**TL;DR:** Ban the summary-only triage pass. Every triage round must end with exactly one new diagnostic action - a command to run, a log to pull, a diff to inspect - and the agent must keep a seen-states list of error signatures. After three passes with no new evidence, it escalates or changes strategy instead of writing a fourth description. Summaries feel like progress but change nothing.

```text
agent loop detected: repeatedly describing the same failed deploy in triage
```

## Steps

1. Write the current hypothesis as one sentence, plus the single command that would prove it wrong. Expected: you have a falsifiable command, not a paragraph of description.
2. Start a seen-states list. After each pass, append the error signature: the first line of the error plus which step failed. Expected: if a new pass produces a signature already in the list, it is declared a loop and no new summary is written.
3. Require every pass to end with exactly one new diagnostic action. Expected: each action returns output you have not seen yet - a new log line, a new command result, a diff.
4. Diff the deploy output between passes instead of re-reading it. Expected: the diff shows whether anything actually changed since the last pass.
5. Cap it: three passes with no new evidence means escalate. Either switch the repair layer (config vs code vs account vs network) or report back with the evidence gathered. Expected: the agent never writes a fourth identical summary.
6. Log the loop break explicitly: what was tried, what the evidence showed, what changed. Expected: the next agent (or the next run) starts from evidence, not from zero.

## Use this when
- triage passes produce descriptions of the failed deploy but no new commands or checks
- the agent's output reads the same on pass 4 as on pass 1
- a deploy failure has been "analyzed" multiple times with no repair attempted
- you need a forcing function that turns observation into action

## Not for this skill when
- the failure is already diagnosed and just needs the fix applied - skip triage, apply it
- each pass genuinely uncovers new evidence - that's working triage, not a loop
- the deploy error changes between passes - the loop is already broken
- a human is driving - this guard is for autonomous runs

## Variant phrasings
- agent keeps summarizing the failed deploy without acting
- triage loop with no new diagnostic steps
- agent describes the same Workers deploy failure every pass
- analysis paralysis on a failed Cloudflare deploy

## Why it happens
Summarization is cheap and reads like progress, so an agent with no forcing function re-derives the same conclusion forever. Triage is a state machine - hypothesis, test, evidence, next hypothesis - and without the seen-states list and the action requirement, there's nothing pushing the state forward.

## Edge cases
- A genuinely novel failure can look like a loop for two passes. The three-pass cap with the layer-switch rule handles this: new layer, new evidence.
- Don't confuse "same root cause, new evidence" with a loop. The signature list tracks evidence, not conclusions.
- If the deploy output is nondeterministic (flaky), diffing misleads. Note the flakiness explicitly and triage the flake, not the deploy.
- Long runs lose the seen-states list if it only lives in context. Persist it to the run's working files.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_ezT4VLgZmBDxJHm39iwq6Q