← All use cases
How it works
How to use /prode-triage
One command turns a raw production ticket into evidence, a surgical fix, a reviewed PR, and a clean Linear handoff. Here's how to drive it.
When to reach for it
Use /prode-triage the moment a Production Engineering ticket lands. It's built for one job and does it end to end.
- A Linear ticket prefixed
PRODE- is assigned or referenced ("fix PRODE-242").
- A Sentry issue under
islamicfinanceguru-limited with a linked PRODE ticket needs investigating.
- A CloudWatch / AWS Batch / RDS CPU alarm maps to a PRODE ticket (the AWS-ops sub-class, no Sentry link).
Not for non-PRODE work (CAR, OXY, TR, COB), greenfield features, or repo hygiene. The "PRODE-only MCP" rule is scoped to exactly these tickets.
How to invoke it
Drop the ticket key into the prompt. The skill fires automatically, or you can name it explicitly.
fix PRODE-242
# or, explicitly
/prode-triage PRODE-242
Golden rule: the agent pulls the Linear ticket and its linked evidence before it reads a single line of code. The ticket is ground truth. Diagnosing from the title alone is how earlier agents fabricated fixes, and the skill exists to stop that.
What it does, in order
- Pull the ticket + evidenceFetches the Linear description, labels, comments, attachments. If it's
sentry-issue or has a Sentry URL, pulls the full event payload first: stack, tags, release, user, counts, first/last seen.
- CloudWatch correlation (backend only)For
invest-api tickets it queries Cur8ApiLogGroup in a narrow window around lastSeen, pulling the logged request body. This is the line between a defensible fix and a guess.
- Classifybug / regression / noise / needs-investigation, plus severity (P0–P3) and a confidence %. Below 70% it stops and asks rather than guessing.
- Worktree + branchA fresh worktree per ticket at
new-wealth-prode-NNN, branch PRODE-NNN/short-slug. Never reused across tickets.
- Linear progress logKeeps a single
## Agent Progress section in the ticket current at every checkpoint. No local notes files. Mandatory comment for every sentry-issue.
- Surgical fixTouches only what it must. Separates the alert from the failure, and never silences a real user-facing error on a critical path (login, auth, payments, onboarding).
- TestsPins the exact production payload shape, adds a negative control that must stay captured, and asserts the ops trail when a
logger.warn is added.
- Opus first-principles reviewMandatory before any PR. A reviewer subagent challenges the diagnosis, scope, edge cases and tests. It has caught a real bug on every shipped PR.
- CI-gated PRDispatches the app's manual unit-test workflow, waits for green, then opens a ready-for-review PR (never a draft) following the repo template with Sentry + CloudWatch evidence.
- Linear handoffSets state to
In Review, re-asserting it after the GitHub integration tries to override it, assigns to Sayf, and confirms the final state before reporting back.
What you get back
- Evidence Sentry + CloudWatch findings, classification, confidence, root cause with file:line citations.
- A reviewed PR surgical diff, tests, Opus review applied, CI green, ready for human review.
- A clean ticket Linear description and status updated, evidence comment posted, assignee set.
After each PR the agent pauses and reports. It won't chain into the next ticket without your go-ahead.
Things to know
- cur8-mob isn't Done on merge. Mobile fixes ship via app release / CodePush, so those tickets hold at
In Review until the release carrying the fix is verified.
- AWS access is read-only. CloudWatch logs, SSM reads, and read-replica
SELECTs only. If it needs a write or hits an MFA deny, it stops and asks you.
- No external IDs in code comments. The trace lives in Linear and the PR body, never in the diff.