diff --git a/dist/index.cjs b/dist/index.cjs
index ed1ef8a..5f8579b 100755
--- a/dist/index.cjs
+++ b/dist/index.cjs
@@ -22464,6 +22464,18 @@ async function collectResourceEntries(sourceRoot, resourceName, entries, sourceE
   }
   await walk(resourceRoot, resourceName);
 }
+async function readBundledCodexMetadata(sourceRoot) {
+  const metadataPath = import_node_path3.default.join(sourceRoot, "agents", "openai.yaml");
+  let metadataStat;
+  try {
+    metadataStat = await fsp.lstat(metadataPath);
+  } catch (error) {
+    if (error.code === "ENOENT") return null;
+    throw error;
+  }
+  assertRealEntry(metadataStat, metadataPath, "file");
+  return fsp.readFile(metadataPath);
+}
 function manifestEntry(entry) {
   if (entry.type === "directory") {
     return { path: entry.relativePath, type: entry.type, mode: entry.mode };
@@ -22515,15 +22527,7 @@ async function buildProjection(host, skill) {
     mode: sourceStat.mode & 511,
     content: sourceContent
   }];
-  if (host === "codex") {
-    entries.push({ type: "directory", relativePath: "agents", mode: 493 });
-    entries.push({
-      type: "file",
-      relativePath: import_node_path3.default.join("agents", "openai.yaml"),
-      mode: 420,
-      content: Buffer.from(generated.openaiYaml)
-    });
-  }
+  let codexMetadata = host === "codex" ? Buffer.from(generated.openaiYaml) : null;
   let packageDigest;
   if (import_node_path3.default.basename(sourcePath) === "SKILL.md") {
     const sourceRoot = import_node_path3.default.dirname(sourcePath);
@@ -22532,12 +22536,24 @@ async function buildProjection(host, skill) {
       throw new Error(`Source skill package not found: ${sourceRoot}`);
     }
     packageDigest = completePackage.digest;
+    if (host === "codex") {
+      codexMetadata = await readBundledCodexMetadata(sourceRoot) ?? codexMetadata;
+    }
     for (const resourceName of RESOURCE_DIRECTORIES) {
       await collectResourceEntries(sourceRoot, resourceName, entries, sourceEntries);
     }
   } else {
     packageDigest = digestEntries(sourceEntries).digest;
   }
+  if (host === "codex") {
+    entries.push({ type: "directory", relativePath: "agents", mode: 493 });
+    entries.push({
+      type: "file",
+      relativePath: import_node_path3.default.join("agents", "openai.yaml"),
+      mode: 420,
+      content: codexMetadata
+    });
+  }
   const rendered = digestEntries(entries);
   const sourceIdentity = resolveSourceIdentity(skill.source);
   return {
diff --git a/docs/swe-compliance/2026-09-02-codex-skill-metadata-sync.md b/docs/swe-compliance/2026-09-02-codex-skill-metadata-sync.md
new file mode 100644
index 0000000..cb91b84
--- /dev/null
+++ b/docs/swe-compliance/2026-09-02-codex-skill-metadata-sync.md
@@ -0,0 +1,70 @@
+## Phase 0: Baseline And Manual Lookup
+
+- Durable issue: `learnrudi/cli#42`, “Preserve bundled Codex metadata during skill sync.”
+- Scope: make Codex native-skill reconciliation preserve a canonical package's `agents/openai.yaml` verbatim, including `policy.allow_implicit_invocation: false`, while retaining generated metadata as the fallback when the canonical file is absent.
+- Base: isolated worktree `/Users/hoff/RUDI/worktrees/cli/issue-42-codex-skill-metadata`, branch `fix/42-codex-skill-metadata-sync`, based on `origin/main` at `aad3a1d38745eaf52b1623fe92b79123ae72b6cb`.
+- Relevant standards: Master Engineering Doctrine; Testing Doctrine and Agent-Assisted Red-Green-Refactor; debugging doctrine; horizontal engineering review standard; repository `AGENTS.md` native-skill lifecycle and generated-artifact requirements.
+- Initial evidence: the Registry's canonical `codex-tasks/agents/openai.yaml` contains the non-implicit policy, while current `src/native-skills/lifecycle.js` always renders `agents/openai.yaml` from `buildCodexSkillFiles`; forced reconciliation therefore replaces the authoritative metadata with the generated subset.
+- Baseline focused test setup initially failed because the isolated worktree had no workspace links (`ERR_MODULE_NOT_FOUND: @learnrudi/utils`). `pnpm install --frozen-lockfile` restored only lockfile-declared workspace dependencies; this is setup evidence, not the behavior-level red result.
+- Risk tier: medium. The change affects persistent, user-visible native host metadata and invocation policy, but is restricted to one projection file and has an explicit fallback.
+- Exit criteria: current owner, failing boundary, scope, invariants, proof commands, and authority are recorded before production edits.
+
+## Phase 1: Scope Lock
+
+- In scope: `src/native-skills/lifecycle.js`, one focused regression in `src/__tests__/unit/native-skill-lifecycle.test.js`, tracked `dist/index.cjs`, and this evidence record.
+- Non-goals: changing canonical package schema or Registry contents; projecting arbitrary `agents/` content to other hosts; changing parsers, dependencies, receipts, CLI syntax, install behavior, release state, live user projections, admin-Mac state, or unrelated saved work.
+- Invariants: an existing canonical Codex metadata file is byte-authoritative; absent metadata uses the existing generated fallback; complete-package validation still rejects symlinks; only Codex receives `agents/openai.yaml`; source packages are never mutated; drift and force semantics remain unchanged.
+- Trust boundaries: canonical package paths and file types are untrusted. The complete-tree inspection must validate the package before authoritative metadata is read, and no symlink may be followed.
+- Designed failure: unsafe or unreadable package metadata fails the projection as the existing lifecycle does for unsafe package content; no partial target promotion is permitted.
+- Horizontal scan: current `main` centralizes native projection in `src/native-skills/lifecycle.js`; `src/commands/skills.js` is an adapter and no competing Codex renderer exists. Disposition: no broader obligation; fix the single owner and preserve the existing portable-host boundary.
+- Authorized actions: public issue, isolated issue branch/worktree, checklist, implementation, independent review, coherent commits, push, and pull request. Stop before merge. Release, product install/activation, live projection changes, admin sync, and unrelated cleanup remain unauthorized.
+- Commit plan: behavior source/test commit; dedicated generated-dist commit; final evidence/checklist commit. Every commit references issue #42.
+- Exit criteria: only the listed behavior and files are admitted, with no dependency or schema changes.
+
+## Phase 2: Red Test
+
+- Behavior: after a Codex projection is drifted and exact force is supplied, reconciliation must restore the canonical `agents/openai.yaml` byte-for-byte, including `policy.allow_implicit_invocation: false`.
+- Red command: `node --test --test-name-pattern='preserves bundled Codex metadata verbatim' src/__tests__/unit/native-skill-lifecycle.test.js`.
+- Red result: failed 0/1 at the strict content assertion. Actual metadata was the generated interface-only YAML; expected metadata also contained the canonical policy with `allow_implicit_invocation: false`. This was the intended behavioral failure, not a setup or syntax failure.
+- Exit criteria: met; the failing assertion and unchanged rerun command were recorded before production edits.
+
+## Phase 3: Implementation
+
+- Implementation rule: after the existing complete-package inspection validates a bundled skill, use its regular-file `agents/openai.yaml` bytes for the Codex projection when present; otherwise use `buildCodexSkillFiles` exactly as before.
+- Boundary rule: do not add `agents` to the portable resource-directory allowlist, because that would project Codex-only metadata to Claude, Gemini, and Antigravity.
+- Implemented boundary: `readBundledCodexMetadata` accepts only a real regular file and is called only after complete-tree inspection. The projection uses canonical bytes when available and retains generated bytes otherwise.
+- Exit criteria: met; the smallest lifecycle change made the unchanged focused test pass without changing source safety or fallback generation.
+
+## Phase 4: Green Tests And Refactor
+
+- Green command: the unchanged Phase 2 command passed 1/1.
+- Adjacent regression command: `node --test src/__tests__/unit/native-skill-lifecycle.test.js src/__tests__/unit/skills-sync.test.js` passed 36/36.
+- Refactor constraint: no unrelated lifecycle restructuring; retain generated metadata as a directly tested fallback.
+- Exit criteria: met; focused and adjacent tests pass, the diff remains narrow, and `git diff --check` passes.
+
+## Phase 5: Full Verification
+
+- Full suite: `pnpm test` passed 788/788 with zero failures.
+- Build: `pnpm build` passed twice; the second build reproduced `dist/index.cjs` at SHA-256 `513d9f274fe28346f18ed55ab82894f3de36d6e2a173119da9c5485c6e3f4a24`.
+- Repository debt gate: `node scripts/agent-debt-runner.mjs --changed-since origin/main --no-log` passed with zero findings.
+- SWE debt scan: the configured `pr-review` scan through `stack:swe-engineering` passed with zero error, warning, or informational findings.
+- Package gate: `npm pack --dry-run --json` passed for six files at version 1.10.26 (328,708 packed bytes; 1,586,154 unpacked bytes).
+- Integration proof: an isolated temporary-root projection of the Registry's actual `codex-tasks` bundle was created, deliberately drifted, and force-reconciled. Result: `metadataMatches=true` and `policyPreserved=true`. No live install or host mutation occurred.
+- Smoke harness correction: the first probe used a string source identity that the command correctly filters as externally owned, so it selected zero skills and the harness later hit `ENOENT`. The rerun used the installed Registry object-source shape and passed; no production change was made for the harness error.
+- Independent review: fresh-context `rudi-code-review` returned Standards pass, Spec pass, Proof pass, and Overall pass with no P0-P3 findings. The reviewer independently reproduced red 0/1, unchanged green 1/1, adjacent 36/36, full 788/788, zero debt, six-file pack, deterministic bundle hash, actual Registry smoke, and unsafe metadata probes.
+- Residual review gate: human review of the metadata-precedence boundary remains required before merge. This issue loop stops before merge.
+- Exit criteria: met for implementation proof; all gates pass with no accepted implementation debt or proof gap.
+
+## Phase 6: Docs, Contracts, And Closure
+
+- Public ledger: issue `learnrudi/cli#42` and its implementation comment point to this checklist and the current lifecycle ownership boundary.
+- Pull request: `learnrudi/cli#43` includes `Fixes #42`, medium risk, invariants, proof, this checklist path, the independent-review verdict, and the stop-before-merge boundary.
+- CI/review: required `quality` CI passed in 37 seconds on the implementation/evidence head. This final ledger-only commit must receive the same remote gate; its result is recorded on the PR and issue rather than creating another self-referential checklist commit.
+- Commit ledger: behavior source/test commit `07e7aa1`; dedicated generated-dist commit `add3f73`; initial evidence commit `e8d6240`; final ledger commit is the commit containing this paragraph.
+- Saved-work preservation: the unrelated Registry worktree remains untouched. The earlier duplicate CLI patch in `/Users/hoff/RUDI/apps/platform/cli` may be removed only after this branch safely contains and publishes the accepted fix.
+- Administrative closeout: Repo Steward records the non-mutating retained-worktree receipt only after the final commit is pushed, so the receipt can bind the exact published HEAD. Its identifier and version are posted to the issue ledger; no cleanup or archive action is authorized.
+- Publication state: branch and PR are published; merge, release, installation, activation, live host reconciliation, admin-Mac synchronization, and branch/worktree cleanup were not performed.
+- Human gate: human review of the metadata-precedence boundary remains required before merge.
+- Accepted debt: none.
+- Proof gaps: none for the implementation or pre-merge delivery. Human merge review is an intentionally outstanding approval gate, not missing implementation proof.
+- Final verdict: PASS for the authorized issue-loop boundary; stop before merge.
diff --git a/src/__tests__/unit/native-skill-lifecycle.test.js b/src/__tests__/unit/native-skill-lifecycle.test.js
index c055412..7a46dc9 100644
--- a/src/__tests__/unit/native-skill-lifecycle.test.js
+++ b/src/__tests__/unit/native-skill-lifecycle.test.js
@@ -75,6 +75,50 @@ test('reconcileNativeSkill creates a complete Codex tree and ownership receipt',
   }
 });
 
+test('forced reconciliation preserves bundled Codex metadata verbatim', async () => {
+  const state = fixture();
+  try {
+    const canonicalMetadata = [
+      'interface:',
+      '  display_name: Demo Skill',
+      '  short_description: Canonical metadata fixture',
+      '  default_prompt: Run the canonical metadata fixture.',
+      '',
+      'policy:',
+      '  allow_implicit_invocation: false',
+      '',
+    ].join('\n');
+    const agentsDir = path.join(path.dirname(state.skill.entryPath), 'agents');
+    fs.mkdirSync(agentsDir, { recursive: true });
+    fs.writeFileSync(path.join(agentsDir, 'openai.yaml'), canonicalMetadata);
+
+    const created = await reconcileNativeSkill({
+      host: 'codex',
+      skill: state.skill,
+      targetRoot: state.nativeRoot,
+      receiptRoot: state.receiptRoot,
+    });
+    fs.writeFileSync(path.join(created.targetDir, 'agents', 'openai.yaml'), 'stale metadata\n');
+
+    const forced = await reconcileNativeSkill({
+      host: 'codex',
+      skill: state.skill,
+      targetRoot: state.nativeRoot,
+      receiptRoot: state.receiptRoot,
+      force: true,
+    });
+
+    assert.equal(forced.action, 'updated');
+    assert.equal(forced.forced, true);
+    assert.equal(
+      fs.readFileSync(path.join(forced.targetDir, 'agents', 'openai.yaml'), 'utf8'),
+      canonicalMetadata,
+    );
+  } finally {
+    fs.rmSync(state.root, { recursive: true, force: true });
+  }
+});
+
 test('managed updates replace the complete tree, prune stale resources, and become idempotent', async () => {
   const state = fixture();
   try {
diff --git a/src/native-skills/lifecycle.js b/src/native-skills/lifecycle.js
index 563cd16..f1d35f7 100644
--- a/src/native-skills/lifecycle.js
+++ b/src/native-skills/lifecycle.js
@@ -288,6 +288,19 @@ async function collectResourceEntries(sourceRoot, resourceName, entries, sourceE
   await walk(resourceRoot, resourceName);
 }
 
+async function readBundledCodexMetadata(sourceRoot) {
+  const metadataPath = path.join(sourceRoot, 'agents', 'openai.yaml');
+  let metadataStat;
+  try {
+    metadataStat = await fsp.lstat(metadataPath);
+  } catch (error) {
+    if (error.code === 'ENOENT') return null;
+    throw error;
+  }
+  assertRealEntry(metadataStat, metadataPath, 'file');
+  return fsp.readFile(metadataPath);
+}
+
 function manifestEntry(entry) {
   if (entry.type === 'directory') {
     return { path: entry.relativePath, type: entry.type, mode: entry.mode };
@@ -350,15 +363,7 @@ async function buildProjection(host, skill) {
     mode: sourceStat.mode & 0o777,
     content: sourceContent,
   }];
-  if (host === 'codex') {
-    entries.push({ type: 'directory', relativePath: 'agents', mode: 0o755 });
-    entries.push({
-      type: 'file',
-      relativePath: path.join('agents', 'openai.yaml'),
-      mode: 0o644,
-      content: Buffer.from(generated.openaiYaml),
-    });
-  }
+  let codexMetadata = host === 'codex' ? Buffer.from(generated.openaiYaml) : null;
   let packageDigest;
   if (path.basename(sourcePath) === 'SKILL.md') {
     const sourceRoot = path.dirname(sourcePath);
@@ -367,12 +372,24 @@ async function buildProjection(host, skill) {
       throw new Error(`Source skill package not found: ${sourceRoot}`);
     }
     packageDigest = completePackage.digest;
+    if (host === 'codex') {
+      codexMetadata = await readBundledCodexMetadata(sourceRoot) ?? codexMetadata;
+    }
     for (const resourceName of RESOURCE_DIRECTORIES) {
       await collectResourceEntries(sourceRoot, resourceName, entries, sourceEntries);
     }
   } else {
     packageDigest = digestEntries(sourceEntries).digest;
   }
+  if (host === 'codex') {
+    entries.push({ type: 'directory', relativePath: 'agents', mode: 0o755 });
+    entries.push({
+      type: 'file',
+      relativePath: path.join('agents', 'openai.yaml'),
+      mode: 0o644,
+      content: codexMetadata,
+    });
+  }
   const rendered = digestEntries(entries);
   const sourceIdentity = resolveSourceIdentity(skill.source);
   return {