fix(anthropic): forward --thinking to extended thinking API by wangwllu · Pull Request #233 · steipete/summarize

wangwllu · 2026-06-03T15:29:00Z

Proof

Latest head 3d7cc96 proof: see comment #4638051618 — outbound request captures for both (a) CLI --thinking xhigh correctly forwarding the thinking block to Anthropic and (b) persisted openai.thinking not leaking into Anthropic requests.

Problem

The Anthropic provider accepts --thinking <effort> (and the matching thinking config field) without error, but never forwards the value to the API. Requests go out with no thinking block, so users who set a reasoning level on anthropic/... models get no extended thinking and don't know it.

Reproduces on main against any Claude model:

summarize "https://example.com" --model anthropic/claude-opus-4-5 --thinking xhigh --verbose

Verbose log shows model=anthropic/claude-opus-4-5 but the outbound request has no thinking field.

Root cause

Two layers:

generateTextWithModelId in src/llm/generate-text.ts reads requestOptions.reasoningEffort for the OpenAI path but never threads it into completeAnthropicText.
completeAnthropicText in src/llm/providers/anthropic.ts calls pi-ai's completeSimple without a reasoning option. Even when threaded through, models built via createSyntheticModel default to reasoning: false, which makes the pi-ai Anthropic adapter silently drop the block (see streamSimpleAnthropic — it gates on model.reasoning).

Fix

Pass requestOptions?.reasoningEffort from generateTextWithModelId into completeAnthropicText, mirroring how the OpenAI path uses it.
In completeAnthropicText, map the OpenAI effort enum (low/medium/high/xhigh; none is treated as off) to a pi-ai ThinkingLevel and forward it as reasoning. When the model came from createSyntheticModel (reasoning: false), opt the model into thinking so the pi-ai adapter does not silently drop the block.

Verification

Tested end-to-end against a custom ANTHROPIC_BASE_URL (Claude Opus 4.7 via a proxy that exposes the synthetic-model code path). The outbound request body (captured with a fetch monkey-patch wrapper) before vs after this PR:

`--thinking`	Before	After
(unset)	`has_thinking: false`	`has_thinking: false` (unchanged)
`low`	`has_thinking: false`	`thinking: {enabled, budget_tokens: 2048}`
`medium`	`has_thinking: false`	`thinking: {enabled, budget_tokens: 8192}`
`high`	`has_thinking: false`	`thinking: {enabled, budget_tokens: 15360}`
`xhigh`	`has_thinking: false`	`thinking: {enabled, budget_tokens: 15360}`

(With max_tokens=16384 the pi-ai SDK caps thinking budget at max_tokens − 1024, hence high and xhigh both land at 15360 — that's SDK behavior, not this patch.)

Summaries still return correctly in all cases.

Notes / scope

Only completeAnthropicText is patched. completeAnthropicDocument builds its request body by hand (no pi-ai dispatch) — extending thinking support there would be a separate, larger change touching the manual payload assembly.
--thinking's CLI help text still reads "OpenAI reasoning effort"; happy to update it in this PR if you'd like the wording to acknowledge the Anthropic path now works.

The Anthropic provider accepted `--thinking <effort>` (and the matching `thinking` config field) without error, but never forwarded the value to the API. Requests went out with no `thinking` block, so users who set a reasoning level got no extended thinking on Claude Opus / Sonnet. Two changes: 1. Pass `requestOptions.reasoningEffort` from `generateTextWithModelId` into `completeAnthropicText`, mirroring how the OpenAI path uses it. 2. In `completeAnthropicText`, map the OpenAI effort enum to a pi-ai `ThinkingLevel` and forward it as `reasoning`. When the model came from `createSyntheticModel` (`reasoning: false` by default), opt the model into thinking so the pi-ai adapter does not silently drop the block. Verified end-to-end against a custom `ANTHROPIC_BASE_URL` (Claude Opus 4.7 via a proxy that exposes the synthetic-model code path): --thinking low -> budget_tokens: 2048 --thinking medium -> budget_tokens: 8192 --thinking high -> budget_tokens: 15360 --thinking xhigh -> budget_tokens: 15360 Before: `has_thinking: false` for every effort level.

clawsweeper · 2026-06-03T15:30:16Z

Codex review: needs maintainer review before merge. Reviewed June 6, 2026, 5:09 AM ET / 09:09 UTC.

Summary
The branch forwards explicit --thinking into Anthropic text generation, scopes OpenAI-only request defaults by provider, and adds focused regression tests for Anthropic reasoning and URL markdown routing.

Reproducibility: yes. Source inspection on current main shows --thinking can be parsed into request options, but the Anthropic text and stream paths do not pass any reasoning option into the pi-ai calls.

Review metrics: 2 noteworthy metrics.

Files changed: 14 files, +487/-37. The patch spans provider dispatch, run configuration, daemon/chat routing, URL markdown conversion, and tests.
Test surface touched: 2 added, 2 changed. Focused tests cover Anthropic reasoning helper behavior, provider-option scoping, run-config mapping, and Anthropic markdown routing.

Merge readiness
Overall: 🐚 platinum hermit
Proof: 🐚 platinum hermit
Patch quality: 🐚 platinum hermit
Result: ready for maintainer review.

Overall follows the weaker of proof and patch quality, so missing proof can cap an otherwise strong patch.

Rank-up moves:

Decide whether Anthropic document-attachment thinking support should be a separate follow-up before describing this as broad Anthropic thinking coverage.

Risk before merge

[P1] The PR intentionally changes outbound Anthropic request payloads and provider option scoping; maintainers should be comfortable with the current-head captures as enough external API and custom ANTHROPIC_BASE_URL proof.
[P1] Anthropic document-attachment requests still build manual payloads without thinking support, so this merge should be treated as text and streaming support unless a follow-up is opened.

Maintainer options:

Merge with scoped compatibility risk (recommended)
Merge after normal maintainer check review, relying on current-head request captures and focused regression tests for the Anthropic payload and OpenAI-default isolation behavior.
Ask for broader Anthropic coverage
Pause merge only if maintainers want document-attachment thinking support or live proxy/API proof beyond the current outbound request captures in this same PR.

Next step before merge

No automated repair is needed; the remaining action is maintainer merge judgment for a compatibility-sensitive provider payload change.

Security
Cleared: The diff changes TypeScript provider/config/test code only and does not add dependencies, workflows, secret handling, downloaded artifacts, or new code execution paths.

Review details

Best possible solution:

Land the provider-scoped text and streaming implementation after normal maintainer review, and track document-attachment thinking separately if full Anthropic coverage is required.

Do we have a high-confidence way to reproduce the issue?

Yes. Source inspection on current main shows --thinking can be parsed into request options, but the Anthropic text and stream paths do not pass any reasoning option into the pi-ai calls.

Is this the best way to solve the issue?

Yes. The provider-scoped merge helper preserves OpenAI defaults while forwarding explicit CLI and per-attempt reasoning, and the Anthropic helper avoids enabling thinking on registered unsupported models.

AGENTS.md: found and applied where relevant.

Codex review notes: model gpt-5.5, reasoning high; reviewed against f40e26b1330e.

Label changes

Label changes:

add proof: sufficient: Contributor real behavior proof is sufficient. The latest PR body/comment includes current-head redacted outbound request captures showing enabled Anthropic thinking for explicit CLI input and no leak from persisted OpenAI thinking defaults.
add rating: 🐚 platinum hermit: Overall readiness is 🐚 platinum hermit; proof is 🐚 platinum hermit and patch quality is 🐚 platinum hermit.
add status: 👀 ready for maintainer look: ClawSweeper has no concrete contributor-facing blocker left for this PR. Sufficient (logs): The latest PR body/comment includes current-head redacted outbound request captures showing enabled Anthropic thinking for explicit CLI input and no leak from persisted OpenAI thinking defaults.
remove status: 📣 needs proof: Current PR status label is status: 👀 ready for maintainer look.
remove rating: 🦪 silver shellfish: Current PR rating is rating: 🐚 platinum hermit, so this older rating label is no longer current.

Label justifications:

P2: This is a normal-priority provider behavior fix with limited blast radius and no evidence of data loss, security bypass, or runtime-wide outage.
merge-risk: 🚨 compatibility: The PR changes request-option scoping and outbound Anthropic request payloads, which can affect existing provider configurations and custom proxy setups.
rating: 🐚 platinum hermit: Overall readiness is 🐚 platinum hermit; proof is 🐚 platinum hermit and patch quality is 🐚 platinum hermit.
status: 👀 ready for maintainer look: ClawSweeper has no concrete contributor-facing blocker left for this PR. Sufficient (logs): The latest PR body/comment includes current-head redacted outbound request captures showing enabled Anthropic thinking for explicit CLI input and no leak from persisted OpenAI thinking defaults.
proof: sufficient: Contributor real behavior proof is sufficient. The latest PR body/comment includes current-head redacted outbound request captures showing enabled Anthropic thinking for explicit CLI input and no leak from persisted OpenAI thinking defaults.

Evidence reviewed

What I checked:

Repository policy read: Full AGENTS.md was read and applied; the review stayed read-only and did not run mutating build or test commands. (AGENTS.md:1, f40e26b1330e)
Current main still lacks Anthropic text reasoning forwarding: completeAnthropicText on current main calls completeSimple without a reasoning option, so the PR is not obsolete. (src/llm/providers/anthropic.ts:76, f40e26b1330e)
Current main still lacks Anthropic stream reasoning forwarding: The Anthropic streaming path on current main also calls streamSimple without a reasoning option. (src/llm/generate-text-stream.ts:304, f40e26b1330e)
Current main stores CLI thinking in OpenAI override state: --thinking is parsed into openaiRequestOptionsOverride on current main, which explains why the PR needed separate provider scoping. (src/run/run-config.ts:113, f40e26b1330e)
Latest PR forwards Anthropic reasoning safely: At PR head, prepareAnthropicReasoning forwards reasoning for supported registered models, drops it for registered unsupported Claude models, and opts synthetic models into reasoning before the text call passes it to completeSimple. (src/llm/providers/anthropic.ts:34, 3d7cc9681194)
Latest PR scopes request options by provider: mergeRequestOptionsForProvider keeps persisted OpenAI defaults and OpenAI-only CLI overrides on OpenAI while forwarding explicit CLI reasoning to the selected provider. (src/llm/model-options.ts:95, 3d7cc9681194)

Likely related people:

steipete: Blame on the Anthropic provider, run config, and request-option merge helper points to the v0.16.3 release commit authored by Peter Steinberger. (role: introduced behavior; confidence: high; commits: fcf8c8e5e98d; files: src/llm/providers/anthropic.ts, src/run/run-config.ts, src/llm/model-options.ts)
Mykhailo: Recent main history touched adjacent model dispatch paths in the Antigravity CLI provider change, including summary-engine and daemon chat routing. (role: recent adjacent contributor; confidence: low; commits: 821e76613ded; files: src/run/summary-engine.ts, src/daemon/chat.ts)

What the crustacean ranks mean

🦀 challenger crab: rare, exceptional readiness with strong proof, clean implementation, and convincing validation.
🦞 diamond lobster: very strong readiness with only minor maintainer review expected.
🐚 platinum hermit: good normal PR, likely mergeable with ordinary maintainer review.
🦐 gold shrimp: useful signal, but proof or patch confidence is still limited.
🦪 silver shellfish: thin signal; proof, validation, or implementation needs work.
🧂 unranked krab: not merge-ready because proof is missing/unusable or there are serious correctness or safety concerns.
🌊 off-meta tidepool: rating does not apply to this item.

Shiny media proof means a screenshot, video, or linked artifact directly shows the changed behavior. Runtime, network, CSP, and security claims still need visible diagnostics.

How this review workflow works

ClawSweeper keeps one durable marker-backed review comment per issue or PR.
Re-runs edit this comment so the latest verdict, findings, and automation markers stay together instead of adding duplicate bot comments.
A fresh review can be triggered by eligible @clawsweeper re-review comments, exact-item GitHub events, scheduled/background review runs, or manual workflow dispatch.
PR/issue authors and users with repository write access can comment @clawsweeper re-review or @clawsweeper re-run on an open PR or issue to request a fresh review only.
Maintainers can also comment @clawsweeper review to request a fresh review only.
Fresh-review commands do not start repair, autofix, rebase, CI repair, or automerge.
Maintainer-only repair and merge flows require explicit commands such as @clawsweeper autofix, @clawsweeper automerge, @clawsweeper fix ci, or @clawsweeper address review.
Maintainers can comment @clawsweeper explain to ask for more context, or @clawsweeper stop to stop active automation.

…hetic models only Addresses the two P1 review findings on steipete#233: 1. Streaming path was uncovered. `--stream auto` enables streaming in TTY, so `summarize` from a terminal would call `streamSimple` without the `reasoning` option and the thinking block would never be sent. Thread the same handling through `streamAnthropic`-style dispatch in `generate-text-stream.ts`. 2. Earlier override flipped `reasoning: true` on ANY model whose pi-ai metadata reported `reasoning: false`. That includes registered Claude 3 / 3.5 models where extended thinking is unsupported, so users with a global `thinking` setting would go from successful no-thinking requests to API rejections. Restrict the override to *synthetic* models (`tryGetModel` miss — typical for custom `ANTHROPIC_BASE_URL` proxies in front of newer Claude versions). Shared helper `prepareAnthropicReasoning` is extracted to keep both dispatch sites in sync. Tests cover the four invariants: - no effort -> model and reasoning untouched - `none` -> off (no reasoning forwarded) - registered supported model -> reasoning forwarded, model metadata preserved - registered unsupported model (`reasoning: false`) -> metadata NOT mutated - synthetic model -> `reasoning: true` flipped so pi-ai forwards thinking Verified end-to-end against a custom `ANTHROPIC_BASE_URL` (Claude Opus 4.7 via proxy) on both `--stream on` and `--stream off`; outbound body now carries `thinking: {type: "enabled", budget_tokens: 15360, display: "summarized"}` in both paths.

wangwllu · 2026-06-03T16:08:29Z

Thanks for the thorough review — addressed both P1 findings in cdcea8f.

Streaming path (generate-text-stream.ts:304): threaded the same handling through the streaming Anthropic dispatch. Both --stream on and --stream off now carry the thinking block in the outbound request body.

Unsupported-model metadata (anthropic.ts): the reasoning: true override is now gated on synthetic models only — i.e. models that miss tryGetModel("anthropic", id), which is what createSyntheticModel produces for custom ANTHROPIC_BASE_URL proxies. Registered Claude 3 / 3.5 models keep their reasoning: false flag so the pi-ai adapter can keep doing the right thing (drop thinking, don't get a 4xx).

Extracted prepareAnthropicReasoning so both dispatch sites stay in sync. Added focused tests covering: no-effort no-op, none alias for off, registered supported model preserves metadata, registered unsupported model NOT flipped, synthetic model gets the override.

Re-verified end-to-end with a fetch wrapper logging outbound bodies:

non-stream synthetic Claude-Opus-4.7 + --thinking high → thinking: {enabled, budget_tokens: 15360} ✅
stream synthetic Claude-Opus-4.7 + --thinking high → thinking: {enabled, budget_tokens: 15360} ✅

Happy to also update the --thinking CLI help text (still reads "OpenAI reasoning effort") if you'd like the Anthropic side acknowledged in the same PR.

wangwllu · 2026-06-06T07:16:33Z

@clawsweeper re-review

Addressed the P1 finding from your last verdict in commit 27c5d61: prepareAnthropicReasoning no longer forwards reasoning for registered Anthropic models whose pi-ai metadata reports reasoning: false (Claude 3/3.5). Added a regression test using claude-3-5-sonnet-20241022 (verified reasoning: false in the pi-ai registry). Synthetic-model branch unchanged.

clawsweeper · 2026-06-06T07:16:36Z

🦞🧹
ClawSweeper re-review requested.

I asked ClawSweeper to review this item again.
Action: item re-review queued (workflow sweep.yml, event repository_dispatch).
Result: the existing ClawSweeper review comment will be edited in place when the review finishes.

Re-review progress:

State: Complete
Detail: The targeted re-review finished, the durable review comment was updated, and the synced verdict was routed.
Run: https://github.com/openclaw/clawsweeper/actions/runs/27055993584
Updated: 2026-06-06T07:24:37.127Z

wangwllu · 2026-06-06T07:46:17Z

@clawsweeper re-review

Addressed your latest P1 (Keep OpenAI thinking defaults from silently affecting Anthropic, src/llm/generate-text.ts:253) in commit d1666f39.

Provider-scope helper mergeRequestOptionsForProvider (src/llm/model-options.ts) drops openai-scoped defaults when the resolved provider is not openai; summary-engine now uses it instead of the unconditional mergeModelRequestOptions. Added 4 regression tests covering:

anthropic + global openai reasoningEffort set + empty attempt opts → reasoningEffort NOT forwarded
anthropic + attempt-scoped reasoningEffort → forwarded
openai + global default → forwarded (no regression)
openrouter + global openai default → not forwarded (same scoping for other non-openai providers)

pnpm check is green (2001 passed / 41 skipped).

clawsweeper · 2026-06-06T07:46:19Z

🦞🧹
ClawSweeper re-review requested.

I asked ClawSweeper to review this item again.
Action: item re-review queued (workflow sweep.yml, event repository_dispatch).
Result: the existing ClawSweeper review comment will be edited in place when the review finishes.

Re-review progress:

State: Complete
Detail: The targeted re-review finished, the durable review comment was updated, and the synced verdict was routed.
Run: https://github.com/openclaw/clawsweeper/actions/runs/27056604077
Updated: 2026-06-06T07:53:01.048Z

wangwllu · 2026-06-06T08:32:27Z

@clawsweeper re-review

Addressed both new P1s from your previous review in commit 3d7cc968:

1. --thinking no longer dropped on Anthropic path — Added a separate cliReasoningEffortOverride field (option (i) in the design space). run-config.ts now routes:

CLI --thinking → cliReasoningEffortOverride (provider-neutral, forwarded to ALL providers)
CLI --fast / --service-tier → stays in openaiRequestOptionsOverride (openai-specific)
Persisted openai.thinking config → openaiRequestOptions global default (openai-only)

mergeRequestOptionsForProvider takes the new field and forwards it for both openai and non-openai dispatch.

2. URL markdown path now provider-scoped — Replaced the unscoped mergeModelRequestOptions in src/run/flows/url/markdown.ts:260 and :368 with mergeRequestOptionsForProvider. Also swept src/daemon/chat.ts:253, :322 for the same reason.

Coverage:

tests/model-options.test.ts — new cases for CLI override-on-anthropic + persisted-default-not-on-anthropic
tests/run.url-markdown-anthropic.test.ts — new file (117 lines) covering markdown provider scoping
tests/run.config.test.ts — updated for the new field

pnpm -s check is green.

clawsweeper · 2026-06-06T08:32:30Z

🦞🧹
ClawSweeper re-review requested.

I asked ClawSweeper to review this item again.
Action: item re-review queued (workflow sweep.yml, event repository_dispatch).
Result: the existing ClawSweeper review comment will be edited in place when the review finishes.

Re-review progress:

State: Complete
Detail: The targeted re-review finished, the durable review comment was updated, and the synced verdict was routed.
Run: https://github.com/openclaw/clawsweeper/actions/runs/27057542901
Updated: 2026-06-06T08:39:21.198Z

wangwllu · 2026-06-06T09:04:30Z

Current-head proof (commit `3d7cc96`)

Per ClawSweeper's latest review, here are the two outbound-request captures requested.

Capture method: one-shot Node script (233-capture.mjs) that monkey-patches globalThis.fetch (the transport the Anthropic SDK uses by default) to short-circuit and record the outbound body, then calls the production completeAnthropicText after composing request options through mergeRequestOptionsForProvider({ provider: "anthropic", … }) — the exact merge src/run/run-config.ts performs at runtime. API key is a fake placeholder; nothing leaves the process.

Artifact 1 — CLI `--thinking xhigh` reaches Anthropic

Run shape: summarize --model anthropic/claude-opus-4-5 --thinking xhigh …
(mergeRequestOptionsForProvider with cliReasoningEffortOverride: "xhigh" resolves to { reasoningEffort: "xhigh" }.)

Outbound request body (redacted)

{
  "url": "https://api.anthropic.com/v1/messages",
  "body": {
    "model": "claude-opus-4-5",
    "messages": [
      {
        "role": "user",
        "content": [
          {
            "type": "text",
            "text": "Say hi.",
            "cache_control": { "type": "ephemeral" }
          }
        ]
      }
    ],
    "max_tokens": 32768,
    "stream": true,
    "thinking": {
      "type": "enabled",
      "budget_tokens": 16384,
      "display": "summarized"
    }
  }
}

Confirms thinking: { type: "enabled", budget_tokens: 16384 } is on the wire — the PR's advertised CLI path now actually works for Anthropic.

Artifact 2 — Persisted `openai.thinking` does NOT leak to Anthropic

Config shape: openai.thinking: "high" in the config file (mirrored as openaiGlobalDefault: { thinking: "high" }), no CLI flag, model anthropic/claude-opus-4-5.

Debug log from the script: mergedRequestOptions: null — mergeRequestOptionsForProvider correctly drops the OpenAI-scoped default before it reaches the Anthropic dispatch.

Outbound request body (redacted)

{
  "url": "https://api.anthropic.com/v1/messages",
  "body": {
    "model": "claude-opus-4-5",
    "messages": [
      {
        "role": "user",
        "content": [
          {
            "type": "text",
            "text": "Say hi.",
            "cache_control": { "type": "ephemeral" }
          }
        ]
      }
    ],
    "max_tokens": 16384,
    "stream": true,
    "thinking": { "type": "disabled" }
  }
}

Note on the thinking: { type: "disabled" } block: pi-ai's streamSimpleAnthropic translates "no reasoning option supplied" into an explicit { thinkingEnabled: false } for reasoning-capable models, which the Anthropic adapter renders as thinking: { type: "disabled" } in the request. That is the API-level opt-out; no budget_tokens and no type: "enabled" block — i.e. extended thinking is not activated, and the persisted openai.thinking: "high" did not leak into the Anthropic request body. (Synthetic / proxy models behind a custom ANTHROPIC_BASE_URL, where model.reasoning is false, would not even emit the disabled marker — pi-ai short-circuits before it.)

wangwllu · 2026-06-06T09:58:33Z

Ready for maintainer review

ClawSweeper verdict on 3d7cc968 (re-reviewed after the latest proof update):

Overall: 🐚 platinum hermit / Patch quality: 🐚 platinum hermit / Proof: 🐚 platinum hermit — "ready for maintainer review"

The patch went through five iterations to converge:

Round	HEAD	Change	ClawSweeper
1	`cdcea8fa`	initial `--thinking` passthrough + streaming + synthetic override	unranked krab (registered-unsupported still leaks)
2	`27c5d619`	drop reasoning for registered + `reasoning: false` models	gold shrimp (cross-provider leak)
3	`d1666f39`	`mergeRequestOptionsForProvider` helper — non-openai drops openai-scoped defaults	platinum patch / silver overall (CLI `--thinking` also dropped)
4	`3d7cc968`	split `cliReasoningEffortOverride` field; sweep all 4 merge call sites (summary-engine, url-markdown, daemon-chat×2)	platinum patch / silver overall (needs current-head proof)
5	`3d7cc968`	(no code change) added redacted outbound-request captures for both PR-advertised paths	platinum across all three axes

The current behavior:

summarize --model anthropic/... --thinking xhigh → outbound contains thinking: { type: "enabled", budget_tokens: 16384 }
Persisted openai.thinking: high in config + no CLI flag → Anthropic outbound has no enabled thinking block
openai.thinking persisted defaults still reach OpenAI providers as before — no regression
--fast / --service-tier CLI flags stay OpenAI-scoped via openaiRequestOptionsOverride

Two outbound-request artifacts attached at #issuecomment-4638051618 (captured via fetch-mock + production completeAnthropicText).

@steipete — tagging since this is your repo. Happy to address any further feedback. PR can be reviewed against current main (the branch was based before the pi-ai 0.78.0 bump but no logical conflicts).

fix(anthropic): drop reasoning for registered unsupported Claude models

27c5d61

clawsweeper Bot added rating: 🦐 gold shrimp Decent PR readiness signal, but merge confidence is limited. and removed rating: 🧂 unranked krab Not merge-ready due to missing proof or serious correctness/safety concerns. labels Jun 6, 2026

fix(anthropic): do not inherit openai-scoped reasoning defaults

d1666f3

fix(anthropic): preserve explicit CLI --thinking when scoping defaults

3d7cc96

clawsweeper Bot added the status: 👀 ready for maintainer look ClawSweeper has no concrete contributor-facing blocker left for this PR. label Jun 6, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix(anthropic): forward --thinking to extended thinking API#233

fix(anthropic): forward --thinking to extended thinking API#233
wangwllu wants to merge 5 commits into
steipete:mainfrom
wangwllu:fix/anthropic-thinking-passthrough

wangwllu commented Jun 3, 2026 •

edited

Loading

Uh oh!

clawsweeper Bot commented Jun 3, 2026 •

edited

Loading

Uh oh!

wangwllu commented Jun 3, 2026

Uh oh!

wangwllu commented Jun 6, 2026

Uh oh!

clawsweeper Bot commented Jun 6, 2026 •

edited

Loading

Uh oh!

wangwllu commented Jun 6, 2026

Uh oh!

clawsweeper Bot commented Jun 6, 2026 •

edited

Loading

Uh oh!

wangwllu commented Jun 6, 2026

Uh oh!

clawsweeper Bot commented Jun 6, 2026 •

edited

Loading

Uh oh!

wangwllu commented Jun 6, 2026

Uh oh!

wangwllu commented Jun 6, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

wangwllu commented Jun 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Proof

Problem

Root cause

Fix

Verification

Notes / scope

Uh oh!

clawsweeper Bot commented Jun 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

wangwllu commented Jun 3, 2026

Uh oh!

wangwllu commented Jun 6, 2026

Uh oh!

clawsweeper Bot commented Jun 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

wangwllu commented Jun 6, 2026

Uh oh!

clawsweeper Bot commented Jun 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

wangwllu commented Jun 6, 2026

Uh oh!

clawsweeper Bot commented Jun 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

wangwllu commented Jun 6, 2026

Current-head proof (commit 3d7cc96)

Artifact 1 — CLI --thinking xhigh reaches Anthropic

Artifact 2 — Persisted openai.thinking does NOT leak to Anthropic

Uh oh!

wangwllu commented Jun 6, 2026

Ready for maintainer review

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

wangwllu commented Jun 3, 2026 •

edited

Loading

clawsweeper Bot commented Jun 3, 2026 •

edited

Loading

clawsweeper Bot commented Jun 6, 2026 •

edited

Loading

clawsweeper Bot commented Jun 6, 2026 •

edited

Loading

clawsweeper Bot commented Jun 6, 2026 •

edited

Loading

Current-head proof (commit `3d7cc96`)

Artifact 1 — CLI `--thinking xhigh` reaches Anthropic

Artifact 2 — Persisted `openai.thinking` does NOT leak to Anthropic