Preserve MCP tool error outcomes

[xeophon]

Overview

Preserves MCP tool-declared outcome semantics across harness execution, typed messages, traces, replay, and provider conversion.

This PR follows #2134 in a native GitHub pull request stack. #2134 owns the shared null/bash MCP client and its session lifecycle and replay-safety behavior.

Details

MCP CallToolResult contains both model-visible content and an isError outcome. The shared MCP client now returns both values, and the harness program carries the outcome into ToolMessage.is_error.

The typed outcome participates in graph identity and trace serialization, so success and failure results with identical content remain distinguishable after persistence and resume.

For chat-compatible providers, the localhost harness handoff uses a private marker that is removed before the provider request while adding a readable failure label to the model-visible content. Anthropic-native tool_result.is_error maps directly into the same typed representation.

---

Note

Medium Risk
Touches message serialization, eval apply_overrides, and graph identity; behavior changes for tool failures on chat providers (prefixed content) but is scoped to error semantics rather than core auth or data paths.

Overview
Adds ToolMessage.is_error so MCP and harness tool failures are typed end-to-end, not only as text in content.

MCP → harness: call_mcp now returns (content, isError) from CallToolResult; the bash/null program sets a private _vf_is_error marker on tool wire messages and passes messages through extra_body so interception keeps that field.

Chat providers: message_to_wire strips the private marker before upstream requests and prefixes model-visible content with Tool execution failed:\n on errors; internal handoff (initial-messages JSON) uses include_internal=True to round-trip the flag. apply_overrides re-serializes messages that still carry the marker so eval overrides don’t leak private fields to the API.

Anthropic: tool_result.is_error maps into the same is_error field on parse.

Traces / replay: message_hash includes error vs success so identical tool bodies dedupe differently after persist/resume.

^{Reviewed by Cursor Bugbot for commit 47ea576. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Preserve MCP tool error outcomes across the message pipeline

• Adds is_error: bool = False to ToolMessage in types.py to carry tool failure status through the message model.
• Updates mcp_client.py so call_mcp returns a (content, is_error) tuple, and program.py propagates that flag via the internal _vf_is_error key on tool messages.
• Updates chat.py to parse, serialize, and rewrite tool error markers: message_to_wire prefixes errored content with a human-visible banner unless include_internal=True, and ChatDialect.apply_overrides strips internal markers before forwarding requests upstream.
• Updates anthropic.py to set is_error when parsing Anthropic tool_result blocks.
• Behavioral Change: tool messages with identical content but different error status now produce different hashes in message_hash.

^{Macroscope summarized 47ea576.}

[macroscopeapp]

Approvability

Verdict: Needs human review

This PR propagates MCP tool error state through the message pipeline, changing function signatures, message formatting (adding 'Tool execution failed:' prefix), and message hash computation. These runtime behavior changes, particularly the hash modification that could affect caching/deduplication, warrant human review.

^{You can customize Macroscope's approvability policy. Learn more.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 1e47ec2. Configure here.}