@mmerterden/multi-agent-pipeline 13.1.0 → 13.3.0

package/pipeline/commands/multi-agent/ios-coding-standard/SKILL.md

@@ -0,0 +1,257 @@
+---
+description: "Audit an iOS module against the shared coding-standard registry (95 stable-ID rules), produce a remediation plan, then hand off to dev/dev-local. Use for a standards pass on a module, or when a review needs rule IDs rather than opinions."
+description-tr: "Bir iOS modulunu paylasilan kodlama-standardi registry'sine (95 sabit-ID kural) gore denetler, duzeltme plani cikarir, sonra dev/dev-local'e devreder. Bir modulde standart gecisi icin, ya da bir review'un gorus yerine kural ID'si istedigi durumda kullan."
+argument-hint: "[module name or path]"
+allowed-tools: Skill, Bash, Read, Edit, Write, AskUserQuestion
+---
+
+# multi-agent ios-coding-standard  -  Module audit → plan → dev handoff
+
+**Input**: $ARGUMENTS  -  optionally a module name or path. When absent, Phase 1 discovers and asks.
+
+This routine is the **procedure**. The rules live in the `ios-coding-standard` skill, whose registry is
+`references/rules.yml` and whose teaching doc is `references/STANDARD.md`. Load that
+skill first; it is installed on every host and carried by the iOS stack plugin, so
+there is one registry rather than a copy per CLI. Never restate a rule here  -  cite its ID. A rule that is not in the
+registry is not a rule; if the audit needs one, propose it as `status: proposed` and say so.
+
+Read-only + planning. This routine never edits source  -  the dev pipeline does.
+
+**Goal.** A developer new to the repo can open any file in the module and understand it without a
+guided tour, and nothing sensitive leaks on the way. Rank every finding by "does fixing this
+shorten the time to first productive PR, or close a real risk?"
+
+Render assistant-facing prose in `outputLanguage`. External payloads (branch names, plan file,
+dev task) stay English.
+
+---
+
+## Phase 1  -  Discover and pick the module
+
+1. Repo root via `git rev-parse --show-toplevel`.
+2. **Discover module roots**  -  any directory with a `Package.swift`, a `Sources/` subtree or an
+   `.xcodeproj`. Sweep the container dirs that exist (`Domains/`, `Packages/`, `Modules/`,
+   `Features/`, `Core/`, `Common/`, top-level module dirs) one level deep, then two if empty.
+3. Record per module: name · path · Swift files · lines · governance docs present · SPM targets.
+   Keep this **module registry** for the whole run  -  `MOD-*` greps every import against it.
+   Classify each module's role, because the role decides which dependency edges are legal:
+   **feature** · **core/common** · **seam** (cross-module contracts/bridges/navigation) ·
+   **composition root** (legitimately knows every module).
+4. If `$ARGUMENTS` resolves to a module, skip the prompt. Otherwise `AskUserQuestion`
+   (single-select, `outputLanguage`), documented modules first, each option showing file/line
+   counts and `docs: yes/no`. If the list exceeds the question limit, group the smallest into an
+   "other" option and ask again  -  never truncate silently.
+5. Scope = the whole module's sources across all its targets, unless the user narrows it.
+
+## Phase 2  -  Resolve what applies to THIS module
+
+### 2a. The registry is the standard  -  in-module prose docs are NOT consulted
+
+The skill's `references/rules.yml` (+ a project `modules/<Module>.yml` overlay) is the single source of truth. **Do not read, cite, or
+derive rules from a module's own `*-CLAUDE.md`, `docs/` set or any other in-repo prose.** They are
+being retired precisely so there is one place a rule can live; consulting them re-creates the
+split this command exists to remove. If such a file exists and contradicts the registry, that is a
+finding against the file, not against the code.
+
+The only in-repo inputs are **code and manifests**: sources, tests, `Package.swift`, `Info.plist`,
+the specs the generator reads, and the module registry from Phase 1.
+
+### 2b. Per-module overlay, then inference
+
+1. **`modules/<Module>.yml`** beside the project's own config carries what code cannot state: the vocabulary
+   bindings, the module's role, prohibitions that currently have **zero** instances, deliberate
+   carve-outs, name locks, and the real verification command. Registry-backed, so its findings may
+   be `blocking`.
+   *Why prohibitions need to be written down:* inference reads dominant patterns, and a rule
+   obeyed everywhere has no counter-example to infer from. A ban at 100% compliance is invisible
+   to a scan  -  it survives only if the overlay states it.
+2. **Repo-level rules** (root `CLAUDE.md`, contributing docs) apply to every module.
+3. **The module's own dominant pattern is its de-facto convention.** Count variants per dimension;
+   the minority instances are the finding. Internal consistency beats conformity to a sibling.
+4. No majority (even split, single instance) → not a finding. Record an Open Question.
+
+Findings from step 3 are tagged `inferred` and capped at **suggestion**. Only registry- or
+overlay-backed rules may be `blocking`.
+
+**When the overlay is missing**, generate a draft from code evidence, print it for confirmation,
+and mark every slot it could not derive  -  a frozen UI target, a name lock, a carve-out and an
+operational constraint all look like ordinary code from the outside. Do not guess them; list them
+as Open Questions.
+
+### 2c. Vocabulary
+
+Bind each slot to a real symbol by grepping the module, then the reference module. **An unbound
+slot disables its rules**  -  never invent a name, never import a sibling's.
+
+`HandlerName` · `EventParam` · `CoordinatorType` · `ResultType` / `ErrorType` / `ErrorFactory` ·
+`DIResolver` / `DIConfigurator` · `RequestModelSuffix` / `ResponseModelSuffix` / `TransportSuffix` ·
+`MapperShape` · `SafeEnumProtocol` / `UnknownCase` · `CopySurface` · `AnalyticsSurface` ·
+`A11yIdentifierSource` · `GeneratedServiceRoot` · `SharedComponentsDir` / `SharedEntitiesDir` ·
+`MicroComponentLibrary` · `FrozenUITarget` · `DesignTokenNamespaces` · `CalendarDayHelper` ·
+`LoggerAPI` · `FileHeaderShape` · `ScreenRoot` · `ScreenRoleSuffixes` · `AsyncStyle` ·
+`CompositionRoot` · `SeamLayer` · `ModuleEntrySurface` · `ConcurrencyPosture` · `LintToolchain` ·
+`CredentialStore` (the module's Keychain wrapper, if any).
+
+Print the resolved table before scanning so a wrong binding is caught early.
+
+### 2d. Sensitive-data inventory  -  required before any SEC rule runs
+
+The `SEC-*` rules are written against the **data classes** in `references/rules.yml →
+sensitive_data_classes`, never against one module's field names. Resolve the module's concrete
+instances of each class: grep entities, request/response models, storage calls and analytics
+events for the values the module actually handles, and map each to its class.
+
+Produce the inventory as a table  -  **class · concrete symbols · where it enters · required
+lifetime (transient / survives-restart / survives-reinstall) · where it is actually stored ·
+where it is logged · where it leaves (network, analytics, pasteboard, another module)**.
+
+The **lifetime column is the one that decides SEC-01**, and it is filled by walking
+`references/rules.yml → persistence_decision`, not by looking at what the code currently does. Compare it
+against the storage column and report both mismatches:
+- required transient but persisted anywhere (including the Keychain) → **over-persistence**;
+- required to survive a restart but stored outside the Keychain → **under-protection**.
+
+Then generate the SEC lint patterns *from this inventory*, so the checks match this module's real
+vocabulary. A class with no instance here yields no finding; a symbol you cannot classify, or one
+whose required lifetime nobody can state, is an Open Question, not a silent pass.
+
+This table is a deliverable in its own right, independent of whether any violation is found.
+
+## Phase 3  -  Scan
+
+Apply the rules in `references/rules.yml` that survived Phase 2 binding, plus the module's own `validation`
+gate. Tag every finding: **rule ID · severity · source (`registry` / `doc` / `inferred`) ·
+onboarding impact (high/medium/low)**.
+
+### Coverage gate  -  on disk, not in memory
+
+Before scanning, write `.<module>-audit-coverage.tsv` to the scratch dir: one row per target
+(every screen under `ScreenRoot`, every shared component, every entity, every file in every
+target) with columns `path · status · rules-applied`. Mark rows as you go.
+
+- Each target is **audited** or **explicitly skipped with a reason** (generated, fixture,
+  vendored). No third state.
+- **The run is incomplete while any row is unmarked**  -  and the file, not a claim, is the proof.
+  Chunk the module if it is too large for one pass; the checklist survives a context reset.
+- No sampling. A partial audit reported as complete certifies the unscanned screens as clean.
+
+### Grep hygiene
+
+Registry `mechanism` patterns are line-based and match doc comments. Filter
+`^[[:space:]]*//`  -  BSD grep does not understand `\s`, so a `\s*//` filter silently lets `/// ...`
+through  -  then open each surviving hit and confirm it is real code. Report pre-filter and
+post-verification counts separately; a table padded with comment matches destroys trust.
+
+### Judgement rules need evidence, not assertion
+
+For any rule marked `enforcement: judgement`, a finding must carry the measurement its `check`
+names  -  a reference count, a call-site count, a consumer count, a file list. **No count, no
+finding.** This is what keeps a subjective rule from becoming an opinion.
+
+Every judgement rule has a worked ✗/✓ pair in `EXAMPLES.md`, keyed by ID. Cite it in the finding's
+proposed fix rather than re-describing the shape  -  the developer applying the fix should be
+reading the same picture the audit used.
+
+### Module-specific overrides
+
+The module's own docs win over the registry, including where they contradict a sibling module.
+Never apply one module's dialect to another. List every override honoured, so the reader sees why
+a registry rule was not raised.
+
+## Phase 4  -  Produce the plan
+
+No findings → report compliant and stop. Otherwise write
+`<ModulePath>/ios-coding-standard-plan-<module>.md` (or the repo's scratch/docs dir if the module
+tree must stay clean), containing, in order:
+
+1. **Coverage ledger**  -  total targets · audited · skipped with reasons. Not 100% accounted for
+   means the plan is not deliverable.
+2. **Summary**  -  counts by severity and source, resolved vocabulary, branch to create.
+3. **Sensitive-data inventory** (Phase 2d)  -  first, because it is the highest-risk content.
+4. **Screen × role matrix** `[STRUCT-02]`  -  screens as rows, manifest roles as columns, cells
+   `✓ / missing / misplaced`. The structural map a newcomer reads first.
+5. **Type placement table** `[STRUCT-05]`  -  type · location · consumer count · correct tier ·
+   move required. Paired with the nested-type list `[STRUCT-01]`: owner · nested type · kind ·
+   reference count · verdict.
+6. **Visibility report** `[VIS-01, VIS-02]`  -  over-exposed declarations per file, non-final
+   classes with no subclass.
+7. **Boundary report** `[MOD-*]`  -  import graph (allowed vs forbidden, with coupling depth) ·
+   manifest graph (declared vs legal for the role, plus dead edges) · **removability delta**
+   ("removing X touches N files: ...") · inbound test (tests compile with no sibling present).
+8. **Testability seam report** `[TEST-*]`  -  types reaching for the environment, logic requiring a
+   view to execute, doubles whose signatures have drifted.
+9. **Change-cost probe** `[FLEX-*]`  -  pick two plausible upcoming changes (a new variant of an
+   existing screen; a new field on a shared entity) and state what each costs in files today.
+10. **Findings table**  -  rule ID · `file:line` · current state · proposed fix · severity · source ·
+    onboarding impact.
+11. **Per-file work**  -  grouped per file, blocking → important → suggestion; within a severity,
+    highest onboarding impact first.
+12. **Behaviour-preservation guard**  -  splitting files, adding MARKs, extracting extensions and
+    moving types must not change behaviour. Anything requiring a behaviour change (a missing
+    analytics event, a wrong result type, a timezone fix, a storage migration) goes in a
+    **separate group**, reviewed and tested on its own, never inside a mechanical split commit.
+13. **Tooling delegation & the ratchet**  -  the section that decides whether any of this survives.
+    - Classify each applied rule by its registry `enforcement`. For every `lint` / `format` /
+      `scan` rule, state the concrete mechanism and whether that toolchain exists in the repo
+      (`LintToolchain`). **When it does not exist, saying so is a finding**, ranked above most
+      individual violations it would have caught.
+    - Recommend **baseline-and-ratchet, never big-bang**: grandfather existing violations, surface
+      only new ones, drain the backlog behind a green build. Order the plan the same way  -  stop
+      the bleeding, then clean up.
+    - Custom lint rule identifiers **are** the registry IDs (`sec_01_no_plaintext_persistence`,
+      `mod_01_no_sibling_import`) so a violation points straight at the rule entry.
+    - **Linting is local-only by decision**  -  the skill's `references/lint-local.sh` with `references/swiftlint.draft.yml`, config
+      and baseline kept outside the repository; no committed config, no build phase, no CI job.
+      State this honestly in the plan rather than describing a gate that does not exist: a
+      pre-PR habit catches less than a wall, so the residual risk is real and belongs in the
+      report. Adopting it project-wide remains an open decision  -  record it as one so it stays
+      visible instead of quietly lapsing.
+    - Never imply the audit substitutes for the linter, or the linter for a gate. The audit finds
+      the debt once; the linter keeps a developer honest; only a gate stops it returning.
+14. **Numeric exit criteria**  -  target and today's measured value side by side: lint violations
+    per 1k lines (<1) · public-surface ratio, public declarations ÷ externally consumed symbols
+    (≈1) · forbidden cross-module imports (0) · removability delta · files over threshold and
+    files >120 lines with no sectioning (0) · screens missing a manifest role (0) · escape-hatch
+    count `[CONC-03]` (flat or falling) · exception-marker count (flat or falling) · dead code (0)
+    · screens with logic and no mirrored tests (0) · unclassified sensitive symbols (0).
+15. **Exception ledger**  -  every `// standard:exception(<ID>)` in the module: rule · reason ·
+    expiry · expired yes/no.
+16. **Open Questions**  -  Phase 2b dimensions with no majority, and unclassifiable symbols.
+17. **Suggested commit slicing**  -  one commit per file-group / concern.
+18. **Scope guard**  -  only this module is touched; definition of done is the module's own
+    verification path.
+
+Then write a **separate one-page onboarding summary** beside the plan, aimed at a developer
+joining next week rather than at whoever does the remediation: the ten always/never lines
+specialised to this module, the screen file manifest, the placement ladder, the sensitive-data
+classes in play, and the verification command. Nobody learns a codebase from a 40-finding audit -
+this page is the actual deliverable, and the audit is how you earn the right to write it
+accurately. Derive it from `references/STANDARD.md section 0`, do not re-invent it.
+
+Show a concise version of the plan to the user too.
+
+## Phase 5  -  Hand off
+
+`AskUserQuestion` (single-select, `outputLanguage`):
+
+| Option | What it does |
+|---|---|
+| `/multi-agent:dev-local` | No worktree; branches from the main development branch, fixes on the current checkout |
+| `/multi-agent:dev` | Opens a worktree; Opus fixes on a fresh branch + PR |
+
+Invoke the chosen command with the plan file as the task input, branching off the repo's main
+development branch (e.g. `chore/<module>-coding-standard`). The dev pipeline applies the fixes,
+verifies, and commits only when asked. This routine ends at the handoff.
+
+## Notes
+
+- **Verification reality is per-module  -  establish it, do not assume it.** Some targets cannot be
+  compiled from the CLI (asset symbol generation, UIKit-dependent packages, a dependency failing
+  earlier in the graph); there, verification is Xcode and a manifest dump validates the manifest
+  only. Never trust a piped build's exit code  -  a pipe reports the tail's status. Record the
+  module's actual verification command in the plan.
+- One module per run. Several modules means several runs and several plans, never one
+  cross-module churn commit.
+- If the audit wants a rule the registry lacks, add it to the skill's `references/rules.yml` as `status: proposed` with a
+  rationale and surface it in Open Questions. Do not enforce an unregistered rule.
+- No AI attribution anywhere (code, commits, PRs). Author = git identity.

package/pipeline/commands/multi-agent/sync/SKILL.md

@@ -59,8 +59,8 @@ Run every step automatically:
 ```
 Step 1: PLATFORM    Detect macOS / Linux / Windows (Git Bash / WSL); export PLATFORM env
 Step 1.5: DETECT    Compare timestamps, find stale targets
-Step 2: COPILOT     Claude Code -> Copilot CLI (instructions + 43 sub-command skills)
-Step 2b: CODEX      Claude Code -> Codex CLI (1 router skill + 43 specs as refs + 8 agent TOML)
+Step 2: COPILOT     Claude Code -> Copilot CLI (instructions + 44 sub-command skills)
+Step 2b: CODEX      Claude Code -> Codex CLI (1 router skill + 44 specs as refs + 8 agent TOML)
 Step 3: REPO        Claude Code -> pipeline repo (genericized, personal data scrub, bash -n on all sh)
 Step 3c: PLUGINS    pipeline shared/external -> multi-agent-plugins marketplace (rebuild knowledge/,
                      bump changed plugins' patch version, commit + push the plugins repo)
@@ -486,12 +486,12 @@ same 43 specs as reference files rather than as peer skills, via Step 2b -  see
 |-------------|-------------|
 | `~/.claude/commands/multi-agent/{cmd}/SKILL.md` | `~/.copilot/skills/multi-agent-{cmd}/SKILL.md` |
 
-**43 commands are synced** (canonical inventory  -  must match `cross-cli-contract.md` section 1; drift = contract violation):
+**44 commands are synced** (canonical inventory  -  must match `cross-cli-contract.md` section 1; drift = contract violation):
 
 ```
 analysis, analysis-resolve, autopilot, build-optimize, channels, create-jira, design-check, dev,
 dev-autopilot, dev-local, dev-local-autopilot, diff-explain, finish, forget, garbage-collect,
-help, issue, jira, kill, language, local,
+help, ios-coding-standard, issue, jira, kill, language, local,
 local-autopilot, log, manual-test, prune-logs, purge, refactor, resume, review, review-issue, review-jira,
 routines, save, scan, search, setup, stack, status, sync, test, testflight-validation, uninstall, update
 ```

package/pipeline/multi-agent-refs/cross-cli-contract.md

@@ -6,12 +6,12 @@
 
 ---
 
-## 1. Command Inventory (43 commands)
+## 1. Command Inventory (44 commands)
 
 ```
 analysis, analysis-resolve, autopilot, build-optimize, channels, create-jira, design-check, dev,
 dev-autopilot, dev-local, dev-local-autopilot, diff-explain, finish, forget, garbage-collect,
-help, issue, jira, kill, language, local,
+help, ios-coding-standard, issue, jira, kill, language, local,
 local-autopilot, log, manual-test, prune-logs, purge, refactor, resume, review, review-issue, review-jira,
 routines, save, scan, search, setup, stack, status, sync, test, testflight-validation, uninstall, update
 ```
@@ -24,7 +24,7 @@ Categories:
 - **Fast modes** (Init -> Dev(Opus) -> Commit -> Report): `dev`, `dev-autopilot`, `dev-local`, `dev-local-autopilot`
 - **Tail modes** (run the pipeline tail over already-done local work): `finish`
 - **Ops commands** (one-shot, no worktree): `status`, `log`, `kill`, `purge`, `uninstall`, `resume`, `review`, `review-jira`, `review-issue`, `analysis`, `analysis-resolve`, `build-optimize`, `channels`, `scan`, `search`, `diff-explain`, `garbage-collect`, `prune-logs`
-- **Local audits** (worktree only to build; no commit, push, PR or channels): `design-check`, `testflight-validation`. `testflight-validation` additionally never invokes `altool --upload-app`  -  a validation run must not be able to ship a build by accident.
+- **Local audits** (worktree only to build; no commit, push, PR or channels): `design-check`, `testflight-validation`, `ios-coding-standard`. `testflight-validation` additionally never invokes `altool --upload-app`  -  a validation run must not be able to ship a build by accident.
 - **Meta-ops**: `setup`, `sync`, `update`, `help`, `refactor`, `test`, `stack`, `manual-test`, `language`
 - **Routines** (user-defined routine registry; the routines they create are local-only and never synced): `save`, `routines`, `forget`
 
@@ -34,7 +34,38 @@ Categories:
 
 Component work is **no longer a bundled pipeline skill set**. Claude Code dispatches `taskType === "component"` to the enabled `ai-<platform>-engineering-toolkit` **marketplace plugin** (`create-component`, fallback `create-ui-component`) via the Skill tool  -  full contract in `$HOME/.claude/multi-agent-refs/component-dispatch.md`. The pipeline no longer ships `pipeline/skills/figma-ios|figma-android|figma-common`; the pipeline-unique component skills (iterate loops, performance harness, validate/review, commit, adapters, wiki) were absorbed into the `ai-ios-engineering-toolkit` plugin so component skills live in one place.
 
-**Intentional cross-CLI divergence.** Copilot CLI has no marketplace-plugin loader, so it retains its own standalone `~/.copilot/skills/figma-*` copies as a frozen fallback. Component/figma dispatch is therefore **not** byte-identical across CLIs, and figma-skill *inventory* parity is **no longer enforced**  -  this is a deliberate exception to the parity rule, not drift.
+**How each host receives the plugin's skills.** Only Claude Code loads the marketplace
+plugin natively. The other two are served by the installer, so all three end up with the
+same active skill set:
+
+| Host | Delivery | Discovery |
+|---|---|---|
+| Claude Code | the marketplace plugin, loaded natively | matched on description |
+| Copilot CLI | `install/copilot.mjs` copies the **enabled** stack plugin's authored skills (`index`, `reference/`, `workflow/`, `tools/`) into `~/.copilot/skills/` | matched on description |
+| Codex CLI | `install/codex.mjs` copies the same set **plus** `skills/shared/external` into `~/.codex/multi-agent-refs/skills/` as reference files | located by path; the AGENTS.md block carries the `ls` / `grep` recipe |
+
+Three rules make that work:
+
+1. **`knowledge/` is never copied to Copilot.** A stack plugin's `knowledge/` tree is
+   generated from this pipeline's `skills/shared/external`, which the installer already
+   lays down there. Copying it again would duplicate ~5 MB the host has.
+2. **Only the ENABLED plugins are delivered.** `fix-bug` and `branch-and-pr` exist in
+   four stack plugins, `component`, `state` and `create-component` in three, and the
+   destination is flat  -  delivering every stack would be last-write-wins, so an iOS
+   repo could end up running the Android `create-component`. The enabled list in
+   `~/.claude/settings.json` is the source of truth, exactly as on Claude Code;
+   `--platform` is the fallback when no such list exists.
+3. **Codex takes refs, not list entries.** Measured on 0.145: one plugin declaring 142
+   skills surfaced 75 and evicted an unrelated user skill, and every plugin-provided
+   skill renders with an empty description, so a block entry buys nothing even when it
+   fits. Parity on that host is **reachability**, not an identical listing.
+
+> This section previously claimed Copilot "retains its own standalone
+> `~/.copilot/skills/figma-*` copies as a frozen fallback". `install/copilot.mjs` pruned
+> exactly those directories, so the fallback did not exist: `create-screen`,
+> `figma-validate`, `figma-review`, `component`, `state` and `navigation` were all
+> absent, and a component task on Copilot had nothing to dispatch to. The contract
+> documented a safety net the installer deleted.
 
 What still MUST match across CLIs for component tasks: the `taskType === "component"` classification, the `state.phases["3"].subphases[]` shape the dispatch layer writes, and `--dev` elision semantics. Resolution/routing is Claude-plugin vs Copilot-local-copy by design.
 

package/pipeline/scripts/build-stack-plugins.mjs

@@ -35,9 +35,14 @@ import {
 import { join } from "node:path";
 
 const args = process.argv.slice(2);
+// Accepts both `--key value` and `--key=value`. The `=` form used to fall through to
+// the default, which meant a run meant for a throwaway checkout silently retargeted
+// the user's real marketplace instead - a wrong-target write, reported as success.
 const getArg = (k, d) => {
+  const eq = args.find((a) => a.startsWith(`${k}=`));
+  if (eq) return eq.slice(k.length + 1);
   const i = args.indexOf(k);
-  return i >= 0 ? args[i + 1] : d;
+  return i >= 0 && args[i + 1] && !args[i + 1].startsWith("--") ? args[i + 1] : d;
 };
 const HOME = process.env.HOME;
 const PLUGINS_REPO = getArg("--plugins-repo", join(HOME, "multi-agent-plugins"));
@@ -102,6 +107,29 @@ const isDir = (p) => {
 };
 const listDirs = (p) => (existsSync(p) ? readdirSync(p).filter((n) => isDir(join(p, n))) : []);
 
+/**
+ * Cheap content fingerprint for a skill directory: every file's relative path and
+ * size. Enough to notice an edit without hashing megabytes on every run, and it is
+ * only used to decide whether a version bump is warranted.
+ */
+function treeFingerprint(dir) {
+  const parts = [];
+  const walk = (d, rel) => {
+    for (const e of readdirSync(d, { withFileTypes: true }).sort((a, b) => a.name.localeCompare(b.name))) {
+      const full = join(d, e.name);
+      const r = rel ? `${rel}/${e.name}` : e.name;
+      if (e.isDirectory()) walk(full, r);
+      else if (e.isFile()) parts.push(`${r}:${statSync(full).size}`);
+    }
+  };
+  try {
+    walk(dir, "");
+  } catch {
+    return "";
+  }
+  return parts.join("|");
+}
+
 // --- 1. enumerate authoring source ---------------------------------------
 if (!existsSync(EXTERNAL)) {
   console.error(`source not found: ${EXTERNAL}`);
@@ -180,13 +208,31 @@ for (const [plugin, want] of Object.entries(desired)) {
   const have = new Set(listDirs(kdir));
   const toAdd = [...want].filter((s) => !have.has(s)).sort();
   const toRemove = [...have].filter((s) => !want.has(s)).sort();
-  const changed = toAdd.length > 0 || toRemove.length > 0;
+  const setChanged = toAdd.length > 0 || toRemove.length > 0;
 
-  if (!DRY && changed) {
+  // Re-copy EVERY wanted skill, not just the newly added ones.
+  //
+  // This used to copy only on a set change, which quietly broke the single-source
+  // claim: editing a skill in shared/external left the plugin serving the old text
+  // forever, because the set had not changed. Found by editing a promoted skill and
+  // seeing the plugin copy stay stale while the generator reported "all plugins up
+  // to date".
+  //
+  // cpSync overwrites, so a full re-copy is idempotent and cheap relative to a
+  // publish. Removals still need the explicit delete.
+  let contentChanged = false;
+  if (!DRY) {
     if (!existsSync(kdir)) mkdirSync(kdir, { recursive: true });
-    for (const s of toAdd) cpSync(join(EXTERNAL, s), join(kdir, s), { recursive: true });
+    for (const s of want) {
+      const from = join(EXTERNAL, s);
+      const to = join(kdir, s);
+      const before = existsSync(to) ? treeFingerprint(to) : "";
+      cpSync(from, to, { recursive: true });
+      if (treeFingerprint(to) !== before) contentChanged = true;
+    }
     for (const s of toRemove) rmSync(join(kdir, s), { recursive: true, force: true });
   }
+  const changed = setChanged || contentChanged;
 
   // regenerate plugin.json skills[]: keep non-knowledge entries, rebuild knowledge set
   const pj = JSON.parse(readFileSync(pjPath, "utf8"));