Context compaction: summarizer failure aborts the turn; re-summarizes on every model call once over threshold

[senamakel]

Part of the v0.58.7 → HEAD TinyAgents-migration audit (see parent issue).

Problem

assemble_turn_harness installs the crate ContextCompressionMiddleware (src/openhuman/tinyagents/mod.rs:1317-1329). Two regressions vs the legacy engine:

1. Summarizer failure now aborts the turn. The crate's before_model does self.summarizer.summarize(...).await?; any provider hiccup maps to TinyAgentsError::Model (summarize.rs:123-131) and fails the run. Legacy (v0.58.7 engine/core.rs ~300): compaction failure was warn! + circuit breaker + fall back to deterministic trim; the turn continued. New hard failure mode on exactly the longest, most valuable threads.
2. Compaction re-runs on every model call once over threshold. The crate middleware rewrites only the per-call request.messages clone; the loop's working transcript never shrinks (crate agent_loop/mod.rs:354 rebuilds the request from messages.clone() each iteration). Once a thread crosses the 90% threshold, every subsequent model call in the turn triggers a fresh full-transcript summarizer LLM call — cost and latency multiply with tool-loop length.

Fix plan

1. Wrap the summarizer we hand the crate in a fault-tolerant adapter: on Err, log a warn with the cause, trip a per-turn circuit breaker, and return a deterministic trim of the input (reusing MessageTrimMiddleware logic) instead of propagating. Turn must never fail because compaction failed.
2. Avoid re-summarizing per call. Options:
   a. Cache the summarization result keyed by (message-count, hash of compacted prefix) inside our adapter, so repeat calls within a turn are free until the transcript grows past the threshold again;
   b. or upstream a crate change to apply compression to the loop's working transcript (with provenance events), then adopt it.
3. Emit a compaction event (what was summarized, token before/after) — the spec's "move context compaction provenance into TinyAgents events" checkbox — so silent compaction is observable.

Acceptance criteria

• Scripted summarizer failure mid-turn: turn completes via deterministic trim; warn logged; no chat_error.
• A long turn over the threshold performs at most one summarizer call per transcript growth step, not one per model call (assert call count via mock provider).

[M3gA-Mind]

Resolved by merged PR #4495 ("TinyAgents migration parity & bug fixes (#4451–#4469)") + follow-up #4504 — both present in main. Part of the completed #4470 cluster; closing as done. (Reopen if a residual remains.)