feat(own-loop): in-memory tool-result prune harness safety net

[jonnyparris]

What

A no-LLM, no-storage harness safety net that keeps the own-loop's in-memory messages array under the context budget when Think-level compaction can't help.

Why

Reproduced this morning by deploying #73 and watching wrangler tail:

• Gemma 4 26B (and Kimi K2.6) make many tool calls in a single first user turn
• Each tool result inflates the own-loop's closure messages array
• Think persists the assistant message only AFTER streamText() finishes, so during the turn realMessages.length === 1
• Compaction has nothing to summarise; runs as a no-op
• Hard-stop fires; session ends with 'context budget exhausted'

PR #73 + PR #74 fixed compaction's reachability and silent-no-op bugs. This PR closes the remaining gap by attacking the actual bytes — the own-loop's in-memory tool-result payloads.

How

New module src/own-loop-prune.ts exporting pruneOversizedToolResults(messages, opts):

• Pure function. No async, no storage.
• Walks tool messages oldest-first, preserves the most recent N (default 2).
• Replaces the value of each large tool-result part with a short placeholder string. Keeps the envelope (toolName, toolCallId) intact so the model's prior tool-call still resolves.
• Stops as soon as estimated tokens drop below target.
• Configurable: targetTokens, preserveRecentToolMessages, minPrunablePayloadChars.

Wired into coding-agent.ts onChatMessage own-loop, between the pre-step Think compaction attempt and the hard-stop:

projectedTokens >= MID_LOOP_COMPACTION_THRESHOLD
  → maybeCompactContext({ force: true })    // PR #73 + #74
  → if still over: pruneOversizedToolResults  // this PR
  → if still over: hard-stop

Tests

test/own-loop-prune-unit.test.ts — 7 cases:

• under-budget no-op
• oldest-largest pruned first
• preserves the most recent N tool messages regardless of size
• envelope (toolName, toolCallId) survives pruning
• skips payloads smaller than minPrunablePayloadChars
• stops pruning once budget back under target
• no-op when there are no tool messages

All 21 existing compaction tests still pass; typecheck clean.

Risk

• Pruned tool results are gone — the model can't re-read them. The placeholder tells the model to re-run the tool if it still needs the data. This trades one ~50k-token tool dump for ~150 chars and the model's discretion to repeat.
• Affects every session that crosses the mid-loop threshold, not just weak orchestrators. Strong models will rarely hit the path; when they do, the same trade-off applies. The default minPrunablePayloadChars: 2000 keeps small results intact.

Linked

• PR fix(compaction): force/truncated bypass the minMessages guard #73 (merged) — force/truncated bypass minMessages
• PR fix(compaction): mid-loop trigger uses force=true; guard slice underflow #74 (merged) — mid-loop force + slice guard
• memory/learnings/dodo-orchestrator-model-matters.md
• memory/workload/ideas-backlog.md → 'Cap single-turn input-token growth' — this is that work

beep-boop-🤖