[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M77HDM3RYH15BAMFH5KXS107":3,"skill-same:01M77HDM3RYH15BAMFH5KXS107":53},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":29,"ownerHtmlUrl":30,"ownerBlogUrl":31,"ownerDescription":6,"updatedAt":32,"install":33,"resources":46},"01M77HDM3RYH15BAMFH5KXS107","field-borrow",null,"High-priority default \"inspired by\" workflow. Invoke proactively whenever an external product, repository, paper, standard, benchmark, release, issue, PR, or field practice could improve fak—not only on explicit borrow requests. Mine code, tests, docs, history, issues, PRs, roadmaps, and provenance; date observations; port\u002Fadapt directly when exact-source licensing permits; explore the spirit of promising incomplete ideas; then dogfood fak self-query to classify PRESENT\u002FPARTIAL\u002FABSENT before shipping or filing witnessed gaps. Use when this named workflow matches the task.","anthony-chaudhary\u002Ffak","anthony-chaudhary","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F18080164?v=4","01M77GSKCAEKV3JS07J40B94DY","Anthony Chaudhary","01M77GSKD26Z7F4FQQD4VA30VX",48,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fanthony-chaudhary\u002Ffak\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Ffield-borrow\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: field-borrow\ndescription: High-priority default \"inspired by\" workflow. Invoke proactively whenever an external product, repository, paper, standard, benchmark, release, issue, PR, or field practice could improve fak—not only on explicit borrow requests. Mine code, tests, docs, history, issues, PRs, roadmaps, and provenance; date observations; port\u002Fadapt directly when exact-source licensing permits; explore the spirit of promising incomplete ideas; then dogfood fak self-query to classify PRESENT\u002FPARTIAL\u002FABSENT before shipping or filing witnessed gaps. Use when this named workflow matches the task.\n---\n\n# field-borrow — borrow from the field, but witness the gap against yourself first\n\n## High-priority inspired-by default\n\nThis is the repository's high-priority default \"inspired by\" workflow. Invoke it proactively\nwhenever a named product, repository, paper, standard, benchmark, release, issue, PR, roadmap,\ndesign discussion, or field practice could improve fak. Do not wait for an explicit research or\nborrowing request: run it before bespoke design when credible prior art may change the choice.\nExternal evidence does not override user scope, security, legal obligations, or coherent shipping.\n\nSix laws govern every pass:\n\n1. **Priority:** consider external prior art early, not as optional cleanup after designing.\n2. **Licensed reuse:** copy or adapt directly when exact-source licensing and technical fit permit;\n   fresh reimplementation is not inherently better.\n3. **Exhaustiveness:** all relevant source classes were checked or explicitly marked unavailable.\n4. **Dating:** every observation carries event time, observation time, source state, and refresh cue.\n5. **Breadth:** mine implementation, docs, tests, history, releases, issues, PRs, discussions,\n   roadmaps, and provenance—not only README claims.\n6. **Spirit:** explore transferable principles and responsible extensions even when upstream has\n   only a partial prototype; label inference separately from shipped fact.\n\n## Exhaustive, dated source ledger\n\nFor a repository-shaped source inspect canonical docs\u002Fdesign\u002FAPI examples; code, tests, fixtures,\nand configuration at a pinned revision; releases\u002Ftags\u002Fchangelog and relevant history\u002Fblame; open\n**and closed** issues; merged, closed, and open PRs plus review discussion; discussions, RFCs,\nADRs, roadmaps, and TODOs; exact-revision LICENSE, per-file headers, NOTICE, vendored\u002Fgenerated\nprovenance, submodules, and contribution terms. For papers\u002Fproducts\u002Fstandards use equivalent\nprimary sources (artifact\u002Fappendix, official release history, normative revision\u002Ferrata). Follow\ncitations to the mechanism's origin. Inaccessible surfaces are omissions, not empty results.\n\nRecord every material observation with:\n\n- `observed_at`: full ISO date\u002Ftime for the research observation;\n- `source_event_at`: commit, merge, release, publication, or update date;\n- source state: released\u002Fshipped, merged-unreleased, proposed\u002Fopen, rejected\u002Fclosed,\n  experimental, or unclear;\n- immutable anchor: `path:line@full-sha`, release\u002Ftag+digest, paper version\u002FDOI, or standards rev;\n- platform context: relevant version, API generation, provider, hardware, plan\u002Ftier, or environment;\n- refresh trigger: release, PR merge, API rollover, benchmark rerun, or `none known`.\n\nNever write timeless “X does Y” for a changing platform. Open issues, PRs, and roadmaps prove\ninterest\u002Fdirection, not shipped behavior.\n\n## License disposition: copy when allowed\n\nRead the license at the exact studied revision and check per-file\u002FNOTICE\u002Fprovenance boundaries.\nClassify every candidate before implementation:\n\n- **DIRECT-PORT:** copying\u002Fmodification is permitted and obligations fit. Prefer copying the\n  smallest coherent implementation or test, preserving notices and source `path@sha`, then adapt.\n- **ADAPT:** reuse is permitted but fak interfaces, dependencies, or constraints require material\n  changes. Identify direct versus rewritten portions and retain attribution.\n- **INSPIRE-ONLY:** terms are absent, unclear, incompatible, proprietary, or behavior-only. Build\n  independently; do not copy expressive code, tests, comments, or assets.\n- **DO-NOT-USE:** provenance or terms are unsafe; exclude and record why.\n\nDo **not** default to inspire-only because clean-room rewriting feels cleaner. Public visibility\nalso does not grant permission. When obligations are ambiguous, use INSPIRE-ONLY or seek review.\n\n## Expansive candidate extraction\n\nExtract all of these before ranking:\n\n- direct mechanisms: algorithms, schemas, tests, UX, operations, and documentation;\n- negative knowledge: reverts, rejected proposals, recurring issue failures, regressions, reviews;\n- emerging direction: open PRs, issues, RFCs, roadmaps, TODOs, and design sketches, marked proposed;\n- **Spirit extensions:** infer the broader principle and adjacent fak applications, using\n  `source fact -> inferred principle -> fak opportunity -> disconfirming check`;\n- combinations across independent sources that produce a stronger design.\n\nBe expansive in discovery and conservative in claims. A promising unimplemented idea is a valid\n**INFERENCE** when paired with a cheap experiment\u002Ffalsifier; it is not parity evidence.\n\n## Bounded-superset portfolio rule\n\nThe target is not \"copy the winner\" and not \"accumulate every feature.\" It is a\n**bounded capability superset**: within fak's declared scope and support budget, users should\nbe able to trust that fak carries the strongest reasonable industry default **and** preserves\nmaterially useful alternatives for cohorts whose constraints differ. fak is egoless about\nwhere those capabilities originated. Its modular leaf\u002Fadapter\u002Fintegration boundaries are an\nadvantage: a useful alternative need not displace the default or contaminate the kernel core.\n\nFor every witnessed capability, record one portfolio disposition:\n\n| Disposition | Use when | Required form |\n|---|---|---|\n| `DEFAULT` | strongest net-true choice for the common supported case | integrated, documented default with a witness against the real alternative |\n| `OPTIONAL-MODULE` | useful for a named user\u002Fjob\u002Fconstraint but not the broad default | isolated leaf, adapter, policy, backend, or integration; explicit opt-in and support owner |\n| `RECIPE` | useful and supportable, but too narrow or environment-specific to ship as maintained code | tested composition\u002Fintegration guide with prerequisites and dated witness |\n| `WATCH` | promising, moment-in-time, immature, or evidence is not strong enough to support yet | dated source + review trigger; no shipped claim |\n| `EXCLUDE` | outside scope, unsafe, license-incompatible, operationally unaffordable, or genuinely dominated | named reason and the evidence\u002Fdisconfirming condition that could reopen it |\n\n`DEFAULT` is not the only successful outcome. Do not reject a capability merely because a\ndifferent technique wins overall, is newer, or is already fak's default. Ask whether a real\nsupported cohort still benefits under a constraint such as hardware, provider, latency,\nquality, privacy, deployment topology, compatibility, cost, or operator preference. If yes,\nprefer an `OPTIONAL-MODULE` or `RECIPE` seam over deletion. Conversely, age, popularity, or\n\"someone might want it\" is not enough: an obviously superseded approach with no surviving\ncohort goes to `EXCLUDE`, and a fast-moving snapshot goes to `WATCH` until revalidated.\n\nEach non-excluded row must name: capability axis; user\u002Fjob\u002Fconstraint served; observed evidence\nand date; relationship to the default; proposed modular seam; incremental build, test,\nsecurity, documentation, and maintenance cost; owner or review trigger; and a witness that\nwould disconfirm usefulness. This is the **reasonable-coverage budget**. Once marginal support\ncost exceeds witnessed user value, stop expanding and register the omission instead of **boiling the ocean**.\n\nA study is coverage-complete only when it reports both:\n\n1. **Default frontier** — where fak's common-case default should lead or change.\n2. **Coverage frontier** — useful supported cohorts still absent even though the default may\n   already be best overall.\n\n## Required completion evidence\n\nIn addition to the PRESENT\u002FPARTIAL\u002FABSENT self-query pass below, completion requires a coverage\nchecklist, dated source ledger, license disposition, candidate matrix (fact versus inference and\nshipped versus proposed), exact fak seam, and either shipped proof or contract-ready filed work.\nPrefer shipping a safe in-scope minimal spine now over reflexively creating backlog. A long link\nlist is not exhaustive research: state what each source changed in the conclusion.\n\n\n## Why this skill exists\n\nfak researches the outward field well — the `docs\u002Fnotes\u002FRESEARCH-*` corpus, `industry-score`,\n`idea-scout`, `sota-check` all point *outward*. But applying a field idea inward has a silent\nfailure mode: an agent reads \"Cognee does GraphRAG\" or \"Zep has bi-temporal memory\" and files\n\"fak should add graph memory\" **without first asking fak whether it already has it** — so the\ntracker accretes gaps fak already closed, or gaps described one abstraction away from the real\nseam. The research was outward-complete but the *witness step* was manual and skipped.\n\nThe law, in one line: **the default answer to \"does fak already have this field-standard\ncapability?\" is _ask fak, don't assume_ — dogfood the self-query surface to witness\npresent \u002F partial \u002F absent, then file only the witnessed gaps, each grounded in a seam.**\n\nThis is the loop that produced #3161 \u002F #3162 \u002F #3163 (edge-aware query, retrieval-eval,\nfreshness rung) under epic #1494: three retrieval-*quality* gaps borrowed from Cognee\u002FZep,\nLongMemEval\u002FLOCOMO, and Zep\u002FMem0 — each witnessed by querying `fak_feature_query` against\nfak's own `internal\u002Fselfquery` before filing, each grounded at `internal\u002Fselfquery\u002Fselfquery.go`\n(the flat lexical ranker at `:566`). See\n[`docs\u002Fnotes\u002FCONCEPT-FIELD-BORROW-QUERY-QUALITY-2026-07-08.md`](..\u002F..\u002F..\u002Fdocs\u002Fnotes\u002FCONCEPT-FIELD-BORROW-QUERY-QUALITY-2026-07-08.md)\nfor that worked instance.\n\n### How this differs from its siblings (do not duplicate them)\n\n| Skill \u002F tool | Direction | Mechanism | Files issues? | Dogfoods the self-query witness? |\n|---|---|---|---|---|\n| `sota-check` | inward, **kernels only** | `fak sota \u003Cop>` prior-art matrix | no | no |\n| `industry-score` | outward | scorecard: coverage + parity-debt over a taxonomy | no | no |\n| `idea-scout` (`tools\u002Fidea_scout.py`) | outward, **automated** | arXiv\u002FGitHub scan, dedup vs **existing issues** (seen-cache · body-stamp · title-Jaccard) | yes (triage queue) | **no** (issue-vs-issue only, never capability-vs-codebase) |\n| **field-borrow** (this) | outward→inward | **dogfood `fak_feature_query` to witness the gap**, then file epic-anchored | yes (grounded, under the right epic) | **yes — this is the point** |\n\nfield-borrow composes with all three: it can be the human-in-the-loop pass over an\n`idea-scout` hit, it feeds witnessed coverage rows to `industry-score`, and once a spine\nships it hands off to `spine-fanout`.\n\n\n## Durable study receipt — query first, record the disposition last\n\nBefore acquiring the external source, run `fak study search \"\u003Csource-or-capability>\" --limit 20`. Use a hit to choose and report **extend**, **recheck**, or **supersede** rather than duplicating the study. Missing or unavailable storage degrades explicitly: report `durable study memory unavailable: \u003Cerror>`, continue with bounded local evidence, and never call that evidence persisted.\n\nAfter the PRESENT\u002FPARTIAL\u002FABSENT witness and disposition exist, write a pinned `fak-study\u002F1` record with `fak study add --file \u003Crecord.json>`. Put the returned `study_...` ID in every downstream GitHub issue body. Verified implementation outcomes append through a new receipt with an outcome and `supersedes` link; they do not rewrite or model-promote the original decision.\n\nPromotion evidence: a fresh invocation discovers the candidate through bounded search and returns the same record ID. Demote or retire this seam if it does not prevent duplicate study or if storage failures can masquerade as persistence. Invalidating assumption: source names and capability terms provide enough lexical signal for bounded rediscovery.\n\n## The pass\n\n### 1 — Name the source and the capability axis (dated)\n\nPick ONE capability from a **named** external system or a dated source: an agent-memory\nproduct (Letta\u002FMemGPT, Mem0, Zep\u002FGraphiti, Cognee, Supermemory), a serving system\n(vLLM\u002FSGLang\u002FTensorRT-LLM), a competitor feature, a benchmark (LongMemEval, LOCOMO), or a\nrow from the `RESEARCH-*` \u002F industry field scans. Write it down with the source, so the\nborrow is falsifiable: *\"Zep\u002FGraphiti — bi-temporal edges: facts carry validity windows;\na newer fact invalidates the old one.\"* One capability per pass; do not batch.\n\n### 2 — Dogfood the self-query WITNESS (the heart of this skill)\n\nAsk fak, three ways, whether it already has the capability — do **not** reason about it\nfrom memory:\n\n- **`fak_feature_query`** (MCP) \u002F `fak capabilities \u003Cintent>` — intent → ranked capability\n  cards (name, summary, `witness` content-address, the call to make).\n- **`fak-dev index docs|leaves|verbs|claims \u003Cterms>`** — the dev self-index over docs, packages,\n  CLI verbs, and the CLAIMS ledger.\n- Also a plain `Grep` for the concept's obvious symbol, to catch a capability the lexical\n  ranker missed (see Honest limits).\n\nClassify the result, and **paste the top cards as the witness**:\n\n- **PRESENT** — a card \u002F verb \u002F leaf directly does it. *Read it and confirm* it truly does\n  (not just a name collision). PRESENT is a valid, complete, **good** outcome — \"fak already\n  has this, here is the card\" ends the pass; do not file.\n- **PARTIAL** — an adjacent surface exists but not this capability (e.g. `fak_index_freshness`\n  covers index↔tree drift but not semantic supersession).\n- **ABSENT** — the query surfaces nothing on-point. An ABSENT claim **must** show the query\n  that returned nothing on-point — that null result is the falsifiable witness.\n\n### 3 — Ground the PARTIAL\u002FABSENT in a seam (file:line)\n\nThe capability cards point at the code. Find the **exact file:line** the capability would\ntouch and quote **one** dogfood result as the in-issue evidence — e.g. the flat lexical\nranker `score()` at `internal\u002Fselfquery\u002Fselfquery.go:566`, or the `FeatureCard` struct at\n`:37` that carries no freshness\u002Fedge field. **No file:line seam, no issue** — a gap you\ncannot locate is not ready to file; keep witnessing until you can point at the code.\n\n### 4 — Dedup, then file grounded (under the right epic)\n\nSearch existing issues **and** the natural parent epic before filing (avoid re-filing what\n`idea-scout` or a human already opened). File under the epic with, in order:\n\n1. the pain, in fak's terms, quoting the seam from step 3;\n2. the **named external inspiration** (dated — a fabricated \"system X does Y\" is the exact\n   citation failure `dos_citation_resolve` exists to catch);\n3. the **dogfood witness** from step 2 (the query + what it returned);\n4. a **concrete first checkable step** (the smallest thing that proves the gap is real and\n   the fix is buildable);\n5. **honest fences** (deterministic \u002F advisory \u002F read-only where they apply).\n\nSet milestone + labels at creation. On this host, run `gh` via PowerShell or the Bash tool\ndeliberately (quote bodies through `--body-file`, not inline heredocs that the shell mangles).\n\n### 5 — Register the trail (or it lives only in a transcript)\n\nRecord the instance in a dated `docs\u002Fnotes\u002FCONCEPT-\u003Ctopic>-\u003Cdate>.md`: source → witness →\nseam → filed issues, with a `Companions:` block. Add its **`INDEX.md`** line in the\n*Notes & research* section (an unindexed dated note is an orphan `fak-dev index freshness`\nflags). Link the children from the parent epic. **This registration — not the issues alone —\nis what makes the research path durable.**\n\n### 6 — (optional) Feed the map\n\nA witnessed PRESENT\u002FPARTIAL\u002FABSENT is a coverage row `industry-score` can absorb; a genuinely\nnovel, spine-shaped gap can seed `spine-fanout` once a working spine ships.\n\n## What \"done\" proves\n\n- Every filed gap carries a **dogfood witness** (the query + its top cards), a **file:line\n  seam**, a **named dated external source**, and a **first checkable step**.\n- No gap was filed that fak already ships (step 2 caught the PRESENT cases).\n- The instance is registered: a dated `docs\u002Fnotes` entry, its `INDEX.md` line, and the epic\n  links its children. `python tools\u002F... ` freshness \u002F `fak-dev index freshness` reads clean for\n  the new note.\n\n## The anti-gaming rule (specific to this surface)\n\n**Never file a capability gap you did not dogfood-witness, and never mark ABSENT without the\nnull query.** The witness is the falsifiable record — the way `sota-check`'s `Prior-art:`\ntrailer is. Two failure modes this rule kills:\n\n- **Assumed-absent.** Filing \"fak lacks X\" from memory when `fak_feature_query` would have\n  returned the card. The fix is the query, pasted.\n- **Fabricated borrow.** \"Zep does bi-temporal invalidation\" when it does not, or citing a\n  system\u002Fpaper you did not read. Name the source, dated; if you could find no prior art, say\n  *that* so the next person can disprove it.\n\nPRESENT is not a failed pass — a witnessed \"we already have this\" is a real result.\n\n## Where this misleads \u002F honest limits\n\n- **The self-query score is lexical.** `rankCards`\u002F`score()`\n  (`internal\u002Fselfquery\u002Fselfquery.go:540,566`) is `strings.Contains` substring matching over\n  name\u002Ftags\u002Fref\u002Fsummary — a poorly-phrased query can miss a card fak *does* have, producing a\n  **false-ABSENT**. Read the top cards, vary the phrasing, and cross-check with a raw `Grep`\n  before trusting an ABSENT. (This limit is itself issue #3162 — the ranker is ungraded.)\n- **ABSENT means \"fak's own index doesn't surface it\", not \"it doesn't exist.\"** The index can\n  be stale or the capability can live under a name you didn't guess.\n- **The witness is a snapshot.** A card added tomorrow changes the verdict (the freshness\n  problem is issue #3163). Re-witness before acting on an old pass.\n- **Composes with, does not replace:** `idea-scout` (automated outward feeder), `industry-score`\n  (the competitive coverage map), `spine-fanout` (post-spine fan-out), `sota-check` (kernel\n  prior-art). If the capability is a **compute kernel**, use `sota-check` instead — this pass\n  is for product \u002F agent \u002F memory \u002F serving capabilities.\n\n## When to run this\n\n- Reading a `RESEARCH-*` \u002F SOTA \u002F competitor note and wondering *\"do we already have this?\"*\n- Before filing any \"fak should add X\" capability-gap issue.\n- After a competitor or a paper ships a capability worth borrowing.\n- On a `\u002Floop` cadence over the `docs\u002Fnotes\u002FRESEARCH-*` corpus, to convert inert research\n  into witnessed, grounded, epic-anchored backlog — the inward, human-curated complement to\n  the daily `idea-scout`.\n","# field-borrow — borrow from the field, but witness the gap against yourself first\n\n## High-priority inspired-by default\n\nThis is the repository's high-priority default \"inspired by\" workflow. Invoke it proactively\nwhenever a named product, repository, paper, standard, benchmark, release, issue, PR, roadmap,\ndesign discussion, or field practice could improve fak. Do not wait for an explicit research or\nborrowing request: run it before bespoke design when credible prior art may change the choice.\nExternal evidence does not override user scope, security, legal obligations, or coherent shipping.\n\nSix laws govern every pass:\n\n1. **Priority:** consider external prior art early, not as optional cleanup after designing.\n2. **Licensed reuse:** copy or adapt directly when exact-source licensing and technical fit permit;\n   fresh reimplementation is not inherently better.\n3. **Exhaustiveness:** all relevant source classes were checked or explicitly marked unavailable.\n4. **Dating:** every observation carries event time, observation time, source state, and refresh cue.\n5. **Breadth:** mine implementation, docs, tests, history, releases, issues, PRs, discussions,\n   roadmaps, and provenance—not only README claims.\n6. **Spirit:** explore transferable principles and responsible extensions even when upstream has\n   only a partial prototype; label inference separately from shipped fact.\n\n## Exhaustive, dated source ledger\n\nFor a repository-shaped source inspect canonical docs\u002Fdesign\u002FAPI examples; code, tests, fixtures,\nand configuration at a pinned revision; releases\u002Ftags\u002Fchangelog and relevant history\u002Fblame; open\n**and closed** issues; merged, closed, and open PRs plus review discussion; discussions, RFCs,\nADRs, roadmaps, and TODOs; exact-revision LICENSE, per-file headers, NOTICE, vendored\u002Fgenerated\nprovenance, submodules, and contribution terms. For papers\u002Fproducts\u002Fstandards use equivalent\nprimary sources (artifact\u002Fappendix, official release history, normative revision\u002Ferrata). Follow\ncitations to the mechanism's origin. Inaccessible surfaces are omissions, not empty results.\n\nRecord every material observation with:\n\n- `observed_at`: full ISO date\u002Ftime for the research observation;\n- `source_event_at`: commit, merge, release, publication, or update date;\n- source state: released\u002Fshipped, merged-unreleased, proposed\u002Fopen, rejected\u002Fclosed,\n  experimental, or unclear;\n- immutable anchor: `path:line@full-sha`, release\u002Ftag+digest, paper version\u002FDOI, or standards rev;\n- platform context: relevant version, API generation, provider, hardware, plan\u002Ftier, or environment;\n- refresh trigger: release, PR merge, API rollover, benchmark rerun, or `none known`.\n\nNever write timeless “X does Y” for a changing platform. Open issues, PRs, and roadmaps prove\ninterest\u002Fdirection, not shipped behavior.\n\n## License disposition: copy when allowed\n\nRead the license at the exact studied revision and check per-file\u002FNOTICE\u002Fprovenance boundaries.\nClassify every candidate before implementation:\n\n- **DIRECT-PORT:** copying\u002Fmodification is permitted and obligations fit. Prefer copying the\n  smallest coherent implementation or test, preserving notices and source `path@sha`, then adapt.\n- **ADAPT:** reuse is permitted but fak interfaces, dependencies, or constraints require material\n  changes. Identify direct versus rewritten portions and retain attribution.\n- **INSPIRE-ONLY:** terms are absent, unclear, incompatible, proprietary, or behavior-only. Build\n  independently; do not copy expressive code, tests, comments, or assets.\n- **DO-NOT-USE:** provenance or terms are unsafe; exclude and record why.\n\nDo **not** default to inspire-only because clean-room rewriting feels cleaner. Public visibility\nalso does not grant permission. When obligations are ambiguous, use INSPIRE-ONLY or seek review.\n\n## Expansive candidate extraction\n\nExtract all of these before ranking:\n\n- direct mechanisms: algorithms, schemas, tests, UX, operations, and documentation;\n- negative knowledge: reverts, rejected proposals, recurring issue failures, regressions, reviews;\n- emerging direction: open PRs, issues, RFCs, roadmaps, TODOs, and design sketches, marked proposed;\n- **Spirit extensions:** infer the broader principle and adjacent fak applications, using\n  `source fact -> inferred principle -> fak opportunity -> disconfirming check`;\n- combinations across independent sources that produce a stronger design.\n\nBe expansive in discovery and conservative in claims. A promising unimplemented idea is a valid\n**INFERENCE** when paired with a cheap experiment\u002Ffalsifier; it is not parity evidence.\n\n## Bounded-superset portfolio rule\n\nThe target is not \"copy the winner\" and not \"accumulate every feature.\" It is a\n**bounded capability superset**: within fak's declared scope and support budget, users should\nbe able to trust that fak carries the strongest reasonable industry default **and** preserves\nmaterially useful alternatives for cohorts whose constraints differ. fak is egoless about\nwhere those capabilities originated. Its modular leaf\u002Fadapter\u002Fintegration boundaries are an\nadvantage: a useful alternative need not displace the default or contaminate the kernel core.\n\nFor every witnessed capability, record one portfolio disposition:\n\n| Disposition | Use when | Required form |\n|---|---|---|\n| `DEFAULT` | strongest net-true choice for the common supported case | integrated, documented default with a witness against the real alternative |\n| `OPTIONAL-MODULE` | useful for a named user\u002Fjob\u002Fconstraint but not the broad default | isolated leaf, adapter, policy, backend, or integration; explicit opt-in and support owner |\n| `RECIPE` | useful and supportable, but too narrow or environment-specific to ship as maintained code | tested composition\u002Fintegration guide with prerequisites and dated witness |\n| `WATCH` | promising, moment-in-time, immature, or evidence is not strong enough to support yet | dated source + review trigger; no shipped claim |\n| `EXCLUDE` | outside scope, unsafe, license-incompatible, operationally unaffordable, or genuinely dominated | named reason and the evidence\u002Fdisconfirming condition that could reopen it |\n\n`DEFAULT` is not the only successful outcome. Do not reject a capability merely because a\ndifferent technique wins overall, is newer, or is already fak's default. Ask whether a real\nsupported cohort still benefits under a constraint such as hardware, provider, latency,\nquality, privacy, deployment topology, compatibility, cost, or operator preference. If yes,\nprefer an `OPTIONAL-MODULE` or `RECIPE` seam over deletion. Conversely, age, popularity, or\n\"someone might want it\" is not enough: an obviously superseded approach with no surviving\ncohort goes to `EXCLUDE`, and a fast-moving snapshot goes to `WATCH` until revalidated.\n\nEach non-excluded row must name: capability axis; user\u002Fjob\u002Fconstraint served; observed evidence\nand date; relationship to the default; proposed modular seam; incremental build, test,\nsecurity, documentation, and maintenance cost; owner or review trigger; and a witness that\nwould disconfirm usefulness. This is the **reasonable-coverage budget**. Once marginal support\ncost exceeds witnessed user value, stop expanding and register the omission instead of **boiling the ocean**.\n\nA study is coverage-complete only when it reports both:\n\n1. **Default frontier** — where fak's common-case default should lead or change.\n2. **Coverage frontier** — useful supported cohorts still absent even though the default may\n   already be best overall.\n\n## Required completion evidence\n\nIn addition to the PRESENT\u002FPARTIAL\u002FABSENT self-query pass below, completion requires a coverage\nchecklist, dated source ledger, license disposition, candidate matrix (fact versus inference and\nshipped versus proposed), exact fak seam, and either shipped proof or contract-ready filed work.\nPrefer shipping a safe in-scope minimal spine now over reflexively creating backlog. A long link\nlist is not exhaustive research: state what each source changed in the conclusion.\n\n\n## Why this skill exists\n\nfak researches the outward field well — the `docs\u002Fnotes\u002FRESEARCH-*` corpus, `industry-score`,\n`idea-scout`, `sota-check` all point *outward*. But applying a field idea inward has a silent\nfailure mode: an agent reads \"Cognee does GraphRAG\" or \"Zep has bi-temporal memory\" and files\n\"fak should add graph memory\" **without first asking fak whether it already has it** — so the\ntracker accretes gaps fak already closed, or gaps described one abstraction away from the real\nseam. The research was outward-complete but the *witness step* was manual and skipped.\n\nThe law, in one line: **the default answer to \"does fak already have this field-standard\ncapability?\" is _ask fak, don't assume_ — dogfood the self-query surface to witness\npresent \u002F partial \u002F absent, then file only the witnessed gaps, each grounded in a seam.**\n\nThis is the loop that produced #3161 \u002F #3162 \u002F #3163 (edge-aware query, retrieval-eval,\nfreshness rung) under epic #1494: three retrieval-*quality* gaps borrowed from Cognee\u002FZep,\nLongMemEval\u002FLOCOMO, and Zep\u002FMem0 — each witnessed by querying `fak_feature_query` against\nfak's own `internal\u002Fselfquery` before filing, each grounded at `internal\u002Fselfquery\u002Fselfquery.go`\n(the flat lexical ranker at `:566`). See\n[`docs\u002Fnotes\u002FCONCEPT-FIELD-BORROW-QUERY-QUALITY-2026-07-08.md`](..\u002F..\u002F..\u002Fdocs\u002Fnotes\u002FCONCEPT-FIELD-BORROW-QUERY-QUALITY-2026-07-08.md)\nfor that worked instance.\n\n### How this differs from its siblings (do not duplicate them)\n\n| Skill \u002F tool | Direction | Mechanism | Files issues? | Dogfoods the self-query witness? |\n|---|---|---|---|---|\n| `sota-check` | inward, **kernels only** | `fak sota \u003Cop>` prior-art matrix | no | no |\n| `industry-score` | outward | scorecard: coverage + parity-debt over a taxonomy | no | no |\n| `idea-scout` (`tools\u002Fidea_scout.py`) | outward, **automated** | arXiv\u002FGitHub scan, dedup vs **existing issues** (seen-cache · body-stamp · title-Jaccard) | yes (triage queue) | **no** (issue-vs-issue only, never capability-vs-codebase) |\n| **field-borrow** (this) | outward→inward | **dogfood `fak_feature_query` to witness the gap**, then file epic-anchored | yes (grounded, under the right epic) | **yes — this is the point** |\n\nfield-borrow composes with all three: it can be the human-in-the-loop pass over an\n`idea-scout` hit, it feeds witnessed coverage rows to `industry-score`, and once a spine\nships it hands off to `spine-fanout`.\n\n\n## Durable study receipt — query first, record the disposition last\n\nBefore acquiring the external source, run `fak study search \"\u003Csource-or-capability>\" --limit 20`. Use a hit to choose and report **extend**, **recheck**, or **supersede** rather than duplicating the study. Missing or unavailable storage degrades explicitly: report `durable study memory unavailable: \u003Cerror>`, continue with bounded local evidence, and never call that evidence persisted.\n\nAfter the PRESENT\u002FPARTIAL\u002FABSENT witness and disposition exist, write a pinned `fak-study\u002F1` record with `fak study add --file \u003Crecord.json>`. Put the returned `study_...` ID in every downstream GitHub issue body. Verified implementation outcomes append through a new receipt with an outcome and `supersedes` link; they do not rewrite or model-promote the original decision.\n\nPromotion evidence: a fresh invocation discovers the candidate through bounded search and returns the same record ID. Demote or retire this seam if it does not prevent duplicate study or if storage failures can masquerade as persistence. Invalidating assumption: source names and capability terms provide enough lexical signal for bounded rediscovery.\n\n## The pass\n\n### 1 — Name the source and the capability axis (dated)\n\nPick ONE capability from a **named** external system or a dated source: an agent-memory\nproduct (Letta\u002FMemGPT, Mem0, Zep\u002FGraphiti, Cognee, Supermemory), a serving system\n(vLLM\u002FSGLang\u002FTensorRT-LLM), a competitor feature, a benchmark (LongMemEval, LOCOMO), or a\nrow from the `RESEARCH-*` \u002F industry field scans. Write it down with the source, so the\nborrow is falsifiable: *\"Zep\u002FGraphiti — bi-temporal edges: facts carry validity windows;\na newer fact invalidates the old one.\"* One capability per pass; do not batch.\n\n### 2 — Dogfood the self-query WITNESS (the heart of this skill)\n\nAsk fak, three ways, whether it already has the capability — do **not** reason about it\nfrom memory:\n\n- **`fak_feature_query`** (MCP) \u002F `fak capabilities \u003Cintent>` — intent → ranked capability\n  cards (name, summary, `witness` content-address, the call to make).\n- **`fak-dev index docs|leaves|verbs|claims \u003Cterms>`** — the dev self-index over docs, packages,\n  CLI verbs, and the CLAIMS ledger.\n- Also a plain `Grep` for the concept's obvious symbol, to catch a capability the lexical\n  ranker missed (see Honest limits).\n\nClassify the result, and **paste the top cards as the witness**:\n\n- **PRESENT** — a card \u002F verb \u002F leaf directly does it. *Read it and confirm* it truly does\n  (not just a name collision). PRESENT is a valid, complete, **good** outcome — \"fak already\n  has this, here is the card\" ends the pass; do not file.\n- **PARTIAL** — an adjacent surface exists but not this capability (e.g. `fak_index_freshness`\n  covers index↔tree drift but not semantic supersession).\n- **ABSENT** — the query surfaces nothing on-point. An ABSENT claim **must** show the query\n  that returned nothing on-point — that null result is the falsifiable witness.\n\n### 3 — Ground the PARTIAL\u002FABSENT in a seam (file:line)\n\nThe capability cards point at the code. Find the **exact file:line** the capability would\ntouch and quote **one** dogfood result as the in-issue evidence — e.g. the flat lexical\nranker `score()` at `internal\u002Fselfquery\u002Fselfquery.go:566`, or the `FeatureCard` struct at\n`:37` that carries no freshness\u002Fedge field. **No file:line seam, no issue** — a gap you\ncannot locate is not ready to file; keep witnessing until you can point at the code.\n\n### 4 — Dedup, then file grounded (under the right epic)\n\nSearch existing issues **and** the natural parent epic before filing (avoid re-filing what\n`idea-scout` or a human already opened). File under the epic with, in order:\n\n1. the pain, in fak's terms, quoting the seam from step 3;\n2. the **named external inspiration** (dated — a fabricated \"system X does Y\" is the exact\n   citation failure `dos_citation_resolve` exists to catch);\n3. the **dogfood witness** from step 2 (the query + what it returned);\n4. a **concrete first checkable step** (the smallest thing that proves the gap is real and\n   the fix is buildable);\n5. **honest fences** (deterministic \u002F advisory \u002F read-only where they apply).\n\nSet milestone + labels at creation. On this host, run `gh` via PowerShell or the Bash tool\ndeliberately (quote bodies through `--body-file`, not inline heredocs that the shell mangles).\n\n### 5 — Register the trail (or it lives only in a transcript)\n\nRecord the instance in a dated `docs\u002Fnotes\u002FCONCEPT-\u003Ctopic>-\u003Cdate>.md`: source → witness →\nseam → filed issues, with a `Companions:` block. Add its **`INDEX.md`** line in the\n*Notes & research* section (an unindexed dated note is an orphan `fak-dev index freshness`\nflags). Link the children from the parent epic. **This registration — not the issues alone —\nis what makes the research path durable.**\n\n### 6 — (optional) Feed the map\n\nA witnessed PRESENT\u002FPARTIAL\u002FABSENT is a coverage row `industry-score` can absorb; a genuinely\nnovel, spine-shaped gap can seed `spine-fanout` once a working spine ships.\n\n## What \"done\" proves\n\n- Every filed gap carries a **dogfood witness** (the query + its top cards), a **file:line\n  seam**, a **named dated external source**, and a **first checkable step**.\n- No gap was filed that fak already ships (step 2 caught the PRESENT cases).\n- The instance is registered: a dated `docs\u002Fnotes` entry, its `INDEX.md` line, and the epic\n  links its children. `python tools\u002F... ` freshness \u002F `fak-dev index freshness` reads clean for\n  the new note.\n\n## The anti-gaming rule (specific to this surface)\n\n**Never file a capability gap you did not dogfood-witness, and never mark ABSENT without the\nnull query.** The witness is the falsifiable record — the way `sota-check`'s `Prior-art:`\ntrailer is. Two failure modes this rule kills:\n\n- **Assumed-absent.** Filing \"fak lacks X\" from memory when `fak_feature_query` would have\n  returned the card. The fix is the query, pasted.\n- **Fabricated borrow.** \"Zep does bi-temporal invalidation\" when it does not, or citing a\n  system\u002Fpaper you did not read. Name the source, dated; if you could find no prior art, say\n  *that* so the next person can disprove it.\n\nPRESENT is not a failed pass — a witnessed \"we already have this\" is a real result.\n\n## Where this misleads \u002F honest limits\n\n- **The self-query score is lexical.** `rankCards`\u002F`score()`\n  (`internal\u002Fselfquery\u002Fselfquery.go:540,566`) is `strings.Contains` substring matching over\n  name\u002Ftags\u002Fref\u002Fsummary — a poorly-phrased query can miss a card fak *does* have, producing a\n  **false-ABSENT**. Read the top cards, vary the phrasing, and cross-check with a raw `Grep`\n  before trusting an ABSENT. (This limit is itself issue #3162 — the ranker is ungraded.)\n- **ABSENT means \"fak's own index doesn't surface it\", not \"it doesn't exist.\"** The index can\n  be stale or the capability can live under a name you didn't guess.\n- **The witness is a snapshot.** A card added tomorrow changes the verdict (the freshness\n  problem is issue #3163). Re-witness before acting on an old pass.\n- **Composes with, does not replace:** `idea-scout` (automated outward feeder), `industry-score`\n  (the competitive coverage map), `spine-fanout` (post-spine fan-out), `sota-check` (kernel\n  prior-art). If the capability is a **compute kernel**, use `sota-check` instead — this pass\n  is for product \u002F agent \u002F memory \u002F serving capabilities.\n\n## When to run this\n\n- Reading a `RESEARCH-*` \u002F SOTA \u002F competitor note and wondering *\"do we already have this?\"*\n- Before filing any \"fak should add X\" capability-gap issue.\n- After a competitor or a paper ships a capability worth borrowing.\n- On a `\u002Floop` cadence over the `docs\u002Fnotes\u002FRESEARCH-*` corpus, to convert inert research\n  into witnessed, grounded, epic-anchored backlog — the inward, human-curated complement to\n  the daily `idea-scout`.\n",{"name":5,"description":7},"main","f565b7b7db63e79649f3a02c08108132a86c5d05",".claude\u002Fskills\u002Ffield-borrow",16,"https:\u002F\u002Fgithub.com\u002Fanthony-chaudhary\u002Ffak","https:\u002F\u002Fanthony-chaudhary.github.io\u002Ffak\u002Fshowcase.html","https:\u002F\u002Fgithub.com\u002Fanthony-chaudhary","https:\u002F\u002Flinkedin.com\u002Fin\u002Fanthonychaudhary","2026-11-15 03:06:00",{"source":8,"skillName":5,"commands":34,"prompt":44,"zipAvailable":45},[35,38,41],{"tool":36,"command":37},"npx","npx skills add anthony-chaudhary\u002Ffak --skill field-borrow",{"tool":39,"command":40},"bunx","bunx skills add anthony-chaudhary\u002Ffak --skill field-borrow",{"tool":42,"command":43},"pnpm","pnpm dlx skills add anthony-chaudhary\u002Ffak --skill field-borrow","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fanthony-chaudhary\u002Ffak\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Ffield-borrow\u002FSKILL.md",true,[47],{"path":48,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":51,"sourceHtmlUrl":52},"contract_test.go","SCRIPT","脚本",3590,"https:\u002F\u002Fgithub.com\u002Fanthony-chaudhary\u002Ffak\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Ffield-borrow\u002Fcontract_test.go",[54,61,70,77,84,89],{"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},"01M78V4059KRP2T88YR6Z05MJQ","serve-receipt","On-demand research protocol + deterministic receipt for serving a Hugging Face model on a concrete GPU target — fan out research lanes (model facts\u002Ffiles, engine support matrix, serving recipes, KV\u002Fstate memory math), compute the tiered feasibility verdict with `fak serve-receipt` (never narrated math), and emit a witnessed launch receipt. Use when asked for an \"initial config\", \"serving receipt\", or \"how do I serve \u003Cmodel> on \u003CGPUs>\".","https:\u002F\u002Fgithub.com\u002Fanthony-chaudhary\u002Ffak\u002Fblob\u002Fmain\u002F.claude\u002Fskills\u002Fserve-receipt\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":65,"licenseFileUrl":66,"sourceHtmlUrl":67,"categories":68,"keywords":69,"publishedAt":19},"01M77HRGKQV2CJ2K6ZFN4YQ32K","caveman","Ultra-compressed communication mode. Cuts output tokens 65% (measured) by speaking like caveman while keeping full technical accuracy. Supports intensity levels: lite, full (default), ultra, wenyan-lite, wenyan-full, wenyan-ultra. Use when user says \"caveman mode\", \"talk like caveman\", \"use caveman\", \"less tokens\", \"be brief\", or invokes \u002Fcaveman. Also auto-triggers when token efficiency is requested.","MIT","https:\u002F\u002Fgithub.com\u002Fanthony-chaudhary\u002Ffak\u002Fblob\u002Fmain\u002Finternal\u002Fsyspromptmmu\u002Fthird_party\u002Fcaveman\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fanthony-chaudhary\u002Ffak\u002Fblob\u002Fmain\u002Finternal\u002Fsyspromptmmu\u002Fthird_party\u002Fcaveman\u002FSKILL.md",[],[],{"skillNo":71,"name":72,"nameZh":6,"description":73,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":74,"categories":75,"keywords":76,"publishedAt":19},"01M77HRAY3Y3BE7APN5PVJF7NV","quality-report","Score the repo quality and emit a report card.","https:\u002F\u002Fgithub.com\u002Fanthony-chaudhary\u002Ffak\u002Fblob\u002Fmain\u002Finternal\u002Fctxplans\u002Ftestdata\u002Frepo\u002F.claude\u002Fskills\u002Fquality-report\u002FSKILL.md",[],[],{"skillNo":78,"name":79,"nameZh":6,"description":80,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":81,"categories":82,"keywords":83,"publishedAt":19},"01M77HR9807KHAX13DG9PW2K2Q","ctx-overlay","Compact the context window and prune the memory store when a session grows too long.","https:\u002F\u002Fgithub.com\u002Fanthony-chaudhary\u002Ffak\u002Fblob\u002Fmain\u002Finternal\u002Fctxplans\u002Ftestdata\u002Frepo\u002F.claude\u002Fskills\u002Fctx-overlay\u002FSKILL.md",[],[],{"skillNo":85,"name":72,"nameZh":6,"description":73,"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},"01M77HR7XQ9X8F2CY9Z4F2ZEZ3","https:\u002F\u002Fgithub.com\u002Fanthony-chaudhary\u002Ffak\u002Fblob\u002Fmain\u002Finternal\u002Fctxplanlint\u002Ftestdata\u002Frepo\u002F.claude\u002Fskills\u002Fquality-report\u002FSKILL.md",[],[],{"skillNo":90,"name":79,"nameZh":6,"description":80,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":91,"categories":92,"keywords":93,"publishedAt":19},"01M77HR5BVKCTR1WNJGRBSDNX7","https:\u002F\u002Fgithub.com\u002Fanthony-chaudhary\u002Ffak\u002Fblob\u002Fmain\u002Finternal\u002Fctxplanlint\u002Ftestdata\u002Frepo\u002F.claude\u002Fskills\u002Fctx-overlay\u002FSKILL.md",[],[]]