artyhoo · artyhoo · Jun 17, 2026 · Jun 17, 2026
diff --git a/docs/meta-factory/prior-art-evaluations.md b/docs/meta-factory/prior-art-evaluations.md
@@ -194,6 +194,8 @@ Each entry is a row in the table at §4 below. The row schema:
 | 121 | Shipped-agent/skill tool-name validity — the **absence** of any tool that fails CI when a shipped CC sub-agent declares a non-canonical tool name. Surveys: own-stack liveness lineage #114 (ESLint guard-liveness gate) + #115 (manual-rule-liveness-prober) + #53 (promptfoo/Inspect/METR behavioural-eval prior art); CC upstream issues #52055 (plugin subagents never receive Grep/Glob in tool schema), #18837 (`allowed-tools` in skill frontmatter not enforced), #8697 — invalid/missing subagent tool-schema is a known real failure class. WebFetch `code.claude.com/docs/en/tools-reference` 2026-06-16 (DeepWiki unavailable in env) established the canonical allow-list. No production tool validates a shipped agent's `tools:` frontmatter against the CC-canonical name set. | M1 gate: deterministic principle test `packages/core/principles/21-shipped-agent-tools-valid.test.ts` parsing every shipped `agents/*.md` `tools:` + shipped-skill `allowed-tools` (install.sh-copied surface only); asserts each entry's base name (scope-stripped, e.g. `Bash(git *)`→`Bash`) ∈ CC-canonical allow-list ∪ `^mcp__<server>__`. Closes #551 (6 shipped agents declared `read_file`/`list_files`/`run_command` → zero-tools CC dispatch → fabrication, while `npm run validate` was green). | 2026-06-16 | 2026-06-16 | BUILD | BUILD (no upstream analog): the survey found no production tool that validates a CC sub-agent's `tools:` frontmatter names and fails CI on a non-canonical entry; CC upstream issues #52055/#18837/#8697 confirm invalid/missing subagent tool-schemas are a real, untooled failure class. Same own-built liveness family as #114 (ESLint guard-liveness) and #115 (manual-rule-liveness-prober) — the framework builds the liveness gate the upstream harness does not provide. **T16 problem-class:** distinct from #53 (behavioural eval — *did an intervention change agent behaviour*) and #115 (manual-rule RED→GREEN behavioural probe) — this M1 gate is a static **FORM-check** (do the declared names *resolve*), explicitly **NOT** a behaviour-check (a valid name ≠ the agent uses the tool; the M2 dispatch-probe is Phase 2, deferred behind a trigger). Problem class = the framework's OWN shipped delivery artefacts (agents/skills) — the `#recursive-self-application-gap` #551 exposed at the delivery layer (thesis applied bottom-up to consumer code, never top-down to shipped sub-agents). Dependency-free (vitest + `node:fs` already installed); zero new deps. CC-scoped by design (only CC's Agent tool reads `tools:`); the agnostic prose body stays harness-neutral (scope decision (i)). | CC ships a built-in validator that fails on non-canonical subagent tool names → flip BUILD→ADOPT; OR the M2 behavioural dispatch-probe (Phase 2) lands, superseding the form-check with a behaviour-check; OR the CC-canonical tool list at `code.claude.com/docs/en/tools-reference` changes → re-verify the allow-list constant. |
 | 122 | `session-report` CC plugin = usage analytics (tokens/cache HTML dashboard); ecosystem session-recap skills (`annikalewis/claude-recap`, Ben Poole) = cross-session memory ("where did we leave off") | Engaging plain-language story-in-acts completion recap for the human (`/story` skill + Stop-hook branch) | 2026-06-16 | 2026-06-16 | BUILD | Composite survey verdict (operative decision = BUILD; reuse nothing). T16 problem-class check: session-report = usage/cost analytics → REJECT (different problem). claude-recap et al. = memory/continuity archaeology for the NEXT session → REFERENCE only: reusable bit is "read ~/.claude/projects/*.jsonl transcript directly", already present (end-of-turn-reminder.sh reads transcript_path). Story-in-acts style is maintainer taste → BUILD (cheap: lang-function + skill + helper + hook branch, no new dep). Survey 2026-06-16: installed companions + SSOT + WebSearch x2. | If a companion ships a human-facing engaging-completion-recap (not analytics, not memory) — re-evaluate ADOPT. |
 | 123 | Coupled internal-English-machinery gate + `AIF_HOOK_LANG`-gated human-facing output localization for an AI-agent harness — surveyed for an upstream that does the *coupled* discipline (not just one half). Search 2026-06-17: DeepWiki MCP unavailable in env (per #121 precedent) → substituted WebFetch (deepwiki.com timed out) + WebSearch ×3 («lint rule enforce English-only code comments», «i18n gate AI agent output language env var locale switch», «CI fail non-English/Cyrillic source comments») and repo-level ×3 («framework gate internal tooling one language while user-facing output locale-switched», «tool fails CI on non-English comments in scripts/hooks», «env-var-gated output-language directive for an AI agent harness»). Adjacent single-purpose tools surfaced: comment-lang-detector (GitHub Action, fails CI on configured-language comments), Cyrillic Check (cyrchk, SVN pre-commit Cyrillic scan), Pulp Fiction (PyPI non-English-comment detector), blocklint (non-inclusive wording). None couples an internal-English gate that *preserves* curated bilingual match-data with an env-gated output-language directive. | The language-discipline guard: `.claude/rules/language-discipline.md` (3-category model — machinery=English / human-facing=`AIF_HOOK_LANG`-gated / category-3 match-data=kept) + `packages/core/principles/22-internal-english.test.ts` (Surface 1 machinery-shell zero-tolerance Cyrillic scan + Surface 2 SKILL.md bodies after frontmatter, allowlisted), reusing the existing `AIF_HOOK_LANG` + `lang/*.sh` pack mechanism. | 2026-06-17 | 2026-06-17 | BUILD | BUILD for the coupled guard; the surfaced detection tools are REFERENCE-class precedent for the detection *sub-pattern* only — not adopted. **T16 problem-class:** Upstream class (comment-lang-detector / Cyrillic Check / Pulp Fiction) = «flag any non-English / configured-language characters in code comments, optionally fail CI» — a blanket language-of-comments check. Our class = «zero-tolerance internal-English over AI-agent *machinery* (hook/script logic, SKILL.md bodies) that *deliberately preserves* category-3 Russian match-data (skill `description` triggers, `AIF_EOT_*` patterns, `lang/ru.sh`) AND drives human-facing prose via `AIF_HOOK_LANG`». Match? No — upstream tools have no allowlist for legitimate-bilingual match-data (would FP-storm on category 3) and no coupling to an env-gated output directive; adopting one adds a CI dependency for what is a few-line dependency-free vitest grep (cost-gate: BUILD is «cheap» per [build-first-reuse-default.md §1.1](../../.claude/rules/build-first-reuse-default.md)). REUSE half already taken: the gate reuses `AIF_HOOK_LANG` + `lang/*.sh` rather than building i18n infra ([language-discipline.md §7](../../.claude/rules/language-discipline.md)). Same own-built liveness/discipline family as #120/#121. Refs: github.com/kubecub/comment-lang-detector; cyrchk.sourceforge.net; pypi.org/project/pulpfiction. | A production tool ships the *coupled* internal-English-gate + env-gated-output-localization discipline → flip BUILD→ADOPT/ADAPT; OR Surface 1's machinery Cyrillic scan is re-evaluated standalone — then comment-lang-detector is the ADAPT candidate to weigh against the dependency-free vitest grep; OR the category-3 keep-list / `lang/*.sh` mechanism changes shape. |
+| 124 | cruft (cruft/cruft, Python, built on Cookiecutter, stable 2026) | Consumer-side **re-sync of templated/vendored files** preserving local edits: `cruft update` diffs from the commit hash in `.cruft.json`; a **`skip` glob list** (`.cruft.json` / `pyproject.toml [tool.cruft]`) preserves consumer-owned files. | 2026-06-17 | 2026-06-17 | ADAPT | Surfaced by the consumer-upgrade-path R-phase sweep (WebSearch ×3, 2026-06-17; DeepWiki MCP unavailable in env per #121/#123 precedent — substituted WebSearch + SSOT consult). **T16 problem-class:** Upstream class = re-sync a Cookiecutter *template* into a project whose files the user is *expected to edit*. Our class = re-sync *framework-owned* artefacts the consumer is *not* expected to edit in place (divergence belongs in `<file>.override.md`). Match? **Partial** — the `skip`-glob (= our consumer-owned boundary) and the `.cruft.json` version stamp (= our optional version-awareness) are the transferable **concepts**; the Python runtime is **not** adopted (same vendoring blocker as #22). ADAPT the concepts into the existing `copy_safe` + `merge_prettierignore` (`install.sh:207-249`) machinery, dependency-free. Recommended path in [research-patches/2026-06-17-consumer-upgrade-path.md §3](research-patches/2026-06-17-consumer-upgrade-path.md). **The final ADOPT-vs-ADAPT dependency call is PARKED for the maintainer** (kickoff §9) — this row records the concept-level evaluation, not a finalised dependency decision. | A Node-consumable re-sync primitive emerges (then re-weigh ADOPT vs the dependency-free ADAPT); OR a consumer needs true 3-way merge of files that are *both* framework-owned and locally-edited (then #22's Copier `update` earns the Python cost); OR the maintainer resolves the §3 dependency fork. |
+| 125 | Copier `update` 3-way merge (copier-org/copier, Python, v9.x) — applied to the **consumer-refresh** capability (distinct from #22's template-render-replay vocabulary) | Idempotent in-place refresh of shipped framework artefacts preserving consumer customisations: 3-way merge (old-render ⊕ user-diff ⊕ new-render), state in `.copier-answers.yml`, conflict markers/`.rej`. | 2026-05-10 (as #22) | 2026-06-17 | ADAPT | Re-evaluation of #22 against the **refresh** capability (not the test-pattern vocabulary #22 adopted). #22's `Trigger to revisit` («Copier ships answer-file-replay primitive consumable from Node without Python interpreter») was checked 2026-06-17 and has **NOT fired** — Copier remains Python-only. The 3-way-merge *core* is largely unneeded for our boundary because framework-owned files are not consumer-edit targets (the `.override.md` escape hatch carries divergence) — so ADAPT the version-stamp concept at most, do not adopt the Python dep. #22 is **not** rewritten (append-only); this row is the refresh-capability-specific evaluation. Recommended ADAPT; dependency adoption (Option B) is the **PARKED** maintainer fork per [research-patches/2026-06-17-consumer-upgrade-path.md §3](research-patches/2026-06-17-consumer-upgrade-path.md). | #22's Node-consumable-primitive trigger fires; OR the recommended dependency-free ADAPT path (research-patch §3 Option A) proves insufficient for a consumer needing partial/3-way merges → ADOPT Copier and re-open the Python-dependency cost. |
 
 ---