Skip to content

thermo-nuclear-agent-triage-burndown.md

Latest commit

 

History

History
87 lines (66 loc) · 3.83 KB

thermo-nuclear-agent-triage-burndown.md

File metadata and controls

87 lines (66 loc) · 3.83 KB
Raw
description Burn down an agent's conversation-triage queue end-to-end: pull every pending ticket, cluster by root cause, fix each cluster (code PR for code, agent-config branch for prompt/procedure), flag existing unmerged work instead of duplicating it, then comment on and resolve every ticket. Usage: /thermo-nuclear-agent-triage-burndown <agent_id>

Thermo-Nuclear Agent Triage Burndown

Batch version of the fix-agent-qa-ticket skill for the whole queue instead of one ticket. Read that skill first — it defines the ticket API host/auth and the branch/procedure/sim-test mechanics this command reuses per cluster.

Args: agent_id (required, $ARGUMENTS). For an agent in an isolated residency environment, swap the host below to https://api.<eu|in|sg>.residency.elevenlabs.io; residency agents don't exist on the central host. Needs $API_KEY (xi-api-key header, host https://api.elevenlabs.io).

1. Pull every pending ticket

GET /v1/convai/agents/$AGENT_ID/conversation-triage-tickets?status=open, then again with status=in_progress — paginate both via cursor/has_more. qa_comment is the primary signal on every ticket — read it first. Only if turn_comments is non-empty, also pull the conversation (GET /v1/convai/conversations/{id}) and read the transcript around each turn_comments[].turn_index, as in fix-agent-qa-ticket step 2. source=manual tickets have no conversation to fetch.

2. Cluster

Group by root-cause theme, not surface wording — same broken tool, same prompt gap, same procedure. Before trusting a cluster's root-cause hypothesis, corroborate it against ≥2 tickets/transcripts in that cluster.

3. Classify each cluster's fix location

  • Codebase bug (backend route, tool handler, frontend) → PR in your codebase.
  • Agent config (prompt/procedure/KB gap) → agent-config branch, per fix-agent-qa-ticket steps 3–6 (branch off main's real tip, draft+publish dance for procedures, sim test to prove it).
  • Not actionable (duplicate, expected behavior, insufficient info) → no fix, just an explanatory comment.

A cluster can need both a code fix and a prompt tweak.

4. Check for in-flight fixes before creating new ones

  • Agent branches: GET .../agents/$AGENT_ID/branches?include_commit_status=true — a non-main branch with commits_ahead > 0 matching a cluster's topic is unmerged prior work. Reuse or flag it for merging instead of duplicating.
  • Code PRs: gh pr list --search "<topic>" for an already-open PR touching the same code.

5. Execute — nothing gets merged by this command

  • Codebase: one small draft PR per cluster, with tests for the touched area. Follow repo conventions (AGENTS.md): draft by default, no auto-merge.
  • Agent config: one branch per cluster with a sim test proving the fix, left unmerged.
  • Flag pre-existing unmerged branches/PRs found in step 4 for the user to merge rather than opening a duplicate.

6. Resolve every ticket

No bulk endpoint — loop per ticket, comment before resolving:

curl -s -X POST "https://api.elevenlabs.io/v1/convai/conversation-triage-tickets/$ID/comments" \
  -H "xi-api-key: $API_KEY" -H "Content-Type: application/json" \
  -d '{"comment": "<root cause>\n\n<PR/branch link, explicitly \"not merged\", or why not-actionable>\n\nWritten by <model>, using <harness>."}'

curl -s -X PATCH "https://api.elevenlabs.io/v1/convai/conversation-triage-tickets/$ID" \
  -H "xi-api-key: $API_KEY" -H "Content-Type: application/json" \
  -d '{"status": "resolved"}'

Leave a ticket open, with a comment explaining why, only when resolving it would paper over a decision a human genuinely has to make — don't silently skip it.

7. Report

List: clusters found, PRs opened (draft, unmerged), branches flagged for merge/reuse, ticket count resolved, and any left open with the reason.