fix(core): replay DeepSeek reasoning_content on all assistant turns#3747
Merged
Conversation
Extend the DeepSeek reasoning_content normalization (introduced in #3729) to assistant turns without tool_calls. The DeepSeek API rejects follow-up requests in thinking mode whenever any prior assistant turn omits reasoning_content, not just turns that carried tool_calls.
Contributor
Code Coverage Summary
CLI Package - Full Text ReportCore Package - Full Text ReportFor detailed HTML reports, please see the 'coverage-reports-22.x-ubuntu-latest' artifact from the main CI run. |
4 tasks
wenshao
added a commit
that referenced
this pull request
May 2, 2026
…rovider (#3788) * fix(core): inject thinking blocks for DeepSeek anthropic-compatible provider DeepSeek's anthropic-compatible endpoint (https://api.deepseek.com/anthropic) rejects follow-up requests with HTTP 400 ("The content[].thinking in the thinking mode must be passed back to the API.") whenever a prior assistant turn carrying tool_use omits a thinking block. The model can legitimately return a tool round without thinking text, so qwen-code stored no thought parts and rebuilt the next request with no thinking block, tripping the API's check. Mirroring the existing OpenAI-side fix (#3729, #3747), the converter now detects DeepSeek by base URL or model name and prepends an empty { type: 'thinking', thinking: '', signature: '' } block to assistant turns missing one. Other anthropic-protocol providers are unaffected. Verified against the live api.deepseek.com/anthropic endpoint: - assistant with tool_use, no thinking → 400 (reproduces #3786) - assistant with tool_use, empty thinking injected → 200 OK Refs #3786 * fix(core): gate DeepSeek thinking-block injection on thinking mode Address PR review feedback: 1. (Critical) Gate empty-thinking injection on the same per-request condition that emits the top-level `thinking` parameter. The previous implementation injected unconditionally on DeepSeek providers, but `buildThinkingConfig()` may omit `thinking` when reasoning=false or `thinkingConfig.includeThoughts=false` — which is exactly what suggestionGenerator / ArenaManager / forkedAgent do. Shipping thinking blocks without enabling thinking mode is a protocol violation that DeepSeek may reject. Move the option from converter constructor to a per-request `convertGeminiRequestToAnthropic` parameter so the generator can compute the gate correctly. 2. (CodeQL) Replace `baseUrl.includes('api.deepseek.com')` with `new URL(baseUrl).hostname` exact-match. The substring check would accept spoofed hosts like `api.deepseek.com.evil.com`. 3. Document the empty-signature workaround inline. 4. Rename the misleading "redacted_thinking" test case. * fix(core): narrow DeepSeek thinking injection to tool_use turns + subdomain test Address PR review round 2: 1. Narrow injection scope to assistant turns containing tool_use. Live verification against api.deepseek.com/anthropic showed plain-text assistant turns without thinking are accepted unchanged — only tool_use turns trigger the HTTP 400. Injecting on every assistant turn unnecessarily bloats replay history with synthetic blocks the API does not require. Existing thinking blocks on any turn are still preserved untouched. 2. Add test coverage for the subdomain hostname branch (us.api.deepseek.com → matches), addressing the gap noted in review. 3. Update existing negative-case tests (non-deepseek / spoofed / reasoning=false / includeThoughts=false) to use tool_use scenarios so they actually exercise the gating logic instead of trivially passing under the narrowed scope. * docs(core): align DeepSeek thinking-injection comments with narrowed scope Address PR review round 3 (copilot-pull-request-reviewer × 3): comments in three locations still described the constraint as applying to "any prior assistant turn", which was true before commit 8721b41 but no longer matches the implementation. Update the doc comment on isDeepSeekAnthropicProvider and the two test-suite header comments to state the actual narrower contract: the API rejects only tool-use turns that omit thinking blocks; plain-text assistant turns are accepted unchanged. Comment-only change; 58 tests still pass. * fix(core): per-request DeepSeek detection + strip thinking when off Address PR review round 4 (copilot-pull-request-reviewer × 2): 1. Stale provider-detection cache (HCia). The constructor cached isDeepSeekProvider once, but Config.setModel() mutates contentGeneratorConfig.model in place. After a runtime /model switch from a non-DeepSeek model to a DeepSeek one on the same auth config, buildRequest() would keep using the stale flag. Move the detection into buildRequest so each call sees the current model. The detector is cheap (URL parse + string compare). 2. Real thought parts leak through when thinking is disabled (HCib). The previous gate only blocked synthetic injection — but the converter still replayed any existing `thought: true` parts in request.contents as thinking blocks. Code paths that disable thinking against a session whose history was built with thinking on (suggestionGenerator / ArenaManager / forkedAgent) would still emit thinking blocks alongside an absent top-level `thinking` config — the same protocol mismatch the gate was meant to avoid. Add a `stripAssistantThinking` converter option, set in buildRequest to `isDeepSeek && !thinking`. The converter strips thinking and redacted_thinking blocks from assistant messages before message construction completes. Mirror behavior is already proven safe by live verification (DeepSeek currently tolerates either shape, but stripping makes the request body internally consistent and robust to future validation tightening). 3 new tests: - converter strips thinking from assistant turns when option set - generator strips real thought parts when reasoning=false - generator reflects runtime model changes (no stale cache) 61 tests pass; lint + typecheck clean. * fix(core): preserve thinking-only assistant turns instead of emitting empty content Address PR review round 5 (copilot-pull-request-reviewer × 2 — code + test): stripThinkingFromAssistantMessages previously replaced message.content with the filtered array unconditionally. For an assistant turn whose only blocks are thinking/redacted_thinking (e.g. a round cut off by max_tokens before any text or tool_use was emitted), this left `content: []` — which Anthropic API rejects. Dropping the message entirely was considered but would break the required user/assistant alternation. Instead, fall back to leaving the original blocks in place when stripping would empty the message. DeepSeek empirically tolerates the residual `thinking-block + no-thinking-config` shape (verified against api.deepseek.com/anthropic in the V2/X scenarios), so leaving the message untouched is the safer choice than emitting invalid structure. Add regression test for the thinking-only turn shape. 62 tests pass; lint + typecheck clean. * fix(core): validate thinking-block signature, rename option, gate output_config Address PR review round 6 — five substantive items: 1. (Critical) Drop non-compliant thinking blocks lacking a `signature` field and replace them with a synthetic one. A `redacted_thinking` block round-tripped through Gemini Part format becomes `{ text: '', thought: true }` (no thoughtSignature) and converts back to `{ type: 'thinking', thinking: '' }` without `signature` — not spec-compliant. The previous `hasThinking` check accepted these as already-satisfying, leaving non-compliant blocks in the wire message. Tighten the check so they're filtered out and the synthetic injection runs. Live verification: DeepSeek currently tolerates both shapes (lenient), but normalizing is defensively correct against future tightening. 2. Rename converter option `ensureAssistantThinking` → `ensureThinkingOnToolUseTurns`. The new name reflects the actual contract (tool-use turns only, not every assistant turn). 3. Honor `thinkingConfig.includeThoughts: false` in `buildOutputConfig`. Previously a per-request opt-out dropped the top-level `thinking` parameter but still emitted `output_config.effort`, leaking a reasoning-shaped field into side queries that don't want it. 4. Add regression test for mixed text + tool_use assistant turns (common shape: model says something, then calls a tool). 5. Add explicit test for the signature-validation path: an existing compliant thinking block (with signature) is preserved untouched. 64 tests pass; lint + typecheck clean. * fix(core): clean up non-compliant thinking blocks on plain-text turns + assert output_config gating Address PR review round 7 (copilot-pull-request-reviewer × 2): 1. (Hp-x) Round-tripped redacted_thinking blocks were left malformed on assistant turns lacking tool_use. The previous structure only ran the cleanup pass when a tool_use block was present (early return on `!hasToolUse`), so plain-text turns kept the non-compliant `{ type: 'thinking', thinking: '' }` shape. Restructure into two sequential steps: a. Drop non-compliant thinking blocks (no `signature`) on every assistant turn — same fallback that avoids `content: []` if the message is thinking-only. b. Inject the synthetic empty thinking block on tool_use turns that still lack a compliant thinking block after step (a). 2. (Hp-7) The includeThoughts=false test asserted that the top-level `thinking` field is suppressed but didn't cover `output_config`, leaving regressions in the new `buildOutputConfig` gate uncaught. Tighten the assertion to also verify `output_config` is absent. 3. New converter test: cleanup runs on plain-text assistant turns too. 65 tests pass; lint + typecheck clean. * test(core): add explicit redacted_thinking injection-path coverage Address PR review round 8 (#30 — copilot reviewer). The converter treats `redacted_thinking` as already satisfying the thinking-block requirement (no synthetic injected), distinguished from a signature-less `thinking` block which is non-compliant and gets dropped/replaced. Existing tests covered the latter path; this adds explicit coverage of the former. processContent doesn't synthesize redacted_thinking from Gemini parts, so the test reaches into the private helper directly. (#31 — subdomain hostname coverage — already exists at line 602.) 66 tests pass; lint + typecheck clean. * fix(core): per-request anthropic-beta + normalize thinking-only turns Address PR review round 9 (copilot-pull-request-reviewer × 2): 1. (Hydz) thinking-only assistant turns (e.g. max_tokens cutoff or round-tripped redacted_thinking) hit the cleanup-empties fallback and kept the original non-compliant `{ type: 'thinking', thinking: '' }` block. The fallback now replaces the message with a synthetic empty thinking block (`signature: ''` included), which keeps the message non-empty AND spec-compliant. 2. (Hyd4) `anthropic-beta` was set once at construction from the global `reasoning` config, so requests with per-request `thinkingConfig.includeThoughts=false` still advertised interleaved-thinking / effort even though the body had dropped the matching fields. Move beta computation to a new `buildPerRequestHeaders` that derives the header from the actual `thinking` / `output_config` fields present in the request body, and pass it via `messages.create(..., { headers })`. The wire shape is now internally consistent. Test updates: - Drop the three constructor-time beta assertions; they no longer apply. - Add four per-request header tests covering: both betas present, only interleaved-thinking, reasoning=false (no betas), and per-request includeThoughts=false (no betas). 67 tests pass; lint + typecheck clean. * fix(core): preserve thinking text by normalizing in place + merge user beta flags Address PR review round 10 (copilot-pull-request-reviewer × 2): 1. (H0oF) The previous cleanup filtered out every thinking block missing a `signature` field. But that shape is the normal output from OpenAI/Gemini/agent-runtime generators, which only set `thought: true` without a signature. Users switching providers mid-session would silently lose preserved thinking text on the first DeepSeek request. Change Step 1 to NORMALIZE in place: when a thinking block has no signature, set `signature: ''` rather than dropping the block. The original `thinking` text is preserved; DeepSeek empirically accepts empty signatures so the wire shape stays valid. 2. (H0oL) `buildPerRequestHeaders()` overwrote `customHeaders['anthropic-beta']` whenever the per-request override fired, regressing the customHeaders escape hatch for unrelated Anthropic beta features. Merge the user's flags into the computed list (deduped) so users can stack their own betas alongside interleaved-thinking / effort. Test changes: - Renamed and rewrote "drops non-compliant... plain-text" test to assert in-place normalization that preserves thinking text. - Updated "replaces a non-compliant thinking block" comment + name to describe the normalization (the assertion was already correct because the test happened to use empty thinking text). - The empty-content fallback in Step 1 is no longer reachable under the new logic, so the dedicated thinking-only-turn test now exercises only the strip path (where it remains relevant). - Added 3 customHeaders[anthropic-beta] tests: merge with computed, passthrough when no thinking/effort, dedupe. 70 tests pass; lint + typecheck clean. * docs(core): align thinking-injection comments with normalize semantics + add stream test Address PR review round 11 (copilot-pull-request-reviewer × 3): 1. (H3Iw) Update the `ensureThinkingOnToolUseTurns` option docstring to describe in-place normalization (preserving thinking text by filling in `signature: ''`) instead of the old drop-and-replace semantics. 2. (H3I9) Same update on the `applyEmptyThinkingToToolUseTurns` helper JSDoc — clarify that signature-less thinking blocks are normalized in place (preserving original text), not dropped. Mention the common case of cross-provider history where non-Anthropic generators only set `thought: true`. 3. (H3I3) Add a streaming test asserting that `generateContentStream()` also attaches the per-request `anthropic-beta` header. The previous coverage only exercised `generateContent()`, leaving the streaming path's separate code path (line 144 in anthropicContentGenerator.ts) unverified. 71 tests pass; lint + typecheck clean. * refactor(core): split DeepSeek thinking option in two + add header coexistence test Address PR review round 12 (copilot-pull-request-reviewer × 2): 1. (H6ws) The single `ensureThinkingOnToolUseTurns` option was misleadingly narrow: the implementation also rewrote non-tool-use turns by normalizing malformed thinking blocks. Future callers could enable it expecting only the tool-use behavior. Split into two precisely-named options: - normalizeAssistantThinkingSignature: fill missing `signature` on every assistant `thinking` block (cross-provider history compat). - injectThinkingOnToolUseTurns: prepend synthetic empty thinking on tool_use turns missing one (issue #3786 trigger). The generator wires both together for DeepSeek when thinking mode is on; either can be used independently if a future caller needs only one pass. 2. (H6w4) Add a test asserting that the per-request `headers` path coexists correctly with `customHeaders`: User-Agent and unrelated customHeaders entries stay in `defaultHeaders` while only the computed `anthropic-beta` rides on the per-request path. Defends against a future regression where header config might be routed through a code path that wipes the constructor defaults. 72 tests pass; lint + typecheck clean. * fix(core): case-insensitive customHeaders[anthropic-beta] merge Address yiliang114 review feedback (#3788). HTTP header names are case-insensitive by spec, and the Anthropic SDK lower-cases them during merge. Previously buildPerRequestHeaders only read the lower-case `anthropic-beta` key from customHeaders, so a user-configured `Anthropic-Beta` or `ANTHROPIC-BETA` would be silently overwritten by the per-request computed value. Replace the direct dict lookup with collectCustomBetaFlags() which walks all customHeaders entries and matches the key case-insensitively. Multiple matching entries (unlikely but possible) are concatenated; the existing dedupe pass handles any duplicates. Add a regression test for both `Anthropic-Beta` and `ANTHROPIC-BETA` key shapes. 73 tests pass; lint + typecheck clean. * docs(core): align thinking-injection docs with normalize-in-place semantics + redacted_thinking strip test Address PR review round 14 (copilot-pull-request-reviewer × 4): 1. (IAl4) PR description still described "dropped here so synthetic injection takes over" but the implementation now normalizes signature-less thinking blocks in place (preserving text). PR description rewritten to describe the two-pass model: normalize-in-place + injection-when-truly-missing. 2. (IAl7) `injectThinkingOnToolUseTurns` option docstring claimed signature-less blocks would be "seen as missing" so the synthetic replaces them. Updated to describe the actual flow: the normalization pass runs first, blocks become compliant in place, the injector then sees them as already-satisfying and prepends nothing. Helper JSDoc on `injectEmptyThinkingOnToolUseTurns` fixed the same way. 3. (IAl8) Strip-path coverage missed `redacted_thinking` blocks. Added regression test that verifies both thinking and redacted_thinking blocks are removed when `stripAssistantThinking` is set. 4. (IAl-) Renamed the converter test suite from "thinking-mode injection + normalization (DeepSeek thinking on)" to "DeepSeek thinking-mode normalization, injection, and stripping" so the title accurately covers all behavior the block exercises (including `stripAssistantThinking` cases later in the same describe). 74 tests pass; lint + typecheck clean. * fix(core): exclude anthropic-beta variants from defaultHeaders to avoid wire duplication Address PR review round 15 (copilot-pull-request-reviewer #1). `buildHeaders()` previously spread the entire `customHeaders` map into the SDK's `defaultHeaders`. After moving anthropic-beta computation to the per-request path, a user-configured mixed-case `Anthropic-Beta` key would survive in defaultHeaders verbatim, while the per-request override added a lowercase `anthropic-beta`. The wire then carried two physical headers for the same logical name — SDK behavior on duplicate headers with different casings is undefined. `buildPerRequestHeaders()` already merges those user flags case-insensitively (commit 0d8b5de), so dropping the entry from defaultHeaders is the right boundary: the per-request path owns the header end-to-end. Other customHeaders entries continue to pass through. Add a regression test asserting no `Anthropic-Beta` (any casing) lands in defaultHeaders while unrelated customHeaders are kept. 75 tests pass; lint + typecheck clean.
DragonnZhang
pushed a commit
that referenced
this pull request
May 8, 2026
…rovider (#3788) * fix(core): inject thinking blocks for DeepSeek anthropic-compatible provider DeepSeek's anthropic-compatible endpoint (https://api.deepseek.com/anthropic) rejects follow-up requests with HTTP 400 ("The content[].thinking in the thinking mode must be passed back to the API.") whenever a prior assistant turn carrying tool_use omits a thinking block. The model can legitimately return a tool round without thinking text, so qwen-code stored no thought parts and rebuilt the next request with no thinking block, tripping the API's check. Mirroring the existing OpenAI-side fix (#3729, #3747), the converter now detects DeepSeek by base URL or model name and prepends an empty { type: 'thinking', thinking: '', signature: '' } block to assistant turns missing one. Other anthropic-protocol providers are unaffected. Verified against the live api.deepseek.com/anthropic endpoint: - assistant with tool_use, no thinking → 400 (reproduces #3786) - assistant with tool_use, empty thinking injected → 200 OK Refs #3786 * fix(core): gate DeepSeek thinking-block injection on thinking mode Address PR review feedback: 1. (Critical) Gate empty-thinking injection on the same per-request condition that emits the top-level `thinking` parameter. The previous implementation injected unconditionally on DeepSeek providers, but `buildThinkingConfig()` may omit `thinking` when reasoning=false or `thinkingConfig.includeThoughts=false` — which is exactly what suggestionGenerator / ArenaManager / forkedAgent do. Shipping thinking blocks without enabling thinking mode is a protocol violation that DeepSeek may reject. Move the option from converter constructor to a per-request `convertGeminiRequestToAnthropic` parameter so the generator can compute the gate correctly. 2. (CodeQL) Replace `baseUrl.includes('api.deepseek.com')` with `new URL(baseUrl).hostname` exact-match. The substring check would accept spoofed hosts like `api.deepseek.com.evil.com`. 3. Document the empty-signature workaround inline. 4. Rename the misleading "redacted_thinking" test case. * fix(core): narrow DeepSeek thinking injection to tool_use turns + subdomain test Address PR review round 2: 1. Narrow injection scope to assistant turns containing tool_use. Live verification against api.deepseek.com/anthropic showed plain-text assistant turns without thinking are accepted unchanged — only tool_use turns trigger the HTTP 400. Injecting on every assistant turn unnecessarily bloats replay history with synthetic blocks the API does not require. Existing thinking blocks on any turn are still preserved untouched. 2. Add test coverage for the subdomain hostname branch (us.api.deepseek.com → matches), addressing the gap noted in review. 3. Update existing negative-case tests (non-deepseek / spoofed / reasoning=false / includeThoughts=false) to use tool_use scenarios so they actually exercise the gating logic instead of trivially passing under the narrowed scope. * docs(core): align DeepSeek thinking-injection comments with narrowed scope Address PR review round 3 (copilot-pull-request-reviewer × 3): comments in three locations still described the constraint as applying to "any prior assistant turn", which was true before commit 8721b41 but no longer matches the implementation. Update the doc comment on isDeepSeekAnthropicProvider and the two test-suite header comments to state the actual narrower contract: the API rejects only tool-use turns that omit thinking blocks; plain-text assistant turns are accepted unchanged. Comment-only change; 58 tests still pass. * fix(core): per-request DeepSeek detection + strip thinking when off Address PR review round 4 (copilot-pull-request-reviewer × 2): 1. Stale provider-detection cache (HCia). The constructor cached isDeepSeekProvider once, but Config.setModel() mutates contentGeneratorConfig.model in place. After a runtime /model switch from a non-DeepSeek model to a DeepSeek one on the same auth config, buildRequest() would keep using the stale flag. Move the detection into buildRequest so each call sees the current model. The detector is cheap (URL parse + string compare). 2. Real thought parts leak through when thinking is disabled (HCib). The previous gate only blocked synthetic injection — but the converter still replayed any existing `thought: true` parts in request.contents as thinking blocks. Code paths that disable thinking against a session whose history was built with thinking on (suggestionGenerator / ArenaManager / forkedAgent) would still emit thinking blocks alongside an absent top-level `thinking` config — the same protocol mismatch the gate was meant to avoid. Add a `stripAssistantThinking` converter option, set in buildRequest to `isDeepSeek && !thinking`. The converter strips thinking and redacted_thinking blocks from assistant messages before message construction completes. Mirror behavior is already proven safe by live verification (DeepSeek currently tolerates either shape, but stripping makes the request body internally consistent and robust to future validation tightening). 3 new tests: - converter strips thinking from assistant turns when option set - generator strips real thought parts when reasoning=false - generator reflects runtime model changes (no stale cache) 61 tests pass; lint + typecheck clean. * fix(core): preserve thinking-only assistant turns instead of emitting empty content Address PR review round 5 (copilot-pull-request-reviewer × 2 — code + test): stripThinkingFromAssistantMessages previously replaced message.content with the filtered array unconditionally. For an assistant turn whose only blocks are thinking/redacted_thinking (e.g. a round cut off by max_tokens before any text or tool_use was emitted), this left `content: []` — which Anthropic API rejects. Dropping the message entirely was considered but would break the required user/assistant alternation. Instead, fall back to leaving the original blocks in place when stripping would empty the message. DeepSeek empirically tolerates the residual `thinking-block + no-thinking-config` shape (verified against api.deepseek.com/anthropic in the V2/X scenarios), so leaving the message untouched is the safer choice than emitting invalid structure. Add regression test for the thinking-only turn shape. 62 tests pass; lint + typecheck clean. * fix(core): validate thinking-block signature, rename option, gate output_config Address PR review round 6 — five substantive items: 1. (Critical) Drop non-compliant thinking blocks lacking a `signature` field and replace them with a synthetic one. A `redacted_thinking` block round-tripped through Gemini Part format becomes `{ text: '', thought: true }` (no thoughtSignature) and converts back to `{ type: 'thinking', thinking: '' }` without `signature` — not spec-compliant. The previous `hasThinking` check accepted these as already-satisfying, leaving non-compliant blocks in the wire message. Tighten the check so they're filtered out and the synthetic injection runs. Live verification: DeepSeek currently tolerates both shapes (lenient), but normalizing is defensively correct against future tightening. 2. Rename converter option `ensureAssistantThinking` → `ensureThinkingOnToolUseTurns`. The new name reflects the actual contract (tool-use turns only, not every assistant turn). 3. Honor `thinkingConfig.includeThoughts: false` in `buildOutputConfig`. Previously a per-request opt-out dropped the top-level `thinking` parameter but still emitted `output_config.effort`, leaking a reasoning-shaped field into side queries that don't want it. 4. Add regression test for mixed text + tool_use assistant turns (common shape: model says something, then calls a tool). 5. Add explicit test for the signature-validation path: an existing compliant thinking block (with signature) is preserved untouched. 64 tests pass; lint + typecheck clean. * fix(core): clean up non-compliant thinking blocks on plain-text turns + assert output_config gating Address PR review round 7 (copilot-pull-request-reviewer × 2): 1. (Hp-x) Round-tripped redacted_thinking blocks were left malformed on assistant turns lacking tool_use. The previous structure only ran the cleanup pass when a tool_use block was present (early return on `!hasToolUse`), so plain-text turns kept the non-compliant `{ type: 'thinking', thinking: '' }` shape. Restructure into two sequential steps: a. Drop non-compliant thinking blocks (no `signature`) on every assistant turn — same fallback that avoids `content: []` if the message is thinking-only. b. Inject the synthetic empty thinking block on tool_use turns that still lack a compliant thinking block after step (a). 2. (Hp-7) The includeThoughts=false test asserted that the top-level `thinking` field is suppressed but didn't cover `output_config`, leaving regressions in the new `buildOutputConfig` gate uncaught. Tighten the assertion to also verify `output_config` is absent. 3. New converter test: cleanup runs on plain-text assistant turns too. 65 tests pass; lint + typecheck clean. * test(core): add explicit redacted_thinking injection-path coverage Address PR review round 8 (#30 — copilot reviewer). The converter treats `redacted_thinking` as already satisfying the thinking-block requirement (no synthetic injected), distinguished from a signature-less `thinking` block which is non-compliant and gets dropped/replaced. Existing tests covered the latter path; this adds explicit coverage of the former. processContent doesn't synthesize redacted_thinking from Gemini parts, so the test reaches into the private helper directly. (#31 — subdomain hostname coverage — already exists at line 602.) 66 tests pass; lint + typecheck clean. * fix(core): per-request anthropic-beta + normalize thinking-only turns Address PR review round 9 (copilot-pull-request-reviewer × 2): 1. (Hydz) thinking-only assistant turns (e.g. max_tokens cutoff or round-tripped redacted_thinking) hit the cleanup-empties fallback and kept the original non-compliant `{ type: 'thinking', thinking: '' }` block. The fallback now replaces the message with a synthetic empty thinking block (`signature: ''` included), which keeps the message non-empty AND spec-compliant. 2. (Hyd4) `anthropic-beta` was set once at construction from the global `reasoning` config, so requests with per-request `thinkingConfig.includeThoughts=false` still advertised interleaved-thinking / effort even though the body had dropped the matching fields. Move beta computation to a new `buildPerRequestHeaders` that derives the header from the actual `thinking` / `output_config` fields present in the request body, and pass it via `messages.create(..., { headers })`. The wire shape is now internally consistent. Test updates: - Drop the three constructor-time beta assertions; they no longer apply. - Add four per-request header tests covering: both betas present, only interleaved-thinking, reasoning=false (no betas), and per-request includeThoughts=false (no betas). 67 tests pass; lint + typecheck clean. * fix(core): preserve thinking text by normalizing in place + merge user beta flags Address PR review round 10 (copilot-pull-request-reviewer × 2): 1. (H0oF) The previous cleanup filtered out every thinking block missing a `signature` field. But that shape is the normal output from OpenAI/Gemini/agent-runtime generators, which only set `thought: true` without a signature. Users switching providers mid-session would silently lose preserved thinking text on the first DeepSeek request. Change Step 1 to NORMALIZE in place: when a thinking block has no signature, set `signature: ''` rather than dropping the block. The original `thinking` text is preserved; DeepSeek empirically accepts empty signatures so the wire shape stays valid. 2. (H0oL) `buildPerRequestHeaders()` overwrote `customHeaders['anthropic-beta']` whenever the per-request override fired, regressing the customHeaders escape hatch for unrelated Anthropic beta features. Merge the user's flags into the computed list (deduped) so users can stack their own betas alongside interleaved-thinking / effort. Test changes: - Renamed and rewrote "drops non-compliant... plain-text" test to assert in-place normalization that preserves thinking text. - Updated "replaces a non-compliant thinking block" comment + name to describe the normalization (the assertion was already correct because the test happened to use empty thinking text). - The empty-content fallback in Step 1 is no longer reachable under the new logic, so the dedicated thinking-only-turn test now exercises only the strip path (where it remains relevant). - Added 3 customHeaders[anthropic-beta] tests: merge with computed, passthrough when no thinking/effort, dedupe. 70 tests pass; lint + typecheck clean. * docs(core): align thinking-injection comments with normalize semantics + add stream test Address PR review round 11 (copilot-pull-request-reviewer × 3): 1. (H3Iw) Update the `ensureThinkingOnToolUseTurns` option docstring to describe in-place normalization (preserving thinking text by filling in `signature: ''`) instead of the old drop-and-replace semantics. 2. (H3I9) Same update on the `applyEmptyThinkingToToolUseTurns` helper JSDoc — clarify that signature-less thinking blocks are normalized in place (preserving original text), not dropped. Mention the common case of cross-provider history where non-Anthropic generators only set `thought: true`. 3. (H3I3) Add a streaming test asserting that `generateContentStream()` also attaches the per-request `anthropic-beta` header. The previous coverage only exercised `generateContent()`, leaving the streaming path's separate code path (line 144 in anthropicContentGenerator.ts) unverified. 71 tests pass; lint + typecheck clean. * refactor(core): split DeepSeek thinking option in two + add header coexistence test Address PR review round 12 (copilot-pull-request-reviewer × 2): 1. (H6ws) The single `ensureThinkingOnToolUseTurns` option was misleadingly narrow: the implementation also rewrote non-tool-use turns by normalizing malformed thinking blocks. Future callers could enable it expecting only the tool-use behavior. Split into two precisely-named options: - normalizeAssistantThinkingSignature: fill missing `signature` on every assistant `thinking` block (cross-provider history compat). - injectThinkingOnToolUseTurns: prepend synthetic empty thinking on tool_use turns missing one (issue #3786 trigger). The generator wires both together for DeepSeek when thinking mode is on; either can be used independently if a future caller needs only one pass. 2. (H6w4) Add a test asserting that the per-request `headers` path coexists correctly with `customHeaders`: User-Agent and unrelated customHeaders entries stay in `defaultHeaders` while only the computed `anthropic-beta` rides on the per-request path. Defends against a future regression where header config might be routed through a code path that wipes the constructor defaults. 72 tests pass; lint + typecheck clean. * fix(core): case-insensitive customHeaders[anthropic-beta] merge Address yiliang114 review feedback (#3788). HTTP header names are case-insensitive by spec, and the Anthropic SDK lower-cases them during merge. Previously buildPerRequestHeaders only read the lower-case `anthropic-beta` key from customHeaders, so a user-configured `Anthropic-Beta` or `ANTHROPIC-BETA` would be silently overwritten by the per-request computed value. Replace the direct dict lookup with collectCustomBetaFlags() which walks all customHeaders entries and matches the key case-insensitively. Multiple matching entries (unlikely but possible) are concatenated; the existing dedupe pass handles any duplicates. Add a regression test for both `Anthropic-Beta` and `ANTHROPIC-BETA` key shapes. 73 tests pass; lint + typecheck clean. * docs(core): align thinking-injection docs with normalize-in-place semantics + redacted_thinking strip test Address PR review round 14 (copilot-pull-request-reviewer × 4): 1. (IAl4) PR description still described "dropped here so synthetic injection takes over" but the implementation now normalizes signature-less thinking blocks in place (preserving text). PR description rewritten to describe the two-pass model: normalize-in-place + injection-when-truly-missing. 2. (IAl7) `injectThinkingOnToolUseTurns` option docstring claimed signature-less blocks would be "seen as missing" so the synthetic replaces them. Updated to describe the actual flow: the normalization pass runs first, blocks become compliant in place, the injector then sees them as already-satisfying and prepends nothing. Helper JSDoc on `injectEmptyThinkingOnToolUseTurns` fixed the same way. 3. (IAl8) Strip-path coverage missed `redacted_thinking` blocks. Added regression test that verifies both thinking and redacted_thinking blocks are removed when `stripAssistantThinking` is set. 4. (IAl-) Renamed the converter test suite from "thinking-mode injection + normalization (DeepSeek thinking on)" to "DeepSeek thinking-mode normalization, injection, and stripping" so the title accurately covers all behavior the block exercises (including `stripAssistantThinking` cases later in the same describe). 74 tests pass; lint + typecheck clean. * fix(core): exclude anthropic-beta variants from defaultHeaders to avoid wire duplication Address PR review round 15 (copilot-pull-request-reviewer #1). `buildHeaders()` previously spread the entire `customHeaders` map into the SDK's `defaultHeaders`. After moving anthropic-beta computation to the per-request path, a user-configured mixed-case `Anthropic-Beta` key would survive in defaultHeaders verbatim, while the per-request override added a lowercase `anthropic-beta`. The wire then carried two physical headers for the same logical name — SDK behavior on duplicate headers with different casings is undefined. `buildPerRequestHeaders()` already merges those user flags case-insensitively (commit 0d8b5de), so dropping the entry from defaultHeaders is the right boundary: the per-request path owns the header end-to-end. Other customHeaders entries continue to pass through. Add a regression test asserting no `Anthropic-Beta` (any casing) lands in defaultHeaders while unrelated customHeaders are kept. 75 tests pass; lint + typecheck clean.
xaelistic
pushed a commit
to xaelistic/qwen-code
that referenced
this pull request
Jun 7, 2026
xaelistic
pushed a commit
to xaelistic/qwen-code
that referenced
this pull request
Jun 7, 2026
…wenLM#3747) Extend the DeepSeek reasoning_content normalization (introduced in QwenLM#3729) to assistant turns without tool_calls. The DeepSeek API rejects follow-up requests in thinking mode whenever any prior assistant turn omits reasoning_content, not just turns that carried tool_calls.
xaelistic
pushed a commit
to xaelistic/qwen-code
that referenced
this pull request
Jun 7, 2026
…rovider (QwenLM#3788) * fix(core): inject thinking blocks for DeepSeek anthropic-compatible provider DeepSeek's anthropic-compatible endpoint (https://api.deepseek.com/anthropic) rejects follow-up requests with HTTP 400 ("The content[].thinking in the thinking mode must be passed back to the API.") whenever a prior assistant turn carrying tool_use omits a thinking block. The model can legitimately return a tool round without thinking text, so qwen-code stored no thought parts and rebuilt the next request with no thinking block, tripping the API's check. Mirroring the existing OpenAI-side fix (QwenLM#3729, QwenLM#3747), the converter now detects DeepSeek by base URL or model name and prepends an empty { type: 'thinking', thinking: '', signature: '' } block to assistant turns missing one. Other anthropic-protocol providers are unaffected. Verified against the live api.deepseek.com/anthropic endpoint: - assistant with tool_use, no thinking → 400 (reproduces QwenLM#3786) - assistant with tool_use, empty thinking injected → 200 OK Refs QwenLM#3786 * fix(core): gate DeepSeek thinking-block injection on thinking mode Address PR review feedback: 1. (Critical) Gate empty-thinking injection on the same per-request condition that emits the top-level `thinking` parameter. The previous implementation injected unconditionally on DeepSeek providers, but `buildThinkingConfig()` may omit `thinking` when reasoning=false or `thinkingConfig.includeThoughts=false` — which is exactly what suggestionGenerator / ArenaManager / forkedAgent do. Shipping thinking blocks without enabling thinking mode is a protocol violation that DeepSeek may reject. Move the option from converter constructor to a per-request `convertGeminiRequestToAnthropic` parameter so the generator can compute the gate correctly. 2. (CodeQL) Replace `baseUrl.includes('api.deepseek.com')` with `new URL(baseUrl).hostname` exact-match. The substring check would accept spoofed hosts like `api.deepseek.com.evil.com`. 3. Document the empty-signature workaround inline. 4. Rename the misleading "redacted_thinking" test case. * fix(core): narrow DeepSeek thinking injection to tool_use turns + subdomain test Address PR review round 2: 1. Narrow injection scope to assistant turns containing tool_use. Live verification against api.deepseek.com/anthropic showed plain-text assistant turns without thinking are accepted unchanged — only tool_use turns trigger the HTTP 400. Injecting on every assistant turn unnecessarily bloats replay history with synthetic blocks the API does not require. Existing thinking blocks on any turn are still preserved untouched. 2. Add test coverage for the subdomain hostname branch (us.api.deepseek.com → matches), addressing the gap noted in review. 3. Update existing negative-case tests (non-deepseek / spoofed / reasoning=false / includeThoughts=false) to use tool_use scenarios so they actually exercise the gating logic instead of trivially passing under the narrowed scope. * docs(core): align DeepSeek thinking-injection comments with narrowed scope Address PR review round 3 (copilot-pull-request-reviewer × 3): comments in three locations still described the constraint as applying to "any prior assistant turn", which was true before commit 5388ded but no longer matches the implementation. Update the doc comment on isDeepSeekAnthropicProvider and the two test-suite header comments to state the actual narrower contract: the API rejects only tool-use turns that omit thinking blocks; plain-text assistant turns are accepted unchanged. Comment-only change; 58 tests still pass. * fix(core): per-request DeepSeek detection + strip thinking when off Address PR review round 4 (copilot-pull-request-reviewer × 2): 1. Stale provider-detection cache (HCia). The constructor cached isDeepSeekProvider once, but Config.setModel() mutates contentGeneratorConfig.model in place. After a runtime /model switch from a non-DeepSeek model to a DeepSeek one on the same auth config, buildRequest() would keep using the stale flag. Move the detection into buildRequest so each call sees the current model. The detector is cheap (URL parse + string compare). 2. Real thought parts leak through when thinking is disabled (HCib). The previous gate only blocked synthetic injection — but the converter still replayed any existing `thought: true` parts in request.contents as thinking blocks. Code paths that disable thinking against a session whose history was built with thinking on (suggestionGenerator / ArenaManager / forkedAgent) would still emit thinking blocks alongside an absent top-level `thinking` config — the same protocol mismatch the gate was meant to avoid. Add a `stripAssistantThinking` converter option, set in buildRequest to `isDeepSeek && !thinking`. The converter strips thinking and redacted_thinking blocks from assistant messages before message construction completes. Mirror behavior is already proven safe by live verification (DeepSeek currently tolerates either shape, but stripping makes the request body internally consistent and robust to future validation tightening). 3 new tests: - converter strips thinking from assistant turns when option set - generator strips real thought parts when reasoning=false - generator reflects runtime model changes (no stale cache) 61 tests pass; lint + typecheck clean. * fix(core): preserve thinking-only assistant turns instead of emitting empty content Address PR review round 5 (copilot-pull-request-reviewer × 2 — code + test): stripThinkingFromAssistantMessages previously replaced message.content with the filtered array unconditionally. For an assistant turn whose only blocks are thinking/redacted_thinking (e.g. a round cut off by max_tokens before any text or tool_use was emitted), this left `content: []` — which Anthropic API rejects. Dropping the message entirely was considered but would break the required user/assistant alternation. Instead, fall back to leaving the original blocks in place when stripping would empty the message. DeepSeek empirically tolerates the residual `thinking-block + no-thinking-config` shape (verified against api.deepseek.com/anthropic in the V2/X scenarios), so leaving the message untouched is the safer choice than emitting invalid structure. Add regression test for the thinking-only turn shape. 62 tests pass; lint + typecheck clean. * fix(core): validate thinking-block signature, rename option, gate output_config Address PR review round 6 — five substantive items: 1. (Critical) Drop non-compliant thinking blocks lacking a `signature` field and replace them with a synthetic one. A `redacted_thinking` block round-tripped through Gemini Part format becomes `{ text: '', thought: true }` (no thoughtSignature) and converts back to `{ type: 'thinking', thinking: '' }` without `signature` — not spec-compliant. The previous `hasThinking` check accepted these as already-satisfying, leaving non-compliant blocks in the wire message. Tighten the check so they're filtered out and the synthetic injection runs. Live verification: DeepSeek currently tolerates both shapes (lenient), but normalizing is defensively correct against future tightening. 2. Rename converter option `ensureAssistantThinking` → `ensureThinkingOnToolUseTurns`. The new name reflects the actual contract (tool-use turns only, not every assistant turn). 3. Honor `thinkingConfig.includeThoughts: false` in `buildOutputConfig`. Previously a per-request opt-out dropped the top-level `thinking` parameter but still emitted `output_config.effort`, leaking a reasoning-shaped field into side queries that don't want it. 4. Add regression test for mixed text + tool_use assistant turns (common shape: model says something, then calls a tool). 5. Add explicit test for the signature-validation path: an existing compliant thinking block (with signature) is preserved untouched. 64 tests pass; lint + typecheck clean. * fix(core): clean up non-compliant thinking blocks on plain-text turns + assert output_config gating Address PR review round 7 (copilot-pull-request-reviewer × 2): 1. (Hp-x) Round-tripped redacted_thinking blocks were left malformed on assistant turns lacking tool_use. The previous structure only ran the cleanup pass when a tool_use block was present (early return on `!hasToolUse`), so plain-text turns kept the non-compliant `{ type: 'thinking', thinking: '' }` shape. Restructure into two sequential steps: a. Drop non-compliant thinking blocks (no `signature`) on every assistant turn — same fallback that avoids `content: []` if the message is thinking-only. b. Inject the synthetic empty thinking block on tool_use turns that still lack a compliant thinking block after step (a). 2. (Hp-7) The includeThoughts=false test asserted that the top-level `thinking` field is suppressed but didn't cover `output_config`, leaving regressions in the new `buildOutputConfig` gate uncaught. Tighten the assertion to also verify `output_config` is absent. 3. New converter test: cleanup runs on plain-text assistant turns too. 65 tests pass; lint + typecheck clean. * test(core): add explicit redacted_thinking injection-path coverage Address PR review round 8 (QwenLM#30 — copilot reviewer). The converter treats `redacted_thinking` as already satisfying the thinking-block requirement (no synthetic injected), distinguished from a signature-less `thinking` block which is non-compliant and gets dropped/replaced. Existing tests covered the latter path; this adds explicit coverage of the former. processContent doesn't synthesize redacted_thinking from Gemini parts, so the test reaches into the private helper directly. (QwenLM#31 — subdomain hostname coverage — already exists at line 602.) 66 tests pass; lint + typecheck clean. * fix(core): per-request anthropic-beta + normalize thinking-only turns Address PR review round 9 (copilot-pull-request-reviewer × 2): 1. (Hydz) thinking-only assistant turns (e.g. max_tokens cutoff or round-tripped redacted_thinking) hit the cleanup-empties fallback and kept the original non-compliant `{ type: 'thinking', thinking: '' }` block. The fallback now replaces the message with a synthetic empty thinking block (`signature: ''` included), which keeps the message non-empty AND spec-compliant. 2. (Hyd4) `anthropic-beta` was set once at construction from the global `reasoning` config, so requests with per-request `thinkingConfig.includeThoughts=false` still advertised interleaved-thinking / effort even though the body had dropped the matching fields. Move beta computation to a new `buildPerRequestHeaders` that derives the header from the actual `thinking` / `output_config` fields present in the request body, and pass it via `messages.create(..., { headers })`. The wire shape is now internally consistent. Test updates: - Drop the three constructor-time beta assertions; they no longer apply. - Add four per-request header tests covering: both betas present, only interleaved-thinking, reasoning=false (no betas), and per-request includeThoughts=false (no betas). 67 tests pass; lint + typecheck clean. * fix(core): preserve thinking text by normalizing in place + merge user beta flags Address PR review round 10 (copilot-pull-request-reviewer × 2): 1. (H0oF) The previous cleanup filtered out every thinking block missing a `signature` field. But that shape is the normal output from OpenAI/Gemini/agent-runtime generators, which only set `thought: true` without a signature. Users switching providers mid-session would silently lose preserved thinking text on the first DeepSeek request. Change Step 1 to NORMALIZE in place: when a thinking block has no signature, set `signature: ''` rather than dropping the block. The original `thinking` text is preserved; DeepSeek empirically accepts empty signatures so the wire shape stays valid. 2. (H0oL) `buildPerRequestHeaders()` overwrote `customHeaders['anthropic-beta']` whenever the per-request override fired, regressing the customHeaders escape hatch for unrelated Anthropic beta features. Merge the user's flags into the computed list (deduped) so users can stack their own betas alongside interleaved-thinking / effort. Test changes: - Renamed and rewrote "drops non-compliant... plain-text" test to assert in-place normalization that preserves thinking text. - Updated "replaces a non-compliant thinking block" comment + name to describe the normalization (the assertion was already correct because the test happened to use empty thinking text). - The empty-content fallback in Step 1 is no longer reachable under the new logic, so the dedicated thinking-only-turn test now exercises only the strip path (where it remains relevant). - Added 3 customHeaders[anthropic-beta] tests: merge with computed, passthrough when no thinking/effort, dedupe. 70 tests pass; lint + typecheck clean. * docs(core): align thinking-injection comments with normalize semantics + add stream test Address PR review round 11 (copilot-pull-request-reviewer × 3): 1. (H3Iw) Update the `ensureThinkingOnToolUseTurns` option docstring to describe in-place normalization (preserving thinking text by filling in `signature: ''`) instead of the old drop-and-replace semantics. 2. (H3I9) Same update on the `applyEmptyThinkingToToolUseTurns` helper JSDoc — clarify that signature-less thinking blocks are normalized in place (preserving original text), not dropped. Mention the common case of cross-provider history where non-Anthropic generators only set `thought: true`. 3. (H3I3) Add a streaming test asserting that `generateContentStream()` also attaches the per-request `anthropic-beta` header. The previous coverage only exercised `generateContent()`, leaving the streaming path's separate code path (line 144 in anthropicContentGenerator.ts) unverified. 71 tests pass; lint + typecheck clean. * refactor(core): split DeepSeek thinking option in two + add header coexistence test Address PR review round 12 (copilot-pull-request-reviewer × 2): 1. (H6ws) The single `ensureThinkingOnToolUseTurns` option was misleadingly narrow: the implementation also rewrote non-tool-use turns by normalizing malformed thinking blocks. Future callers could enable it expecting only the tool-use behavior. Split into two precisely-named options: - normalizeAssistantThinkingSignature: fill missing `signature` on every assistant `thinking` block (cross-provider history compat). - injectThinkingOnToolUseTurns: prepend synthetic empty thinking on tool_use turns missing one (issue QwenLM#3786 trigger). The generator wires both together for DeepSeek when thinking mode is on; either can be used independently if a future caller needs only one pass. 2. (H6w4) Add a test asserting that the per-request `headers` path coexists correctly with `customHeaders`: User-Agent and unrelated customHeaders entries stay in `defaultHeaders` while only the computed `anthropic-beta` rides on the per-request path. Defends against a future regression where header config might be routed through a code path that wipes the constructor defaults. 72 tests pass; lint + typecheck clean. * fix(core): case-insensitive customHeaders[anthropic-beta] merge Address yiliang114 review feedback (QwenLM#3788). HTTP header names are case-insensitive by spec, and the Anthropic SDK lower-cases them during merge. Previously buildPerRequestHeaders only read the lower-case `anthropic-beta` key from customHeaders, so a user-configured `Anthropic-Beta` or `ANTHROPIC-BETA` would be silently overwritten by the per-request computed value. Replace the direct dict lookup with collectCustomBetaFlags() which walks all customHeaders entries and matches the key case-insensitively. Multiple matching entries (unlikely but possible) are concatenated; the existing dedupe pass handles any duplicates. Add a regression test for both `Anthropic-Beta` and `ANTHROPIC-BETA` key shapes. 73 tests pass; lint + typecheck clean. * docs(core): align thinking-injection docs with normalize-in-place semantics + redacted_thinking strip test Address PR review round 14 (copilot-pull-request-reviewer × 4): 1. (IAl4) PR description still described "dropped here so synthetic injection takes over" but the implementation now normalizes signature-less thinking blocks in place (preserving text). PR description rewritten to describe the two-pass model: normalize-in-place + injection-when-truly-missing. 2. (IAl7) `injectThinkingOnToolUseTurns` option docstring claimed signature-less blocks would be "seen as missing" so the synthetic replaces them. Updated to describe the actual flow: the normalization pass runs first, blocks become compliant in place, the injector then sees them as already-satisfying and prepends nothing. Helper JSDoc on `injectEmptyThinkingOnToolUseTurns` fixed the same way. 3. (IAl8) Strip-path coverage missed `redacted_thinking` blocks. Added regression test that verifies both thinking and redacted_thinking blocks are removed when `stripAssistantThinking` is set. 4. (IAl-) Renamed the converter test suite from "thinking-mode injection + normalization (DeepSeek thinking on)" to "DeepSeek thinking-mode normalization, injection, and stripping" so the title accurately covers all behavior the block exercises (including `stripAssistantThinking` cases later in the same describe). 74 tests pass; lint + typecheck clean. * fix(core): exclude anthropic-beta variants from defaultHeaders to avoid wire duplication Address PR review round 15 (copilot-pull-request-reviewer QwenLM#1). `buildHeaders()` previously spread the entire `customHeaders` map into the SDK's `defaultHeaders`. After moving anthropic-beta computation to the per-request path, a user-configured mixed-case `Anthropic-Beta` key would survive in defaultHeaders verbatim, while the per-request override added a lowercase `anthropic-beta`. The wire then carried two physical headers for the same logical name — SDK behavior on duplicate headers with different casings is undefined. `buildPerRequestHeaders()` already merges those user flags case-insensitively (commit 414546d), so dropping the entry from defaultHeaders is the right boundary: the per-request path owns the header end-to-end. Other customHeaders entries continue to pass through. Add a regression test asserting no `Anthropic-Beta` (any casing) lands in defaultHeaders while unrelated customHeaders are kept. 75 tests pass; lint + typecheck clean.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
reasoning_contentnormalization to all prior assistant turns, not just ones withtool_calls.reasoning_contenton tool-calling assistant turns. Live testing againstapi.deepseek.comshows the API also rejects follow-up requests when an assistant turn withouttool_callsis missingreasoning_content— same HTTP 400 (The reasoning_content in the thinking mode must be passed back to the API), just from a different turn. The guard was too narrow.reasoning_contentis still preserved.Validation
Before(A series of failed requests)
After(Request succeed)
Proof of response with tool_calls but no reason_content
Scope / Risk
reasoning_content. The previous PR (fix(core): inject reasoning_content on DeepSeek tool-call replays #3729) already validated the empty-string approach against the live DeepSeek endpoint for tool-calling turns; this change extends the same fallback to non-tool-calling turns.Testing Matrix
Testing matrix notes:
Linked Issues / Bugs
Relates to #3695, #3729