fix(compaction): mid-loop trigger uses force=true; guard slice underflow

[jonnyparris]

What

Two small follow-ups uncovered by deploying #73 and watching wrangler tail on a real failed session.

1. Mid-loop trigger now passes force: true

The mid-loop compaction trigger in onChatMessage()'s own-loop was the only one of the three triggers (loop-entry, mid-loop, pre-step) not passing { force: true } to maybeCompactContext(). Because maybeCompactContext early-returns when lastInputTokens === 0 (no persisted assistant message yet) unless force is set, the mid-loop call was a silent no-op on the very first user turn — exactly the scenario the own-loop budget check was designed for. The 'mid-loop compaction triggered' log line still fired so the failure was invisible.

2. Slice underflow guard in maybeCompactContext

maybeCompactContext built messagesToCompact via realMessages.slice(0, compactCount) and then indexed messagesToCompact[compactCount - 1], which throws TypeError when compactCount > realMessages.length (e.g. force: true on a turn with only 1 persisted message). Now guarded with an early return + log line, and the toId lookup uses messagesToCompact.length - 1.

Why

Reproduction (Gemma 4 26B on prompt 'Check whether the Dodo auto-nudge feature is working as intended'):

own-loop: step complete  step=0  budgetUsage=10%
own-loop: step complete  step=1  budgetUsage=19%
own-loop: step complete  step=2  budgetUsage=29%
own-loop: step complete  step=3  budgetUsage=38%
own-loop: step complete  step=4  budgetUsage=48%
own-loop: step complete  step=5  budgetUsage=58%
own-loop: mid-loop compaction triggered  step=6  cumulative=118659
…end of stream, compactionCount=0, totalTokenInput=197535

The log line lied. maybeCompactContext() returned without compacting because lastInputTokens === 0 (first turn, no persisted assistant yet) and force wasn't passed.

Caveat

This does not fully unblock weak orchestrator models on the first turn. By design, compaction needs >=2 real persisted messages to have something meaningful to summarise. The deeper issue (single assistant turn ballooning to 200k tokens before any tool result is persisted, because Think persists only after streamText() completes the whole generator) is captured in memory/learnings/dodo-orchestrator-model-matters.md as a future investigation.

Even with this fix, expect Gemma 4 26B as orchestrator to keep failing — the underlying issue is the model not the safety nets. PR is still worth landing because the silent-noop is misleading and the slice-underflow is a real crash risk under force: true.

Tests

• All 21 existing compaction tests pass (compaction-policy-unit, compaction-pipeline, compaction-e2e)
• Typecheck clean

Linked context

• PR fix(compaction): force/truncated bypass the minMessages guard #73 (force/truncated bypass minMessages) — landed on main
• PR fix(explore): attach recovery hint to failure summaries #72 (explore failure hint) — open
• PR feat(ui): surface skills and tool catalog in sidebar #71 (skills/tools UI surfacing) — open

beep-boop-🤖