Skip to content

t1k:unity:base:lib-candidates

FieldValue
Modulebase
Version5.116.6
Effortmedium
ToolsBash, Read

Keywords: adopt candidates, contribution candidates, game code worth contributing, lib candidates, lib-candidates scan, library candidates, library-grade code

/t1k:unity:base:lib-candidates

Unity Library Candidates: detect game code that belongs in a library

Section titled “Unity Library Candidates: detect game code that belongs in a library”

Answers “which first-party types in this game are generic enough to live in a studio library, and which did the game rebuild although a library already ships it?” without spending a token on the question. The deterministic half lives here (scripts/lib-candidates.cjs); judging a candidate is the detached t1k-corpus-scout hop, and the PR is t1k-lib-contribute’s. The CLI contract is the Unity port of the Cocos stdlib-candidates.cjs (state schema v2); a core Stop hook drives it, a human or agent can too.

Read-only on the game: it never edits Assets/. State lives in ~/.claude/session-state/lib-candidates/<sha1(project)>.json (plus a per-project type-index cache).

ClassMeaningNext step
contributelibrary-grade, and no library type matches by name + public-member overlapcorpus-scout confirms, then t1k-lib-contribute (rule: docs/library-first-contribute-back.md)
adoptan equivalent already exists (overlap >= libCandidate.adoptThreshold)the game consumes the library type instead of keeping its copy

An adopt row carries localPath, libraryRepo, libraryPackage, equivalentType, equivalentPath, overlap: the one-line frame core prints needs exactly these.

Terminal window
CLI=.claude/skills/t1k-unity-base-lib-candidates/scripts/lib-candidates.cjs
bun $CLI scan --project <root> --all # first run: ranks the existing tree as `baseline`
bun $CLI scan --project <root> --files Assets/A.cs,B.cs # changed files -> `pending`
bun $CLI status --project <root>
bun $CLI brief --project <root> [--max N] # claims a batch, prints the scout prompt
bun $CLI ingest --project <root> --batch <id> --result <file>
bun $CLI report --project <root> [--format md|json] [--include-skip]
bun $CLI ack --project <root> --ids a,b --decision approved|dismissed|consumed|reset
bun $CLI purge --project <root>

Every command prints one JSON document (exit 0; 2 usage; 1 internal), except report which prints markdown by default. scan prints scanned, candidatesFound, added[], rejected{reason:n}, wallMs, and an equivalence block naming every library source it indexed. If equivalence.sources is empty every candidate is contribute by default, not by proof: say that, never “no equivalent exists”. unsearched[] names the registry libraries with no local clone (no libRoots / not cloned), and decompiledExport: true flags a project laid out like a raw decompiled export (candidates from decompiled files are rejected per file, never offered).

Flags: --config <file> (else T1K_UNITY_CONFIG, else t1k-config-unity.json found upward from the script), --registry <file>, --lib-roots a,b (or T1K_LIB_ROOTS; directories holding clones named like the registry repos), --no-write (scan without touching state).

All of: a top-level non-partial class/struct/interface/record, public or internal; outside Tests/Editor/Plugins/ThirdParty/Generated, vendor directories and vendor namespaces (config data); no package.json ancestor; >= 3 public members; 40..1500 code lines; the coupling budget; and a generic name score read from libCandidate.nameTokens (a type with no token hit needs zero coupling and >= 6 members). Exact rules, the module exemption, rejection reasons and the stage-1 score formula: references/detector.md.

Every candidate also carries facts.provenance (kind: submodule | tracked | ignored | untracked | unknown, plus ownerRepo/mount for a submodule mount) — whose repository the file belongs to. The scout reads it BEFORE judging portability: a framework mirror mounted at Assets/_Game/<Framework>/ is another repository’s code, and an ignored/untracked file cannot be contributed back as-is. It is a fact, never a filter: nothing is rejected or re-ranked for its provenance (references/detector.md → “Git provenance”).

baseline (parked by scan --all) -> pending (new or changed file) -> triaging (brief) -> triaged (ingest; verdict CONTRIBUTE / REUSE / SKIP / unverified) -> approved | dismissed | consumed. brief feeds at most maxPerBatch baseline types per UTC day, so a 3,000-file game cannot flood the steward. A dismissed fingerprint is never re-proposed unless the file changed and its rank rose. Full schema, verdict grammar and the contract core relies on: references/state-and-contract.md.

Config (all in t1k-config-unity.json, this kit owns them)

Section titled “Config (all in t1k-config-unity.json, this kit owns them)”

features.autoLibTriage, libContribute.mode (auto default), and libCandidate.*: nameTokens, minPublicMembers, minLoc, maxLoc, maxGameTypeRefs, zeroTokenMinMembers, adoptThreshold, moduleMaxShare, dataOnlyRankPenalty, maxPending, maxPerBatch, batchSize, maxWaitHours, vendorPathSegments, vendorNamespaceRoots, libRoots. Thresholds are uncalibrated first guesses; the replay on real projects tunes them. A new generic word is a config edit, never a code change.

Fleet-wide view (which types several games copied, ranked, on demand only): references/harvest.md.

Runs inline in the main session or an inline sub-agent Bash call: every subcommand is a deterministic script invocation that prints one JSON document. The judgment step (stage 2) is the detached t1k-corpus-scout hop that brief describes; this skill never spawns it.

  • Do not treat contribute as a verdict: it is a hypothesis until the scout (stage 2) confirms it.
  • Do not read rejected counts as precision; they are filter hits, not judgments.
  • Do not run it against a worktree you do not own: it reads every .cs under Assets/.