Repository navigation
chore: apply prompt-audit fixes from harness-config to kindle-dedrm - #16333
Conversation
Apply the verified prompt-audit edits to the harness-config, harness-memory, harness-ops, implementation, improvement, instruction-placement, javascript-format and kindle-dedrm plugins: capital-letter emphasis lowered (ADR 0098), stale paths and facts corrected, restated upstream facts replaced by pointers, and copied version gates given an as-of date and recheck trigger. The kindle-dedrm sync-prep.sh hunk is left out: its test asserts the upgrade-refusal wording the audit edit changed. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Codex Review Summary: ✅ CompletedThis 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. |
Lane run report
Failed GitHub API calls
|
|
claude-security-review reviewed Security review (incremental, f319416 → 3772157)I found no security issues in this push. What I read. The scope is one file, Instruction-surface check. Nothing is deleted, narrowed or softened. The push only brings back emphasis on a rule that was already there, so no protected class loses a check. I did not open the protected-class register because there was no removal to check against it. Other checks.
Prior findings and mode. The lane's earlier post at f319416 had 0 findings, so nothing is carried forward. This review was blocking only ( critical-findings: 0 |
|
claude-review reviewed Code-review lane, incremental review (f319416 → 3772157): no findings. This review was blocking only, because this lane already reviewed an earlier head. The change since the last review is one line in Checked with my own tools (nothing run):
Closing issue: none. The PR context names no related issue, so there are no acceptance criteria to check. Manifests: this change touches no Prior findings: the only prior post is the summary comment, which has no line anchor and was a 0-finding review, so there is nothing to restate. Verifier dropped 0 of 0 candidates. |
…t-dispatch reattach-slice.test.sh pins "Major divergence (fundamental assumption wrong) still STOPS" inside the re-attach slice; the audit edit lowercased it and failed Test Bash (2). Revert that line to main's text. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Post-merge mutation testingKilled 3, survived 4, timeout 0, no coverage 0, invalid 0, not run 0.
Surviving mutants
The run's summary has the full report. |
No related issue: prompt-audit findings from an owner-requested full-read audit
Summary
Group 5 of 10 from a full-read prompt audit (Anthropic's prompt-audit guide) of the marketplace's instruction text, covering the harness-config, harness-memory, harness-ops, implementation, improvement, instruction-placement, javascript-format and kindle-dedrm plugins. Each edit was checked by a second agent, and fix-agent edits were reviewed again.
The change is instruction text: skill bodies, references, READMEs, and comments or printed notes inside scripts. Two script edits change behavior:
plugins/instruction-placement/scripts/verify-load.shnow passes--disallowedTools Agentto its probe session, andplugins/harness-ops/skills/audit-native-overlap/scripts/overlap.pyprints the real fingerprint length (FINGERPRINT_LEN, 32) in a validation message that said 16.Fix
CLAUDE_CONFIG_DIR, the handoff-notes location in the implementation README,/harness-ops:inventoryadded to the harness-ops README).Left out of the audit patch: the
plugins/kindle-dedrm/skills/manage/scripts/sync-prep.shhunk, becausesync-prep.test.shasserts the upgrade-refusal wording it changed (if Kindle prompts to upgrade, REFUSE). The test stays; that file is unchanged.Verification
inventory.test.sh(133),permission-merge.test.sh(92),automode-entry-diff.test.sh(76),automode-block-lint.test.sh(76),audit.test.sh(37),changelog-status.test.sh(96),lane-launcher.test.sh(512),restart-consumer.test.sh(213),verify-load.test.sh(18),mechanics-sources.test.sh(12),check-drift.test.sh,sync-prep.test.shafter the revert,overlap.test.sh(205) and the install_state unittest suite (91).scripts/sync-shared-copies.sh --check: all 904 registered copies match.scripts/check-changelog-fragments.sh --checkand--check-required origin/main: 8 fragments valid, one per changed plugin.markdownlint-cli2over the touched plugins: 0 issues. The pre-pushcheck-affectedhook passed..githubsuite and model-graded evals, which CI and the review lanes run.Related
N/A
🤖 Generated with Claude Code