An approved playbook change is shown in full before dispatch, then applied by a subagent only after explicit approval using a rollback snapshot, sanctioned service/admin writes, surgical edits, and post-write verification.Not checked
prod
The skill classifies every transcript-supported issue into playbook modifications, deeper investigations, or feature requests, records evidence and who raised it, assigns difficulty, and renders an inline difficulty diagram.Not checked
prod
A meeting-ended event can parse the customer and date from the meeting title, fire a Claude Code routine with the transcript, and post the resulting difficulty map and triage link to Slack while ignoring bot-originated Slack messages.Not checked
prod
Each deeper-investigation item is automatically sent to a parallel Sherlock subagent for evidence-based root-cause analysis, with findings written to a per-item file and no database, Slack, monitor, or pull-request writes.Not checked
prod
The skill persists the complete triage in customer/date-scoped triage.md and updates each item with checkbox status, decision, and follow-up as the user walks through the list.Not checked
prod
Each playbook-modification item receives an exact proposed change, target agent and blast radius, plus separate record-only behavior replay and read-only persistence proof, without applying the change.Not checked
prod
Libra found 1 production surface on prod but could not judge any of them yet. 1 surface had no requests at all. Libra checks hourly for 3 days after each deploy.
What
Adds a triage-enablement Claude Code skill. Given a customer and an enablement-call date, it:
finds the meeting and pulls the full transcript from the meetings bucket (the recorder notes alone overstate and rename things), stating the source at the top
triages into playbook changes / needs investigation / feature requests, ranked by difficulty
runs one Sherlock-style investigation subagent per investigation item, automatically and in parallel
gives every playbook change a package: proposed diff, agent, blast radius, and a dry run proven not to persist
holds interim findings and sends one rollup message at the end; the rollup is inline in chat, with no need to open a file
applies approved changes through subagents only (dry run first, snapshot, surgical patch, verify, rollback file)
writes a plain rep-facing wrap-up message when the user is done
Only .claude/skills/triage-enablement/SKILL.md is added. Nothing else changes.
Notes for review
The skill writes per-customer output to docs/enablement-triage/<customer>/<date>/. That folder holds transcript quotes and deal names, so it should not be committed.
The automatic meeting-ended trigger is described in the skill but not built here.
Verify
Run /triage-enablement <customer> <date> against a recent enablement call and check that the source line, the rollup format and the per-change packages come out as described.
### Issue 1
.claude/skills/triage-enablement/SKILL.md:29
The skill directs transcript lookup to `cedar-aws-staging-meetings-619071350399`, but the configured staging bucket is `cedar-aws-staging-meetings-619071350399-us-east-1`. Even when a meeting has a valid transcription key, this lookup targets the wrong bucket and can make the triage fall back to recorder notes instead of the full transcript.
### Issue 2
.claude/skills/triage-enablement/S
Show production surfaces and changed-file mappingHide production surfaces and changed-file mapping
Production surfaces
Surface
Requests
Errors
p95
Users
Verdict
/v1/claude_code/routines/<id>/fire
0 → 0
0 → 0 (0%)
not measured
0
No traffic
No requests recorded since this deploy.
Changed files → surfaces
.claude/skills/triage-enablement/SKILL.mdno production surface mapped
 Incorrect meetings bucket name ▶
 Customer artifacts are not ignored ▶
 Evaluator cannot prove dry runs ▶