iris-agent

Tune Iris auto-compaction and microcompaction settings with evidence, not vibes. Use when changing trigger thresholds, keep-recent tail, microcompaction watermark or cache-timing posture, summary reserve, or hover-level defaults — or when a compaction cost/quality claim needs proof. Covers the knob inventory, the deterministic-twin-first loop, the cache-economics math, the live campaign harness, gates, and reporting conventions.

Cut and verify a public iris-agent release (crates.io + prebuilt binaries). Use when asked to release iris, cut/tag a new version, merge the rolling release PR, or verify/finish a release that failed midway. Canonical runbook is docs/RELEASING.md; this skill adds the operational hardening learned from live cuts.

Build or extend a token-efficiency benchmark for an Iris tool result (ADR-0036 rule 5, "reduction is measured"). Use when adding or changing any output-reduction path — bash filters, windowing, summaries, handle offload — when a savings claim needs proof, or when extending the ADR-0037 corpus. Covers real-fixture capture, the shared bench_support helpers, minimum bars, verbatim-survival needles, overhead bounds, and the committed report.

Documentation and codemap maintenance specialist. Use when generating architectural codemaps, updating README/docs from code, mapping imports/exports, validating documentation freshness, or ensuring docs match the current codebase.

Repo documentation style guide for Iris. Use when writing or editing any repo doc — README, PRODUCT/DESIGN, architecture, TUI specs, features, roadmap, ADRs, codemaps, troubleshooting, or code comments. Enforces the terse field-manual voice, progressive disclosure, measured claims, placeholder examples, and per-doc-type templates.

Design, implement, review, or debug Iris's terminal UI from the current repository sources. Use for work under src/ui; Screen state; pager or inline rendering; transcript, tool, composer, session-bar, settings, picker, modal, or overlay surfaces; palette, symbols, layout, focus, input, golden frames, and TUI conformance. Read the live design language, ADRs, source, and tests instead of copied design-system snapshots.

Diagnose a saved Iris provider-transport fallback immediately after the TUI says OpenAI exhausted WebSocket recovery and switched to SSE. Use only when the user explicitly invokes $provider-stream-diagnostics.