compaction.reserved silently ignored for models without limit.input (e.g. GLM-5.2) — revival of #13980

[winstern1998-commits]

Summary

compaction.reserved is silently ignored for any model whose limit lacks an input field. This was reported in #13980, which was closed by the stale bot without being fixed — the current code still ignores reserved on this code path. Filing this to revive it with a fuller analysis and to tie together several related reports.

Environment

• opencode: 1.17.14 (source pinned to commit 203885d24; overflow.ts identical on current HEAD)
• Model: zhipuai/glm-5.2 (limit = { context: 1000000, output: 131072 }, no input)

Reproduction

Config:

"compaction": { "auto": true, "reserved": 648928 }

Use a built-in model that has no limit.input (e.g. zhipuai/glm-5.2). Observe that compaction does not trigger at input - reserved; it triggers at context - maxOutputTokens = 1000000 - 32000 = 968000. The reserved value has zero effect.

Root cause

packages/opencode/src/session/overflow.ts — usable() has two mutually exclusive branches, and reserved only participates in one:

const reserved =
  input.cfg.compaction?.reserved ??
  Math.min(COMPACTION_BUFFER, ProviderTransform.maxOutputTokens(input.model, input.outputTokenMax))
return input.model.limit.input
  ? Math.max(0, input.model.limit.input - reserved)          // branch A: reserved applies
  : Math.max(0, context - ProviderTransform.maxOutputTokens(input.model, input.outputTokenMax))  // branch B: reserved ignored

Models without limit.input (which is common — most models.dev entries only declare context + output) fall into branch B, where reserved is computed but never used. maxOutputTokens = min(limit.output, OUTPUT_TOKEN_MAX=32000) (from provider/transform.ts), so for glm-5.2 the threshold is 1000000 - 32000 = 968000 regardless of reserved.

The config schema (packages/core/src/v1/config/provider.ts) does support limit.input on custom providers, but there is no way to override limit.input for a built-in provider's model — a custom provider with the same name as a built-in does not replace the built-in entry. So users of built-in providers cannot work around this in config.

Expected vs actual

	Value
Expected threshold	limit.input - reserved (reserved takes effect)
Actual threshold	context - maxOutputTokens = 968000 (reserved silently ignored)

Impact

Silent failure: the user sets reserved, gets no error, but the setting never takes effect. This is a UX trap — there's no signal that the config is being ignored for their model.

Proposed fix

Make reserved apply in branch B as well:

: Math.max(0, context - ProviderTransform.maxOutputTokens(input.model, input.outputTokenMax) - reserved)

This makes reserved a consistent "trigger N tokens earlier" knob regardless of whether limit.input is declared. Default reserved is min(COMPACTION_BUFFER=20000, maxOutputTokens), so the default threshold for branch-B models shifts earlier by ~20K — safer (triggers before the hard limit) and consistent with branch A's behavior.

Open semantic question for maintainers: should reserved be subtracted from limit.input (branch A) or from context - maxOutputTokens (branch B)? These differ when limit.input != context - maxOutputTokens. The two branches should agree on the reference point.

Related issues

• Bug: compaction.reserved config is ignored for models without explicit 'input' limit #13980 — the original report of this exact bug; closed by stale bot, unfixed. This issue revives it.
• fix(compaction): output-budget reservation capped at 20K for limit.input models, risking overflow #32656 (open) — sibling problem in the same file, branch A: reserved capped at 20K via min(COMPACTION_BUFFER, …), below real max output. Same class of "reserved not behaving as expected" bug.
• [BUG]: Compaction overflow threshold too low when limit.input is set; compaction LLM overflow causes retry loop #30805 / PR fix: compaction context loss, prune re-attach, overflow threshold, missing config key #31891 (closed/merged) — fixed the limit.input branch only; the no-limit.input branch (this issue) was missed.
• opencode-go/glm-5.2 routed to 262K backend (DigitalOcean) despite 1M advertised context, overflow not classified #36481 (open) — GLM-5.2 specifically, but about backend routing / 262K cap rather than reserved being ignored. Cross-referenced as same-model.
• context % under-reports for split-window models (divides by limit.context instead of limit.input) #32119 (open) — limit.input vs limit.context semantics; relevant to the open question above.
• Context bloat occurs severely. I have configured automatic session compression and reserved tokens, yet automatic compression never triggers normally. This issue exists in every version of OpenCode. Please fix this problem as soon as possible. #24683 (open) — same end-user symptom (reserved set, compaction never triggers).

Complementary upstream

The models.dev registry also omits limit.input for these models (systematic across the zhipuai provider). I'll file that separately against models.dev. But the opencode-side fix is needed regardless, since reserved should not be silently dropped for any model.