agentty 0.2.12 — Zed-native ACP parity, a smarter retrieval pipeline, and a hardened composer

A dense release: ACP shell parity with Zed's own agent, a retrieval quality upgrade with a funnel you can actually see, and a composer editing sweep that closes six real bugs, one of them data loss. Full detail in the changelog.

ACP: agentty now looks native inside Zed

Running agentty as an ACP agent used to feel like a second-class citizen next to Zed's own agent. Four gaps closed:

• Slash-command menu. Session start now emits available_commands_update, so Zed's composer / menu is populated with /compact, /new, and every installed skill as /<skill-name>.
• Model picker. A model config_option_update advertises the provider's full model catalog, so you can switch models from Zed's per-session dropdown instead of relaunching with -m.
• Sign-in prompt. When agentty has no credentials, initialize advertises an authMethod, so Zed renders a proper "Sign in" affordance instead of erroring on the first turn.
• Live thread titles. The first message of a session pushes a derived title via session_info_update, and restored sessions re-announce it on load, so Zed's thread sidebar always shows something meaningful instead of a blank or stale name.

Tool cards got the same treatment. A completed read now renders as a fenced, tab-numbered excerpt (1\talpha, 2\tbeta, …) matching claude-code-acp's gutter view. All non-diff tool output is markdown-escaped so literal *, backticks, #, |, and < in file/grep/command output render verbatim instead of being parsed as bold/headers/tables/HTML. Two rendering bugs are fixed too: a bash run through the live-terminal fast path used to show nothing once the terminal widget was released — the completion now also carries the captured stdout/stderr as a fenced block — and write now announces the same before/after diff shape as edit instead of a distinct "new file" affordance. Every card that used to render as a bare title now has a body, on both first run and session replay.

Retrieval: rag-cpp's measured-best pipeline, and you can see why

search_docs/search_code upgrade to adaptive convex (TM2C2) fusion in place of plain reciprocal-rank fusion — rag-cpp's benchmarks show it beating RRF on NDCG because it preserves the score distribution RRF discards, and the adaptive variant shifts weight toward whichever retriever (lexical vs. dense) is more confident on a given query. Two new refinement stages ride along: near-duplicate dedup (folds paraphrase/boilerplate copies so the model isn't re-reading the same passage twice) and relevance autocut (trims the low-relevance tail at the score knee, so a query with three strong answers returns three, not a fixed *k* padded with weak matches).

Retrieval used to be a black box — a ranked list and a terse (mode: hybrid+ctx, reranked) label. Every result is now headed by a readable funnel that walks the real per-stage counts the engine recorded:

hybrid: 47 candidates ↳ reranked top 30 ↳ dedup 30→24 ↳ stitch: merged 3 adjacent ↳ autocut 24→8 ↳ top-8

And it's cheaper, not just clearer: five frugality levers on the output path — a relevance floor, score-proportional water-filling of the token budget, CRAG-confidence-scaled total budget, an LLMLingua-style extractive compression pass for oversized prose passages, and a tighter cap on unprompted proactive context — cut retrieval output ~13% (1,845 → 1,605 estimated tokens/query on a 161-file corpus) with identical ranking quality (recall@10 1.000, MRR 0.968, nDCG@10 0.976). Every lever is on-by-default and tunable via env var; AGENTTY_RAG_MEASURE=1 agentty rag-bench quantifies any one of them on your own corpus. See Retrieval for the full knob list.

Composer: a data-loss bug and five papercuts, fixed

A sweep across the composer's reducer, view, and widget layers:

1. Editing a peeked queued message could silently delete the wrong one. After Alt+↑ loaded a queued message for editing, the peek index survived an ordinary keystroke, so Enter sometimes removed a different queue slot than the one on screen. A stray edit now cleanly drops the peek and becomes the live draft.
2. Undo now rewinds word-by-word, not character-by-character. Ctrl+Z used to undo one character at a time, so a long sentence blew the whole 64-deep undo history. Consecutive typing now coalesces into one undo unit, broken on whitespace and any non-typing op.
3. The live token/word/line counters were wrong with an attachment present — they counted the short chip caption instead of the payload, so a 400-line paste read as "1 line, ~10 tok." The counter now measures the expanded attachment body, i.e. what actually goes to the model.
4. Ctrl+←/→ word motion now steps over a run of punctuation as one unit ()))) used to take four presses).
5. The / command palette opens on any line-leading slash, not only a completely empty composer, matching shell muscle memory.

Kill-to-end-of-line is back too, rebound to Alt+K (Ctrl+K is claimed by the command palette), pairing with Ctrl+U (kill-to-start) the same way Alt+D pairs with Ctrl+W. The full composer keymap is now documented in the README and the keybindings page.

Also in this line

• Thread list startup on a large history went from ~1000 ms to <1 ms (~1400×) via a small threads/index.json metadata sidecar, self-healing if missing or corrupt.

Get it

curl -fsSL https://agentty.org/install.sh | sh
# or
brew tap 1ay1/tap && brew install agentty

See the full changelog for every detail.