A task agent's replayed turns now carry the native reasoning lane the
model produced (Anthropic thinking blocks + signatures, OpenAI Responses
reasoning items, Gemini thought_signature blocks, vLLM/llama.cpp parsed
reasoning text) instead of being rebuilt from content + tool_calls with
the reasoning dropped — restoring reasoning continuity across the
agent's own multi-turn tool loop on every provider lane.
The prerequisite is the id half: replace legalize_tool_call_ids with
restore_provider_tool_ids, a lowering pass that maps the session-minted
sub-tool ids back to the provider's own ids on the transient wire copy
(from the per-run mint map, never by string-splitting). The native
tool_use block is replayed verbatim — its id and signature untouched —
and the top-level mirror and tool_result agree with it on every request.
The minted id stays the sole internal key (registry, DOM, recall,
cancel ledger), #820 unchanged.
Chat-Completions lane: non-streaming create_completion now surfaces
reasoning/reasoning_content as CompletionResult.reasoning (the twin of
the streaming reasoning_delta extraction), and the agent seam runs the
Phase 5 vLLM reasoning-field replay against the agent's own provider
and alias. The native lane is finalized by a shared helper
(_finalize_provider_blocks) so the main loop and the sub-harness cannot
drift; replay honors the per-model replay_reasoning_to_model flag on
every lane, and llama.cpp stays capture-only, matching the main loop.
Sub-agent tool ids were namespaced {parent}::{provider_id} — unique
across concurrent agents but not across turns within one agent. A local
provider reissuing "call_0" every response minted the same id twice, so
the live card's DOM row lookup collapsed distinct calls onto one row
while FIFO recall kept them apart: two views of one trajectory disagreed
on identical input (the bug-3 id-consistency defect). When the provider
also reuses the PARENT call id, sequential runs repeated the collision
one level up.
Mint {parent}::r{run}s{step}::{provider_id} at the single rewrite point:
a session-monotonic run tag (lock-allocated; runs start concurrently on
the 4-wide task pool) plus a per-run step tag make each id unique within
the session, and every consumer — nesting registry, error flags, DOM
data-call-id, recall projection, cancel ledger — keys on that one id.
The FIFO pairing helper stays as honest pairing for un-minted input
(unparented runs, direct construction), with its rationale rewritten.
The agent wire seam (_run_agent's _api_call) also runs the same two
validity passes the main loop already ran — sanitize_tool_call_arguments
(a documented vLLM deepseek_v4 renders malformed args and 400s; agents
hit the same backends) and legalize_tool_call_ids (projects the long,
::-containing ids to plain tokens, call/result pairing preserved). The
id projection is DEFENSIVE hardening, not a fix for an observed break:
the ids replay fine on the lenient anthropic-compatible deployment (the
prior ::-containing format ran reliably), it just keeps an agent's
self-built history valid on a hypothetically stricter backend. Applied
at the agent seam only — main-loop assistant turns carry a provider-
native block lane whose id must stay byte-identical to the mirrored
tool_calls, so the projection cannot run there without desyncing them.
Follow-ups: parent-level card aliasing under a reused parent id; the same
id hygiene for the main conversation loop / native lane.
`tool_args_preview` feeds `stream.tool_args_malformed` (WARNING) and
`wire.tool_args_legalized` (DEBUG), and tool arguments are model/user
controlled — they can carry secrets (a token in a bash command, a password in a
connection string) or raw CR/LF that break log lines. Route the preview through
`output_guard.redact_credentials` over the full value first (before the 120-char
cap, so a secret straddling the cut isn't half-shown past the pattern's reach),
then collapse every control char to a space, mirroring `audit._scrub_string`.
Addresses the PR review comments.
A tool call whose `arguments` is not a JSON-object string (an unterminated
string from a non-`length` truncation, or an empty `""` from a no-arg call)
was committed verbatim and replayed on every subsequent send. Strict renderers
that re-parse arguments at render time (vLLM's `deepseek_v4`
`_postprocess_messages` runs `json.loads` on them) reject the whole request
with HTTP 400, wedging the conversation. The only prior guard dropped partial
tool calls on `finish_reason == "length"`; a `stop`/`tool_calls` finish reason
carrying invalid JSON slipped through, and its synthetic "retry" result kept it
from being an orphan, so the orphan-repair pass never touched it.
Add `sanitize_tool_call_arguments`, a wire-neutral legalize pass in lowering
(fold, legalize, repair), normalizing any non-JSON-object `arguments` to `{}`
on the transient wire copy only. The canonical trajectory keeps the raw model
output, so a wedged session self-recovers on its next send. A
`wire_valid_arguments` predicate is shared with a non-destructive
`stream.tool_args_malformed` warning at the stream accumulator, which surfaces
the model-quality problem at production time.
Convert `lowering.py` to structlog so the new pass emits structured events.
- Turn.effect_status also catches TypeError: a corrupt non-string meta value
(e.g. a dict that survived into the column) would otherwise crash a consumer
on access, since EffectStatus(non-str) raises TypeError, not ValueError.
Degrade to None, mirroring the meta decoders (Copilot review).
- test_lowering: the wire-repair synth now carries the _effect_status side
channel (stripped before the provider wire) — assert it.
- test_session_mcp_dispatch_error: the _capture stub swallows the new status
kwarg via **_ so it stays signature-compatible with _report_tool_result.
The canonical-Turn migration left lowering's fold/drop/repair passes Turn-typed even
though they convert to dicts internally and feed dicts to the translators, so
_prepare_wire_messages round-tripped the whole history Turn->dict->Turn ~7-8x per send
(even on the no-op early-return paths). Make fold_system_turns / drop_empty_user_turns
/ repair_wire_messages dict-native (list[dict]->list[dict]); _prepare_wire_messages now
threads the dict projection _full_messages already produced straight through, with no
Turn round-trip. self.messages stays the canonical Turn trajectory. export.py is
simplified (it converted to dicts immediately after repair anyway). Equivalence-
preserving — test_wire_payload_golden stays byte-identical.
repair_wire_messages / fold_system_turns / drop_empty_user_turns take and
return list[Turn] — the neutral lowering layer (A representation + B validity)
now speaks the canonical type. Their intricate content-merge / orphan-detect
internals run over the dict projection (reading Turn content blocks would only
duplicate turn_to_dict's content logic), so each bridges
dicts_from_turns ↔ turns_from_dicts at its boundary; byte-identical.
ChatSession._prepare_wire_messages lifts the wire dicts into Turns, runs the
lowering passes, and lowers the result back to the dict projection the provider
translators (the C layer) consume — the dict bridge now lives in the wire layer,
not in _full_messages. Export runs the same repair, reordered before the
non-canonical reasoning-content attach (a key the Turn model does not carry).
The provider translators keep their dict input by design: they are the format
layer that emits provider bytes, the vLLM reasoning-attach is a non-canonical
wire concern that sits between lowering and the provider on dicts, and feeding
the converters the lowered projection is equivalent to — and simpler than —
threading Turn content through them. Wire harness byte-identical; full
non-live suite green (7130).
Synthesizing a cancellation result for an assistant tool_call with no
matching tool result was triplicated across the translators: Anthropic's
verbatim-replay (pc_tool_ids) and rebuild branches, and sanitize_messages
for the OpenAI-compatible lanes (Chat, Responses, Google). The Anthropic
pc_tool_ids branch was also the sole repairer of a native tool_use orphan.
Lift it to one neutral policy — lowering.repair_wire_messages — run once in
ChatSession._prepare_wire_messages before the translator. It reads tool_calls
only, which is sound because the native/tool_calls mirror is enforced at save
(normalize_native_for_save): a verbatim-replay orphan is caught via its
mirrored top-level call. The translators become pure format translation and
carry no orphan synthesis.
The neutral cancellation turn carries is_error=True; Anthropic renders it on
the tool_result block, the OpenAI-compatible tool message has no such field
so sanitize_messages drops it (the C-layer translation of the flag).
sanitize_messages keeps one orphan synth of its own: a back-filled empty-id
tool_call (local servers that omit ids) is id-less when the upstream repair
runs and so invisible to it, so that lane owns its cancellation — preserving
the pre-refactor behavior for local servers.
reconstruct's load-time strip and the runtime-cancel persist-synth are
unchanged. Proven byte-identical against the per-provider wire-payload golden
harness (including a new native_orphan fixture); the harness applies the same
send-side repair the session does.
Patrick Buckley