observe

name

observe

description

Feedback step that closes the loop from production back into the harness — take a signal you BRING to it (a pasted stack trace, log snippet, error text, or an issue/monitoring URL), triage it into a located, reproducible failure, and route it into the existing /fix lane. Bring-a-signal only — no polling, no credentials, no new MCP. It locates the failing area and shapes a repro SEED; it does NOT write the finished failing test — that is /fix's job. Opt-in; invoked explicitly. Use when production surfaced something and you want it fixed under the harness's gates.

/observe — triage the signal, locate the area, route to /fix

Goal: turn a raw production signal into something the harness can act on. The user brings the signal (the harness does not watch anything); /harness-claude:observe grounds it in real code — which component, what failure, what would reproduce it — and hands that package to /harness-claude:fix, which runs the reproduce-first discipline. This is the step that makes the pipeline a loop rather than a line: …/ship → /deploy → (prod) → /observe → /fix → ….

Opt-in. Invoked explicitly. Bring-a-signal only — no polling, no credentials, no new MCP; you paste/point at the signal. Git boundary: /harness-claude:observe is read-only — it never commits, pushes, or branches; the fix is committed (if at all) only later through /harness-claude:ship.

When /harness-claude:observe, when /harness-claude:fix

• /harness-claude:observe is for a prod signal you don't yet know how to reproduce in code — it triages and locates, then routes. Start here when all you have is a stack trace / log / alert.
• /harness-claude:fix is the lane itself — reproduce-as-a-failing-test-first, root-cause, minimal fix. If you already have a red test or a known repro, skip observe and go straight to fix.

How it works

parse the brought signal → locate the failing area (mgrep/graph) → shape a repro SEED → hand to /fix

Do this

1. Parse the brought signal. From the pasted stack trace / log / error / issue URL, extract a structured failure description: the service or component, the error message or pattern, the conditions it appears under, and any data visible in the signal. No polling, no credentials, no new MCP — you work only from what was handed to you. (AC-O1)
2. Locate the failing area in real code. Use mgrep / the knowledge graph to trace the signal to its origin — the function, module, or path the error points at. Ground the signal in the codebase rather than guessing; name the suspected location. (AC-O2)
3. Shape a repro SEED — not the finished test. Describe the smallest input/state likely to trigger the failure (or concrete manual repro steps when automation genuinely isn't possible — the same carve-out as /harness-claude:fix step 1). Do NOT write the finished failing test here — capturing the bug RED is /harness-claude:fix's non-negotiable first step; duplicating it blurs the boundary. (AC-O3)
4. Hand off to /harness-claude:fix — then stop. Pass the package: the structured failure description (1), the located area (2), and the repro seed (3). /harness-claude:fix runs reproduce → root-cause → minimal fix-plan and rides the existing review/verify/ship gates. /harness-claude:observe does not duplicate any of that, and stays read-only throughout. (AC-O4)

Security & trust

• A brought signal (stack trace, log line) may contain PII or secrets. Do not echo raw log content into any persisted artifact; carry forward only the structured failure description needed by /harness-claude:fix.
• Read-only: /harness-claude:observe never writes to the repo. The git boundary holds.

Notes

• Self-contained: a single skill that leads and reuses existing tools — mgrep / the graph (locate) and /harness-claude:fix (the hand-off). No new agent, no new dependency, no new MCP server, no new runtime.
• Active monitoring is deferred (YAGNI). Polling Sentry / Datadog / CloudWatch would need an MCP + credentials + a running poller — out of scope for v1. The intake is a clean seam: a future source adapter could push signals in, but it is built only when a backend actually exists.
• The skill is the lead — there is no separate observe agent (same pattern as /harness-claude:fix, /harness-claude:orchestrate, /harness-claude:operate).

Exit criterion

The brought signal was parsed into a structured failure, the failing area was located in real code, a repro seed (or manual steps, with the reason automation wasn't possible) was shaped, and the package was handed to /harness-claude:fix — never a finished RED test produced here, and never a commit made.