fix(engine): bound tool results in model context on replay; 1h cache TTL

[mgoldsborough]

Problem

A small number of long-lived conversations were driving the overwhelming majority of token cost. Root cause, traced end to end:

• Tool results are capped at MAX_TOOL_RESULT_CHARS (50K) only inside the live engine loop (engine.ts).
• The full output is persisted to tool.done.output, and the history reconstructor replays that full value verbatim (event-reconstructor.ts) with no cap.
• So on every subsequent run, the model is re-fed the entire payload — more than it even saw live. Long threads balloon toward the 1M context ceiling, and with prompt caching the giant prefix gets re-written into the cache on every turn (cache-write was ~80% of spend; write:read ratio ~2:1, backwards).

tool.done.output was a single field serving two masters: the UI/record (wants the full payload) and model replay (wants a bound).

Fix

Separate the two concerns:

• output (full) — UI/display + conversation record. Unchanged.
• modelOutput (bounded) — exactly what the model saw on the live turn. New optional field on ToolDoneEvent, persisted only when it differs from output.

A single shared boundToolResultForModel() (in content-helpers.ts, next to extractTextForModel) produces the bound. The engine and the reconstructor both call it, so the model's live view and replayed view are byte-identical. Legacy events without modelOutput fall back to bounding output at read time, so existing conversations are fixed too. The bound is pure/deterministic → the replayed prompt prefix stays stable and cacheable. It trims on line boundaries (never mid-record) and preserves the inline-UI pointer behavior for new events (legacy UI-tool events without a persisted modelOutput are line-trimmed on replay — still bounded, strictly better than the prior full-payload replay).

Framed precisely: this is a replay-fidelity fix — today replay gives the model more than the live run did. It happens to also eliminate the runaway context growth.

Also: 1h prompt-cache TTL

Both Anthropic cache breakpoints move from the 5-minute default to ttl: "1h". Agentic turns pause far longer than 5 minutes (user steps away; automation waits on I/O), so the prefix constantly lapses and the next turn re-writes the whole thing at the write rate. 1h keeps the prefix alive across those gaps, converting full re-writes into cheap reads.

Tests

• boundToolResultForModel: under-limit passthrough, line-boundary trim, single-huge-line hard-slice fallback, inline-UI pointer, determinism, limit<=0 unbounded.
• Reconstructor replay: large legacy result bounded; modelOutput replayed verbatim; small result unchanged; determinism; UI metadata still carries full output.
• Updated existing cache-control assertions (now ttl: "1h") and the prompt-injection truncation-notice assertion (new marker wording). Security property (injection beyond the bound is dropped) preserved.

tsc --noEmit clean, biome clean on src/, full unit suite green except one unrelated pre-existing missing-dep (dompurify) in an automations-UI test this PR doesn't touch.

Rollout note

The model bound takes effect immediately for all conversations (new events store modelOutput; legacy events are bounded on read). The 1h TTL is fleet-wide. No tenant config change required. Expected to cut the dominant cache-write cost substantially and stop context from ballooning on long threads.