Index of every page on this site: llms.txt
All skills
deepseek-ai avatar

/dsh-trim-cot-leakage

@2e39e70
by DeepSeekdeepseek-ai/deepseek-harness285k stars
35,399
Run it yourself

Use when auditing or fixing prose that reads like a leaked reasoning transcript — dead design-session citations such as (decision N), audit item codes, or §N of uncommitted drafts; change narration such as "used to", "no longer", "this cut"; stack or review vantage ("a later PR in this stack", "rejected in review"); reviewer-addressed justifications; control-flow narration; or hedged planning residue in comments, JSDoc, docs, or Agent Notes.

Use this Skill: https://skilld.dev/gh/deepseek-ai/deepseek-harness/dsh-trim-cot-leakage

Nothing lands on disk. Nothing to clean up.

Fork this Skill

Edit a local copy. It keeps the author and licence.

SKILL.md

≈118 tokens for metadata: the name and description. ≈1.8k when used: this file. ≈4.4k more on demand in 2 files.

Description uses 5.9% of example budget

Before choosing a Skill, your Agent reads its name and description. All available Skills share that space.

  • A shorter description leaves more room for other Skills. This entry exceeds our 1% size suggestion.

In our Claude Code example, all Skill names and descriptions share 8,000 characters. This Skill uses ≈468 characters, or 5.9%.

The 1% threshold is a size suggestion. Longer descriptions can still fit.

Your model, settings, and other Skills decide how much text your Agent can read.

Example settings and source

The example uses a 200k-token context and default Claude Code settings. The count includes the name, description, separators, and when_to_use when present. Codex also counts local file paths.

Skit's source and limits: Codex 0.160.1, Claude Code 2.1.292.

Trimming Chain-of-Thought Leakage

Chain-of-thought leakage is prose whose vantage is the authoring session rather than the repository: it cites artifacts only that session could see, narrates the change instead of the state, or argues with a reviewer who has left. The fix is never deletion alone when a passage carries factual clauses — restate each so it stands at HEAD, then delete the transcript around it; a passage carrying none (an audit code, control-flow narration) is deleted outright. REQUIRED BACKGROUND: dsh-prose-standard owns the complete-proposition rule this skill applies; the archived committed-artifact-citations note records the citation rule's historical rationale. It is guidance, not a script.

The one test

For every suspect passage ask: could a reader at HEAD, with no access to any session transcript, PR thread, or uncommitted draft, resolve every reference and verify every claim? If no, restate the surviving facts from the repository's vantage and delete the rest. If yes, it is not leakage, however historical it sounds — but resolvability only clears this skill's bar: on current-state surfaces (READMEs, docs, JSDoc) a resolvable change story is still change narration, and class 3 routes it to its sanctioned home.

Taxonomy

  1. Dead design-session citations — (decision 7), (audit C2), design §4.7, plan §1.4, phase labels (T4, W3, P-I), "the design ledger", "(B ruling)". If the decision has a committed owner, cite it by name and path; otherwise delete the citation and restate its factual clause to stand alone.
  2. Stack and PR vantage — "a later PR in this stack", "this PR adds", "the previous commit". State the shipped mechanism or the extension point; deferred work moves to a TODO marker or an issue reference.
  3. Change narration and version stamps — "used to", "no longer", "the old X", and indexical stamps ("v1", "this cut", "today", "now" contrasting with a past state). State the present behavior; a fixed regression becomes a present-tense counterfactual ("without X, Y happens"), never repo history ("used to Y").
  4. Review choreography — "Rejected in review:", "the reviewer confirmed", draft ordinals ("v5 of this note"), round attributions. Keep the surviving decision and rationale as plain fact; delete who said it when.
  5. Reviewer-addressed justification — "the cast is safe — it simply…", "this is correct because…". A comment arguing its own correctness addresses a reviewer, not a maintainer. State the invariant that makes the code safe, or delete the comment if the code shows it.
  6. Restatement and derivation transcripts — control-flow narration ("first we X, then we Y"), test walkthroughs, proofs of obvious branches. Delete; keep only a non-obvious contract or invariant.
  7. Hedges and planning residue — "probably fine for now", "should be enough", deferrals with no marker. Promote to TODO/FIXME or restate as the actual bound; delete the hedge.
  8. Authoring-language slips — untranslated working-language fragments (端, 设计稿, ---- 私有 ---- separators) in prose whose language is otherwise English, or the reverse in a zh counterpart. Translate or delete.

What is not leakage

Unaided citation passes fail in both directions by deleting durable references and keeping dead ones. Apply these keep rules as written; examples calibrates each:

  • Issue references — #1470, TODO(name):, "issue #N owns the follow-up" resolve at HEAD; keep them on any surface, including READMEs. Do not relocate them to Agent Notes.
  • Merged-PR and issue citations inside Agent Notes and postmortems — sanctioned evidence per the documentation standard's change-story routing.
  • Suppression justifications — oxlint-disable … -- reason, coverage-ignore reasons, empty-catch explanations are required prose; fix a false reason, never delete it.
  • Counterfactual-present regression pins — "without X, Y happens", "a naive X would…".
  • Measured bounds — "(measured: 512 nests ≈ 0.15s)" calibrating a constant; the evidence word "measured" is load-bearing.
  • Runtime old/new states — "the old connection drains before the new one accepts" is runtime lifecycle, not change history.
  • Historical stage names inside a note's change-story sections — "the first cut shipped X" is current-state-safe there; indexical stamps ("this cut") stay banned everywhere.
  • External references that resolve outside the repo by design — standards sections (RFC 9110 §10.1.5), Figma frame names; the §-ban covers uncommitted internal drafts, not external standards or committed docs that own their §-numbering.
  • Project voice and genre forms — "we" as project voice; a note's Alternatives-considered section.

Workflow

  1. Scope and exclusions per dsh-prose-standard: require an explicit scope; never touch vendor/ or .agents/notes/archived/. Recorded fixtures and snapshots are derivatives, not prose targets: change the owning source or scenario and regenerate them only when an authorized behavior change requires new evidence.
  2. Audit read-only first: run the recall batteries (with --hidden so .agents/ is searched), calibrating each probe against a known positive and a near-miss negative before trusting its output, then judge every hit semantically. The batteries are probes, not the definition — each review round of the original purge found cases the batteries missed, so also read the densest prose in scope (module JSDoc, READMEs, Agent Notes) without a pattern in hand.
  3. Fix owner-first per surface: generated catalogs → trace every consumer, fix the source JSDoc or generator template, then regenerate all derivatives; type-equivalence fences → fix the source JSDoc, then re-paste both bilingual pages (verify-type-equiv pins them); bilingual prose → update the counterpart minimally and re-record it through the lightweight routine; bilingual fences → copy the corrected verbatim block byte-for-byte into both sides per dsh-doc-writing, then re-record the pair; model- or user-visible strings → route through dsh-prose-standard and change only with owning behavior evidence, otherwise leave unchanged and report the deferral.
  4. Before deleting anything, enumerate the passage's propositions (prose-standard) and check the overcorrection traps: trims that flip an obligation into an endorsement, promote a hypothetical to a shipped feature, delete a true fact, or drop the evidence source.
  5. Verify: re-run the batteries expecting only sanctioned keeps, this skill's own directory, and the owning note's quoted evidence; confirm every remaining citation resolves at HEAD; run the gates for touched surfaces (doc-sync for docs, verify-type-equiv, verify-translation-pairing).

Source: SKILL.md on GitHub

No rule matched.

skilld matched fixed text patterns in SKILL.md and file names. Patterns miss obfuscated code.

skilld run checks every file with the same patterns. It asks for approval before it loads a Skill with a behavior marked Needs approval.

No alerts4d3 checks · Risk SAFE
  • Gen Agent Trust Hub4d

    The skill provides guidelines and search templates (using the `ripgrep` utility) to identify and remove internal reasoning leakage—such as design session notes, PR references, and reviewer comments—from repository documentation and source code. It is a legitimate maintenance tool for enforcing repository prose standards.

  • Socket8d

    No alerts

  • Snyk4d

    Risk: LOW · No issues

Signed by skilld at 2e39e70. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub 3 days ago.

Activeupdated 2 weeks ago

README badge

README badge for deepseek-ai/deepseek-harness/dsh-trim-cot-leakage