fix: preserve thinking block signatures and fix compaction headroom asymmetry by gnadaban · Pull Request #14393 · anomalyco/opencode

gnadaban · 2026-02-20T07:55:17Z

Issue for this PR

Closes #13286
Related: #10634, #8089, #12621, #8185

Type of change

Bug fix
New feature
Refactor / code improvement
Documentation

What does this PR do?

Two compounding bugs cause sessions to crash with 'thinking' or 'redacted_thinking' blocks in the latest assistant message cannot be modified when compaction fires for models with extended thinking enabled (e.g. Claude Opus on Bedrock).

Bug 1 — Thinking block signature stripping (message-v2.ts)

toModelMessages() compares the current model against each historical message's original model via a differentModel guard. When they differ, all providerMetadata is stripped from message parts. This removes the cryptographic signature that Anthropic/Bedrock requires on thinking blocks, causing the API to reject the messages.

The trigger is switching model variants mid-conversation within the same provider — e.g. changing from claude-opus-4-6 to claude-opus-4-7 on Bedrock, or any model mismatch between a message's recorded model and the current active model. When this happens, the Bedrock adapter receives a reasoningText block without a signature field and rejects it with signature: Field required.

Note: in the default config (no custom compaction model), compaction uses the same providerID + modelID as the conversation, so differentModel does not fire during compaction. The bug manifests when switching model variants mid-conversation.

The fix narrows the guard from differentModel (providerID + modelID) to differentProvider (providerID only). Provider metadata is namespaced (e.g. { bedrock: { signature: "..." } }), so passing it to a different model variant within the same provider is always safe. Stripping is still applied when crossing provider boundaries (e.g. Bedrock → OpenAI) to avoid leaking provider A's metadata to provider B.

Bug 2 — Compaction headroom asymmetry (overflow.ts)

isOverflow() had an asymmetric buffer calculation:

With limit.input: reserved = Math.min(20_000, maxOutputTokens) → only 20K headroom
Without limit.input: reserved = maxOutputTokens → full 32K headroom

This meant models with limit.input could grow ~12K tokens larger before compaction triggered, making context overflow more likely.

The fix removes the 20K cap and uses maxOutputTokens() (itself capped at 32K via OUTPUT_TOKEN_MAX) for both paths. Also fixed the non-input path to respect config.compaction.reserved — previously it hardcoded maxOutputTokens(), ignoring the user's configured override from #12924.

How did you verify your code works?

74 tests pass across 2 test files (message-v2, compaction), 0 failures
New tests cover: same-model preservation, same-provider-different-model preservation, cross-provider stripping, text/reasoning/tool parts, undefined metadata
LSP diagnostics clean on all changed files
Typecheck passes across all packages
Manually verified with lower token limits that Opus 4.6 and Sonnet 4.6 on AWS Bedrock compacts successfully

Screenshots / recordings

N/A — not a UI change.

Checklist

I have tested my changes locally
I have not included unrelated changes in this PR

github-actions · 2026-02-20T07:56:00Z

The following comment was made by an LLM, it may be inaccurate:

Potential Related PRs Found:

PR fix(provider): preserve redacted_thinking blocks and fix signature validation #12131 - "fix(provider): preserve redacted_thinking blocks and fix signature validation"
- Related to thinking block preservation and signature handling, which is a core part of the current PR
PR fix: compaction bugs #13946 and #13980 #14245 - "fix: compaction bugs Bug: opencode run exits after compaction when compaction model's token usage exceeds overflow threshold #13946 and Bug: compaction.reserved config is ignored for models without explicit 'input' limit #13980"
- Addresses compaction-related issues, overlapping with the compaction headroom asymmetry fix in the current PR
PR fix(opencode): prevent context overflow during compaction #11453 - "fix(opencode): prevent context overflow during compaction"
- Related to compaction overflow prevention, potentially overlapping with the headroom fix

These PRs address similar domains (thinking block signatures and compaction behavior), though they may be addressing different specific bugs or prior versions of the same issues.

gnadaban · 2026-02-20T07:59:41Z

Re: related PRs flagged by bot

Checked all three:

fix(provider): preserve redacted_thinking blocks and fix signature validation #12131 — Overlaps with the thinking block fix here, but tackles it differently (removes trimEnd() on thinking text + reorders blocks in normalizeMessages()). Our fix removes the differentModel guard that strips providerMetadata/signatures. Both address the same symptom from complementary angles — possible merge conflict in message-v2.ts but the changes are compatible.
fix: compaction bugs #13946 and #13980 #14245 — Their fix for Bug: compaction.reserved config is ignored for models without explicit 'input' limit #13980 (context - reserved in the non-input fallback path) is identical to what this PR does on that same line. Whichever merges first makes the other a no-op for that specific change.
fix(opencode): prevent context overflow during compaction #11453 — No overlap. Adds truncateForCompaction() for when compaction itself overflows — complementary to the headroom fix here.

…rategy UI Root cause fix (from PR anomalyco#14393): - Always pass providerMetadata for reasoning parts (removed differentModel guard) - Always pass callProviderMetadata for tool parts - Fix asymmetric compaction buffer (use maxOutputTokens consistently) Configurable thinking strategy (none/strip/compact): - Settings > General: Thinking Strategy dropdown - Context tab: Always-visible strategy selector - Error card: Retry buttons for thinking block errors - Processor: Auto-compact on thinking error with compact strategy Default 'none' preserves original behavior.

Dulani · 2026-04-17T08:28:08Z

Confirming this bug is reproducible with Claude Opus 4.7 (released April 15, 2026) on Amazon Bedrock.

Error appears at message 63:

undefined: The model returned the following errors: 
{"type":"error","error":{"type":"invalid_request_error",
"message":"messages.63.content.0.thinking.signature: Field required"}}

Environment:

Model: anthropic.claude-opus-4-7 via Amazon Bedrock
OpenCode version: 1.4.7
Extended thinking: enabled
Trigger: Multi-turn conversation with thinking blocks from earlier turns

The model works fine in AWS Console (fresh requests), but fails in OpenCode when replaying conversation history with cached thinking blocks.

This PR's fix (removing the differentModel guard that strips providerMetadata) appears to be the correct approach. Happy to help test or debug the CI failures if needed.

gnadaban · 2026-04-17T12:18:20Z

Just rebased, there were hundreds of commits apparently on dev.
Once confirmed it still works as intended, this should be reviewed and merged ASAP, it's needed by people.

rekram1-node · 2026-04-17T14:55:25Z

talking to someone on discord about this, not sure if it is you in discord but if not ill move discussion here.

rekram1-node · 2026-04-17T15:32:11Z

    }

    if (msg.info.role === "assistant") {
-      const differentModel = `${model.providerID}/${model.id}` !== `${msg.info.providerID}/${msg.info.modelID}`


Killing this logic will break switching models mid convo for a variety of cases, we cannot do this, also your PR says compaction uses a different model? Well that'd only happen if you manually specify a different model and even then, this logic should prevent a separate model from causing issues

Good point, what do you think of the revised change?

…symmetry Two compounding bugs caused sessions to crash with 'thinking blocks cannot be modified' when compaction fired for models with extended thinking: 1. toModelMessages() stripped providerMetadata (including cryptographic signatures) from message parts when the current model differed from the original. Anthropic's API requires signatures to be byte-identical. Fix: always pass providerMetadata through — the API handles filtering. 2. isOverflow() used an asymmetric buffer when limit.input was set (capped at 20K via COMPACTION_BUFFER) vs the full maxOutputTokens on the non-input path. This caused compaction to trigger too late. Fix: use maxOutputTokens (capped at 32K) for both paths. Also fixed the non-input path to respect config.compaction.reserved.

This was referenced Feb 20, 2026

fix(core): repair interleaved text/tool-call parts in assistant messages #14456

Closed

feat: Switch compaction model from session model #14665

Closed

gnadaban force-pushed the bugfix/fix-thinking-block-compaction branch 2 times, most recently from fd795c0 to c045a8f Compare February 24, 2026 12:17

gnadaban force-pushed the bugfix/fix-thinking-block-compaction branch from c045a8f to 04021bb Compare March 2, 2026 12:20

renbou mentioned this pull request Mar 2, 2026

fix(provider): preserve redacted_thinking blocks and fix signature validation #12131

Closed

This was referenced Mar 9, 2026

normalizeMessages() removes empty text parts between reasoning blocks, invalidating Anthropic thinking block signatures #16748

Open

fix(provider): skip empty-text filtering for assistant messages in normalizeMessages (#16748) #16750

Draft

gnadaban force-pushed the bugfix/fix-thinking-block-compaction branch 2 times, most recently from b898bf0 to cef2b52 Compare March 11, 2026 09:52

github-actions Bot mentioned this pull request Mar 11, 2026

fix(provider): resolve bedrock compaction failures for Claude models #17043

Closed

gnadaban force-pushed the bugfix/fix-thinking-block-compaction branch from cef2b52 to a49278b Compare March 16, 2026 08:38

gnadaban force-pushed the bugfix/fix-thinking-block-compaction branch 2 times, most recently from 0bf6a56 to be50ae3 Compare April 7, 2026 12:06

github-actions Bot mentioned this pull request Apr 8, 2026

fix(opencode): preserve tool context through compaction and prompt loops #21425

Closed

6 tasks

chan1103 mentioned this pull request Apr 10, 2026

fix: replace empty text in reasoning messages to preserve thinking block positions #21860

Open

3 tasks

github-actions Bot mentioned this pull request Apr 16, 2026

fix(opencode): preserve reasoning providerMetadata across model switches #22818

Closed

opencode-agent Bot force-pushed the dev branch from 2f2ed11 to 9f201d6 Compare April 16, 2026 21:55

gnadaban force-pushed the bugfix/fix-thinking-block-compaction branch from be50ae3 to 8f53ae2 Compare April 17, 2026 12:16

gnadaban force-pushed the bugfix/fix-thinking-block-compaction branch from 8f53ae2 to bb23cf6 Compare April 17, 2026 14:23

rekram1-node reviewed Apr 17, 2026

View reviewed changes

github-actions Bot mentioned this pull request Apr 17, 2026

fix(opencode): preserve reasoning providerMetadata across model switches #23104

Draft

3 tasks

gnadaban force-pushed the bugfix/fix-thinking-block-compaction branch from bb23cf6 to 626e8a5 Compare April 20, 2026 09:02

gnadaban requested a review from rekram1-node April 20, 2026 09:31

github-actions Bot mentioned this pull request Apr 20, 2026

fix: preserve thinking/redacted_thinking blocks in Anthropic message transforms #23574

Closed

6 tasks

gnadaban force-pushed the bugfix/fix-thinking-block-compaction branch 2 times, most recently from c8af25d to 07ba2f7 Compare April 22, 2026 10:14

rekram1-node force-pushed the dev branch from 1026791 to ac26394 Compare April 23, 2026 04:27

rekram1-node requested a review from adamdotdevin as a code owner April 23, 2026 04:27

PershingSquare mentioned this pull request Apr 23, 2026

Bedrock Claude Opus 4.7 replays thinking blocks through interleaved OpenAI-compatible transform #23926

Open

rekram1-node and others added 2 commits April 23, 2026 14:43

feat: update codex plugin to support 5.5 (anomalyco#23789)

745cc65

gnadaban force-pushed the bugfix/fix-thinking-block-compaction branch from 07ba2f7 to 611ebdb Compare April 23, 2026 12:43

github-actions Bot mentioned this pull request Apr 29, 2026

fix: preserve Anthropic thinking block signatures when switching Claude models #25003

Closed

3 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

fix: preserve thinking block signatures and fix compaction headroom asymmetry#14393

fix: preserve thinking block signatures and fix compaction headroom asymmetry#14393
gnadaban wants to merge 2 commits intoanomalyco:devfrom
gnadaban:bugfix/fix-thinking-block-compaction

gnadaban commented Feb 20, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Feb 20, 2026

Uh oh!

gnadaban commented Feb 20, 2026

Uh oh!

Dulani commented Apr 17, 2026

Uh oh!

gnadaban commented Apr 17, 2026

Uh oh!

rekram1-node commented Apr 17, 2026

Uh oh!

rekram1-node Apr 17, 2026

Uh oh!

gnadaban Apr 20, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

gnadaban commented Feb 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Issue for this PR

Type of change

What does this PR do?

How did you verify your code works?

Screenshots / recordings

Checklist

Uh oh!

github-actions Bot commented Feb 20, 2026

Potential Related PRs Found:

Uh oh!

gnadaban commented Feb 20, 2026

Uh oh!

Dulani commented Apr 17, 2026

Uh oh!

gnadaban commented Apr 17, 2026

Uh oh!

rekram1-node commented Apr 17, 2026

Uh oh!

rekram1-node Apr 17, 2026

Choose a reason for hiding this comment

Uh oh!

gnadaban Apr 20, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

gnadaban commented Feb 20, 2026 •

edited

Loading