bug(runtime): safe-boundary continuation silently drops prior-turn history from the provider request #5907

What happened

A long session whose provider-reported context usage had exceeded the model's reported window was interrupted mid-turn (stop button) and later resumed through the safe-boundary continuation path. The continuation replayed the interrupted turn's full event prefix correctly, but the rebuilt provider request silently excluded all prior completed-turn history — no compaction checkpoint, no summarization call, no projection transitions, and no user-facing indication. The model continues the turn with amnesia of everything that happened before the interrupted turn.

Expected: the continued turn's provider request contains the full session history (as the pre-abort requests did) — or, if reduction at this seam is intended, it goes through a visible, durable compaction (summary + checkpoint), never a silent drop.

Evidence from ~/.config/Maka/workspaces/default/runtime.sqlite:

Provider request sizes (usage_model_call_attempts)

304,452 − 81,600 = 222,852 ≈ the runtime's own estimate of the prior-turn history alone: contextBudget.estimatedTokens = 230,549 over keptEvents = 1007, keptTurns = 2 (runtime_fact recorded at abort time). The continued turn's requests never re-include the prior history across 10+ subsequent steps.

Turn extents (runtime_session_turn_extents)

Every legitimate reduction channel is ruled out: no summarization call (every post-resume attempt is call_kind = main), no history-compact checkpoint events in runtime_events, and no new model_projection_transition records (only 2 exist, both from the first turn, matching its prunedToolResults: 2). The durable ledger is intact — the 1,007 events exist in runtime_events; they simply never enter the continued turn's provider requests.

How to reproduce

1. Build a session with substantial completed-turn history (observed: ~230k estimated tokens across 2 turns).
2. Start a new turn and let it run until provider-reported input approaches or exceeds the model's context window (observed: 304k on a 262k-window model).
3. Abort the turn mid-flight (stop button).
4. Resume the session so the runtime continues the interrupted turn via safe-boundary continuation (invocation_opened with source.kind = continuation).
5. Compare input_tokens of the last pre-abort attempt with the first post-resume attempt in usage_model_call_attempts; the drop equals the size of the prior-turn history.

Observed once in the wild on a real session; not yet reduced to a scripted minimal reproduction.

Environment

• Maka version or commit: 1e80e3b885 (runtime-host running from source checkout)
• OS and version: Linux x64
• Surface: TUI + Runtime Host
• Node.js version: 26.3.0
• Connection/model: kimi-coding-plan / k3-256k (fetched contextWindow 262,144); modelOverrides empty, so no compactionThreshold is declared and proactive history compaction is not armed — which rules out compaction as the cause of the drop.

Logs, screenshots, or additional context

• Same resume area as bug(runtime): safe-boundary resume rejects Code Mode exec calls #5876 (safe-boundary resume rejects Code Mode exec calls).
• Hit while dogfooding feat(desktop): offer authoritative turn resume from the composer send slot #5903 (authoritative turn resume from the composer send slot): the session was developing that feature, and the interrupted turn was later resumed in the TUI through the safe-boundary continuation path.
• User-visible symptom: the TUI ctx segment follows the latest settled request's inputTokens, so it falls from 262k/262k 100% to ~82k/262k 31% right after resume — looking exactly like a successful "compression" that never happened.
• Reporter-local forensics ids: session cb1d809b-4cf5-473e-811f-4802c2c19d46, source turn 582a5323-0e80-4876-a3be-45276ead74d4, continuation turn 093a6dd1-0720-441e-8774-bb014590baef, continuation invocation edf3ca60-610c-4bb1-bbd2-a50dab41fe87.

Drafted with AI-agent assistance (Maka) from local runtime forensics; reviewed and submitted by @me2seeks.