fix(session): drop oldest turns when compaction head exceeds model context - #51121

Thanks for your contribution!

This PR doesn't have a linked issue. All PRs must reference an existing issue.

Please:

1. Open an issue describing the bug/feature (if one doesn't exist)
2. Add Fixes #<number> or Closes #<number> to this PR description

See CONTRIBUTING.md for details.

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

Thanks for updating your PR! It now meets our contributing guidelines. 👍

Until this PR (or some other fix) is merged, a simple workaround can help: select a different model which has a larger context window than the current one. Then compression will work. After the successful compression, the original model can be used again to continue the work.

Reviewed against a live reproduction on opencode 1.18.32 (stable). Environment: provider ollama-cloud (https://ollama.com/v1), model deepseek-v4.1-flash (1,048,576 context / 393,216 max output).

Reproduction. Auto-compaction fired at a 666,763-token live context and the summarizer call was rejected: The prompt is too long: 1122556, model maximum context length: 1048576. The session then showed Session too large to compact and was unusable.

Mechanism (both steps verified).

1. The provider admits a request on input tokens only; the output budget is not counted against the window. Probed directly: prompt 800,031 + max_tokens 393,216 → accepted; prompt 1,050,041 + max_tokens 1 → rejected (The prompt is too long: 1050041); max_tokens 2,000,000 → separately rejected (exceeds maximum output tokens). So compaction.reserved and max_tokens do not affect admission.
2. The head is effectively the whole history. I reconstructed the full serialized head with the exact serialize() rules and measured it through the real provider: 1,128,448 input tokens, vs the logged crash prompt 1,122,556 (0.5% apart — the difference is the prompt template + prior summary). select() budgets only the retained tail; the default tail is tiny (min(15_000, …)), so with the newest turn (~43k) above it the head is everything older. That is the missing budget this PR adds.

The approach is right. Dropping oldest complete turns bounds the head. On my session the head was ~1.03M tokens while the newest single turn was only ~43k, so the head, not the tail, is where the budget was missing.

Three gaps worth closing before merge:

1. Irreducible head. The loop stops at i < all.length - 1 and returns input.messages unchanged when start === 0. A single-turn head (or a newest turn larger than the budget) is returned oversized and the provider still rejects it — the same wedge survives. Suggest splitting the newest turn via splitTurn, or allowing an empty head (summary-only) with a clear log.
2. Budget under-count. estimate() measures toModelMessagesEffect(messages), but the request is built from serialize() text (all reasoning included, tool output capped at 2,000 chars) plus previousSummary and plugin context. SUMMARIZER_PROMPT_RESERVE = 8_000 will not absorb a large previousSummary. Suggest estimating the actual built nextPrompt and iterating until estimate(builtPrompt) <= usable − summaryOutput.
3. Cost. Each iteration re-estimates a fresh slice (O(turns²)); pre-computing per-turn sizes once makes it linear.

Style nit per CONTRIBUTING: prefer an immutable local over let.

Settings that mitigate until this lands (opencode.json): set limit.input so compaction.reserved is honored; keep reserved large (it fires compaction earlier, when the serialized head is smaller); keep compaction.prune: true; and raise compaction.preserve_recent_tokens so a larger tail is retained verbatim and the summarized head shrinks. On my session preserve_recent_tokens: 500000 brought the reconstructed head from ~1,128,448 down to 634,064 real tokens — comfortably under the input window. (Turning compaction onto a larger-context model, as @stweil noted, also works.) This is a stopgap: preserve_recent_tokens caps the tail, so as the session grows the head grows again — only dropping oldest turns (this PR) bounds it permanently.

Assisted-by: opencode ollama-cloud/deepseek-v4.1-flash (analysed on a live system)

This branch has been upgraded to V2. We're no longer taking PRs for V1. If you think this is still relevant, please port it over and open it for V2.

— from 𝕺𝖕𝖊𝖓𝕮𝖔𝖉𝖊