compaction: misleading error when compaction summary exceeds the model's context window #50187

Activity

1. pm25coder commented on Sep 21, 2026

   

   pm25coder

   on Sep 21, 2026

   More actions

   Confirmed and extended — I traced both gates in packages/core/src/session/compaction.ts on main and ran the arithmetic locally.

   Both gates size against the configured limit, so neither can fire here

   compactAfterOverflow takes the window from the configured limits only (compaction.ts:179) and then guards the prompt it just built:

   if (Token.estimate(summaryPrompt) > context - summaryOutput) return false   // compaction.ts:190

   compactIfNeeded reads the same number (compaction.ts:234) before handing off (compaction.ts:242). With your config — limit.context: 65536, an LM Studio slot of 8192, generation.maxTokens unset — output falls back to 0 (compaction.ts:181, 236), so:

   gate	threshold	vs. the served window
   auto-compaction (line 237-239, DEFAULT_BUFFER = 20000)	45,536 est tokens ≈ 182 KB	5.6×
   summary guard (line 190, SUMMARY_OUTPUT_TOKENS = 4096)	61,440 est tokens ≈ 246 KB	7.5×

   I ported the two functions verbatim and ran them under node 22, reading the constants out of the tree rather than retyping them:

   auto gate  (compaction.ts:237-239) fires above 45536 est tokens = 182144 chars = 5.56 x the served window
   summary guard (compaction.ts:190)  admits up to 61440 est tokens = 245760 chars = 7.50 x the served window
   summary request asks the provider for maxTokens = 4096 of its 8192 token window
   -> after reserving 4096 for output, input may be about 4096 tokens = 15.00 x over the guard's own headroom
   
   auto gate needs about 171484 chars of conversation ( 5.23 x the served window in chars )
   summary guard needs about 275390 chars ( 8.40 x )
   

   So with a stale limit.context the guard never gets a chance to refuse before the provider does. Two details in that arithmetic are worth naming:

   • The summary call reserves summaryOutput = min(output || 4096, 4096) for output (compaction.ts:189, 208) — half of an 8192-token window — yet the guard subtracts it from the configured 65536. It starts from roughly 15× the input budget that actually exists.
   • Token.estimate is round(length / 4) over the JS string (util/token.ts:5), applied to the serialized transcript, not to provider tokens.

   Why the raw provider error arrives with nothing around it

   In packages/core/src/session/runner/llm.ts, recoverOverflow is wired to this very function (compaction.compactAfterOverflow, llm.ts:377), and the retry is taken only when the failure classifies as an overflow (llm.ts:292). When the summary stream itself errors, compactAfterOverflow sets failed = true (compaction.ts:213) and returns a plain false (compaction.ts:221) — after which the runner republishes the original provider error verbatim (if (overflowFailure) yield* publish(overflowFailure), llm.ts:296). The return value is a bare boolean, so no caller can tell "there was no head left to summarize" apart from "the summary request overflowed as well". That is the mechanical reason a user sees exceed_context_size_error with no explanation attached.

   The information your expected message needs is already on the event

   The repo's own classifier already matches your strings (packages/llm/src/provider-error.ts:18). I ran the pattern list from that file over the messages in your report:

   TRUE  | exceed_context_size_error: request (8588 tokens) exceeds the available context size (8192 tokens), try increasing it
   TRUE  | request (22254 tokens) exceeds the available context size (8192 tokens), try increasing it
   false | Rate limit exceeded, please retry after 30 seconds.
   

   So the overflow is recognised, the event carrying it is in hand at llm.ts:296, and its message already contains both numbers. Printing your suggested sentence needs no new plumbing — the wording path just has to stop treating this case as an output-token problem.

   Suggested order of fixes

   1. Cheapest, and it removes the misleading half of the report on its own: when the failed summary stream is an overflow, surface the provider's own numbers/limits instead of the generic message.
   2. Return a reason from compactAfterOverflow instead of false (e.g. no-head / summary-too-large / summary-failed), so the runner can say "the model's window (8192 tokens) is too small to compact this conversation — raise it or start a new session".
   3. Size the summary budget against the effective input window rather than the configured one: min(configured, observed) minus summaryOutput, and clamp summaryOutput to a fraction of that window instead of min(output || 4096, 4096) against a window it never checked. The cap quoted in this very error is enough to seed a per-model cache.

   Scope note: this is source reading plus a local harness porting the two functions and the classifier — I had no LM Studio server to drive, so treat the multiples above as gate arithmetic, not as a replay of your session.

   One loose end: I could not find that banner string anywhere. It is not in the tree on main, and it is not in the build you are running either — I unpacked opencode-windows-x64@1.14.42 (the platform package behind the opencode-ai wrapper) and searched the 158,887,816-byte binary; the compaction strings it ships are the " Compaction " / " Auto Compaction " box titles and the two ContextOverflowError messages ("Conversation history too large to compact - exceeds model context limit" and "Session too large to compact - context exceeds model limit even after stripping media"). If you still see that exact wording, it is worth pinning which build produced it — the sizing defect above is independent of it.
2. Tarquinen commented on Oct 20, 2026

   

   Tarquinen

   on Oct 20, 2026

   Member

   More actions

   Thanks for the report! OpenCode 2.0 is now out and development is focused on V2, so we're closing issues filed against OpenCode 1. Please upgrade (opencode upgrade --major, details: https://anoma.ly/notes/opencode-2-0), and if this is still relevant on 2.0, reopen this issue or open a new one with your 2.0 version.

   — from 𝕺𝖕𝖊𝖓𝕮𝖔𝖉𝖊
3. Tarquinen

   closed this as not plannedon Oct 20, 2026