluded from both shortcuts, for that same reason.** A\n**collision** ask — a path the target version newly ships where the operator already has\na file — carries the file Rule 2 protects, so a batch \"keep base\" must not replace it as a\nside effect of accepting ordinary skill updates. A **split-case** ask — a `SKILL.md` with\n`create` rows beside it under the same skill's `references\u002F` — needs the two-copies and\nnew-home disclosure below, which neither batch answer can give. Itemize both one by one\ntoo; the batch shortcuts are for asks whose two outcomes an operator can weigh from the\npath alone.\n\nFor a content `ask` (anywhere outside `inspire_kb\u002F`), ask per file: **mine** \u002F **base** \u002F\n**merge**.\n- `mine` \u002F `base` resolve exactly like the batch choice, scoped to that file.\n- **`merge` is the one case where your judgement is legitimate.** Read both the\n  operator's file and the matching file under `${CLAUDE_PLUGIN_ROOT}\u002Fbase\u002F...`, write a\n  combined file that preserves the operator's intent, save it at the operator's path, then\n  resolve that path as `mine` in Step 3 (the file now on disk — the merged one — is the\n  one to keep). Nowhere else in this flow should you edit file content: everywhere else\n  the resolution is \"pick one side,\" and the script performs it.\n\n**The split case — an `ask` on a `SKILL.md` with `create` rows under the same skill's\n`references\u002F`.** Read that pairing off the grouped **stderr** report, which lists both\nkinds of row: the plan JSON's `verdicts` is a tally, not a path list, and update-mode\nstdout carries only `ask`, so neither one tells you which paths are the `create` side.\nThe pairing means the monolith's content has a new home: the operator's divergent edit\nmay belong in one of the new reference files rather than in `SKILL.md`. Say so when you\nask, and name the shape of keeping mine — a **kept monolith\nwith the new references beside it**, two live copies of the same guidance — so the choice\nis made with that in view.\n\nTaking the split is a `merge` whose combined content lands at a *different* path than the\nconflicted one, and **the sequencing is load-bearing: place content only after\n`--mode update` has finished.** That run reclassifies from disk before it writes, so a\nfile you create at a `create` path stops being a creation — it becomes an `ask` (\"new in\nthis release, and you already have a different file here\"), the unresolved default keeps\n*your* file, and the reference INSPIRE ships never lands at all. So: resolve the\n`SKILL.md` itself as `--take-base`, let the run install the new references, then write the\noperator's content into the appropriate one and tell them where it went. Never pass\n`--take-base` \u002F `--take-mine` for a `create` path: it matches no `ask` row, and the script\nwarns that the resolution \"matched nothing\" while changing nothing.\n\nFor a retirement `ask` (any path under `inspire_kb\u002F` — see the batch note above for the\ndiscriminator), the choice is **keep** \u002F **retire** only. There is no `merge`: there is no\nbase-side file to merge against either way — the index seeds left `base\u002Fkb\u002F` in this same\nrelease, and the ADR index (`01_adr\u002F_index.md` in the layered layout) never shipped one to begin with. **keep**\nleaves the file on disk untouched (the default if left unresolved); **retire** deletes it.\nHand both back the same way, as `--take-mine` \u002F `--take-base` in Step 3.\n\nLeaving a path unresolved is a valid choice, not an error: an unresolved `ask` defaults to\n**keeping the operator's file** — `materialize.sh` does this itself if you pass nothing\n(see Rule 2).\n\n## Step 3 — Apply\n\nRe-run with `--mode update`, plus one `--take-base \u003Cpath>` or `--take-mine \u003Cpath>` per\nresolved `ask` path — paths come straight back from the plan's `ask` array\n(project-relative, no leading `\u002F`, no `..`; `--skip` is accepted as a deprecated alias for\n`--take-mine`):\n\n```bash\n\"${CLAUDE_PLUGIN_ROOT}\u002Fscripts\u002Fmaterialize.sh\" \\\n  --mode update \\\n  --plugin-root \"${CLAUDE_PLUGIN_ROOT}\" \\\n  --project-root \"$(git rev-parse --show-toplevel)\" \\\n  --source-root \"\u003Cfrom stack.md>\" --prototype-root \"\u003Cfrom stack.md>\" \\\n  [--take-base \u003Cpath>]... [--take-mine \u003Cpath>]...\n```\n\nThe two flags mean different things depending on which kind of `ask` they resolve. For a\ncontent `ask`, `--take-base \u003Cpath>` replaces the file with the new base version and\n`--take-mine \u003Cpath>` keeps the operator's. For a retirement `ask`, there is no base file to\ntake, so `--take-base \u003Cpath>` retires it instead — it deletes the file — and\n`--take-mine \u003Cpath>` keeps it (`hops\u002F0.7.0.sh`'s `_h7_settle`: an explicit `--take-mine`\noutranks everything, an explicit `--take-base` outranks the verdict and calls the hop's own\n`rm`, and anything left unmatched falls through to `hop_ask`). Either kind left unresolved\ndefaults to keeping the operator's file — `_apply_resolutions` for a content `ask`, the\nsame fallthrough for a retirement one — doing nothing is always the safe choice.\n\nRead `source_root` and `prototype_root` from `inspire_kb\u002F00_bootstrap\u002Fstack.md` — update\nnever re-asks; those were settled at init and changing them is a Shape change owned by\n`\u002Finspire-bootstrap stack`.\n\nRun with `--dry-run` first if the operator wants to see the exact file list before\nwriting — the decision logic is identical in both, only the write is skipped, so the\npreview cannot diverge from what happens next.\n\n**On exit 2, relay the failure and tell the operator to re-run — that is now the correct\nrecovery, not a hopeful guess:**\n- a hop that already moved a path is a silent no-op the next time (a missing source is\n  treated as \"already done,\" never an error);\n- a failed layout hop stops the run *before* `.inspire.lock` is rewritten, so the\n  version still reads the old one. That is deliberate and it is what makes the re-run\n  work: a lock claiming a migration that did not happen would make the next run score\n  the project as already migrated and never retry the hop. Tell the operator the\n  version was left as it was — it is not a second failure to explain away;\n- the content merge converges: a file already replaced with the new base version hashes\n  identical to that base version, so a second run reads it as already matching and leaves\n  it alone rather than reapplying or re-flagging it.\n\nThere is nothing to hand-patch and no different flags to pass. Re-running\n`\u002Finspire:update` with the same command is the whole recovery.\n\n## Step 4 — Report\n\nRelay the final report and every entry in the JSON's `warnings` array verbatim — these\nare conditions the run could not fix on the operator's behalf (a `.gitignore` excluding\nthe runtime is the common one, which silently keeps the update out of the commit\nentirely). Skill edits need no session restart: `.claude\u002Fskills\u002F` already existed at\nsession start, so edits to files inside it load live. **The agents class is the\ncaveat.** When this run created `.claude\u002Fagents\u002F` for the first time — every project\ncoming from 0.7 or earlier — say plainly that a directory which appears mid-session may\nnot be discovered until the next one, so an agent the project adds there can need a fresh session\nbefore `\u002Fagents` lists it. Then remind the operator to review the git diff and commit.\n\n**A layered KB.** When a warning says `inspire_kb\u002F` is still on the layered layout\n(`01_adr` … `05_screens`), say plainly that the migration is **mandatory**: from 0.15.0\nthe skills and validators read the specification only from `inspire_kb\u002F01_spec\u002F`, and\nthis update moved nothing in the KB, because it never does. The next step is\n`\u002Finspire:migrate`, after the update is committed. Skip Steps 5, 8 and 9 — they act on\n`01_spec\u002F` — and run them after the migration instead.\n\n**Wikilinks.** When a warning counts wikilinks in `inspire_kb\u002F`, say that from 0.17.0\na KB link is a relative Markdown link to the target file, that this update changed no\nlink, and that `links-resolve` reports each wikilink until it is converted. The next\nstep is `\u002Finspire:migrate`, after the update is committed: on a KB already on the\nconcept layout it converts the links and moves nothing.\n\n## Step 5 — Backfill the UI specification\n\nA release can add parts of the screen layer an older vault never wrote — the token\nfile, the supported viewports, the `layouts\u002F` tier, a screen's `## Render` table.\nMaterializing ships the rules and the validators for them; it never writes product\ncontent. This step closes the gap through the skills that own those files, so the\noperator ends the update with a specification the new release can read, not a list\nof warnings.\n\n1. **Detect.** Run, from the project root:\n\n   ```bash\n   .inspire\u002Fbin\u002Fdesign-system-coherence.sh inspire_kb\u002F01_spec\n   .inspire\u002Fbin\u002Fscreen-presentation.sh inspire_kb\u002F01_spec\n   ```\n\n   No finding from either, or no `inspire_kb\u002F01_spec\u002F_design-system\u002Fdesign-system.md` at\n   all (a project with no screen layer yet, or one still on the layered layout, which\n   `\u002Finspire:migrate` moves first): skip to Step 6. Findings that only count\n   `generated` or `undecided` values are not gaps — they are decisions still open,\n   and this step leaves them to the operator.\n2. **Ask once**, in one `AskUserQuestion` with two questions:\n   - **The reference** the gaps are measured from, per\n     `.claude\u002Fskills\u002F_references\u002Fui-reference.md` § Which reference. Options: **the\n     prototype** (recommended, offered first, when `prototype_root` in\n     `inspire_kb\u002F00_bootstrap\u002Fstack.md` is not `none` and holds a project), **another\n     reference** (the operator types a URL, a path or a description in \"Other\"),\n     **no reference — generate by common sense**, and **skip for now**.\n   - **The narrowest supported width**, only when `design-system.md` carries no\n     `## Supported viewports`: laptop (1024px) and up, tablet (768px) and up, phone\n     (390px) and up. A product decision, never inferred.\n3. **Run the two backfills in order**, passing the answers: first\n   `\u002Finspire-bootstrap design-system backfill` (the token file and the three\n   sections — every later row cites them), then `\u002Finspire-screens backfill` (the\n   layouts, the `**Layout:**` lines, the `## Render` tables, the component entries).\n   Each follows its own reference file; this skill writes none of it.\n4. **Relay both reports.** When either generated a value, say so plainly and\n   first: how many values, in which files, that nobody has decided them, and that\n   `\u002Finspire-bootstrap design-system` confirms them one by one. A generated value\n   is a proposal, and the operator must know which ones they are before they\n   commit.\n\n**Skip for now** leaves the vault exactly as Step 3 left it, and says so: the two\nvalidators keep reporting every gap as a warning at every lifecycle, never a block,\nand running the two backfills later closes them the same way.\n\n## Step 6 — Nudge\n\nFor each kept **skill** edit (an `ask` resolved to `mine`, or a plain `keep` verdict —\nnever a kept `inspire_kb\u002F` retirement ask: that is the operator declining to retire a seed\nfile, not a skill edit, and it earns no lesson-capture suggestion): tell the operator the\ncustomization belongs in `\u002Finspire-lesson note`. Say plainly that\nnothing is remembered by this flow — the same file will be classified as `ask` (or\n`keep`) again on the next upgrade, and that recurrence is deliberate: it is what stands in\nfor the lesson-reconciliation half (D6) that does not exist yet.\n\nFor each kept **validator** edit: there is no capture mechanism for it — a lesson has no\ncategory for bash. Say it belongs upstream: open an issue or a PR against INSPIRE.\n\n## Step 7 — Trust report\n\nIf `.inspire\u002Fbin\u002Ftrust.sh` exists, offer — never run unprompted — `.inspire\u002Fbin\u002Ftrust.sh\nreport`. What its groups mean is explained at\n`.claude\u002Fskills\u002F_references\u002Ftrust-stamps.md`; point there rather than restating it.\n\n## Step 8 — Single-file ADRs\n\nIf `.inspire\u002Fbin\u002Fadr-split.sh` exists, run `.inspire\u002Fbin\u002Fadr-split.sh --list`. It writes\nnothing. When it lists files, say how many ADRs are still single files, and offer —\nnever run unprompted — `\u002Finspire-adr migrate`, which splits each one into a decision and\na rationale. The KB is the operator's (`adr-upgrade-path` D10), so the upgrade itself\nnever splits them.\n\n## Step 9 — Heading keys\n\nIf `.inspire\u002Fbin\u002Fheading-keys.sh` exists, run `.inspire\u002Fbin\u002Fheading-keys.sh --list`. It\nwrites nothing. When it lists files, say how many KB artifacts still name their\nsections by the English name (`## Fields`) and **recommend** rewriting them into the\n`{ID}: {label}` shape (`_references\u002Foutput-language.md`): the KB then has one heading\nshape, the one every skill now writes, and its headings read in `output_language`. Present it as the\nrecommended option, not as one of two equal choices. The KB is the operator's\n(`adr-upgrade-path` D10), so rewrite only on their yes:\n\n1. Run `.inspire\u002Fbin\u002Fheading-keys.sh`. Each rewritten heading carries its section id,\n   and keeps its label or takes the English name as one: `## Fields` becomes\n   `## ENT4: Fields`.\n2. If `output_language` in `inspire_kb\u002F00_bootstrap\u002Fproject.md` is not English, translate\n   every English label the script wrote, in the files it reported. Change only the text\n   after `: `. The id stays verbatim.\n3. Run `.inspire\u002Fbin\u002Freview.sh inspire_kb` and report anything it finds.\n\nThe validators still read the English name as its section, so a project that declines\nkeeps validating.\n\n## What this skill is, and what is missing from it\n\nUpdate has two halves, and only one of them can be a script:\n\n| Half | Nature | Where it lives | Status |\n|---|---|---|---|\n| Version compare, manifest-driven content classification, the hop chain, copying, settings merge, lock rewrite | **deterministic** | `materialize.sh` | **complete** |\n| Re-deriving each skill as `new base + surviving lessons` | **semantic** — a judgement about meaning, not a file operation | this skill | **not built** |\n\nThe second half is `adr-runtime-lifecycle-and-lessons` D6: classify every lesson against the\nnew base (absorbed \u002F untouched \u002F partial \u002F contradicted), then rebuild each affected skill\nfrom `theirs + apply(surviving lessons)`. It cannot be scripted — it is a semantic merge over\nprose, which is exactly why that ADR rejected textual three-way merge for skills.\n\nUntil it exists, this skill's honest behaviour is to **ask rather than approximate**: a\nconflicted file (an `ask` path) is never silently reconciled. Step 2 puts every one of them\nto the operator — mine \u002F base \u002F merge — and `merge` is the only point in this whole flow\nwhere the agent's own judgement about file content is legitimate; everywhere else the\nresolution is \"pick one side,\" and the script performs it. The deterministic half is\ngenuinely useful on its own — it brings validators, hooks and untouched skills forward\nsafely, migrating even a pre-0.3 project in the process — and it is the substrate the\nsemantic half will call.\n\n**Refining that second half is future work**, specified by D6 of the ADR linked at the\ntop of this file and tracked by the operator rather than promised here. Nothing in this\nskill waits on it: everything above is what the current release actually does.\n\n## Rules\n\n1. **All runtime file operations go through `materialize.sh`.** No copying, hashing\n   or settings editing in prose. The one KB write this flow causes is Step 5's, and\n   it goes through the skills that own those files, never through this one. Content classification is `--mode plan` (`--mode drift-check` is\n   accepted as a deprecated alias).\n2. **Never silently overwrite an operator's file.** An `ask` row left unresolved defaults\n   to keeping the operator's version — doing nothing is how work survives. Only\n   `--take-base` moves a conflicted path to the new base version; only a `merge` you write\n   yourself combines the two.\n\n   **The guarantee's exact scope: a path the target version does not ship.**\n   Classification compares disk against the SOURCE version's manifest\n   (`plugin\u002Fmanifests\u002F\u003Cversion>.json`) and the current `base\u002F` — not against any lock;\n   the lock's old `files` map is gone. A file the *project* authored at a path no\n   manifest lists and the new base does not carry — e.g.\n   `.claude\u002Fskills\u002Finspire-code-nestjs\u002Freferences\u002Fmystack.md` — always classifies as `keep`\n   (\"yours — INSPIRE never shipped this\") and is never touched or deleted, even though it\n   sits inside a skill directory INSPIRE owns. This is the fix for the pre-0.3 behaviour\n   (`install.sh` used to `rm -rf` a whole owned entry, destroying exactly these files) —\n   already shipped, not a roadmap item.\n\n   **Where the target version does ship that path, the guarantee is narrower — and a\n   skill's `references\u002F` is exactly such a place.** INSPIRE ships reference files under\n   `inspire-*\u002Freferences\u002F`, and 0.7.0 ships a batch of new ones, so a file the project\n   authored there can **collide** with a newly shipped one. It then classifies `ask` —\n   \"new in this release, and you already have a different file here\" — and an unresolved\n   `ask` keeps the operator's file, exactly as everywhere else. It is still never silently\n   touched or deleted; what changes is that the operator gets a question where before there was\n   nothing to ask. Project content still belongs in the KB, in a `98_lessons` node, or in\n   the project's `CLAUDE.md` — not inside an `inspire-*` directory — but a file left there\n   by mistake survives rather than being silently destroyed.\n\n   `inspire_kb\u002F` is out of scope for *classification* — content classification never runs\n   against it at all, because no shipped manifest lists a KB path in the first place. KB\n   seeding is a separate, purely additive path that runs in both `init` and `update` (see\n   Step 1): a path already on disk is never replaced, only a missing one is added.\n   `--mode update` is **not**, though, flatly non-destructive under `inspire_kb\u002F` as of\n   0.7.0: a versioned hop may retire a KB seed file the project no longer needs to keep —\n   a proven-derivable one is deleted silently, and any divergence from the derivation lands\n   in `ask` rather than a silent delete (Step 2's retirement branch). Two further exceptions\n   are purely additive and never clobber: `seed_design_system` creates\n   `01_spec\u002F_design-system\u002Fdesign-system.md` (never in a KB still on the layered layout), and `seed_claude_md` creates `CLAUDE.md`, **only if\n   absent**.\n3. **No writes happen before Step 2's decisions are gathered.** `--mode plan` itself\n   writes nothing, whatever the operator decides — that is asserted by the test suite, not\n   just documented here.\n4. **Never downgrade.** A plugin older than the project's detected version is an error, not\n   an update — enforced by `version_cmp` in the script, in both `--mode plan` and\n   `--mode update`.\n5. **No lesson reconciliation here.** Classification, archiving and skill rebuilding are\n   D6, and unbuilt. This skill does not read `98_lessons\u002F`. Do not approximate it — telling\n   the operator an edit was kept is truthful; silently merging it is not.\n6. **Do not re-ask the product roots.** Read them from `stack.md`; changing them is a Shape\n   change owned by `\u002Finspire-bootstrap stack`.\n",{"name":5,"description":7},"main","bf5124558fc1203b1fe5aea2d2ae38cfba34939b","plugin\u002Fskills\u002Fupdate",1,"https:\u002F\u002Fgithub.com\u002FGenomcore\u002Finspire","https:\u002F\u002Finspire.openbims.dev","https:\u002F\u002Fgithub.com\u002FGenomcore","https:\u002F\u002Fgenomcore.com","2026-11-13 22:27:32",{"source":8,"skillName":5,"commands":33,"prompt":43,"zipAvailable":44},[34,37,40],{"tool":35,"command":36},"npx","npx skills add Genomcore\u002Finspire --skill update",{"tool":38,"command":39},"bunx","bunx skills add Genomcore\u002Finspire --skill update",{"tool":41,"command":42},"pnpm","pnpm dlx skills add Genomcore\u002Finspire --skill update","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FGenomcore\u002Finspire\u002Fblob\u002Fmain\u002Fplugin\u002Fskills\u002Fupdate\u002FSKILL.md",true,[],[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":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":51,"categories":52,"keywords":53,"publishedAt":18},"01M79DMAENSXPJJPWTRA9CMBV0","migrate","Move this project's knowledge base from the layered layout (01_adr, 02_modules, 03_features, 04_domain, 05_screens and the surface roster) to the concept layout under inspire_kb\u002F01_spec\u002F, where every artifact has a typed extension, and convert its [[wikilinks]] to relative Markdown links. Run once, on purpose, after \u002Finspire:update reports the layered layout or counts wikilinks. Rebuilds nothing that is already built.","https:\u002F\u002Fgithub.com\u002FGenomcore\u002Finspire\u002Fblob\u002Fmain\u002Fplugin\u002Fskills\u002Fmigrate\u002FSKILL.md",[],[],{"skillNo":55,"name":56,"nameZh":6,"description":57,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":58,"categories":59,"keywords":60,"publishedAt":18},"01M79DM6HZ7YN9M3WGXVN64AW1","init","Install the INSPIRE runtime into this repository: materialize the skills, validators, hooks and knowledge-base skeleton, register the hooks, and freeze the runtime version. Run once per project. Use when setting up INSPIRE in a new or existing repo.","https:\u002F\u002Fgithub.com\u002FGenomcore\u002Finspire\u002Fblob\u002Fmain\u002Fplugin\u002Fskills\u002Finit\u002FSKILL.md",[],[],{"skillNo":62,"name":63,"nameZh":6,"description":64,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":65,"categories":66,"keywords":67,"publishedAt":18},"01M79DM28JSC3GWEJ82346VGW6","inspire-workspace","Workspace-level validation: the global pre-PR review (orchestrating module + cross-module checks) and vault-structure validation. Use for the required pre-merge gate and top-level structure checks. ADR lifecycle lives in \u002Finspire-adr; the task tracker in \u002Finspire-task.","https:\u002F\u002Fgithub.com\u002FGenomcore\u002Finspire\u002Fblob\u002Fmain\u002Fplugin\u002Fbase\u002Fskills\u002Finspire-workspace\u002FSKILL.md",[],[],{"skillNo":69,"name":70,"nameZh":6,"description":71,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":72,"categories":73,"keywords":74,"publishedAt":18},"01M79DKYPKA8TKJCW1GF0W2WAN","inspire-task","Task tracker: create \u002F update \u002F close \u002F list \u002F show tickets in inspire_kb\u002F99_tracker\u002F. A plain-file kanban (one Markdown file per ticket, no external tool) for tracking work, drift, and skill-feedback. Use to open, advance, close, or query tickets.","https:\u002F\u002Fgithub.com\u002FGenomcore\u002Finspire\u002Fblob\u002Fmain\u002Fplugin\u002Fbase\u002Fskills\u002Finspire-task\u002FSKILL.md",[],[],{"skillNo":76,"name":77,"nameZh":6,"description":78,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":79,"categories":80,"keywords":81,"publishedAt":18},"01M79DKTXQWFBW2P6VMQW3DYWX","inspire-spike","Vertical spikes: register and harvest external functional-prototype repos. A spike is a narrow, deep, throwaway build in its OWN repo that answers 'can we build it as we think?'. Use to register such a repo (name\u002Freference it, set up its learnings hygiene) and to capture its learnings into the vault as a gap analysis against the current project.","https:\u002F\u002Fgithub.com\u002FGenomcore\u002Finspire\u002Fblob\u002Fmain\u002Fplugin\u002Fbase\u002Fskills\u002Finspire-spike\u002FSKILL.md",[],[],{"skillNo":83,"name":84,"nameZh":6,"description":85,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":9,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":86,"categories":87,"keywords":88,"publishedAt":18},"01M79DKQDXVS3TJ6GG6KXYTMXF","inspire-screens","Create and validate screens (a screen is the UI spec for one view). A screen declares its own bindings — data sources, dispatches, navigation, states — and names a shared layout and shared components as dependencies. Use when designing or reviewing an application's screens.","https:\u002F\u002Fgithub.com\u002FGenomcore\u002Finspire\u002Fblob\u002Fmain\u002Fplugin\u002Fbase\u002Fskills\u002Finspire-screens\u002FSKILL.md",[],[]]