core: compaction trusts a measured prompt from a different model behind the same provider #61372

Activity

1. opencodejr commented on Nov 3, 2026

   

   opencodejrBot

   on Nov 3, 2026

   Contributor

   More actions

   Thanks for the precise report. I reproduced this on the latest dev (3ba597f). With your assertion added next to the other-provider case in session-compaction.test.ts, SessionCompaction.estimateContext returns 89000 instead of 21. So a token count measured for the previous model behind the same provider is still used as the current prompt size. due decides automatic compaction on that estimate, while measuredPrompt refuses the same count because the model differs.

   The script below adds the assertion, runs the test and reverts the file.

   Reproduction script (repro.sh)

   #!/usr/bin/env bash
   # Reproduces #61372 against an opencode checkout (default /workspace/repo, dev branch, deps installed).
   set -euo pipefail
   REPO="${1:-/workspace/repo}"
   cd "$REPO/packages/core"
   python3 - <<'PY'
   p = "test/session-compaction.test.ts"
   s = open(p).read()
   a = '    expect(SessionCompaction.estimateContext({ ...grown, messages: [foreign] })).toBe(21)\n'
   b = a + ('    const swapped = { ...assistant, model: { ...assistant.model, id: Model.ID.make("other-model") } }\n'
            '    expect(SessionCompaction.estimateContext({ ...grown, messages: [swapped] })).toBe(21)\n')
   assert a in s, "anchor assertion not found"
   if "other-model" not in s:
       open(p, "w").write(s.replace(a, b))
   PY
   trap 'git checkout -- test/session-compaction.test.ts' EXIT
   # Expected: 21 (same as the other-provider case); actual: 89000 (previous model's count reused)
   bun run script/test.ts test/session-compaction.test.ts -t "auto compaction estimates current content"

   Run it: bash repro.sh
2. opencodejr

   added

   reproduced

   on Nov 3, 2026
3. jimmyallinai commented on Nov 3, 2026

   

   jimmyallinai

   on Nov 3, 2026

   Contributor

   More actions

   I'll take a look at this one. Plan: first reproduce with the swapped-model assertion in packages/core/test/session-compaction.test.ts, then align the estimatePrompt anchor check with the stricter measuredPrompt criteria (model id, variant, agent) so a count measured for a different model behind the same provider is no longer trusted, and run the session-compaction tests plus the neighbouring core tests and typecheck before opening a PR.
4. r266-tech commented on Nov 3, 2026

   

   r266-tech

   on Nov 3, 2026

   Contributor

   More actions

   I have an implementation in #61379 and updated it to match usage by provider, model, variant, and agent, including a native-window regression. This covers the anchor cases from your plan, so we can avoid parallel implementation. The visible-text estimate may still undercount dense content until the provider reports overflow.