release: 0.18.0

[yuanhao]

Twelve commits since v0.17.0. Six breaking changes, so a minor bump.

Breaking

• CostConfig::new and ContextTier::new take two rates; cache rates move to with_cache_read / with_cache_write
• CostConfig::context_tier: Option<ContextTier> → context_tiers: Vec<ContextTier>, kept sorted
• CostConfig and ExecutionLimits are now #[non_exhaustive] — build with new() / Default::default() plus with_*
• ExecutionLimits gained max_consecutive_identical_tool_calls (on by default, Some(3))
• AgentLoopConfig gained tool_output_sink

ExecutionTracker also lost struct-literal construction; it is recorded in the CHANGELOG.

Headline

• Retrievable tool output — truncation stashes the full text and the marker names a key the model can fetch
• Loop detection — steers on the third consecutive identical call, stops on the next
• Price drift audit against models.dev, gating every shipped rate
• Context tiers in CostConfig (no preset uses one; see gpt_5_5's docs for why)
• Fixed: tools with no parameters were uncallable on Anthropic — pre-existing, affects released versions

Verification

Tests	619 passed, 0 failed
Clippy	clean under -Dwarnings
Feature combos	default / openapi / gasp all green
Live price audit	26 compared, 0 drifted
Live smoke	6/6 Sonnet 5, 5/5 DeepSeek
Long-horizon	transcript valid across compaction, 67.2% cache hit rate
Packaging	cargo package builds
Docs	reference updated to the new API, mdbook build clean

Known, not blocking

#150 — LlmCompaction takes the deterministic fallback on tool-heavy workloads, root-caused to compact_headroom_turns driving the target ratio to its floor. Byte-identical in v0.16.5 and v0.17.0, so holding this release protects nobody; the fix is a defaults change that wants measurement.

Release commit is exactly two files — Cargo.lock is gitignored.

🤖 Generated with Claude Code