Build an evidence-first agent that investigates production tickets, prepares safe fixes, and stops before irreversible actions.
The suite is a ticket-solving system, not a single prompt. It separates orchestration, evidence gathering, memory, implementation, and approval so each run can be reviewed.
Owns the Linear ticket, decides what evidence is needed, and chooses stop, handoff, or PR.
Read bounded sources such as Sentry, AWS, the DB replica, vendors, releases, and code paths.
Combines facts, source links, timestamps, confidence, gaps, classification, and approval state.
Stores operational memory that routes future investigations but never replaces current evidence.
Stop writes, risky deploys, broad suppression, PR merge, and final ticket closure until a human approves.
Paste this into a coding agent to start a PRODE run. It gives the agent the right shape before it touches code: read Linear, inventory tools, gather live evidence, then decide.
Use for a dry run, ticket investigation, or runner implementation task.
You are running the PRODE production engineering workflow. Start from the Linear ticket, not the source code. 1. Read the Linear ticket, labels, comments, attachments, linked Sentry/AWS/request IDs, and timestamps. 2. Inventory available tools: Linear, Sentry, AWS/CloudWatch, GitHub, production DB read replica, vendor/support sources, browser/emulator, and repo skills. 3. Read only relevant PRODE Wiki pages. Treat them as routing memory, not evidence. 4. Route read-only specialists for the evidence sources required by the ticket. 5. Build one evidence bundle: sources, facts, confidence, gaps, classification, and approval state. 6. If confidence is below 70%, evidence is missing, specialists conflict, or approval is needed, stop and write the handoff. 7. Only after the evidence gate may you read source code, patch the smallest safe path, verify, ask for review, open a PR, and update Linear. Never perform production writes, deploy-risk changes, broad Sentry suppression, Linear Done/Cancelled transitions, or PR merge without fresh explicit human approval.
.agents/skills/prode-triage/SKILL.md.The runner owns durability. The coordinator owns the ticket decision. Specialists answer bounded evidence questions. Humans approve irreversible actions.
Lease the Linear ticket, read the ticket snapshot, and collect linked production evidence.
Inventory tools, read relevant memory, and route read-only specialists.
Build the evidence bundle, set confidence, classify, and choose stop, handoff, or PR.
Patch only after the evidence gate, verify, open a PR, update Linear, and monitor after human merge.
Build the suite from small, inspectable pieces. Each piece has one job and a clear boundary.
Owns the ticket decision, evidence gate, classification, PR handoff, and Linear updates.
Defines read-only roles for Sentry, AWS, DB, vendor, support, release, reproduction, code path, and review.
Persists leases, phase state, idempotency keys, approval waits, artifact pointers, and post-merge monitoring.
Stores reviewed operational memory that can route investigation but cannot prove the current incident.
Keep secrets, direct identifiers, copied rows, payloads, and customer-sensitive data out of git and Linear.
Checks that a run is reproducible from Linear, specialist outputs, evidence bundle, PR, and monitor result.
The agent should not rely on conversation memory. Give it structured resources it can read and write in every run.
Available MCPs, connectors, CLIs, credentials, blocked tools, and selected routes for this ticket.
Current sources, facts, confidence, gaps, classification, and approval state.
Question asked, sources queried, findings, confidence, gaps, and next pivot.
Lease, phase, idempotency namespace, artifact pointers, approval records, PR URL, and terminal outcome.
Redacted candidate learnings from completed runs. Humans review before promotion to the PRODE Wiki.
Keep ADRs and workflow files readable by agents without requiring browser-only context.
Specialists are not mini-coordinators. They answer one evidence question using read-only tools and return a structured result.
Read issues, events, tags, releases, breadcrumbs, and recurrence signals.
Read logs, metrics, service health, queue age, task state, and account identity.
Run narrow, read-only aggregate checks. No row copying, locks, writes, migrations, or backfills.
Check provider status, webhook IDs, support reports, screenshots, expected behavior, and customer-safe summaries.
Compare deploy windows, feature flags, app versions, local reproduction, staging, browser, and emulator evidence.
Read source after the evidence bundle exists. Review the final diff before handoff.
Each example has the same rhythm: intake, evidence, decision, outcome. Not every run should produce a PR.
Sentry shows onboarding API 500s after a deploy.
CloudWatch creates a ticket for queue saturation.
A user-visible count disagrees with backend totals.
SELECT checks using a read-only role.A burst of request-aborted errors appears after mobile clients close a long-running screen.
These rules are the operating contract. They make automation useful without letting it become unsafe.
Current production evidence must exist before source investigation and implementation.
PRODE Wiki pages can route the investigation, but only current evidence can classify the ticket.
Sentry, AWS, vendor, support, release, browser, and DB specialists run with read-only boundaries.
Production writes, deploy-risk changes, broad telemetry suppression, PR merge, and Linear Done/Cancelled are human-owned.
FOR UPDATE, the run stops and asks for approval.
The page is the front door. The ADRs remain the source of detailed rationale and tradeoffs.
PRODE Wiki Operational Memory
Explains the Karpathy-style operational memory layer and why memory cannot replace current evidence.
PRODE Engineering Agent Suite
Explains the coordinator, specialists, runner contract, safety gates, and rollout decision.
Use these checks before calling the suite ready.
A low-risk ticket produces a reproducible evidence bundle without reading source code first.
Retrying a run does not duplicate Linear comments, branches, PRs, specialist artifacts, or candidate learnings.
Production writes, merge, deploy-risk actions, Sentry suppression, and Linear Done/Cancelled all pause for exact human approval.
A reviewer can reconstruct the run from Linear, tool inventory, specialist outputs, evidence bundle, PR, and monitor result.