Stop an interrupted tool call wedging a harness node forever

[dovvnloading]

Problem

backend/harness/loop.py:575-581 appends the assistant turn — tool_calls and all — and persists it to the transcript before invoking any tool. An interruption in that window (a Stop, a timeout, a provider fault, the process dying) leaves the transcript ending on an assistant turn whose calls were never answered.

Measured against the real transcript writer and loader — a Stop before either tool ran:

reloaded history:
   user       tool_calls=-        id=-
   assistant  tool_calls=c1,c2    id=-

  requested tool_calls : ['c1', 'c2']
  answered             : []
  UNANSWERED           : ['c1', 'c2']
  history ends on      : assistant WITH tool_calls

loop.py:459 then appends the follow-up user message after that, so the provider is handed assistant(tool_calls=[c1,c2]) → user("try again"). Every major provider rejects that shape: OpenAI requires a tool message per tool_call, Anthropic a tool_result per tool_use.

The malformed turn is on disk, so every subsequent follow-up fails identically. The node is wedged permanently, and the only visible symptom is a provider error the user cannot act on.

drop_leading_orphan_tools already handles the mirror case — a history that opens mid tool-sequence, from a tail cut, with a docstring saying "the rule lives in one place". This is the other end of the same rule, and it was missing.

Change

close_unanswered_tool_calls synthesizes a result for every requested call that has none, so a reloaded history always satisfies the one-result-per-call contract.

Two deliberate choices:

Repaired on load, not when the run lands. That fixes transcripts already sitting on disk from before this existed — users with an already-wedged node recover on the next follow-up — and covers interruption paths that never reach a landing handler at all (a crash, a kill).

A synthetic result rather than dropping the assistant turn. The turn is a real record of what the agent decided to do; dropping it would make the transcript lie about the run. The synthetic content says plainly that the tool was interrupted.

Test plan

• 5 new tests: a fully interrupted turn gets synthetic results for both calls; a partially answered turn (a Stop between two calls) keeps the real result and fills only the gap; every requested call ends up answered — the invariant providers actually enforce, asserted directly; a healthy transcript is untouched; and an assistant turn with no tool_calls is not disturbed.
• Full suite: 3205 passed, 19 skipped. ruff check . clean.

🤖 Generated with Claude Code