Skip to content

docs(skills): slim sdk-integrations and sdk-vcr-workflows, fix dead spec link - #1110

Merged
Abhijeet Prasad (AbhiPrasad) merged 1 commit into
mainfrom
abhi/run-skill-health-check-b2f390ad
Nov 12, 2026
Merged

Abhijeet Prasad (AbhiPrasad) merged 1 commit into
mainfrom
abhi/run-skill-health-check-b2f390ad

Conversation

@AbhiPrasad

Copy link
Copy Markdown
Member

Summary

Ran the skill-health skill against 30 days of coding-agent traces for this repo (2026-10-13 to 2026-11-12, 164 sessions). Two skills stood out. Agents often never saw their full text, and sdk-integrations had a broken spec link.

sdk-integrations (23.3 KB → 11.1 KB)

  • Dead spec link. ## Spec pointed to braintrust-spec/docs/instrumentation-guide.md, which returns 404 (it moved to skills/instrumentation-spec/ in July). 26 of 42 sessions tried the dead URL; only 13 found the new location. The link now goes to skills/instrumentation-spec/SKILL.md, with a gh api command that fetches the guide.
  • Partial reads. Only 23 of 74 loads received every section. Codex bundles SKILL.md with other files in one command, and that output is cut off at about 10k tokens. Span design, metrics, attachments, span origin and the streaming-lifecycle review moved into references/, each linked with a note on when to open it. The rules were moved, not deleted.
  • Missing step. 26 of 42 sessions needed a follow-up asking to "combine redundant tests or remove low value tests". A required consolidation step is added before handoff: parametrize near-duplicate tests, drop overlapping ones, run make check-unused-cassettes, and report the test count before and after.
  • Removed references to sdk-wrapper-migrations, which isn't in this repo.

sdk-vcr-workflows (21.0 KB → 6.4 KB)

  • Only 5 of 20 loads received every section. The Discovery Engine setup and the recording-failure/debugging detail moved into references/. The "Repo Rules" section, which repeated AGENTS.md, is gone.
  • Fixed a stale wrappers/claude_agent_sdk/ path and the sdk-wrapper-migrations reference.

Both skills plus AGENTS.md now come to about 7.5k tokens, which should fit in a single batched read.

Expected impact

Failure cases are in the skill-health-sdk-integrations (52 rows) and skill-health-sdk-vcr-workflows (12 rows) datasets in bt-coding-agent-insights. Targets:

  • sdk-integrations: full-read share 31% → ≥80%. Fetches of docs/instrumentation-guide.md 26/42 sessions → 0. Test-consolidation follow-ups 26/42 → ≤20% of sessions.
  • sdk-vcr-workflows: full-read share 25% → ≥80%.

Re-check once each skill has about 10 sessions on the new version.

Test plan

  • Pre-commit passes
  • Every references/ link resolves
  • The new spec fetch command returns the guide

…pec link

Datasets: skill-health-sdk-integrations, skill-health-sdk-vcr-workflows (BT Staging)
Expect: sdk-integrations full-read share 31% -> >=80%; sessions fetching
docs/instrumentation-guide.md 26/42 -> 0; "combine redundant tests" follow-ups
26/42 -> <=20% of sessions. sdk-vcr-workflows full-read share 25% -> >=80%.
@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Nov 12, 2026 •

Copy link
Copy Markdown
Codex Review Summary: ✅ Completed

This comment shows the latest Codex review activity on this pull request.

Review Status Commit Review trigger
📝 Code Review ✅ Completed 2026-11-12T20:07:36.115119Z d19b81d PR opened
ℹ️ About Codex in GitHub

Your 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.

@AbhiPrasad
Abhijeet Prasad (AbhiPrasad) merged commit 5f3869e into main Nov 12, 2026
95 of 96 checks passed
@AbhiPrasad
Abhijeet Prasad (AbhiPrasad) deleted the abhi/run-skill-health-check-b2f390ad branch November 12, 2026 20:09
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