compaction.buffer: flat token count can't serve mixed model windows (no percentage or per-model form) #59994

What

compaction.buffer accepts only an absolute token count (NonNegativeInt). Its default is proportional — limit.input - max(15% of limit.input, 16_000) — but the moment buffer is set explicitly that proportional behaviour is discarded and one absolute number applies to every model in the config.

Why that's a problem

A single role commonly mixes model windows of very different sizes (e.g. a 131K local model and a 1M cloud model in the same opencode.jsonc). One flat reserve cannot serve both:

A buffer large enough to protect a slow local model's summarization request pushes that model down to ~52% of its window; a buffer small enough to be negligible for the cloud model leaves the local summarizer no headroom. There is no value that is correct for both.

Evidence from a real deployment

Raising buffer 10,000 → 30,000 fixed timeouts on one single-GPU local model. Because the key is global, it silently moved every model's trigger on that role: the local 131K model dropped from ~84% of its usable input to ~52% of its window. Nothing in the config said 52% — the flat buffer did. Diagnosing it required reading calculateCeiling() and the config backup history.

Related: a flat buffer also interacts badly with a per-model limit.input pin, because the trigger is limit.input - buffer while the UI gauge divides by limit.context. Two models with identical buffer values show completely different "percent used" at the moment they compact.

Proposal

Any one of these would remove the footgun:

1. Accept a percentage, applied per model against limit.input:

   "compaction": { "buffer": "10%" }

2. Per-model override, mirroring the existing limit block:

   "provider": { "x": { "models": { "m": {
     "limit": { ... },
     "compaction": { "buffer": 10000 }
   } } } }

3. At minimum, warn when buffer is set on a config whose models span very different windows. Today an unknown key under compaction is silently dropped, so a typo'd percentage key looks configured and does nothing.

Repro

{
  "compaction": { "auto": true, "buffer": 30000, "keep": { "tokens": 15000 } },
  "provider": {
    "local": { "models": { "small": { "limit": { "context": 131072,  "input": 98304,  "output": 32768  } } } },
    "cloud": { "models": { "large": { "limit": { "context": 1000000, "input": 980000, "output": 131072 } } } }
  }
}

Compaction then fires at 68,304 for small (52% of its window) and 950,000 for large (95% of its window).

Version

v2.1.4 (also present on dev): packages/schema/src/config/compaction.ts declares buffer: NonNegativeInt; RESERVE_SHARE = 0.15 and RESERVE_MIN = 16_000 are module constants in packages/core/src/session/compaction.ts, not config keys