Use when changing PR body rendering, home-path redaction, artifact path publication, pipeline-attestation markers, or pre-push attestation.
65%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
Passed
No findings from the security scan
Fix and improve this skill with Tessl
tessl review fix ./.agents/skills/pr-publication-safety/SKILL.mdHome-Path Redaction in Published PR Content (security)
internal/safepath is the one owner of home-directory redaction, the path analogue of internal/safeurl. RedactText rewrites the process's own home plus /home/<user>, /Users/<user>, and C:\Users\<user> to ~, unconditionally and for every occurrence. Add new shapes there rather than scrubbing paths at a call site. Candidate resolution must stay free of filepath.IsAbs/VolumeName and of any reliance on filepath.Clean's separator normalisation: those answer for the build platform, and on Windows IsAbs discards the POSIX-rooted HOME that Git Bash, MSYS2, and Cygwin set - silently disabling redaction instead of failing. Regression: TestUsableHomeCandidate_AcceptsBothPlatformSpellings, TestHomeCandidates_AreSeparatorSpellingIndependent.redactPRContent in pr.go owns the publication redaction boundary: safepath.RedactWorktrees (the run worktree sctx.WorkDir, then any run worktree under nmWorktreesRoot()) before safepath.RedactText, the same order as evidence publication, since home redaction would stop the worktree roots matching. PRStep.buildPRContent uses it for ordinary drafts; template creation and live author-preserving updates use it through composeOwnedPRContent before stamping their byte-integrity guard. Every source that can reach a PR body - agent prose, extracted intent, findings, fix summaries, step errors, artifact path, artifact captions, and captured output embedded from evidence files - is covered there, so a new rendering path cannot reintroduce the leak. Ordinary drafts redact after length caps (the placeholder never grows a path); template composition redacts before its integrity guard and fail-closed size check. pr_ownership.go owns that marked-appendix contract: never feed live author text through heading-based stripping/clamping, and keep rebindOwnedPRAttestation in the pre-push restamp path so its integrity guard remains valid. Regressions: pr_template_test.go, pr_ownership_test.go.artifacts[].path description in testFindingsSchema (common.go) must not solicit absolute paths, and must not forbid them either. The renderer's allowlist is the worktree or the run's evidence directory and a path under neither is dropped, while the evidence directory defaults under the operator's home - so soliciting more just re-supplies what the boundary has to strip, and a blanket "never report a home directory path" clause makes an obedient agent drop its own evidence. Publication safety is the pr.go boundary's job; the schema only stops soliciting paths from elsewhere on the machine. Regressions: TestTestFindingsSchema_DoesNotSolicitAbsolutePaths, TestTestFindingsSchema_KeepsEvidenceDirectoryPathsReportable.test.evidence.store_in_repo) is the second boundary: evidence.Publish runs its text files through the same worktree and home redaction (hashEvidenceFile, owned by the test-evidence-storage skill). One other public surface deliberately does NOT share this rendering and is not covered: agent-authored commit subjects (commitAgentFixes -> Commit.RenderFixMessage) and their trailers (withCommitTrailers -> Commit.RenderTrailers, whose agent and model values are reduced to one safe token), which reach the remote through Push. Keep the internal/safepath package doc honest about that scope.require-no-mistakes (.github/actions/require-no-mistakes/verify.py) binds the FIRST marker in the RAW body to the PR head, so a foreign copy placed earlier fails a PR the pipeline did produce - and a code fence is no defense, because that scan is raw text. Step agents embed foreign markers routinely, by capturing a generated PR body as evidence.head_sha in the current PR body (restampPublishedAttestation) and does not send a title. It never inserts a marker that was not already there. Hosts without a PR content reader skip the restamp instead of failing the push. Regressions: TestCIStep_PublishRepairRebindsAttestationAcrossRepairPushes, TestCIStep_PublishRepairDoesNotMintAttestation, TestCIStep_PublishRepairSkipsRestampWithoutReader, TestRestampPRAttestation_PreservesContentEditedWhilePreparingRewrite, TestUpdatePROmitsTitleWhenEmpty.repository_overrides.<url>.pr.update_text is keep or regenerate (default, also the absent key). keep is machine-local only. On create it uses the ordinary drafter and stores the narrative above an owned appendix (publishKeptDraft). On an existing PR it reuses the template preservation path: the title is omitted unless pr.title_format is set, the bytes before the appendix stay, and the appendix or minimal risk line plus the attestation are rebuilt for the current head. An unowned attestation still fails closed. regenerate stays on the ordinary rewrite. Regressions: TestPRUpdateTextKeepOnUpdate, TestPRUpdateTextKeepOnCreate, TestPRUpdateTextRegenerateRewrites, TestPRUpdateTextKeepRefusesUnownedAttestation, TestRepositoryPRUpdateText.axi run --pr-body-file, runs.pr_body) is author text, published on create through buildSuppliedPRContent as the narrative above an owned appendix. When freshly supplied on an existing PR, updateOwnedPRWithNarrative replaces the author narrative and refreshes the appendix; a concurrent pre-write author edit refuses the replacement. Inherited text on a plain rerun/retry never does so. runs.pr_body_fresh and runs.pr_title_fresh persist that distinction across recovery, and runs.pr_text_applied_at (markSuppliedPRTextApplied, set on create and after a fresh existing-PR update) makes later PR-step passes in the same run treat the text as inherited, so a CI-repair restart never reverts a maintainer's edit on an owned or kept body; the ordinary regenerate rewrite still sends the supplied title (keepAppliedTitle) instead of a drafted one. Do not clear the fresh flags instead: conflictingActiveRunPRText reads them for reattach. The daemon includes those flags in RunInfo, and conflictingActiveRunPRText refuses attaching a fresh request to a run with matching inherited text. ProbeRunPRTextResult.ExistingPRUpdates refuses an old create-only daemon before a supplied-text run; OK alone does not establish update support. It never goes through stripGeneratedSections or a drafting turn; it does go through neutralizeAttestationMarkers and the composeOwnedPRContent redaction boundary. steps.ValidateRunPRBody (CLI and daemon) refuses reserved markers, controls, and bidi overrides before it is stored. Regressions: TestPRSuppliedTextCreatesOwnedBody, TestPRFreshSuppliedTextUpdatesExistingPR, TestPRInheritedTextIsNotAppliedToAnExistingPR, TestPRSuppliedDescriptionRefusesConcurrentAuthorEdit, TestPRFreshSuppliedTextAppliesOncePerRun, TestPRFreshTitleSurvivesARegenerateRestartInTheSameRun, TestPRStep_SuppliedBodyIsRedactedBeforePublishing.appendGeneratedSectionsToCleanBodyWithinLimit plus the two intent paths), never per render path. pipelineMD alone carries the real marker and is left intact; BuildPipelineSummaryFor neutralizes its own step-detail blocks, which quote agent text. A first attempt put this in escapePipelineFoldMarkers - per-render-path - and shipped three live foreign markers to #831 anyway. Regressions: TestPRStep_ForeignAttestationsInEveryComponentDoNotShadowTheRealOne (all components at once), plus the per-component guards in pr_test.go.redactPRContent also runs neutralizePRControls (pr_controls.go): published bytes must be a fixed point of gh pr view --json, whose sanitizer rewrites C0/C1 controls and the literal text \u00XX of them to caret notation, or the post-create body check and the appendix digest fail with a false "edited" refusal (#1435, #1454). composeOwnedPRContent neutralizes the appendix BEFORE escapePRAppendixMarkers so removing a byte cannot splice a marker together after the escape ran. parsePROwnedBody still accepts an appendix whose digest matches only after restoreCaretControls (PRs stamped by older versions). internal/pipeline/fakecli/ghsanitize is the shared port of that sanitizer; do not hand-write another. Regressions: internal/pipeline/steps/pr_controls_test.go.internal/safepath/redact_test.go, internal/pipeline/steps/pr_homepath_test.go.Pre-Push Pipeline Attestation
publishRunHead must write an existing PR attestation for the proposed head before pushing it on every supported provider with raw content reads. docs/src/content/docs/reference/repo-config.md (pr.template) owns provider-specific ownership/budget and visible Bitbucket metadata caveats. The protocol has single-publisher scope because the daemon enforces one active run per repository branch; do not add cross-publisher coordination without a separate requirement. Detailed behavior is owned by docs/src/content/docs/reference/pipeline-steps.md; local rationale lives in attestHeadBeforePush and publishRunHead. Regressions: TestPushStep_AttestsHeadBeforePush, TestPushStep_AttestationWriteFailureAbortsBeforePush, TestPushStep_UnavailableSCMLeavesStaleAttestationFailingClosed, TestPushStep_PushFailureAfterAttestationLeavesBodyAhead, and TestCIStep_PublishRepairRebindsAttestationAcrossRepairPushes.kunchenguid/no-mistakes
3d2d4bf
If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.