Skip to content

chore(knowledge): remove the system-prompts entry and empty the docpage-digest queue - #1874

Merged
kyle-sexton merged 1 commit into
mainfrom
chore/docqueue-system-prompts-complete
Aug 1, 2026
Merged

chore(knowledge): remove the system-prompts entry and empty the docpage-digest queue#1874
kyle-sexton merged 1 commit into
mainfrom
chore/docqueue-system-prompts-complete

Conversation

@kyle-sexton

Copy link
Copy Markdown
Contributor

The platform.claude.com/docs/en/release-notes/system-prompts slice completed. It is the last entry in the doc queue — after this merges, the queue section holds only the deferred task-budgets trigger entry, which was never queued.

What ran

The largest slice this pipeline has processed: 18 digests over a 2,548-line source, 659 claim rows, 561 api-only.

  • Seven correction rounds and five dual-verification rounds, each on SHA-256-pinned bytes with the frozen-tree rule enforced — both arms re-hash before and after auditing, and no correction runs while a verifier holds the tree.
  • Final state: both arms VERDICT: PASS, with a confirming round after the last corrections.
  • Final gates: 1,895 recorded command/count pairs replay with 0 mismatches, 0 <corpus> placeholders, 659/659 rows parse with N. Quote (line count equal to - Tag: count in all 18 digests, every stated tally equal to its anchored grep -c, check-quotes.py clean.

What the run found worth carrying forward

  • Narrow-slug false absences remained the dominant defect — a basis testing one literal while the corpus documents the same concept in other words. Rows now disclose near-misses by name and line, each produced by a command recorded in that row.
  • Prose is unparsed by every gate. Summary paragraphs, Implications bullets and Open questions carried stale counts and false statements across several rounds; two sweeps (one whitespace-normalized with an offset map back to physical lines) closed the class.
  • A contested tag-ordering question is escalated, not decided — whether selecting a positive tag requires first establishing harness applicability. It is written up as a candidate profile amendment with both readings, the evidence, and a recommendation, for the dispositions interview.

Residual MINOR findings are disclosed in the slice's interview-handoff.md as Open questions, per the run's stopping rule.

What this PR changes

  • Removes the system-prompts entry and the now-empty "Supplementary references" heading from the profile's Doc queue.
  • Bumps the knowledge plugin 0.10.130.10.14.
  • Adds the matching CHANGELOG entry.

No linked issue

Related

🤖 Generated with Claude Code

…ge-digest queue

The platform.claude.com/docs/en/release-notes/system-prompts slice completed:
18 digests over a 2,548-line source, 659 claim rows, 561 api-only. It is the
largest slice this pipeline has run and the last entry in the queue.

Both verification arms reached PASS on SHA-256-pinned bytes, with a final
confirming round after the last corrections.

The "Supplementary references" heading is removed with it, since the entry was
the last one under it. Only the deferred task-budgets trigger entry remains,
which was never queued — so the Doc queue section is now empty.

Residual MINOR findings are disclosed in the slice's interview-handoff.md as
Open questions for the batched dispositions interview. One contested
tag-ordering question is escalated there as a candidate profile amendment
rather than decided in-run.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@claude

claude Bot commented Aug 1, 2026

Copy link
Copy Markdown

Claude finished @kyle-sexton's task in 32s —— View job


Security Review — PR #1874

  • Gather context (diff, changed files)
  • Review each changed file for security issues
  • Post findings

Scope reviewed (via git diff origin/main...HEAD):

  • plugins/knowledge/.claude-plugin/plugin.json — version bump only (0.10.130.10.14), no other fields touched (no permission, hook, MCP, or trust-boundary changes).
  • plugins/knowledge/CHANGELOG.md — new changelog entry, prose only.
  • plugins/knowledge/skills/docpage-digest/context/anthropic-docs-profile.md — removes a completed doc-queue entry (a markdown bullet with a URL reference) and its now-empty "Supplementary references" heading.

No security issues found.

This PR is a pure documentation/metadata change: a plugin version bump, a changelog entry, and removal of a markdown reference line from a static context file consumed as reference prose by the docpage-digest skill. None of the changed content is executable — no scripts, hooks, workflow YAML, .mcp.json entries, or permission grants are touched, so there's no code-execution, injection, credential-handling, or access-control surface to assess here.

@kyle-sexton
kyle-sexton merged commit 4bebb89 into main Aug 1, 2026
55 of 57 checks passed
@kyle-sexton
kyle-sexton deleted the chore/docqueue-system-prompts-complete branch August 1, 2026 21:42
kyle-sexton added a commit that referenced this pull request Aug 2, 2026
…d pipeline amendments (#1875)

Applies the evidence-forced profile and pipeline amendments the
docpage-digest campaign produced, and nothing else. `knowledge` 0.10.14
-> 0.10.15.

## What this ships

Eleven normative rules across `SKILL.md` and
`context/anthropic-docs-profile.md`, each mapping to a triage row marked
`evidence-forced`:

- **PA-A** row-local reachability — a cited `file.md:NN` counts as
disclosed only when a command recorded in that same row produces it,
plus the `| wc -l` and sampled-set corollaries.
- **PA-B** a mechanical gate reports only what it parsed, and only the
fields it checks; unrecognized input becomes a hard error *before* the
gate is made required.
- **PA-C** every correction round leaves an applied record, and its "New
findings" section is a required input to the next round's brief; a
verdict lands in `verification/` or it did not happen.
- **PA-D** commands are replayable in every pipeline artifact, not just
digest rows.
- **PA-E** reconcile the digest set against itself before Phase 5.
- **PA-G** every non-zero result names its match sites; a sampled hit
set states that scope at the row.
- **PA-R** absence-establishing fetches use the raw `.md` channel via
`curl` and record the retrieved length.
- Plus the frozen-tree rule (no tree moves until every dispatched arm
has reported) and the `api-only` exact-command-and-raw-count basis. Both
were forced by the record but had no triage row, because the triage
never ingested the process-failure file; rows `PA-AL` and `PA-AM` were
added to close that classification gap.

## Verification

Two independent verifier arms, one adversarial, audited whether any
amendment here is really a judgment call. Both concluded no: every rule
maps to an `evidence-forced` row, nothing from the judgment-amendments
file is present, and the escalated Decision-A ordering amendment (PA-AI)
is untouched.

Both arms then failed the branch on its *supporting evidence* — five
parentheticals that misattributed an instrument, compressed defects
spanning three jobs into "a single round", reported a
written-but-unconsumed record as "left no record", blended a run-9
figure into a run-10 slice, and asserted a reconstruction the record
does not describe. All five are corrected against primary sources and
re-verified by two fresh arms.

Rules were proven untouched mechanically rather than by eye: stripping
every balanced parenthetical span and normalizing whitespace yields
byte-identical remainders against the pre-correction commit (15,210 and
5,838 chars).

No linked issue

## Related

- Follows the three merged doc-queue PRs #1872, #1873, #1874, which
emptied the digest queue.
- Triage and judgment classification live in the campaign's memory tier
under `.work/_harness-snapshots-2026-07-31/`, which is deliberately not
committed.

---------

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant