fix(chat): surface the full error in a copyable block

[Dani Akash (DaniAkash)]

Problem

Chat uses the Vercel AI SDK useChat. When a turn fails, the classifier built the error envelope's details from the short constructed error.message and capped it at 500 characters. The gateway's real, actionable reason (OpenRouter-style gateways wrap it in the response body / metadata.raw) was captured at the fetch boundary as APICallError.responseBody / data.raw but never forwarded, so the user saw only a generic sentence. Errors the AI SDK could not classify carried no details at all, leaving nothing to copy or act on.

Team decision: any error that cannot be fully expressed through the AI SDK's masked stream should be shown verbatim, after scrubbing secrets, so users are empowered to fix it.

Change

• Classifier: forward the full upstream response body (or structured data.raw), pretty-printed as JSON, into details on every path, redacted and bounded at a larger cap so the whole body survives to be copied. Errors that fall through to the generic path now carry the scrubbed raw message too.
• Error card: render that detail always expanded under a Full error heading with a copy control. The copy targets the full string, not the DOM, so the entire error is copied even when the block clips it to a scroll area on screen. The classified title and message (credits exhausted, session expired, rate limited, and so on) stay on top, unchanged.
• The block is omitted when the detail only repeats the message, so a clean one-line error is not duplicated.

No wire or schema change: this reuses the envelope's existing details field. The side panel and new tab share the card component, so both surfaces are covered.

Tests

• Classifier: full response body preserved and pretty-printed in details; falls back to data.raw; secrets redacted and a pathological body bounded; a plain error surfaces its scrubbed message as copyable detail.
• Card: the full error renders always expanded with a copy control and no toggle; the raw specifics a generic message hid are visible; the block is suppressed when it only repeats the message; the classified message stays above it.
• Type-check and lint clean across both packages.

Follow-up

If credit exhaustion is ever returned as HTTP 200 with the error embedded in the SSE stream (rather than a 4xx the custom fetch already intercepts), the stream body would need inspecting too. The current 4xx handling covers the HTTP-error path, so that is a separate, verifiable follow-up.

[github-actions]

✅ Tests passed: 1211/1214

Ran 8 of 15 suites (7 not affected by this change).

Suite	Passed	Failed	Skipped
✅ server-agent	298/298	0	0
✅ server-api	163/163	0	0
✅ server-tools	254/254	0	0
✅ server-browser	10/10	0	0
✅ server-integration	10/10	0	0
✅ server-lib	120/120	0	0
✅ server-root	38/41	0	3
✅ agent	318/318	0	0
⏩ claw-app	n/a	n/a	not affected
⏩ claw-onboard	n/a	n/a	not affected
⏩ build	n/a	n/a	not affected
⏩ release	n/a	n/a	not affected
⏩ claw-server-rust	n/a	n/a	not affected
⏩ claw-server-rust-quality	n/a	n/a	not affected
⏩ claw-mcp	n/a	n/a	not affected

View workflow run

[greptile-apps]

Greptile Summary

This follow-up expands chat error details so users can inspect and copy scrubbed upstream responses while retaining the classified error message.

• Raises the detail-size bound and prefers full response bodies or structured raw provider diagnostics.
• Broadens credential redaction across several token formats and sensitive JSON fields.
• Replaces the details toggle with an always-expanded, scrollable Full error block and clipboard control.
• Adds server and client coverage for formatting, redaction, truncation, rendering, and copying behavior.

Confidence Score: 3/5

This PR should not merge until arbitrary upstream diagnostics are scrubbed without exposing unsupported credential formats in the error card and clipboard.

The reply from the unnamed thread participant states that the scrubber was broadened, but the current finite regex list still forwards Basic authorization and cookie/session credentials unchanged, providing a concrete residual disclosure despite that fix claim.

Files Needing Attention: packages/browseros-agent/apps/server/src/agent/chat-error.ts

Security Review

The attempted secret-scrubbing fix remains incomplete: common credentials such as Basic authorization and cookie/session values can pass through the finite regex list into the always-visible and copyable upstream error details. How this was verified: The complete redaction list lacks Basic-auth and cookie patterns while responseBody and data.raw flow through it directly to the rendered and copied details string.

Important Files Changed

Filename	Overview
packages/browseros-agent/apps/server/src/agent/chat-error.ts	Forwards richer upstream diagnostics and broadens redaction, but the replacement scrubber still permits common credential representations into client-visible details.
packages/browseros-agent/apps/app/screens/sidepanel/index/ChatError.tsx	Formats, displays, and copies the full details string through the shared error card as intended.
packages/browseros-agent/apps/server/tests/agent/chat-error.test.ts	Adds substantial coverage for upstream detail selection, formatting, truncation, and several supported secret formats.
packages/browseros-agent/apps/app/screens/sidepanel/index/ChatError.test.tsx	Covers the always-expanded Full error presentation, copy control, duplicate suppression, and preservation of the classified message.

Prompt To Fix All With AI

### Issue 1
packages/browseros-agent/apps/server/src/agent/chat-error.ts:76-80
**Unsupported credentials remain exposed**

If an upstream diagnostic contains a credential such as `Authorization: Basic <base64>` or `Cookie: session=<opaque value>`, this finite regex scrubber leaves it unchanged before the response body is displayed and copied from the Full error block, exposing the credential to the user and any pasted bug report.

**How this was verified:** The complete redaction list has no Basic-auth or cookie-value rule, while `responseBody` and `data.raw` flow through it into the rendered and copied details string.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

_{Reviews (2): Last reviewed commit: "fix(chat): broaden secret scrubbing for ..." | Re-trigger Greptile}

[Dani Akash (DaniAkash)]

Greptile (@greptileai)