[Broccolito]

Summary

Three defects the final regression run found on main after #403-#407, in the local providers' stream retry. Each was reproduced independently before it was fixed.

1. A continued reply lost the space where it joins the cut-off text.
   • Bug: a reply cut at a word boundary and then continued read "The quick brown foxjumps over the lazy dog." This reproduced with a real GPT-5.5 proxied through a mock llama-server.
   • Cause: recover_leaked_tool_calls (providers/formats/text_tool_calls.rs) trimmed the leading whitespace of the first text of every completion, continuations included.
   • Fix: the agent now marks a provider call that continues the model's own cut-off text (CONTINUES_A_REPLY), and recovery keeps that call's leading whitespace. The mark is read from the shape of the history: the model's text followed by an agent-only user row. That shape also matches a few fresh replies after another agent-only note (self-critique or done-gate feedback, a supervision prompt). Those keep a leading space too, which is harmless. Whitespace alone is still no reply.
2. Stop during a retry left a "continue the reply" order in front of the next message.
   • Bug: stopping while the retry of a partial reply waited left the retry note agent-visible, and the person's next message went out merged behind it. With GPT-5.5, "test" was answered with " jumps over the lazy dog."
   • Fix: a Stop or failed stream while a retry waits now hides that note from the model in place, with a compare-and-set of its metadata; nothing is deleted. This happens on every surface:
     • the daemon runner and Crew, in settle_interrupted_turn and record_turn_stopped;
     • the CLI and ACP, which call Agent::retire_unanswered_retry_note once the turn's token is cancelled.
   • A note whose retry ran, or one followed by an agent-visible row, stays.
   • The pattern predates this cycle: before fix(agent): keep a retried turn answering the user, not the error #405, every retry left a similar note.
3. The GPU out-of-memory notice said "try again" right before "Trying again (1/1)…". It now follows fix: post-merge edge cases in stream retry, serve behind a proxy, and script artifacts #407's rule for the compute-failure notice. This wording was already in v1.93.0.

Test plan

• New tests: text_tool_calls.rs (a continuation keeps its joining whitespace), llamacpp_sidecar.rs (an_out_of_memory_failure_does_not_ask_the_person_to_try_again), and five retire tests in tests/output_recovery_agent_loop.rs (Stop, failed stream, the reply loop's own stop, a client that settles no turn, a client that drops the stream). Every new test fails with its fix undone, checked by mutation.

• Ran: cargo fmt --check, ./scripts/clippy-lint.sh, cargo test -p biorouter --lib -- providers:: agents::, --test output_recovery_agent_loop, and the steer and stack binaries (soft_interrupt_agent_loop, turn_abort_tests, subagent_delegation, steer_always_lands).

• Live against a mock llama-server, on the desktop's HTTP path, the CLI and ACP: each reproduction fails with the pre-fix binary and passes with this build.

• Known and not changed here:

  • a stopped delegated subagent settles through its own path and keeps the note;
  • a steer that lands after the note keeps it visible.

  Both are narrower than the reported Stop path.

🤖 Generated with Claude Code