## TL;DR

Do not paste multi-megabyte Playwright traces into an agent's context. Upload the trace as a CI artifact and give the agent the artifact link plus a short structured summary: which test failed, the failing action, the console errors. The agent heals from the summary; a human opens the full trace in the viewer when needed.

## Error

```text
agent's playwright trace attachments too large for context: selective upload
```

Trace files attached to every failure eat the agent's context window, slowing healing loops and crowding out the actual error.

## Steps

1. Measure: check trace sizes from a recent run and compare against the agent's context budget. Expected output: a number (e.g. traces are 40MB per failure) that justifies the policy.
2. Change the reporter config: keep traces on first retry only, or on failure only, instead of always. Expected output: far fewer trace files per run.
3. Upload traces as CI artifacts with a retention window (e.g. 7 days), and pass the agent only the artifact URL plus a generated summary. Expected output: the agent's prompt contains a link and a summary, not megabytes of trace JSON.
4. Generate the summary automatically: failing test name, last action before failure, console errors, and the failing selector. Expected output: a consistent short brief the agent can act on.
5. Confirm the healing loop still works: the agent fixes failures from summaries, and humans can still open full traces from the artifact store. Expected output: same fix rate, smaller prompts, traces available on demand.

## When to use

- Trace attachments are pushing the agent near its context limit.
- Every healing iteration is slow because the prompt carries full traces.
- You want traces kept for human debugging but not fed to the agent raw.

## When not to use

- The debugging task genuinely needs the trace timeline (then a human opens the trace viewer; the agent is the wrong tool).
- Traces are small (a few KB); the plumbing is not worth it.
- You have no artifact store; fix storage first, then add selective upload.

## Variant phrasings

### Video recordings too large for the agent's context

Same treatment: artifact the video, summarize the failure, link the file.

### Screenshot spam on every retry

Keep screenshots for the final failure only; earlier retries get text logs.

### HAR files attached to every API test failure

Artifact the HAR, give the agent the failing request/response pair extracted from it.

## Why it happens

The default Playwright config retains traces generously, and agent harnesses naively attach everything to the prompt. Traces are designed for the interactive viewer, not for language-model consumption: most of their bytes are DOM snapshots and network payloads the agent never uses.

## Edge cases

- The summary generator can miss the real cause (e.g. a failed request hidden among 200 passing ones); include failed requests explicitly, not just the last action.
- Retention expiry can delete a trace an open investigation still needs; pin retention for quarantined tests.
- Artifact URLs behind auth need the agent to have read access; otherwise the link is useless to it.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_dzbJ_1F6kHrfmxS7XU8uOQ