fix(provider): preserve assistant message content when reasoning blocks present#21370
Open
edevil wants to merge 1 commit intoanomalyco:devfrom
Open
fix(provider): preserve assistant message content when reasoning blocks present#21370edevil wants to merge 1 commit intoanomalyco:devfrom
edevil wants to merge 1 commit intoanomalyco:devfrom
Conversation
…ks present normalizeMessages() unconditionally filtered empty text parts from all message roles, including assistant. When Anthropic adaptive thinking (Opus 4.6, Sonnet 4.6) emits an empty text part between two reasoning blocks, removing it shifts thinking block positions and invalidates the cryptographic signatures, causing the API to reject with: 'thinking blocks in the latest assistant message cannot be modified' Skip the empty-text filter for assistant messages that contain reasoning blocks. These messages must be replayed verbatim since thinking block signatures encode positional context. Assistant messages without reasoning blocks continue to be filtered normally. Closes anomalyco#16748
Contributor
|
The following comment was made by an LLM, it may be inaccurate: Potential Duplicate Found:
This PR appears to be directly related or a potential duplicate. It addresses the same issue (#16748) with the same root cause (empty-text filtering in |
Contributor
Author
|
It seems this error message predates adaptive thinking so it must be something else. |
Contributor
Author
|
I managed to confirm with a past session that this does indeed fix the issue in question. |
fairyhunter13
added a commit
to fairyhunter13/opencode
that referenced
this pull request
Apr 10, 2026
Port anomalyco/sst upstream dev's proven sync architecture to permanently
fix the flaky "blank subagent view" bug where ctrl+x down into a running
subagent showed an empty content area until completion.
sync.tsx:
- Remove `force` parameter from session.sync()
- Always mark fullSyncedSessions after fetch (drop time.completed gate)
- Remove redundant session.created handler (session.updated upserts)
- Match upstream/dev byte-for-byte in sync-relevant sections
routes/session/index.tsx:
- Task component: createEffect -> onMount with empty-messages guard
- Route-level sync: drop force=true argument
Preserved defensively (ahead of upstream):
- worker.ts + server/routes/event.ts still use GlobalBus.on("event", ...)
as defense-in-depth on top of InstanceState shared PubSub
- worker.ts WASM stdout/stderr capture to suppress Aborted() TUI corruption
- worker.ts per-directory filter in GlobalBus.on handler
provider.ts:
- Raise FLOOR_CEILINGS.critical from 4 to 10 for non-Anthropic providers
- Expand isAnthropicParent to also match parentModelID containing "claude"
(covers claude-via-proxy cases)
transform.ts:
- Preserve assistant messages with reasoning blocks verbatim to avoid
Anthropic 400 "thinking blocks cannot be modified" (ref upstream PR anomalyco#21370)
- Enable 4-BP cache strategy for GitHub Copilot Claude models
llm.ts:
- Use 5m TTL for short-lived subagent BP2 cache breakpoint; 1h for long-lived
processor.ts:
- Remove duplicate SessionSummary.summarize() call that triggered full-history
DB hydration on every LLM step in multi-step tool loops (ref upstream anomalyco#21762)
Tests:
- task-escalation.test.ts: update FLOOR_CEILINGS critical assertions to
match new ceiling (critical > opus, toBe(10)); expand coverage
- cache-max-efficiency-e2e.test.ts: new E2E test validating Anthropic prompt
cache hit rate across multi-turn conversations
Skill:
- Add opencode-prompt-cache-maximization skill (CLAUDE.md)
Submodule:
- Bump cmd/opencode-search-engine to include idle CPU perf improvements
Root cause: our prior fullSyncedSessions + time.completed guard raced
with Task.createEffect's fire-and-forget sync(id, force=false) -- first
fetch at T=0 returned 0 messages, session was never marked synced,
and no re-sync trigger existed, leaving the view blank until the
subagent finished.
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
1 participant
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.
Issue for this PR
Closes #16748
Type of change
What does this PR do?
normalizeMessages()intransform.tsunconditionally filters out empty text parts from all message roles, including assistant. When Anthropic adaptive thinking (Opus 4.6, Sonnet 4.6) emits a whitespace-only text part between two reasoning blocks,processor.tstrims it to"", and thennormalizeMessagesremoves it. This shifts the block arrangement from[thinking, text, thinking, text, tool_use]to[thinking, thinking, text, tool_use], invalidating the positionally-sensitive cryptographic signatures on the thinking blocks. The Anthropic API then rejects with: "thinking blocks in the latest assistant message cannot be modified".The fix is one line: skip the empty-text filter for assistant messages that contain reasoning blocks. These messages must be replayed verbatim since thinking block signatures encode positional context. Assistant messages without reasoning blocks (e.g. plain text responses, compaction summaries) continue to be filtered normally.
The empty-text filter was added on Jan 5 (
c285304a) before adaptive thinking existed (Feb 13,0d90a22f9), so it was correct at the time. The bug is an emergent interaction with Opus 4.6's signature-sensitive thinking blocks.How did you verify your code works?
[reasoning(sig1), text(""), reasoning(sig2), text("answer"), tool_use]— asserts all 5 parts are preservedScreenshots / recordings
N/A — backend logic change, no UI impact.
Checklist