…tory (tinyhumansai#1871)

Claim A confirmed: a single blank completion from an ordinary (non-tool)
model response reaches AttemptOutcome::Empty. The previous comment calling
this path unreachable was wrong; corrected.

Claim B confirmed and fixed: when run_with_steer retried a failed blank
turn by calling agent.turn(message) a second time, the user message
appeared twice in the provider request. Two interacting bugs caused it:

1. After a failed turn, OpenHuman's record_failed_turn appends a
   '[turn failed]' assistant note to history AND persists the session
   transcript to disk.

2. On the retry, history was truncated back to its pre-attempt length.
   For the first turn ever (history_len_before_first == 0) this left
   history empty. The second turn() call saw empty history, fired
   try_load_session_transcript(), reloaded the just-persisted transcript,
   re-added the user message, and then pushed its own copy on top.

Fix: in the AttemptOutcome::Empty arm, always set
suppress_transcript_autoload = true on the retry's TurnOverrides
alongside the existing scope-preservation overrides. This prevents
try_load_session_transcript() from running on the retry, so regardless
of whether history_len_before_first is zero the retry sees exactly one
copy of the user message.

Also adds truncate_history_to() to vendor/openhuman's agent accessor
surface (minimal, no fork) and raises the recursion_limit in lib.rs to
256 to give rustc headroom for the deeper async generator types this
change introduces.

Regression tests (built_in_tests_part11):
- a_single_blank_script_reaches_the_empty_retry_arm: proves Claim A
- chat_only_turn_keeps_its_reduced_scope_on_the_empty_retry: overrides preserved
- a_normal_turn_keeps_its_full_toolbelt_on_the_empty_retry: control test
- an_empty_retry_does_not_duplicate_the_user_message_in_history: Claim B

All four tests fail on the unfixed code and pass after this commit.

Fixes tinyhumansai#1871

Co-authored-by: Cursor <[email protected]>

…rface

The empty-retry arm in run_with_steer called agent.truncate_history_to()
which did not exist in the vendored openhuman at the current pin.
Bumps the submodule pointer to the commit that adds the method.

Co-authored-by: Charan Rathore <[email protected]>

…er bump

The previous commit added truncate_history_to() to the vendored openhuman
and bumped the submodule pointer. CI cannot fetch a commit that exists only
in the local checkout — git submodule update would fail for any fresh clone.

This commit:
1. Reverts the vendor/openhuman pointer back to 93e9938f3 (the original pin).
2. Replaces agent.truncate_history_to(clean_len) with a save/clear/reseed
   sequence using the two public APIs that already exist on Agent:
     - clear_history() — removes all rows
     - seed_resume_from_messages(pairs, "") — sets cached_transcript_messages
       so the retry's turn absorbs the saved rows back via the normal resumed-
       prefix path, without hitting the durable-transcript file that
       record_failed_turn just wrote.

The retained rows (history[..clean_len]) are extracted as (role, content)
pairs.  After absorb_resumed_transcript_prefix the absorbed rows are all
Chat(ChatMessage) variants, so the filter_map captures them with full
fidelity for the common cases (cold start, chat-only turns, resumed
sessions).  Non-Chat rows (AssistantToolCalls / ToolResults from genuinely
warm multi-tool turns) would be dropped, but an EmptyProviderResponse on
such a turn is an extremely rare edge case — the retry still runs correctly,
just with less structural context.

suppress_transcript_autoload is set only when did_reseed is false (no rows
survived the filter), which means cached_transcript_messages was not set —
without the flag the retry's autoload guard would pick up the just-written
failed transcript and recreate the duplicate user message.

Co-authored-by: Charan Rathore <[email protected]>

Resolve the mod.rs conflict by adopting main's retry arm wholesale: at the
current OpenHuman pin (bd8ea4a) the session runtime is transactional, so a
blank first attempt commits nothing to history or the durable transcript and
the clear+reseed rollback this branch carried has nothing to roll back (and
no API left to run against). Rewrite the part11 regression suite for that
contract and widen the scripted-provider capture to per-message roles so the
tests can pin system-role and typed tool-row retention.

…tTarget

CI on the merged branch caught both: the capture assertion clones the
recorded calls, and upstream main added ChatTarget::message_seq since the
tests were written.

CI round two: every E0277 was the {calls:?} formatting in the part11
assertions. CapturedRole already derives Debug.

The workspace was seeded under "acme" but registration used a fresh uuid,
so the agent resolved an empty workspace and file_read had nothing to
read. Generate the unique id first and carry it through seeding, build,
and registration (tinysweeper 4107158658).

A closure returning &str borrowed from its argument hits the inferred-
lifetime wall on the CI toolchain; a plain fn elides per call.

The test for a single blank script now uses `scripted_agent_with_capture` instead of `scripted_agent`, and checks the captured provider calls via the capture object rather than the returned usages. This aligns the test with the new capture-based assertion pattern used elsewhere in the test suite.

Auto-committed-on: dragonfly
Co-authored-by: Medulla <[email protected]>

Updated the test assertion in built_in_tests_part11.rs to properly validate behavior when an unexpected input is provided, ensuring the test correctly reflects the intended error handling logic.

Auto-committed-on: dragonfly
Co-authored-by: Medulla <[email protected]>