# Integration implementation report

This is the implementer's local checkpoint for the additive integration on baseline `d1c0d036302a939387dc04b90792b1d2cef48a59`. Independent implementation QA is a separate gate owned by a fresh reviewer. The authoritative final independent verdict and completion status are the retained `/private/tmp/codinary/unified-agentic-20261028T220833Z/review-<n>.json` and `/private/tmp/codinary/unified-agentic-20261028T220833Z/summary.md`; this report records implementer observations without predicting that verdict. No commit, push, deployment, publication, global installation or real human specification approval occurred.

## Delivered behavior

The seven existing skills remain canonical. One requirements skill, grilling, and all seven EGM design/documentation commands bring the registry to 15. Codinary remains the only delivery owner. Its original skill content, roles/reviewer schema, fix budget, waivers, review-only/release semantics and pre-existing user cleanup/resource edits are preserved; only optional design-intake/planning pointers and references are added.

Canonical EGM resources retain all nine personas, five complete rubrics, command assets/templates, runtime adapters and bounded execution. The Agentic marketplace registers nine thin Claude wrappers once, with explicit read-only reviewer tools and omitClaudeMd declarations. Native enforcement is unverified on Claude. General fresh-agent fallback supports selected skill copies. No upstream installer/plugin identity or competing delivery engine is imported.

Registry parsing, dependency/alias/path/cycle/reference checks and generated metadata/bundles prevent discovery drift. Single selected EGM copies include all resources, Node API/CLI and Agentic/Matt/EGM notices. Existing local copy/link, remote tar, init/npm and plugin packaging remain. Update replaces explicitly selected skills; uninstall preserves unrelated skills/dependencies/configuration.

The optional handoff validates strict version-1 data, actual repository origin/HEAD, requirement/criterion traceability, preset-derived review coverage, exact design/report hashes and a detached human Ed25519 receipt. Approval requires a separately caller-pinned external verification key/current identity/version/hash; project metadata cannot manufacture approval. Lifecycle transitions return new JSON without storage mutation. Invalid explicit handoff stops intake without downgrade. Direct requests stay independent. Full signing representation, severe-finding guards and stable-snapshot limitations are in handoff-contract.md and security-model.md.

## Verification evidence and limits

Full local evidence is retained at `/private/tmp/codinary/unified-agentic-20261028T220833Z/evidence/`; handoff-1.md and source-state.json identify the original implementation checkpoint. The subsequent global fix cycle 1 addresses review-1's trust-path, preset-type, dependency-depth and extensionless package-resolution findings; handoff-fix-1.md and source-state-fix-1.json identify that updated source and verification. These corrections await fresh independent review. Commands are reproducible from this repository. npm uses a task-local cache because the user's default cache has an observed permission failure; the cache was not modified.

| Check | Observed implementation result | Evidence kind |
| --- | --- | --- |
| Protected baseline | 81/81 protected hashes/content checks pass, including exact initial roles/verification/platform references and initial Codinary body after removing only the new pointer | Characterization |
| Existing skills installer | All checks pass, including selection, remote path, update/uninstall/link and manifest agreement | Executable integration |
| Existing initializer/npm launcher | All checks pass | Executable integration |
| Initializer golden characterization | 70/70 scenarios match; seven languages/templates preserved | Characterization |
| Codinary recovery harness | 10 passing tests; seeded broken fixture files remain unchanged | Controlled regression |
| New integration/security/package/workflow suite | API/actual CLI boundaries, synthetic independent signing, denial cases, selected-copy and package consumers; current fix-cycle counts/results are in handoff-fix-1.md | Executable integration/security |
| Registry/provenance | 15 skills; metadata/bundles valid; 351 exact shipped source/destination digest mappings | Static plus filesystem validation |
| Shellcheck, ruff | Pass for repository-required script/Python scopes | Automated readiness |
| Claude strict marketplace validation | Pass locally | Static plugin validation |
| TypeScript, Go and Python existing projects | Actual isolated init/install CLI preserves user files; installed intake CLI launches | Executable integration; not native compiler/agent delivery certification |
| npm package | Actual pack/extract/API/CLI/selected EGM consumer passes; package dry-run retained | Packaging/runtime consumer |

The new scenarios cover medium/complex approved intake projection, blocked/stale intake, supplied-state/transition guards, and read-only validator behavior. Existing controlled recovery tests protect delivery role/fix/evidence contracts. These are controlled cases, not claims that separate live agents completed small-bug, medium-feature, review-only or blocked delivery journeys. The actual main integration receives independent source QA separately. No unsupported behavior was simulated as native execution.

## Actual design and runtime reviews

The original proposal received all five design stages using pinned upstream assets; original Stage9 and its one allowed re-audit both returned FAIL. Raw reports and dispositions are retained under `/private/tmp/codinary/unified-agentic-20261028T220833Z/design-review/`, and implementation-addendum.md records concrete RA1–RA6 corrections. The cap was honored. Implemented corrections are submitted to independent source QA; they are not a fabricated design-approved verdict or real human approval.

A separate actual installed-copy Stage6 comprehension probe used final generated skill/persona/runtime/stage inputs under `installed-probe/`. Evidence is retained at `/private/tmp/codinary/unified-agentic-20261028T220833Z/runtime-probe/{comprehension_report.md,manifest.json,input-fingerprints.json}`. A fresh child used fork_turns:none, read the full synthetic target and persona, returned both full paraphrases and three real findings, and did not know a marker seeded only in the parent conversation. This controlled sample supports conversation isolation for that probe only. Read-only behavior was followed by instruction, not host-enforced permissions. It does not prove filesystem/memory isolation or hostile-host protection.

A separate complete five-stage adapted installed-package review-only fixture is recorded under `/private/tmp/codinary/unified-agentic-20261028T220833Z/runtime-probe/full-review/`; its manifest/reports are authoritative for actual completion and findings. This implementation checkpoint does not predict that fixture outcome or treat it as another capped integration-architecture re-audit. Native Claude/Gemini/Antigravity/Cursor/Copilot reviewer execution, tool enforcement and cross-host product verification remain unverified. Static installation or upstream runtime claims do not establish them. No UI surface changed, so visual/browser/native-device checks are not applicable to this CLI/framework integration. No resource waivers were invented.

## Final bounded correction checkpoint

Independent review 2 closed all four first-cycle findings and identified R2-1 in the Bash installer. Global fix cycle 2 validates cached dependency height and every selected/default/all installation graph before destination writes, including unterminated metadata lines. The installer stays self-contained for piped old-ref archives, preserves installation order and explicit uninstall recovery, and requires no Node helper. Actual CLI RED/GREEN evidence and the final matrix are recorded in `evidence/fixer-2/`, `response-2.json` and `handoff-fix-2.md` under the retained run directory. Fresh independent review 3 remains required; this checkpoint does not claim its verdict.

The completed [five-stage native fixture](/private/tmp/codinary/unified-agentic-20261028T220833Z/runtime-probe/full-review/run_manifest.json) records nine fresh reviewers, 13 findings and failed synthetic document readiness without approval. The separate [native delivery fixture report](/private/tmp/codinary/unified-agentic-20261028T220833Z/native-delivery-fixtures/summary.md) records A's passing small-bug implementation/independent QA, B's completed review-only run with a failing known-broken product, and C's blocked explicit unapproved handoff without fallback or mutation. These attributed fixtures do not establish other native-host support or enforced read-only isolation. Original design readiness and its one allowed re-audit remain FAIL.

## Provenance and maintenance

Pinned Matt revision: `59db609eb1e18aaf59b63284d0e52255aff0c63c`, MIT. Pinned EGM revision: `f0ba151b252c9abf1318d9fe857a43265d06de89`, Apache-2.0. Existing Matt adaptations' historical import revision remains unknown; no false provenance is asserted. Agentic retains its MIT license. Complete notices and modified-file declarations accompany canonical and generated distributions. EGM is not an officially supported Google product.

Edit canonical resources once, run registry generate/check, deliberately update provenance digests after source changes, and rerun relevant checks. No automatic upstream tracking, signer, mutable approval catalog or new agent runtime engine is introduced. Retained installed-probe and synthetic-handoff-fixture contain test data and public verification material only, no private signing key. Temporary per-test projects are removed by scoped teardown; run evidence/caches/probes remain for independent QA and resumption.

## Caveman and Security Audit extension (2026-10-29)

The separate [integration plan and decisions](caveman-security-audit.md) adds two standalone specialists to the same canonical registry/distribution architecture. [Verification evidence](caveman-security-verification.md) records current tests, behavioral probes and runtime limitations. Historical Codinary reviews and fix counts above are retained as historical evidence; this extension does not revise their outcomes or claim full native audit support.