Greptile Summary

This PR adds a targeted fix for JSON parse errors from the Anthropic SDK's SSE streaming layer leaking as raw error messages to Discord channels. It introduces isStreamingJsonParseError() in errors.ts and wires it into both formatAssistantErrorText() and sanitizeUserFacingText() (error context only), replacing the raw SyntaxError text with a user-friendly message. Five new unit tests lock in the behaviour for three error format variants and the sanitizeUserFacingText guard.

• New isStreamingJsonParseError() helper uses /\b(?:expected|unexpected)\b.+\bin json\b.+\bposition\b/i to detect V8 JSON.parse error strings from the SDK's SSE parser.
• Check is inserted in formatAssistantErrorText() after the existing isInvalidStreamingEventOrderError guard and before the role-ordering check — appropriate placement in the chain.
• In sanitizeUserFacingText() the check is correctly gated behind if (errorContext), so normal assistant text containing the word "JSON" is never affected.
• The regex is intentionally narrow (requires all three keywords: expected/unexpected, in json, position) to minimise false-positive risk. Node v22 V8 error messages for other JSON.parse failure modes (e.g. "Unexpected end of JSON input" or the newer "X is not valid JSON" format) will not match, but those are out of scope for issue JSON parse error leaked to Discord when streaming tool calls with long CJK text #59076.
• All five new tests pass; build verification confirmed by author.

Confidence Score: 5/5

Safe to merge — the change is purely additive, scoped to error formatting paths, and all tests pass.

No P0 or P1 issues found. The regex is appropriately scoped, the guard is only active inside error-context paths, and five new tests lock in the intended behaviour. All remaining observations are at most P2 style notes.

No files require special attention.

_{Reviews (1): Last reviewed commit: "fix: suppress raw JSON parse errors from..." | Re-trigger Greptile}

Thanks for chasing this leak down. Routing the fix through src/agents/pi-embedded-helpers/errors.ts is the right scope, and reusing the existing user-facing sanitization layer is better than patching this in a channel-specific path.

That said, I do not think the current implementation is robust enough yet. isStreamingJsonParseError() is still matching specific JSON.parse() message shapes:

• it depends on wording like expected|unexpected
• it requires in JSON
• it requires position

That is a pretty fragile contract for a user-facing fix, especially since JSON.parse() error text varies not only across Node/V8 versions, but also across different malformed-input cases on the same runtime. For example, Node 22 also emits messages like Unexpected end of JSON input and Unexpected non-whitespace character after JSON at position 4, and those would still fall through to raw user-visible text.

So my concern is not the scope of the fix, but the level it is implemented at: this currently looks more like a patch for the observed strings than a robust fix for the underlying error class.

I think the best path would be one of:

1. Preferably classify this earlier as a structured "invalid/malformed streaming fragment" error and then reuse that classification here.
2. If that is too large for this PR, broaden the matcher to cover the wider JSON.parse() SyntaxError family and add regression coverage for at least a truncated-input case plus a trailing-garbage case.

As written, I do not think we can confidently say this fixes the general leak class yet.

ProjectClownfish pushed a narrow repair to this branch so the original contributor path can stay canonical.

Source PR: #59118
Validation: pnpm -s vitest run src/agents/pi-embedded-helpers.formatassistanterrortext.test.ts src/agents/pi-embedded-helpers/errors.test.ts; pnpm check:changed
Contributor credit is preserved in the branch history and PR context.

Codex review: needs maintainer review before merge.

What this changes:

This PR classifies malformed Anthropic SSE JSON fragments as a stable transport error, renders friendly assistant/TUI/channel-facing text, adds regression tests, and adds a changelog entry.

Maintainer follow-up before merge:

This is already an open implementation PR with maintainer repair commits; the remaining action is human review, exact-head CI/rebase judgment, and landing rather than a separate automated fix PR.

Best possible solution:

Land the focused transport-boundary classification and shared rendering fix after maintainer review and exact-head validation, while keeping wider retry/failover work separate from this leak-suppression PR.

Acceptance criteria:

• pnpm test src/agents/pi-embedded-helpers.formatassistanterrortext.test.ts src/agents/pi-embedded-helpers/errors.test.ts src/agents/anthropic-transport-stream.test.ts src/tui/tui-event-handlers.test.ts src/tui/tui-formatters.test.ts src/tui/tui.test.ts
• pnpm check:changed

What I checked:

• Current main parses Anthropic SSE JSON directly: parseAnthropicSseBody() still yields JSON.parse(data) for normal frames and tail data, so malformed SSE payloads can still throw a raw SyntaxError before classification on current main. (src/agents/anthropic-transport-stream.ts:561, a972c9ec4547)
• Current formatter lacks malformed-fragment classification: formatAssistantErrorText() only has the invalid streaming event-order special case; no malformed_streaming_fragment, MALFORMED_STREAMING, or isStreamingJsonParseError symbol exists on main, and unclassified short errors still fall through to the raw text return. (src/agents/pi-embedded-helpers/errors.ts:1080, a972c9ec4547)
• Channel-facing reply path remains live: When the last assistant message has stopReason === "error", buildAgentReplyPayloads() calls formatAssistantErrorText() and pushes the returned text as an isError reply item, so the formatter fallback can still reach chat/channel surfaces. (src/agents/pi-embedded-runner/run/payloads.ts:216, a972c9ec4547)
• Main tests do not lock this leak class: The current formatter tests cover invalid streaming event order but not malformed Anthropic SSE JSON fragments or the proposed stable sentinel path. (src/agents/pi-embedded-helpers.formatassistanterrortext.test.ts:344, a972c9ec4547)
• PR head is current and focused: Remote PR head is c0bd453c5bb938447486d29aa0fb7f49baf29395; the fetched diff is limited to agent transport/error formatting, shared error rendering, TUI rendering/tests, a whitespace-only transport shared change, and the changelog. (c0bd453c5bb9)
• Security review pass: The current PR diff does not touch GitHub workflows, dependency manifests/lockfiles, package metadata, lifecycle scripts, permissions, secrets handling, downloaded artifacts, generated/vendor code, or publishing surfaces; it changes TypeScript error classification/rendering and tests. (c0bd453c5bb9)

Likely related people:

• steipete: Current-line blame in the checked-out main points to Peter Steinberger's recent main history across the Anthropic transport, assistant error formatter, shared error formatter, TUI error rendering, and channel payload path involved in this leak. (role: recent main maintainer / adjacent owner; confidence: medium; commits: 1fb58ca5eee0, a972c9ec4547; files: src/agents/anthropic-transport-stream.ts, src/agents/pi-embedded-helpers/errors.ts, src/agents/pi-embedded-runner/run/payloads.ts)
• hxy91819: Reviewed the original approach, called out the fragile JSON.parse string matching, and authored later branch commits that move classification to the transport boundary and prevent sentinel leaks through TUI/raw formatter paths. (role: reviewer and recent branch repair author; confidence: high; commits: 9b358ca8eb0d, 5e2716775d40, c0bd453c5bb9; files: src/agents/anthropic-transport-stream.ts, src/agents/pi-embedded-helpers/errors.ts, src/agents/pi-embedded-helpers/sanitize-user-facing-text.ts)
• vincentkoc: The PR discussion says vincentkoc pushed a narrow repair to keep the contributor path canonical, and the branch contains a commit preserving provider diagnostics while suppressing streaming parse leaks. (role: branch repair author / adjacent error-format maintainer; confidence: high; commits: 1735b05fb940; files: src/agents/pi-embedded-helpers/errors.ts, src/agents/pi-embedded-helpers/sanitize-user-facing-text.ts, src/shared/assistant-error-format.ts)

Remaining risk / open question:

• Exact-head CI/test proof for c0bd453c5bb938447486d29aa0fb7f49baf29395 still needs maintainer verification before merge; this read-only review did not run tests.
• The broader retry/failover behavior remains covered by Recurring Anthropic streaming JSON parse errors (Sonnet 4.5 / Opus) — failover often fails to recover #60593 and should not be treated as solved by this PR unless maintainers explicitly expand the scope.

Codex review notes: model gpt-5.5, reasoning high; reviewed against a972c9ec4547.

This assigned pull request has been automatically marked as stale after being open for 27 days.
Please add updates or it will be closed.

Merged via squash.

• Prepared head SHA: b8b3686445b2f7f2bce9e29974b3089873190c73
• Merge commit: 0544c6d493ff89887b3480aaf9fa7e326d9357fc

Thanks @singleGanghood!