fix(cli): trigger auto-compaction at the configured context threshold by rakshith1928 · Pull Request #13340 · Kilo-Org/kilocode

rakshith1928 · 2026-08-23T21:04:41Z

Issue

Context

Auto-compaction was firing on message send long before the configured context threshold — with threshold_percent: 80, compaction hit around 40% of the context bar, inconsistently. Two compounding causes in the preflight check:

Wrong base for the percentage. It applied threshold_percent to model.limit.input || model.limit.context. For models with a separate input limit (GPT-5 family: 272k input of a 400k window), "80%" was really 54% of the displayed bar.
Inflated estimate as the trigger value. It compared that limit against a char-based estimate of the whole payload (chars/4 × 1.3 + tool schemas), whose overshoot varies with content type — hence the randomness. Since the check defers during tool loops, it landed on the next user message: the reported "triggers on send" pattern.

Implementation

`packages/opencode/src/kilocode/session/overflow.ts`

limit() now applies threshold_percent to model.limit.context — the denominator the UI uses. Math.min(usable, cap) is unchanged, so the usable input limit remains the hard ceiling.
measure() additionally returns tail: the 1.3×-inflated estimate of messages added after the last assistant reply, i.e. the content the provider has not counted yet. Tool schemas stay out of tail since the baseline already includes them.
shouldCompact() projects the next request as reported + tail, where reported is the provider-reported count from the last finished turn. Without a baseline (first turn, post-compaction summary, zero-usage report) the full outgoing estimate decides alone. The safety factor stays everywhere — dense CJK/code still over-count rather than under-count, which matters since fix(cli): preserve output budget for encrypted reasoning #13349 normalizes encrypted reasoning out of local estimates, so neither the raw estimate nor the stale baseline alone sees reality.

On models whose input limit sits below their context window (GPT-5: 272k of 400k), the reserved input ceiling fires before a high threshold — 80% triggers around 63% of the displayed bar. That's the model's physical wall, not policy; lower thresholds stay predictable.

`packages/opencode/src/session/llm.ts`

Passes reported: input.reportedContextTokens and tail: usage.tail into the preflight decision — both values already existed on StreamInput/measure(), no new plumbing.

Tests (test/kilocode/session-overflow.test.ts) cover the false-negative regressions and the effective ceiling below.

Screenshots / Video

before	after

Config:

  "compaction": {
    "auto": true,
    "threshold_percent": 10,
    "prune": true
  },

How to Test

Manual/local verification

bun test ./test/kilocode/session-overflow.test.ts from packages/opencode/: 36 pass, including new regressions for dense CJK payloads, a large new message on top of a high reported baseline with encrypted reasoning history, the input-limited effective ceiling, and zero-usage reports falling back to the estimate path. The false-negative cases were written failing before the fix .

Reviewer test steps

Set compaction.threshold_percent: 80; use a model with an input limit below its context window (e.g. GPT-5).
Chat until the bar reads 40-60%, send a message — no compaction.
Continue past the effective limit — compaction fires on the next send (~80% where input equals context; reserved ceiling earlier on input-limited models).
Optional: paste a large file at low usage — still compacts before the request goes out.

Blocked checks and substitute verification

Checklist

Issue linked above, or exception explained
Tests/verification described
Screenshots/video included for visual changes, or marked N/A
Changeset considered for user-facing changes
I personally reviewed the diff and can explain the changes, including any AI-assisted work.