fix(transport): strip empty/null tool_calls on assistant messages

[TurgutKural]

Summary

Strict OpenAI-compatible providers (notably onerouter / Qwen and DeepSeek v4) reject an assistant message that carries tool_calls: [] (empty array) — or tool_calls: null — with a non-retryable HTTP 400:

InternalError.Algo.InvalidParameter: Empty tool_calls is not supported in message.

The pre-API sanitizer in agent.agent_runtime_helpers.sanitize_api_messages already strips these, but only on the conversation_loop path. Auxiliary / custom-provider routes can reach the wire without passing through that sanitizer, so a stale empty tool_calls array aborts the entire session with a 400 that the retry loop cannot recover from.

This PR normalizes the same invariant at the transport layer (ChatCompletionsTransport.convert_messages) so every route agrees.

Root cause

convert_messages already walks tool_calls to strip Codex scaffolding fields (call_id, response_item_id, extra_content), but it never handled the case where the array itself is empty/invalid. An empty tool_calls on an assistant message slipped through to the wire.

Fix

In agent/transports/chat_completions.py::convert_messages:

• Detection: flag needs_sanitize=True when an assistant message carries "tool_calls" that is an empty list or explicit null.
• Strip: on the per-call copy (never the stored history), drop the tool_calls key when empty/null. Real calls (non-empty arrays) are preserved and still get their Codex scaffolding stripped as before.
• Safety: only assistant messages are normalized — a stray empty tool_calls on other roles is left untouched, and copy-on-write semantics are preserved (input list is never mutated in place).

Evidence

• Observed in production: a custom provider session against https://llm.onerouter.pro/v1 (model qwen/qwen3.8-max-preview:free) failed with the exact error above, repeated under the same request id — i.e. the same invalid payload was resent every attempt, confirming the sanitizer was being bypassed on that route.
• agent.agent_runtime_helpers.sanitize_api_messages (commit a7932d86c, [Bug]: repair_message_sequence creates empty tool_calls array, triggering DeepSeek v4 HTTP 400 (follow-up to #56980) #58755) already drops tool_calls: [] on the conversation_loop path; this change extends the same guarantee to the transport layer.

Tests

New file tests/agent/transports/test_chat_completions_empty_tool_calls.py (10 cases):

• empty list dropped, null dropped
• real calls preserved
• mixed batch: only the empty-assistant entry is stripped
• user-role empty tool_calls left untouched
• empty array with Codex fields still stripped (call preserved, scaffolding removed)
• clean list is identity (no copy)
• empty array triggers copy-on-write (input not mutated)
• cross-provider parity (qwen, deepseek-v4, gpt-4o)

Full run: tests/agent/transports/ + tests/run_agent/test_message_sequence_repair.py + sanitizer suites → 502 passed, no regressions.

Backward compatibility

Pure additive normalization at the transport layer. No schema changes, no behavior change for well-formed payloads. Strict providers that previously 400 now receive a valid message; permissive providers already ignored the empty array, so their behavior is unchanged.

[alt-glitch]

This was generated by AI during triage.

Related to the open empty-tool_calls family (nearest #68942; origin #58755). This patches convert_messages for null and assistant-role handling; #68942 also covers the run_agent sanitizer. The competing open fixes need a maintainer consolidation decision rather than a duplicate call.

[TurgutKural]

Rebased onto current upstream/main (2d40494). The transport-layer fix is orthogonal to #68942 — this PR covers empty-list/null tool_calls at the transport layer (all routes), while #68942 covers the run_agent sanitizer path. Ready for maintainer review.

[teknium1]

Thanks for the focused copy-on-write transport hardening. Current main still leaves an assistant tool_calls: [] or null unchanged in agent/transports/chat_completions.py:197-211, so the transport-level guard is a valid addition.

Problems

• The claimed auxiliary/custom-provider coverage is not delivered by this transport-only change. agent/auxiliary_client.py:7225-7228 builds raw request kwargs from messages, and its async call path directly invokes the SDK at agent/auxiliary_client.py:8801; neither uses ChatCompletionsTransport.convert_messages.
• tests/agent/transports/test_chat_completions_empty_tool_calls.py:77 describes an empty-array case, but the fixture is a non-empty list at line 83 and correctly asserts preservation.

Suggested changes

• Either scope this to transport defense-in-depth, or add the normalization at the actual auxiliary request boundary and test that wire path.
• Rename or correct the contradictory Codex-fields test.

The normal conversation and iteration-summary paths already use the shared sanitizer (agent/conversation_loop.py:1640, agent/chat_completion_helpers.py:2167), so the added coverage should identify a real additional boundary. This is an automated hermes-sweeper review.

[TurgutKural]

Rebased onto current upstream/main (7965462). The CI failure was a pre-existing vercel sandbox test issue (ImportError: lazy installs disabled) now fixed on main — not related to this PR's changes.

[TurgutKural]

Good catch — renamed test_empty_array_with_codex_fields_still_dropped → test_nonempty_array_codex_fields_stripped and updated the comment to accurately describe the fixture (one tool call with codex scaffolding markers stripped, call preserved). Pushed in b94ea97.

[GottZ]

This was generated by AI during triage.

Summary

Seventy PRs reference a broad issue complex spanning Feishu recipient routing, provider/model catalogs, message-sequence repair, strict-provider payload normalization, HOME handling, delegation limits, reasoning configuration, and unrelated vision work. The target #72591 adds transport-level empty/null tool_calls normalization; the central unresolved dedup defect is instead addressed most directly by #67933, while several other sub-clusters are already implemented on main.

Related pull requests

• fix(feishu): infer receive_id_type from ID prefix in _send_raw_message #7694 best fix — (+51/-1) — n/a: Best existing Feishu fix for fix(feishu): _send_raw_message hardcodes receive_id_type="chat_id", breaks delivery to open_id (ou_) targets #7685/Feishu: send to user open_id fails with [230001] invalid receive_id #9980, but it targets the retired adapter; despite the keep_open review, only its on_ → union_id branch and request-level test remain salvageable in the active plugin.
• fix(feishu): infer receive_id_type from chat_id prefix #7769 partial — (+8/-1) — n/a: Duplicates fix(feishu): infer receive_id_type from ID prefix in _send_raw_message #7694's prefix routing without tests; despite its keep_open review, only the on_ branch remains salvageable after the adapter migration.
• Fix custom provider API mode detection for OpenAI-compatible endpoints #9155 best fix — (+198/-15) — n/a: Best existing diff for [Bug]: 'GptVParam.Messages' Error:Field validation for 'Messages' failed on the 'required' tag #9064, whose stale api_mode failure is already fixed by 0e4c879; the keep_open review preserves only hostname hardening in hermes_cli/providers.py.
• fix: auto-detect Feishu receive_id_type from ID prefix #10040 partial — (+71/-5) — n/a: Duplicates the Feishu routing fix in fix(feishu): infer receive_id_type from ID prefix in _send_raw_message #7694/fix(feishu): infer receive_id_type from chat_id prefix #7769 against the retired adapter and bundles an already-implemented whitespace sanitizer; its keep_open review supports salvaging only on_ routing into the plugin.
• fix: detect receive_id_type dynamically in Feishu adapter #10123 [closed] partial — (+42/-1) — n/a: Duplicate Feishu prefix helper, closed as stale; it remains relevant only as an earlier implementation of the fix(feishu): infer receive_id_type from ID prefix in _send_raw_message #7694 behavior.
• fix: Feishu send to user open_id fails with [230001] invalid receive_id #10502 [closed] partial — (+21/-1) — n/a: Duplicate Feishu prefix helper, closed as stale and superseded by the active-adapter implementation plus the remaining union-ID gap.
• feat(kimi): add k2.6-code-preview support across kimi-coding, kimi-coding-cn, and moonshot #10675 [closed] best fix — (+59/-3) — n/a: Best existing focused feat: add k2.6-code-preview support for kimi-coding / kimi-coding-cn / moonshot #11168 diff for K2.6 catalog, validation, provider-ID, and context metadata support, but it was closed pending readiness.
• feat(kimi): add k2.6-code-preview support across kimi-coding, kimi-coding-cn, and moonshot #10751 [closed] best fix — (+361/-19) — n/a: Best existing broader feat: add k2.6-code-preview support for kimi-coding / kimi-coding-cn / moonshot #11168 diff, closed after mixing K2.6 support with OAuth, headers, aliases, and sampling policy beyond the issue's cause.
• fix: preserve None content on tool-call-only assistant messages (#11906) #11924 [closed] partial — (+275/-7) — n/a: Closed broad three-layer empty-content normalization for [Bug]: Assistant messages with empty content cause HTTP 400 on Anthropic-compatible proxies #11906, now superseded by shared pre-send and adapter-side handling.
• fix: preserve None content for assistant messages to avoid Anthropic HTTP 400 #11945 [closed] related — (+4/-1) — n/a: Closed duplicate of fix: preserve None content in assistant messages (#11906) #12470 that preserves None but then recreates empty content on tool-call turns.
• fix: preserve None content in assistant messages (#11906) #12470 [closed] partial — (+17/-4) — n/a: Same storage-layer [Bug]: Assistant messages with empty content cause HTTP 400 on Anthropic-compatible proxies #11906 fix as fix(agent): normalize empty assistant content to None for Anthropic compatibility #12737; despite its keep_open review, the documented maintainer design blocks rewriting byte-stable history.
• fix(agent): normalize empty assistant content to None for Anthropic compatibility #12737 [closed] partial — (+83/-1) — n/a: Closed storage-layer duplicate for [Bug]: Assistant messages with empty content cause HTTP 400 on Anthropic-compatible proxies #11906; the keep_open review records that adapter/pre-send normalization, not persisted-history rewriting, is the accepted direction.
• feat: replace kimi-k2.5 with kimi-k2.6 on OpenRouter and Nous Portal #13148 [merged] related — (+281/-3) — n/a: Merged reference implementation for Kimi K2.6 aggregator catalogs, though much of its diff also contains unrelated redaction work.
• feat: add kimi-k2.6 to kimi native provider lists #13152 [merged] related — (+7/-3) — n/a: Merged reference implementation adding Kimi K2.6 to native provider and picker lists.
• feat(models): add kimi-k2.6 to opencode-zen/go, alibaba, HuggingFace #13156 [closed] related — (+10/-4) — n/a: Closed as superseded by merged feat: replace kimi-k2.5 with kimi-k2.6 on OpenRouter and Nous Portal #13148/feat: add kimi-k2.6 to kimi native provider lists #13152 and later provider catalog updates.
• feat: add Kimi K2.6 model support #13380 [closed] related — (+72/-2) — n/a: Closed as already implemented by feat: replace kimi-k2.5 with kimi-k2.6 on OpenRouter and Nous Portal #13148/feat: add kimi-k2.6 to kimi native provider lists #13152; its only unique benchmark file had no repository destination.
• fix(compressor): strip orphaned tool messages at compression boundary #20886 [closed] fixes — (+168/-7) — n/a: Closed because compression-boundary alignment is already implemented by d132e34 with regression coverage.
• fix(agent): merge consecutive assistant tool_calls before orphan-tool drop (#29148) #29168 [closed] fixes — (+189/-1) — n/a: Best existing Parallel tool calls stored as separate assistant messages break DeepSeek session replay #29148 contributor diff, closed after its consecutive-assistant repair was incorporated with attribution into merged fix(agent): merge consecutive assistant messages before API replay (#29148, #49147) #55603.
• fix(agent): preserve tool_name on sanitizer-injected stub tool results #29185 related — (+7/-6) — n/a: The keep_open review finds its tool_name persistence claim cannot hold on the API-copy path and would leak a strict-provider-invalid field.
• fix(agent): surface delegate_task truncation to the model and document the per-turn cap #30472 [closed] best fix — (+130/-0) — n/a: Best existing [Bug]: Truncated delegate_task is unkown to the model. #30405 fix, closed voluntarily; it documents the delegate cap and surfaces dropped-call feedback to the model.
• fix(chat): sanitize empty assistant content before requests #31175 [closed] best fix — (+10/-0) — n/a: Best existing early shared-boundary fix for [Bug]: Assistant messages with empty content cause HTTP 400 on Anthropic-compatible proxies #11906/Empty assistant content with tool_calls causes HTTP 400 from strict OpenAI-compat upstreams #31583/Empty-content assistant messages with tool_calls break DeepSeek API (HTTP 400) #63200, closed after equivalent empty-message repair landed via refactor: single owner for empty-content wire repair (salvages #71439) #73071.
• fix(agent): normalize empty assistant content to None when tool_calls present #31582 best fix — (+153/-0) — n/a: Best existing preserve-and-normalize fix for [Bug]: Assistant messages with empty content cause HTTP 400 on Anthropic-compatible proxies #11906/Empty assistant content with tool_calls causes HTTP 400 from strict OpenAI-compat upstreams #31583/Empty-content assistant messages with tool_calls break DeepSeek API (HTTP 400) #63200; its keep_open review confirms the shared sanitizer location and non-mutating scope.
• fix(agent): normalize empty tool-call content for strict chat APIs #31615 fixes — (+96/-3) — n/a: Overlaps fix(agent): normalize empty assistant content to None when tool_calls present #31582/fix(chat): sanitize empty assistant content before requests #31175 at transport and summary boundaries; its keep_open review requires porting to current copy-on-write paths.
• fix(agent): merge consecutive assistant tool_calls messages on session replay #34510 [closed] fixes — (+36/-1) — n/a: Closed duplicate of fix(agent): merge consecutive assistant tool_calls before orphan-tool drop (#29148) #29168 whose repair kernel was later consolidated into merged fix(agent): merge consecutive assistant messages before API replay (#29148, #49147) #55603.
• fix(agent): enforce tool-result adjacency after compression #36157 [closed] best fix — (+65/-1) — n/a: Best existing source-layer Bug: Context compression produces invalid tool_call/tool message ordering → HTTP 400 crash #20883 adjacency fix, closed after the issue was resolved on main by boundary alignment.
• fix(vision): normalize CgBI image replay payloads #38060 related — (+433/-6) — n/a: Independent CgBI image normalization work; its keep_open review blocker was addressed, but it is unrelated to fix(transport): strip empty/null tool_calls on assistant messages #72591's message-schema cause.
• fix(agent): preserve real HOME for session tools #38107 [closed] best fix — (+186/-56) — n/a: Best existing Agent session HOME points to wrong path — tools look in hermes profile instead of user's actual home #36144 diff, closed because 723c233 already implements the HOME policy on main.
• fix(kimi): surface K2.7 Code in native picker #46309 [merged] partial — (+52/-12) — n/a: Merged reference implementation surfacing K2.7 Code in the native Kimi picker despite lagging live catalogs.
• fix: merge split assistant messages to prevent DeepSeek 400 on role alternation #49162 [closed] fixes — (+58/-0) — n/a: Closed after its split content/tool_calls repair was incorporated with attribution into merged fix(agent): merge consecutive assistant messages before API replay (#29148, #49147) #55603.
• fix: normalize null content on tool/user messages for strict providers #51812 [closed] related — (+31/-0) — n/a: Separate null-content normalization and error-classifier change, not the empty-tool_calls dedup-output defect targeted here.
• fix(kimi): add kimi-k2.7-code to kimi-coding-cn and moonshot model lists #53532 best fix — (+4/-2) — n/a: Best existing kimi-k2.7-code missing from kimi-coding-cn provider model list #53520 diff, but its keep_open review requires endpoint availability evidence and removal of ineffective setup.py edits.
• fix(agent): merge consecutive assistant messages before API replay (#29148, #49147) #55603 [merged] best fix — (+301/-1) — n/a: Merged best fix for Parallel tool calls stored as separate assistant messages break DeepSeek session replay #29148/DeepSeek 400: assistant message with content+tool_calls split into 2 messages in API payload #49147, combining consecutive assistant turns while preserving distinct rounds and Codex interim state.
• fix(agent): strip orphan tool_calls in repair_message_sequence (#56980) #56990 [closed] partial — (+111/-0) — n/a: Closed because stripping orphan tool calls would destructively rewrite history while the outgoing sanitizer already inserts valid stubs.
• fix(agent): close orphaned assistant(tool_calls) in repair_message_sequence (#56980) #57013 [closed] partial — (+173/-9) — n/a: Closed/superseded orphan-close attempt whose design gap was restated with stronger provenance in fix(agent): make repair_message_sequence self-contained for orphaned assistant(tool_calls) #57041/fix(agent): make repair_message_sequence self-contained for orphaned assistant(tool_calls) #57055.
• fix(agent): make repair_message_sequence self-contained for orphaned assistant(tool_calls) #57041 [closed] best fix — (+198/-9) — n/a: Best existing repair_message_sequence orphan-close pass doesn't use canonical call_id||id extractor #57036/repair_message_sequence: orphaned assistant(tool_calls) not self-contained (only patched downstream by sanitize_api_messages) #57039 repair-self-containment diff, closed by its author before a rebuilt continuation.
• fix(agent): make repair_message_sequence self-contained for orphaned assistant(tool_calls) #57055 [closed] best fix — (+198/-9) — n/a: Best existing [Bug]: repair_message_sequence fails to prevent HTTP 400 "insufficient tool messages following tool_calls" on DeepSeek v4 after long tool-call sessions #56980/[Bug]: repair_message_sequence not self-contained for orphaned tool_calls — continuation of #56980 #57053 continuation, but the filed incident was not reproducible and the main request path already sanitizes orphans.
• fix(agent): match tool results on call_id||id in pre-request repair (#58168) #58179 [merged] best fix — (+125/-4) — n/a: Merged best fix for Bug: Context compaction produces invalid message sequences (orphaned tool messages), breaking sessions permanently #58168, matching tool results against both call_id and id before strict-provider replay.
• fix(agent): deduplicate tool_call_id across pre-API sanitizers (#58327) #58350 [merged] best fix — (+139/-0) — n/a: Merged best fix for [Bug]: Context compression breaks tool message chain, causing "role 'tool' must be a response to a preceding message with 'tool_calls'" on strict providers #58327, deduplicating assistant call IDs and tool results at both pre-API chokepoints.
• fix(agent): dedup duplicate tool_call_id after compaction (#58327) #58362 [closed] fixes — (+175/-1) — n/a: Closed duplicate of merged fix(agent): deduplicate tool_call_id across pre-API sanitizers (#58327) #58350, retaining only a stricter representation preference beyond the implemented 400 fix.
• fix(agent): drop empty tool_calls arrays in pre-API sanitizer (#58755) #58768 [closed] fixes — (+95/-0) — n/a: Closed after its empty/malformed tool_calls normalization and tests were incorporated with authorship into merged fix(agent): drop empty tool_calls arrays in pre-API sanitizer (#58755) #59110.
• fix: strip empty tool_calls before sending to strict APIs #58953 [closed] partial — (+8/-1) — n/a: Closed duplicate of fix(agent): drop empty tool_calls arrays in pre-API sanitizer (#58755) #58768/fix(agent): drop empty tool_calls arrays in pre-API sanitizer (#58755) #59110 using narrower persistence and strict-sanitizer sites.
• fix(agent): drop empty tool_calls arrays in pre-API sanitizer (#58755) #59110 [merged] best fix — (+95/-0) — n/a: Merged best fix for [Bug]: repair_message_sequence creates empty tool_calls array, triggering DeepSeek v4 HTTP 400 (follow-up to #56980) #58755, non-destructively dropping pre-existing empty or malformed tool_calls at the shared pre-API chokepoint.
• fix: normalize empty tool-call assistant content #59427 [closed] partial — (+55/-0) — n/a: Closed duplicate of fix(agent): normalize empty tool-call content for strict chat APIs #31615's transport-level empty-content normalization for strict Bedrock-backed gateways.
• fix(agent): drop empty text content on assistant tool-call turns in pre-API sanitizer #59714 [closed] fixes — (+92/-0) — n/a: Closed after its Bedrock premise was shown already handled by adapter conversion; the keep_open review required a concrete uncovered non-Bedrock path.
• fix(compression): dedup duplicate tool_call_ids to prevent HTTP 400 on strict providers #60270 [closed] partial — (+120/-0) — n/a: Closed contaminated compressor-level duplicate-ID branch, superseded by the shared pre-API fix in fix(agent): deduplicate tool_call_id across pre-API sanitizers (#58327) #58350.
• fix(compression): dedup duplicate tool_call_ids to prevent HTTP 400 on strict providers #60398 [closed] partial — (+120/-0) — n/a: Closed as already implemented by dba585c via fix(agent): deduplicate tool_call_id across pre-API sanitizers (#58327) #58350.
• fix(agent): normalize empty content to None for assistant messages with tool_calls #63452 [closed] partial — (+110/-0) — n/a: Closed duplicate of fix(agent): normalize empty assistant content to None for Anthropic compatibility #12737's builder-layer empty-content normalization, while fix(agent): normalize empty assistant content to None when tool_calls present #31582 provides the preferred pre-send mechanism.
• fix(agent): strip empty-content+tool_calls assistant messages before API call (#63200) #64087 [closed] related — (+666/-11) — n/a: Closed contaminated remove-message approach; its keep_open review found that deleting the assistant call record creates dangling tool results.
• fix(agent): strip reasoning replay fields for strict chat-completions providers #64296 related — (+269/-3) — n/a: Separate strict-provider reasoning-field replay fix whose keep_open review requirements were reportedly addressed; it does not fix empty tool_calls.
• fix(sanitize): drop tool_calls key when dedup removes all calls #64345 fixes — (+48/-1) — n/a: Implements the Bug: sanitize_api_messages dedup pass writes tool_calls: [] — HTTP 400 on DeepSeek/OpenAI #64335 dedup-output drop-key fix but is a larger/stale alternative to best-fix fix(agent): drop tool_calls key when dedup removes every call #67933.
• fix(agent): prevent empty tool_calls arrays to fix DeepSeek V4 HTTP 400 errors #64843 fixes — (+4/-1) — n/a: Same core drop-key repair; despite its keep_open review confirming the defect, it lacks the required all-duplicates regression test.
• fix(agent): prevent empty tool_calls arrays after dedup for DeepSeek V4 #65211 [closed] fixes — (+4/-1) — n/a: Closed exact duplicate of fix(agent): prevent empty tool_calls arrays to fix DeepSeek V4 HTTP 400 errors #64843/fix(sanitize): drop tool_calls key when dedup removes all calls #64345 for the post-dedup empty-array branch.
• feat(models): add moonshotai/kimi-k3 to Nous Portal + OpenRouter curated lists, retire kimi-k2.x #65913 [merged] related — (+9/-18) — n/a: Merged catalog update replacing aggregator Kimi K2.x entries with Kimi K3, unrelated to the strict-message sanitizer defect.
• fix(agent): uniquify duplicate tool-call ids to keep call/result pairing lossless #67777 [merged] partial — (+173/-0) — n/a: Merged upstream prevention that uniquifies reused tool-call IDs before dispatch, reducing but not eliminating the need for replay sanitization.
• fix(agent): drop tool_calls key when dedup removes every call #67933 best fix — (+38/-1) — n/a: Best existing Bug: sanitize_api_messages dedup pass writes tool_calls: [] — HTTP 400 on DeepSeek/OpenAI #64335 fix: the minimal diff drops tool_calls after all calls are deduplicated and adds the exact cross-message regression; its keep_open review confirms the premise.
• fix(agent,transport): drop empty tool_calls array for strict API compatibility #68942 [closed] partial — (+16/-0) — n/a: Closed because a7932d8 already strips pre-existing empty arrays through the shared sanitizer; it does not address arrays created later by dedup.
• fix(agent): fill empty user/assistant content in pre-API sanitizer to prevent strict-provider 400s #69177 [closed] related — (+129/-1) — n/a: Closed after refactor: single owner for empty-content wire repair (salvages #71439) #73071 implemented its broader empty user/assistant turn repair at the shared sanitizer.
• fix(agent): prevent empty tool_calls arrays from reaching strict providers (#69280) #69351 [closed] best fix — (+70/-13) — n/a: Best existing [Bug]: Switching to DeepSeek fails with "messages[N].tool_calls: empty array" after context compaction #69280 diff, closed because the wire guarantee already exists in a7932d8; its unrelated profile-skill changes also prevent reuse as-is.
• fix(agent): clamp GLM-5.2 ultra reasoning effort #69917 best fix — (+26/-1) — n/a: Best existing did a branch, failed #69855 GLM-5.2 clamp, but its keep_open review requires applying the same normalization to direct summary requests.
• fix(agent): clamp 'ultra' reasoning_effort to 'max' for all models (#70058) #70079 best fix — (+78/-30) — n/a: Best existing reasoning_effort: "ultra" rejected by GLM API (HTTP 400), triggers silent fallback to Claude Sonnet #70058 policy candidate, but its keep_open review blocks the syntactically invalid unrelated Desktop and placeholder files and requires summary-path coverage.
• fix(agent): drop tool_calls key when all tool_calls are deduplicated (#70126) #70176 [closed] fixes — (+7/-3) — n/a: Closed duplicate of fix(sanitize): drop tool_calls key when dedup removes all calls #64345/fix(agent): drop tool_calls key when dedup removes every call #67933 that also bundled unrelated transport and CLI changes.
• fix(agent): drop tool_calls key when all tool_calls are deduplicated (#70126) #70181 [closed] best fix — (+1/-1) — n/a: Best existing focused [Bug]: sanitize_api_messages dedup writes tool_calls: [], breaking Qwen-series tool-calling sessions on v0.18.2+ #70126 diff, closed after providing the same one-line drop-key correction as fix(agent): drop tool_calls key when dedup removes every call #67933.
• fix(transports+agent): GLM reasoning_effort ultra + drop empty tool_calls key (#70058, #70126) #70182 fixes — (+2/-2) — n/a: Combines reasoning_effort: "ultra" rejected by GLM API (HTTP 400), triggers silent fallback to Claude Sonnet #70058 and [Bug]: sanitize_api_messages dedup writes tool_calls: [], breaking Qwen-series tool-calling sessions on v0.18.2+ #70126 incorrectly; its keep_open review notes that dedup remains ineffective when no calls survive and GLM scope is overbroad.
• fix(sanitizer): drop tool_calls key when all calls deduped, not set empty array #70781 [closed] related — (+14/-1) — n/a: Closed duplicate of fix(sanitize): drop tool_calls key when dedup removes all calls #64345, with only a placeholder-content edge case beyond the canonical drop-key repair.
• fix(agent): strip empty-content+tool_calls assistant messages before API call #71787 [closed] related — (+321/-14) — n/a: Closed remove-message resubmission superseded by main's substitution-based empty-message repair, which preserves assistant/tool linkage.
• fix(transport): strip empty/null tool_calls on assistant messages #72591 related — (+175/-0) — n/a: Transport defense-in-depth correctly strips assistant tool_calls values of [] or null; despite the keep_open review, its claimed auxiliary coverage is unsupported because auxiliary requests bypass this transport.
• fix: strip empty tool_calls before API calls to avoid HTTP 400 #72619 [closed] partial — (+54/-0) — n/a: Closed after the author found stale installed code caused the reproduction; it also bundled unrelated Feishu parsing.
• fix(agent): strip empty-content+tool_calls assistant messages before API call #73135 [closed] related — (+321/-14) — n/a: Closed resubmission of the remove-message approach, superseded by main's repair_empty_non_final_messages substitution mechanism.
• Fix: drop empty tool_calls array + content-null leak in sanitize_api_messages #74906 related — (+196/-1) — n/a: Fresh comprehensive variant that drops dedup-created empty arrays and reruns empty-message repair; its keep_open review confirms both reproduced defects and exact tests.
• fix(sanitize): drop tool_calls key when dedup removes all calls (fixes empty-array 400 class) #77377 fixes — (+147/-111) — n/a: Fresh current-main resubmission of fix(sanitize): drop tool_calls key when dedup removes all calls #64345's drop-key fix, but it bundles a substantial unrelated delegate-tool documentation refactor that should be split.

Duplicates

Feishu: #7694/#7769/#10040/#10123/#10502; empty assistant content: #11924/#11945/#12470/#12737/#31175/#31582/#31615/#59427/#59714/#63452/#64087/#71787/#73135; consecutive assistants: #29168/#34510/#49162 superseded by #55603; duplicate IDs: #58350/#58362/#60270/#60398; pre-existing empty tool_calls: #58768/#58953/#59110/#68942/#72591/#72619; dedup-created empty tool_calls: #64345/#64843/#65211/#67933/#70176/#70181/#70182/#70781/#74906/#77377. Kimi catalogs overlap across #10675/#10751/#13148/#13152/#13156/#13380/#46309/#53532/#65913, while GLM reasoning overlaps #69917/#70079/#70182.

Suggested consolidation

Keep #72591 open with a salvage path limited to transport defense-in-depth: retain its copy-on-write []/null normalization and corrected non-empty-call test, but revise the auxiliary/custom-provider claim or add normalization and a wire-level test at the actual auxiliary request boundary, explicitly following the keep_open review on #72591. For the canonical post-dedup defect, author action on best-fix #67933: rebase its focused source hunk and regression onto current main; #64345, #64843, #65211, #70176, #70181, #70182, #70781, and the core hunk of #77377 can then close as duplicates, while #74906 should remain separate only if maintainers want its additional post-dedup empty-content repair. Split #77377's unrelated delegate-tool refactor before further review so the sanitizer fix and documentation work can be evaluated independently.

Cross-PR triage: Reviewed 70 pull requests and 27 issues in this complex. Diffs were read for 63 of 70 PRs (rest unavailable); Assessment working set: 486 kB of PR diffs, 253 kB of issue/PR text, 99 kB of discussion (158 comments), 111 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.

[TurgutKural]

Noted — this sweep's relevant line is the relationship to #67933. They are complementary, not competing:

• fix(agent): drop tool_calls key when dedup removes every call #67933 fixes the sanitizer/dedup path (agent_runtime_helpers.sanitize_api_messages): when dedup removes every call, drop the tool_calls key instead of leaving [].
• fix(transport): strip empty/null tool_calls on assistant messages #72591 fixes the transport layer (agent/transports/chat_completions.convert_messages): strips empty/null tool_calls arrays on assistant messages at the API boundary, covering all routes regardless of how the empty array was produced.

Different files, different layers, no overlap. Both remain open; a maintainer consolidation decision can land them independently (either order works — each fixes a distinct production path).

[TurgutKural]

Response to the scope point from GottZ consolidation sweep: the "auxiliary / custom provider routes" claim in the review comment was narrowed (a3148fd2). Actual scope: this layer is only the last boundary for requests serialized through this transport; auxiliary clients that build fully separate payload paths never pass through this layer and are out of scope. Behavior is unchanged — only the claim text was corrected.

[teknium1]

Merged via #86654 — your three commits were cherry-picked onto current main with your authorship preserved in git log. Thank you! The salvage combines your transport-boundary strip with the dedup-pass fix (#64345) and the repair-merge fix (#77944) so the whole empty-tool_calls class is closed at every layer.