Skills that Improve Development loops run: review pull request history for agent context and Tessl verifier improvements, and read recent agent runs to record recurring problems as project findings.
77
97%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
Low
Low-risk findings worth noting
After all three area coverage gates pass, deduplicate retained candidates against all project findings without a state filter: tessl finding list --project --limit 100 --json. Redirect each JSON page to raw storage with stderr separate, then parse the saved file. If parsing fails, preserve the error and retry a smaller page size; a local parse failure does not establish unavailable access.
The list envelope is data[], with fields under attributes. Read compact analysis and recommendation summaries as well as IDs, titles and states. Follow links.next using its decoded page[after] as --after. Inspect likely matches with tessl finding view --json and validate its shape rather than assuming the list envelope applies. Findings about the same subsystem, affected files or correction PR deserve comparison even if their titles differ.
Reuse matches already containing the evidence. Append genuinely new evidence only to an open match. Preserve ignored findings and report resolved recurrences for an editor. Do not change recommendations, reopen, resolve or ignore findings. Compare recommendation components as well as whole findings. When a distinct candidate includes an incidental change already recommended elsewhere, omit that duplicate change and cite the existing finding; related target files alone do not make the core problems identical. Add the deduplication result to each candidate note before preparing a deliverable payload.
A create payload has schemaVersion: 1, exactly one observations entry with observedAt and an array of evidence strings, and nonempty string fields title, analysis, recommendation and expectedImprovement. Write the title as one plain statement of the current gap, under 100 characters, naming the area and what is missing; keep mechanisms, file lists, PR numbers and history for analysis. Inside the recommendation string, state Context only, Verifier only or Paired, then give the corresponding concrete specification from the candidate decision. For a pair, keep both components linked to the same failure and avoid unrelated changes. Keep the existing API shape: recommendation remains a string, never an object or new top-level fields. Validate shape and evidence links, including rehearsals.
Create with tessl finding create --project --file <payload.json> --json. Append with tessl finding update --file <payload.json> --json, supplying only the single observations entry. With run credentials omit loopId/runId; the service supplies trusted provenance. Read back writes and verify project and observation run provenance. Reread before retrying an uncertain write.
In an explicitly requested rehearsal, save payloads and the report locally instead of publishing. Read existing findings when available; otherwise mark candidates not deduplicated and not ready to publish. Missing publication access does not block investigation.