Skip to content

Commit 8f53ae2

Browse files
gnadabangnadaban
committed
fix: preserve thinking block signatures and fix compaction headroom asymmetry
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.
1 parent c57c531 commit 8f53ae2

4 files changed

Lines changed: 202 additions & 29 deletions

File tree

‎packages/opencode/src/session/message-v2.ts‎

Lines changed: 6 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -689,7 +689,6 @@ export const toModelMessagesEffect = Effect.fnUntraced(function* (
689689
}
690690

691691
if (msg.info.role === "assistant") {
692-
const differentModel = `${model.providerID}/${model.id}` !== `${msg.info.providerID}/${msg.info.modelID}`
693692
const media: Array<{ mime: string; url: string }> = []
694693

695694
if (
@@ -711,7 +710,7 @@ export const toModelMessagesEffect = Effect.fnUntraced(function* (
711710
assistantMessage.parts.push({
712711
type: "text",
713712
text: part.text,
714-
...(differentModel ? {} : { providerMetadata: part.metadata }),
713+
providerMetadata: part.metadata,
715714
})
716715
if (part.type === "step-start")
717716
assistantMessage.parts.push({
@@ -747,7 +746,7 @@ export const toModelMessagesEffect = Effect.fnUntraced(function* (
747746
input: part.state.input,
748747
output,
749748
...(part.metadata?.providerExecuted ? { providerExecuted: true } : {}),
750-
...(differentModel ? {} : { callProviderMetadata: providerMeta(part.metadata) }),
749+
callProviderMetadata: providerMeta(part.metadata),
751750
})
752751
}
753752
if (part.state.status === "error") {
@@ -760,7 +759,7 @@ export const toModelMessagesEffect = Effect.fnUntraced(function* (
760759
input: part.state.input,
761760
output,
762761
...(part.metadata?.providerExecuted ? { providerExecuted: true } : {}),
763-
...(differentModel ? {} : { callProviderMetadata: providerMeta(part.metadata) }),
762+
callProviderMetadata: providerMeta(part.metadata),
764763
})
765764
} else {
766765
assistantMessage.parts.push({
@@ -770,7 +769,7 @@ export const toModelMessagesEffect = Effect.fnUntraced(function* (
770769
input: part.state.input,
771770
errorText: part.state.error,
772771
...(part.metadata?.providerExecuted ? { providerExecuted: true } : {}),
773-
...(differentModel ? {} : { callProviderMetadata: providerMeta(part.metadata) }),
772+
callProviderMetadata: providerMeta(part.metadata),
774773
})
775774
}
776775
}
@@ -784,14 +783,14 @@ export const toModelMessagesEffect = Effect.fnUntraced(function* (
784783
input: part.state.input,
785784
errorText: "[Tool execution was interrupted]",
786785
...(part.metadata?.providerExecuted ? { providerExecuted: true } : {}),
787-
...(differentModel ? {} : { callProviderMetadata: providerMeta(part.metadata) }),
786+
callProviderMetadata: providerMeta(part.metadata),
788787
})
789788
}
790789
if (part.type === "reasoning") {
791790
assistantMessage.parts.push({
792791
type: "reasoning",
793792
text: part.text,
794-
...(differentModel ? {} : { providerMetadata: part.metadata }),
793+
providerMetadata: part.metadata,
795794
})
796795
}
797796
}

‎packages/opencode/src/session/overflow.ts‎

Lines changed: 6 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -3,8 +3,6 @@ import type { Provider } from "@/provider"
33
import { ProviderTransform } from "@/provider"
44
import type { MessageV2 } from "./message-v2"
55

6-
const COMPACTION_BUFFER = 20_000
7-
86
export function isOverflow(input: { cfg: Config.Info; tokens: MessageV2.Assistant["tokens"]; model: Provider.Model }) {
97
if (input.cfg.compaction?.auto === false) return false
108
const context = input.model.limit.context
@@ -13,10 +11,11 @@ export function isOverflow(input: { cfg: Config.Info; tokens: MessageV2.Assistan
1311
const count =
1412
input.tokens.total || input.tokens.input + input.tokens.output + input.tokens.cache.read + input.tokens.cache.write
1513

16-
const reserved =
17-
input.cfg.compaction?.reserved ?? Math.min(COMPACTION_BUFFER, ProviderTransform.maxOutputTokens(input.model))
18-
const usable = input.model.limit.input
19-
? input.model.limit.input - reserved
20-
: context - ProviderTransform.maxOutputTokens(input.model)
14+
// Reserve headroom so compaction triggers before the next turn overflows.
15+
// maxOutputTokens() is capped at 32K (OUTPUT_TOKEN_MAX) regardless of the
16+
// model's raw output limit, so this is never excessively aggressive.
17+
// Users can override via config.compaction.reserved if needed (#12924).
18+
const reserved = input.cfg.compaction?.reserved ?? ProviderTransform.maxOutputTokens(input.model)
19+
const usable = input.model.limit.input ? input.model.limit.input - reserved : context - reserved
2120
return count >= usable
2221
}

‎packages/opencode/test/session/compaction.test.ts‎

Lines changed: 11 additions & 11 deletions
Original file line numberDiff line numberDiff line change
@@ -351,20 +351,20 @@ describe("session.compaction.isOverflow", () => {
351351
),
352352
)
353353

354-
// ─── Bug reproduction tests ───────────────────────────────────────────
355-
// These tests demonstrate that when limit.input is set, isOverflow()
356-
// does not subtract any headroom for the next model response. This means
357-
// compaction only triggers AFTER we've already consumed the full input
358-
// budget, leaving zero room for the next API call's output tokens.
354+
// ─── Headroom reservation tests ──────────────────────────────────────
355+
// These tests verify that when limit.input is set, isOverflow()
356+
// correctly reserves headroom (maxOutputTokens, capped at 32K) so
357+
// compaction triggers before the next API call overflows.
359358
//
360-
// Compare: without limit.input, usable = context - output (reserves space).
361-
// With limit.input, usable = limit.input (reserves nothing).
359+
// Previously (bug), the limit.input path only subtracted a 20K buffer
360+
// while the non-input path subtracted the full maxOutputTokens — an
361+
// asymmetry that let sessions grow ~12K tokens too large before compacting.
362362
//
363363
// Related issues: #10634, #8089, #11086, #12621
364364
// Open PRs: #6875, #12924
365365

366366
it.live(
367-
"BUG: no headroom when limit.input is set — compaction should trigger near boundary but does not",
367+
"no headroom when limit.input is set — compaction should trigger near boundary",
368368
provideTmpdirInstance(() =>
369369
Effect.gen(function* () {
370370
const compact = yield* SessionCompaction.Service
@@ -390,7 +390,7 @@ describe("session.compaction.isOverflow", () => {
390390
)
391391

392392
it.live(
393-
"BUG: without limit.input, same token count correctly triggers compaction",
393+
"without limit.input, same token count correctly triggers compaction",
394394
provideTmpdirInstance(() =>
395395
Effect.gen(function* () {
396396
const compact = yield* SessionCompaction.Service
@@ -410,15 +410,15 @@ describe("session.compaction.isOverflow", () => {
410410
)
411411

412412
it.live(
413-
"BUG: asymmetry — limit.input model allows 30K more usage before compaction than equivalent model without it",
413+
"asymmetry — limit.input model does not allow more usage than equivalent model without it",
414414
provideTmpdirInstance(() =>
415415
Effect.gen(function* () {
416416
const compact = yield* SessionCompaction.Service
417417
// Two models with identical context/output limits, differing only in limit.input
418418
const withInputLimit = createModel({ context: 200_000, input: 200_000, output: 32_000 })
419419
const withoutInputLimit = createModel({ context: 200_000, output: 32_000 })
420420

421-
// 170K total tokens — well above context-output (168K) but below input limit (200K)
421+
// 181K total tokens — above usable (context - maxOutput = 168K)
422422
const tokens = { input: 166_000, output: 10_000, reasoning: 0, cache: { read: 5_000, write: 0 } }
423423

424424
const withLimit = yield* compact.isOverflow({ tokens, model: withInputLimit })

‎packages/opencode/test/session/message-v2.test.ts‎

Lines changed: 179 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -57,6 +57,17 @@ const model: Provider.Model = {
5757
release_date: "2026-01-01",
5858
}
5959

60+
const model2: Provider.Model = {
61+
...model,
62+
id: ModelID.make("other-model"),
63+
providerID: ProviderID.make("other"),
64+
api: {
65+
...model.api,
66+
id: "other-model",
67+
},
68+
name: "Other Model",
69+
}
70+
6071
function userInfo(id: string): MessageV2.User {
6172
return {
6273
id,
@@ -359,7 +370,90 @@ describe("session.message-v2.toModelMessage", () => {
359370
])
360371
})
361372

362-
test("omits provider metadata when assistant model differs", async () => {
373+
test("preserves reasoning providerMetadata when model matches", async () => {
374+
const assistantID = "m-assistant"
375+
376+
const input: MessageV2.WithParts[] = [
377+
{
378+
info: assistantInfo(assistantID, "m-parent"),
379+
parts: [
380+
{
381+
...basePart(assistantID, "a1"),
382+
type: "reasoning",
383+
text: "thinking",
384+
metadata: { openai: { signature: "sig-match" } },
385+
time: { start: 0 },
386+
},
387+
] as MessageV2.Part[],
388+
},
389+
]
390+
391+
expect(await MessageV2.toModelMessages(input, model)).toStrictEqual([
392+
{
393+
role: "assistant",
394+
content: [{ type: "reasoning", text: "thinking", providerOptions: { openai: { signature: "sig-match" } } }],
395+
},
396+
])
397+
})
398+
399+
test("preserves reasoning providerMetadata when model differs", async () => {
400+
const assistantID = "m-assistant"
401+
402+
const input: MessageV2.WithParts[] = [
403+
{
404+
info: assistantInfo(assistantID, "m-parent", undefined, {
405+
providerID: model2.providerID,
406+
modelID: model2.api.id,
407+
}),
408+
parts: [
409+
{
410+
...basePart(assistantID, "a1"),
411+
type: "reasoning",
412+
text: "thinking",
413+
metadata: { openai: { signature: "sig-different" } },
414+
time: { start: 0 },
415+
},
416+
] as MessageV2.Part[],
417+
},
418+
]
419+
420+
expect(await MessageV2.toModelMessages(input, model)).toStrictEqual([
421+
{
422+
role: "assistant",
423+
content: [{ type: "reasoning", text: "thinking", providerOptions: { openai: { signature: "sig-different" } } }],
424+
},
425+
])
426+
})
427+
428+
test("preserves text providerMetadata when model differs", async () => {
429+
const assistantID = "m-assistant"
430+
431+
const input: MessageV2.WithParts[] = [
432+
{
433+
info: assistantInfo(assistantID, "m-parent", undefined, {
434+
providerID: model2.providerID,
435+
modelID: model2.api.id,
436+
}),
437+
parts: [
438+
{
439+
...basePart(assistantID, "a1"),
440+
type: "text",
441+
text: "done",
442+
metadata: { openai: { assistant: "meta" } },
443+
},
444+
] as MessageV2.Part[],
445+
},
446+
]
447+
448+
expect(await MessageV2.toModelMessages(input, model)).toStrictEqual([
449+
{
450+
role: "assistant",
451+
content: [{ type: "text", text: "done", providerOptions: { openai: { assistant: "meta" } } }],
452+
},
453+
])
454+
})
455+
456+
test("preserves tool callProviderMetadata when model differs", async () => {
363457
const userID = "m-user"
364458
const assistantID = "m-assistant"
365459

@@ -375,16 +469,97 @@ describe("session.message-v2.toModelMessage", () => {
375469
] as MessageV2.Part[],
376470
},
377471
{
378-
info: assistantInfo(assistantID, userID, undefined, { providerID: "other", modelID: "other" }),
472+
info: assistantInfo(assistantID, userID, undefined, {
473+
providerID: model2.providerID,
474+
modelID: model2.api.id,
475+
}),
476+
parts: [
477+
{
478+
...basePart(assistantID, "a1"),
479+
type: "tool",
480+
callID: "call-1",
481+
tool: "bash",
482+
state: {
483+
status: "completed",
484+
input: { cmd: "ls" },
485+
output: "ok",
486+
title: "Bash",
487+
metadata: {},
488+
time: { start: 0, end: 1 },
489+
},
490+
metadata: { openai: { tool: "meta" } },
491+
},
492+
] as MessageV2.Part[],
493+
},
494+
]
495+
496+
expect(await MessageV2.toModelMessages(input, model)).toStrictEqual([
497+
{
498+
role: "user",
499+
content: [{ type: "text", text: "run tool" }],
500+
},
501+
{
502+
role: "assistant",
503+
content: [
504+
{
505+
type: "tool-call",
506+
toolCallId: "call-1",
507+
toolName: "bash",
508+
input: { cmd: "ls" },
509+
providerExecuted: undefined,
510+
providerOptions: { openai: { tool: "meta" } },
511+
},
512+
],
513+
},
514+
{
515+
role: "tool",
516+
content: [
517+
{
518+
type: "tool-result",
519+
toolCallId: "call-1",
520+
toolName: "bash",
521+
output: { type: "text", value: "ok" },
522+
providerOptions: { openai: { tool: "meta" } },
523+
},
524+
],
525+
},
526+
])
527+
})
528+
529+
test("handles undefined metadata gracefully", async () => {
530+
const userID = "m-user"
531+
const assistantID = "m-assistant"
532+
533+
const input: MessageV2.WithParts[] = [
534+
{
535+
info: userInfo(userID),
536+
parts: [
537+
{
538+
...basePart(userID, "u1"),
539+
type: "text",
540+
text: "run tool",
541+
},
542+
] as MessageV2.Part[],
543+
},
544+
{
545+
info: assistantInfo(assistantID, userID, undefined, {
546+
providerID: model2.providerID,
547+
modelID: model2.api.id,
548+
}),
379549
parts: [
380550
{
381551
...basePart(assistantID, "a1"),
382552
type: "text",
383553
text: "done",
384-
metadata: { openai: { assistant: "meta" } },
385554
},
386555
{
387556
...basePart(assistantID, "a2"),
557+
type: "reasoning",
558+
text: "thinking",
559+
time: { start: 0 },
560+
},
561+
{
562+
...basePart(assistantID, "a3"),
388563
type: "tool",
389564
callID: "call-1",
390565
tool: "bash",
@@ -396,7 +571,6 @@ describe("session.message-v2.toModelMessage", () => {
396571
metadata: {},
397572
time: { start: 0, end: 1 },
398573
},
399-
metadata: { openai: { tool: "meta" } },
400574
},
401575
] as MessageV2.Part[],
402576
},
@@ -411,6 +585,7 @@ describe("session.message-v2.toModelMessage", () => {
411585
role: "assistant",
412586
content: [
413587
{ type: "text", text: "done" },
588+
{ type: "reasoning", text: "thinking", providerOptions: undefined },
414589
{
415590
type: "tool-call",
416591
toolCallId: "call-1",

0 commit comments

Comments
 (0)