fix(agent): compaction must not hoist an answered steer to the newest-input slot

[WhichPaths]

The bug

Stage-2 compaction partitioned history by item type: every non-tool item sitting after the first tool item went into a tailNonTool bucket that both rebuild paths re-appended after all surviving pairs.

const newHistory = [
  ...stage1History.slice(0, leadingNonToolEnd),
  summaryMarker,
  ...survivingTailPairs,
  ...tailNonTool,        // <- everything non-tool, moved to the end
]

So compaction reordered, not just deleted. That matters more than it looks, because of two things that are true together:

• The only non-tool items that ever appear mid-history are role:'user' items the runtime injects itself — the mid-turn steer render, the status-required nudge, and the completion-rejection nudge.
• There is no assistant message in history for the model to see its own reply in. Assistant text is deliberately never added as a history item (synthesizing one with no item_id / part_id / responseId corrupts history shape — the comment in turn.ts says so and names the earlier bug).

Put together: the only in-context evidence that a steer was already handled is that the function_call / function_call_output pairs answering it sit after it. The tail-move inverted exactly that.

Reproduced

Against the real module, on the history shape turn.ts builds:

before: [0] user  book the team sync for 4pm and send the agenda
        c1 c2 c3 pairs  (create 4pm, draft agenda, reply "booked 4pm")
        [7] user  [Mid-turn update — new message from Alice] actually make it 5pm
        c4 c5 pairs     (move to 5pm, reply "moved to 5pm")

after:  [0] user  book the team sync for 4pm …
        [1] user  [auto-compaction · 3 earlier pairs summarized]
        [2..5]    c4 c5          ← the answer, now BEFORE the question
        [6] user  [Mid-turn update] actually make it 5pm     ← LAST ITEM

The steer's own text is written as an imperative fresh arrival — "Re-assess your current plan in light of these and continue" — so sitting in the newest-input slot it reads as a live instruction. The agent moves the event a second time and posts a second reply. turn.ts assigns the compacted array back to history, so it re-asks on every remaining hop, not once.

Anything side-effecting the steer asked for — send an email, post to a channel, move a calendar event — gets done twice. The same shape applies to the status-required and completion-rejection nudges: a satisfied nudge re-presented last is a live instruction again.

How often: it needs a mid-turn steer, at least one more hop of work after it, and stage-2 compaction firing (past 75% of the context window, with stage-1 truncation not enough on its own). Steering is a headline feature and long tool-heavy turns are exactly the ones that cross 75%, so this is a recurring bug on busy agents rather than a corner case. It cannot happen when compaction stops at stage 1.

The fix

Rebuild in original order. partitionHistory no longer collects a tail bucket — it exists only to decide which call_id groups get dropped. Both rebuild sites now walk the original array and skip the dropped groups:

function tailWithoutDroppedGroups(history, leadingNonToolEnd, droppedCallIds) {
  const kept = []
  for (let i = leadingNonToolEnd; i < history.length; i++) {
    const item = history[i]
    const key = groupKeyOf(item)
    if (key !== null && droppedCallIds.has(key)) continue
    kept.push(item)
  }
  return kept
}

Surviving items keep their positions and their relative order. Documented as invariant 4 in the module header: compaction only deletes, never reorders.

This also stops compaction re-grouping interleaved items. One hop pushes fc1, fc2, out1, out2; that is the shape every uncompacted hop already puts on the wire, so handing it back unchanged is strictly closer to known-good than the fc1,out1,fc2,out2 the bucket rebuild produced. The existing invariants still hold — a function_call still immediately precedes its own function_call_output, and the leading seed messages are untouched.

Tests

Three added, no existing test changed — 100 lines added, 0 removed:

• stage 2 (drop path): an ANSWERED mid-turn steer keeps its position — never hoisted to the newest-input slot
• stage 2 (summary path): an ANSWERED mid-turn steer keeps its position
• stage 2: surviving tool items keep their original relative order (no regrouping by call_id)

against main:   not ok 41, not ok 42, not ok 43        # pass 40  # fail 3
with the fix:                                          # pass 43  # fail 0

The 40 that were already there stay green, including the one that pins how a malformed no-call_id item is handled — that item's original position happens to be after the surviving pairs, so order-preserving and tail-moving agree there.

Also green: tsc --noEmit, biome lint ., all three scripts/guard-*.mjs, and the unit suite (1399 tests, 0 failures).