plat 002

← Pulse platform issue index

PLAT-002 — nested tool failures remain semantically successful

Coordination	Value
Assigned agent	Claude Code
Ticket state	runtime_reverify
Last synchronized	2026-08-04

Claim this ticket in this file before implementation. During active work, update this fragment rather than the shared index; synchronize the index once at handoff, review, or completion.

• Priority: P0
• Owner: tool bridge, timing telemetry, and terminal status
• Source findings: HARNESS-NESTED-ERROR-STATUS-PRECEDENCE (Upwork) and HARNESS-TOOL-ENVELOPE-ISERROR-2026-08-03 (Build-in-public), plus HARNESS-TIMING-EMBEDDED-TOOL-ERROR (Social Media, observed before the canonical implementation landed)
• Problem: the outer transport succeeds while the nested tool payload says ERROR, carries a non-zero nested exit code, or reports an HTTP failure. Stored traces still set IsError=false and errored_count=0.
• Impact: retries, alerts, validation, reviewers, and terminal status can treat a real failed operation as clean.
• Important distinction: tool_failures_invisible_in_backend_logs.md fixed visibility with [TOOL_ERROR] logs and red UI rendering. It did not by itself make the canonical runtime/timing result an error.
• Implementation (2026-08-03): mcpagent/toolerr now has a narrow canonical classifier separate from the broad log-only suspect detector. The CLI stream adapter emits ToolCallErrorEvent instead of ToolCallEndEvent for nested failure envelopes, and saved CLI conversation history sets IsError=true. Sequential and parallel in-process tool paths use the same classifier. Problem-reporting/query tools are excluded from payload promotion so a returned domain row such as status=failed is not confused with transport failure.
• Verification: fixtures pass for nested ERROR, non-zero shell exit, permission denial, HTTP 4xx, success=false, and MCP isError; negative controls pass for prose discussing errors and historical failed DB rows. The real post-build timing artifact still needs to prove errored_count changes.
• Current workaround: agents and reviewers parse nested stdout/content and apply explicit error precedence themselves.
• Acceptance: fixtures for nested ERROR, HTTP failure, permission denial, and non-zero shell exit all set canonical error state, increment error counts, and prevent an unrecovered parent execution from being clean. Text merely discussing an error remains a success.