docs(knowledge): admit anthropic.com/engineering and add cross-slice synthesis artifact target - #1892
Conversation
The profile scoped this publisher to platform.claude.com, code.claude.com and claude.com/blog. Anthropic's engineering posts are first-party and are the stated best-practices channel for topics the corpus already wants a page from, so the narrower list forced every such page to enter by exception. Engineering pages are in scope by default now. Declining and then admitting them one at a time is the same coverage with worse bookkeeping — the exemption list grows, and nothing in the profile records why. Two standing costs are real and stay unwritten here rather than being absorbed silently. The vendor-blog attestation bullet still names `claude.com/blog` literally, so it does not reach the new property; rescoping it would decide whether engineering deep-dives are marketing-adjacent vendor voice, which is a fresh judgment, not a consequence of this one. And unlike both docs properties, this one publishes no machine-readable page index, so page selection and absence checks against it have no instrument. Both are named in the CHANGELOG so a consumer meets them before the first engineering slice, not during it. Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
The Artifact targets taxonomy named three dispositions, and a cross-model synthesis artifact matches none of them: it is not per-model, not an audit rule row, and not graduation of one slice. Content deferred to such a pass therefore had no route out of the pipeline at all — the digest fan-out is barred from reaching across units by design, and no later stage exists to pick it up. Four units in one slice deferred content into that gap. Without the fourth entry, a disposition the taxonomy cannot express reads as out of scope. It is not; it is a hole. The added paragraph says so directly so a future handoff does not re-derive the distinction per finding. The target is deliberately named without a host. Which repository or seam a cross-slice artifact lands in is a separate decision no run has taken, and the one candidate a reader might assume — a model-adaptation reference tree — is scoped to model-adaptation content, which this is not. Pinning a host here would take that decision by implication. Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
The fourth target's parenthetical already carries the whole amendment — the three shapes it is not, and the undecided host. The paragraph following it argued why the target should exist, which is a decision record, not routing contract, and the profile is the consumer-facing surface. The rationale survives in the commit that added the target and in the release note, both of which a reader reaches when they want it. Keeping a third copy in the contract is where drift starts. Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
Claude finished @kyle-sexton's task in 2m 4s —— View job Review complete
Net diff verified. The three commits ( Version bump and CHANGELOG are consistent. Finding: new property admitted without Fetch-channel guidance
This matters here specifically because the CHANGELOG entry's own framing is "the two standing costs are stated ... rather than silently absorbed" — but a third cost (no fetch-channel instrument for the new property) went undisclosed rather than named alongside the other two (vendor-blog attestation scope, no machine-readable page index). This isn't necessarily a blocker — the profile may intend the first real engineering-page run to establish the channel empirically (the docs-page bullet already models "re-verify per doc — precedent, not a guarantee"). But as written, a consumer hitting this property today has strictly less guidance than for blog, and that gap isn't named the way the other two are. No other correctness or CLAUDE.md-alignment issues found — the change stays within the "prose body" scope (skill context file, not frontmatter/manifest contract fields), so the fresh-docs mandate's live-fetch requirement doesn't apply here. |
|
Claude finished @kyle-sexton's task in 1m 5s —— View job Security review complete
No security issues found. This PR is a pure documentation/prose change to a knowledge-plugin skill context file, plus a mechanical version bump and matching changelog entry:
I considered this file as an agent-instructions surface (it's loaded as skill context and does direct fetch behavior via Nothing else in the diff touches GitHub Actions workflows, permissions, or executable code, so there's no applicable finding for this PR. |
![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: 4db932a377
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".
|
Claude has reviewed this PR 2 times. The lane skips further automatic reviews after 5; deleting this comment resets the count. |
…prefix Review finding on #1892: SKILL.md selects the publisher profile by matching the URL's host against the profile and proceeds profile-less on no match, and every live engineering link in this repo uses www.anthropic.com — so the bare anthropic.com declaration would never match a real run. Declare www-insensitive host matching at the property list. Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
Claude finished @kyle-sexton's task in 2m 7s —— View job Code review complete
Net diff verified against branch base ( The Codex host-matching finding is fixed. Version bump and CHANGELOG stay consistent with the style of the prior Finding: Fetch channel section still has no bullet for the new property (carried over, unresolved)
Not a blocker — the docs-page bullet already models "re-verify per doc — precedent, not a guarantee," so the profile could reasonably defer this to the first real engineering-page run. But as written it's an undisclosed gap, not a disclosed one, which cuts against the PR's own framing ("costs... stated... rather than silently absorbed"). No other correctness or CLAUDE.md-alignment issues found. This stays a "prose body" edit (skill context file, not frontmatter/manifest contract fields) per the fresh-docs mandate's scope discriminator, so the live-fetch requirement doesn't apply. Security review is out of scope for this lane per REVIEW.md's split. |
|
Claude finished @kyle-sexton's task in 1m 42s —— View job Security review complete
No security issues found. This PR (through the latest commit,
I looked specifically at whether the new host-matching prose changes the security posture, since this file drives agent fetch behavior (
Nothing in the diff touches GitHub Actions workflows, permissions, secrets, or executable code. |
…ster (#1895) ## Summary Repopulates the docpage-digest Anthropic profile's doc queue with the doc-corpus campaign's adopted enqueue roster (knowledge 0.10.18 → 0.10.19): - **14 pages queued across seven groups** — thinking troubleshooting first (on the corrected rationale: the harness carries a parallel troubleshooting surface, so the page's transfer is demonstrated, not conjectured), then the last thinking page, the retention/ZDR slice (one slice, two lanes, three pages — org-level policy, the one queued topic with compliance weight), one Agent SDK page with the wider-scope boundary stated in the group heading, the models freshness pair, three Claude Code companion docs in digestion order, two blog posts, and one engineering post (the property admitted in 0.10.18). - **5 pages deferred with triggers** — each carrying its named re-arm condition, including a context-window deferral that states its discharged check's result rather than leaving the trigger dormant. - The retention/ZDR enqueue is recorded as still open to an owner veto; the engineering entry's contingency is disclosed as discharged (its prerequisite property edit was the 0.10.18 merge). Independently verified (fresh-context, rationale withheld): SHIP, with the two findings (deferred-block position; contingency disclosure) applied in `6395841eb3` and confirmed CLOSED. All 14 queued URLs return HTTP 200 live; queue counts recounted at the bytes by producer and verifier independently. No linked issue ## Related - #1892 — the property admission and artifact-target amendment this queue builds on (0.10.18) - #1887 — DOC bundle 2, where the queue's prior entries were retired on slice completion 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_011UarawwEnZQu7cB6i7WatJ --------- Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
1 participant
Summary
Implements two adopted doc-corpus campaign decisions in the docpage-digest publisher profile (
plugins/knowledge/skills/docpage-digest/context/anthropic-docs-profile.md):anthropic.com/engineeringas a fourth covered property in the profile's property list. The two standing costs are stated in the CHANGELOG rather than silently absorbed: the vendor-blog attestation bullet remains literally scoped toclaude.com/blog(rescoping it is a separate judgment), and the new property contributes no machine-readable page index.Bumps knowledge plugin 0.10.17 → 0.10.18 with CHANGELOG entry. Independently verified against the adoption records with author rationale withheld: net branch diff = exactly the two adopted amendments, nothing more; merge-tree clean against current main.
No linked issue
Related
🤖 Generated with Claude Code
https://claude.ai/code/session_011UarawwEnZQu7cB6i7WatJ