Recreate MSDMD as a native-first metadata skill - #102
Conversation
Replace block-only coverage doctrine with native-source ingestion, source-scoped provenance, conflict preservation, explicit reader support, and information coverage. Add convention catalogue and acceptance contract; synchronize README, index and generated discovery adapter. Runtime collector remains block-only and is explicitly documented as such.
|
@codex review Review exact head |
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 83bd040f40
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
|
@codex address that feedback Repair the four live P2 findings without widening into native-reader implementation: reconcile dependent MSDMD application contracts with native-first metadata so block-only requirements no longer contradict the foundation; retain the shipped block collector/visualizer as runnable in skills.json while marking native ingestion separately unimplemented; qualify or explicitly bound current block-collector edge identity so valid cross-file id reuse is not misrepresented; update the owning LLMS declaration and regenerate llms.txt. Preserve existing block compatibility, source CHECKS/CONTRACTS ownership, and the explicit hmmm that native readers/schema/acceptance fixtures are not yet implemented. Re-run exact drift/index/adapter/CI gates, then request exact-head review. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 83bd040f40
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Align eight application contracts and their discovery descriptions with native-first information coverage. Restore scoped discovery of the shipped block collector while keeping native ingestion contract-only. Explicitly document current bare-ID and duplicate-validation limitations instead of claiming runtime guarantees. Regenerate LLM guidance and Codex adapters from their owning sources, and add six editorial regression checks. Collector/parser/schema runtime is unchanged. Native readers, qualified edge identities and duplicate-ID diagnostics remain explicit hmmm implementation boundaries.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 324610f7f2
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| These rules replace blanket requirements to duplicate native information in | ||
| MSDMD blocks or to label missing blocks as missing information. The dependent |
There was a problem hiding this comment.
When an agent starts new-module work, the governing AGENTS.md still explicitly requires a MODULE_BUILD block, while CLAUDE.md:38-48,251 and ORG_DISTRIBUTION.md:168-169 retain equivalent block-first instructions. Because these entrypoints are read before the revised skill, native planning declarations will still be duplicated despite this supersession claim. The dependent SKILLs are now updated, but a fresh repo-wide search confirms these higher-level contracts remain stale; update them in the same change.
AGENTS.md reference: AGENTS.md:L120-L121
Useful? React with 👍 / 👎.
| That legacy footer is not evidence of an exhaustive native-source inventory; | ||
| source-owned definitions and disclosed reader limitations still govern. |
There was a problem hiding this comment.
Calling this footer legacy does not prevent llms/build.py:239 from emitting it, and the committed llms.txt:25 still tells consuming agents that the generated projection is the single source of truth and that unstated information does not exist. In the newly supported native-first scenario, that directly contradicts this skill's rule that source-owned instruction metadata remains authoritative, so consumers reading only the generated artifact can discard valid native declarations; change the generator template and regenerate the output rather than merely documenting the contradiction.
Useful? React with 👍 / 👎.
| 2. **Discover.** Inventory files and applicable metadata conventions using the | ||
| catalogue. Include tests, manifests, documentation, extensionless files, and | ||
| permitted reports. Record exclusions, inaccessible paths, and unsupported | ||
| files; do not silently exclude them from the coverage denominator. |
There was a problem hiding this comment.
For production coverage, ORG_DISTRIBUTION.md:117-120 directs consumers to docs/runner-config-guidance.md, but that guide defines the default included scope around executable files with supported comment markers (docs/runner-config-guidance.md:9-14) and permits whole categories to be skipped without requiring them in an exclusion ledger (:16-25). For repositories whose metadata lives in manifests, CODEOWNERS, documentation, extensionless files, or unsupported formats, following the official guide can therefore omit exactly the eligible sources this workflow says must remain accounted for and produce false coverage results. Update the guide to distinguish the block-only helper scan from native discovery and require every exclusion to remain visible.
Useful? React with 👍 / 👎.
| 7. **Publish once.** Emit one versioned collection with declarations, provenance, | ||
| relationships, conflicts, reader coverage, and diagnostics. Documentation, | ||
| inventories, graphs, and audit tools consume that collection, not a second | ||
| independently maintained metadata system. |
There was a problem hiding this comment.
ORG_DISTRIBUTION.md:113 designates skill-lib_msdmd.ts as this repository's root collection point, but its canonical msdmd_foundational_contract entry still summarizes MSDMD as block syntax plus the old visible-gap doctrine (skill-lib_msdmd.ts:40-47) and does not identify the new convention catalogue or native-first coverage contract. Consumers of this machine-readable collection therefore continue to receive the superseded contract even though this workflow says inventories and graphs consume the single published collection; update the collection point alongside the owning skill or explicitly mark it stale.
Useful? React with 👍 / 👎.
| **Implementation boundary:** at the reviewed source revision | ||
| `22c2c5702d14fb4b0faeb717777ecab2665770a1`, `collect.py` implements MSDMD-block | ||
| collection and `collection.ts` represents block-origin declarations. The native | ||
| reader, provenance, conflict, and information-coverage requirements below are | ||
| contracts for implementation, not capabilities supplied by this document. |
There was a problem hiding this comment.
For users discovering the library through the Claude marketplace, .claude-plugin/marketplace.json still advertises “self-declared module metadata with runnable drift gates” and claims that every assertion ships with a checker. After this patch, the primary contract is native-first and the native reader, provenance, conflict, and information-coverage requirements are explicitly unimplemented, so the installation listing presents a materially different capability from the bounded implementation described here. Update the marketplace description to distinguish the runnable block-only helpers from specified-only native ingestion.
Useful? React with 👍 / 👎.
1 participant
Requested change
Recreate MSDMD to consume existing code and tooling metadata without requiring equivalent MSDMD blocks. User subsequently authorized repairing reviews and merging after clean tests.
Delivered
Review repair
Repair commit: 324610f
Parent reviewed head: 83bd040
Base: 22c2c57
All five identified review threads have a commit-linked response and are resolved: dependent-contract contradictions, erased runner discovery, overstated qualified identities, stale LLM-facing doctrine, and overstated duplicate detection.
File plan and validation
All 22 repair-file Git blob hashes match the locally tested bytes. The selected local suite passes 6/6. Four planted editorial regressions were each rejected (runner status, helper identity disclosure, native LLMS definition and compulsory OWNERS rule); original bytes were restored and the six tests passed again.
Full repository verification ran in GitHub Actions, not in the partial local checkout:
Usage
Read msdmd/SKILL.md and the convention catalogue for inputs actually present. Existing python -m msdmd.collect and msdmd.visualize remain block-only inventory/prototype tools. Native-first contracts do not turn an unsupported reader into absent information or a clean audit.
hmmm
Native readers, a native-capable versioned collection schema, qualified edge identities, duplicate-ID validation and their executable acceptance fixtures remain implementation work. The 20 catalogue acceptance cases are specifications, not passing native runtime tests. Resolving the two overstated-guarantee reviews repairs documentation and scope, not the underlying identity limitations. The existing strict ratios result also retains 48 hmmm observations; a passing gate does not mean every repository uncertainty is resolved.