[bug] native direct preflight aborts the whole tool batch on one malformed context call #310

New issue

Closed

opened 2026-09-02 17:25:37 +00:00 by nsaspy · 0 comments

nsaspy commented

2026-09-02 17:25:37 +00:00

Owner

Problem

In native direct mode, one malformed context-tool call causes the entire model-returned tool batch to fail before any valid sibling call can execute.

Observed in Auto-Dig integration runs #308, #309, and #310. In run 33268628425, the model returned three calls in one response:

context_peek
context_search
context_peek

A schema rejection on context_peek.selector.index caused:

phase=native_call
kind=malformed_arguments
message=native call batch failed preflight
context_calls=0
tool_calls=0

So even independently valid calls in the same batch never run, and the direct loop terminates after the first model response.

Expected behavior

Malformed tool arguments should be represented as a tool-level failure the model can recover from. One bad call should not necessarily kill the entire direct-mode trajectory or suppress unrelated valid calls in the same batch.

Required fix

Validate calls independently during native preflight.
Preserve stable ordering and call IDs.
Execute valid calls when safe even if a sibling call is malformed, or return structured per-call errors back into the loop so the model can correct them.
Do not terminate the whole direct run solely because one recoverable tool call has malformed arguments.
Keep hard-fail behavior for genuinely unrecoverable protocol/runtime corruption.
Add regression tests with a mixed batch containing at least one valid context call and one malformed context call.
Verify accounting remains correct for model calls, tool calls, context ops, and iterations.

The concrete schema mismatch that currently triggers this path is tracked separately in #312.

Consumer evidence: lost-rob0t/starintel-gpt-auto-dig#2095, workflow run 33268628425, Prolog-RLM ref eb2411afdaf981a0a121f55b7ddf23dc9fcef812.

## Problem In native direct mode, one malformed context-tool call causes the entire model-returned tool batch to fail before any valid sibling call can execute. Observed in Auto-Dig integration runs #308, #309, and #310. In run `33268628425`, the model returned three calls in one response: - `context_peek` - `context_search` - `context_peek` A schema rejection on `context_peek.selector.index` caused: ```text phase=native_call kind=malformed_arguments message=native call batch failed preflight context_calls=0 tool_calls=0 ``` So even independently valid calls in the same batch never run, and the direct loop terminates after the first model response. ## Expected behavior Malformed tool arguments should be represented as a tool-level failure the model can recover from. One bad call should not necessarily kill the entire direct-mode trajectory or suppress unrelated valid calls in the same batch. ## Required fix - Validate calls independently during native preflight. - Preserve stable ordering and call IDs. - Execute valid calls when safe even if a sibling call is malformed, or return structured per-call errors back into the loop so the model can correct them. - Do not terminate the whole direct run solely because one recoverable tool call has malformed arguments. - Keep hard-fail behavior for genuinely unrecoverable protocol/runtime corruption. - Add regression tests with a mixed batch containing at least one valid context call and one malformed context call. - Verify accounting remains correct for model calls, tool calls, context ops, and iterations. ## Related The concrete schema mismatch that currently triggers this path is tracked separately in #312. Consumer evidence: lost-rob0t/starintel-gpt-auto-dig#2095, workflow run `33268628425`, Prolog-RLM ref `eb2411afdaf981a0a121f55b7ddf23dc9fcef812`.

nsaspy closed this issue

2026-09-02 17:25:37 +00:00

nsaspy referenced this issue

2026-09-02 17:25:37 +00:00

[bug] context_peek schema accepts selector.index but native validator rejects it #311