agentty 0.2.11 — a cost-aware agent: cheaper compaction, cheaper subagents, real OAuth for MCP

Two releases packed into one push: v0.2.10 shipped a from-scratch-green Windows build and per-user MSI install; v0.2.11 is a session's worth of work making agentty economical — compaction, subagent fan-out, and the model catalog all got cheaper or more correct without changing how you use the tool. Full detail, as always, in the changelog.

Compaction and subagents now cost less

Auto-compaction used to fire at a fixed absolute margin (context_max - 17k) — inconsistent across window sizes (98% on a 1M window, 91% on 200K) and it forced a summarization pass long before a big-window model actually needed one. It now fires at a percent of the window (default 90%, clamped to always leave 20K tokens of output headroom), tunable from the command palette's new Compaction depth entry (75% Aggressive → 90% Balanced → 95% Deep, persisted to settings.json).

More important than the trigger: the summarization request itself now runs on the cheapest capable model on your active provider instead of the flagship model you're chatting with. Compacting used to mean a full ~context-max-sized flagship-priced input on every trigger; it now costs a Haiku-class summary.

The same idea applies to task fan-out: read-only subagent roles (explorer, reviewer) route to the cheapest model that still clears a capability floor (tool support, dispatchable, non-Weak tier) — tester, coder, and general keep the parent model since they mutate the workspace. A single-model or Opus-only account sees no behavior change; a multi-model account gets exploration for a fraction of the cost.

1M-context models, done Claude-Code-faithfully

Signed in with Claude Pro/Max OAuth, the model picker now offers a (1M context) row right below every Sonnet/Opus/Haiku 4+ model — verified against Claude Code's own shipped binary: the base window is always 200K, and 1M is an explicit, entitlement-gated picker row, never auto-detected from account tier. Picking it widens the tracked context window and sends Anthropic's extended-context beta on your behalf; nothing else changes about how you use the model. A raw API key doesn't get the row (the beta is account-tier gated) — it's OAuth-only, matching reality.

The picker sorts on strength, not a hand-written list

The old sort ordered models by a hardcoded family bucket — Opus, Sonnet, Haiku, Fable, Mythos, in that fixed position — so Fable/Mythos (Anthropic's newest flagship-tier lane, same strength class as Opus, just a different codename) always sank to the bottom of the picker no matter how new or capable it actually was. The new agentty::model_picker_less comparator is the single source of truth for picker ordering: tier descending (Flagship — Opus and Fable/Mythos — always leads), then newest generation/revision within a tier, then a family tie-break for stable grouping among peers. Any future family name sorts by what it is.

MCP gets real interactive OAuth

agentty mcp-login <server> now walks an OAuth-gated MCP server end to end: RFC 9728 protected-resource discovery, dynamic client registration (or a CIMD https:// client-id, or a pre-registered one — whichever the server supports), a PKCE (S256) browser flow on an ephemeral loopback callback, and a sealed, auto-refreshed token at ~/.agentty/mcp_tokens/<server>.json (chmod 600). mcp-logout clears it, mcp-status lists every configured server and its authorization state. The whole path implements the 2026-07-28 MCP auth hardening: RFC 9207 issuer validation (closing the AS mix-up attack), application_type=native DCR, issuer-bound credentials, and CIMD support — all in the dependency-free, unit-tested mcp-cpp layer, Windows-verified loopback server included.

Provider ingress, fully unified

Every ingress concern shared across agentty's four transports — Anthropic, OpenAI-compatible, Ollama-native, and ChatGPT/Codex's Responses API — now lives in exactly one shared helper: Retry-After backoff parsing, strict UTF-8 scrubbing, the leaked-tool-call sniffer, all three token-usage wire shapes, and OpenAI-family auth headers. No more per-transport copies that could silently drift; fixed two real bugs found while unifying (Ollama was silently ignoring server backoff; two transports accepted invalid UTF-8 the canonical scrubber now rejects).

repo_map won't leak sibling projects

The ranked codebase skeleton repo_map returns could surface source from a submodule, a vendored checkout, or an unrelated repo living under or alongside your workspace. The walk now stops at any nested repository boundary — including a submodule's gitlink file, not just a directory — and re-asserts every accepted file is genuinely inside the workspace before it becomes a graph node.

Also in this line (v0.2.10)

• Windows installs with no admin/UAC prompt. The MSI is now perUser: installs to %LocalAppData%\Programs\agentty, edits only your PATH, no elevation required. The winget manifest matches (Scope: user).
• Windows builds green end-to-end again. Retrieval and the jetalloc allocator degrade gracefully on MSVC (no-op fallback) rather than blocking the whole release; the macOS build no longer trips configuring a GPU backend agentty's CPU-only retrieval never uses.
• Package channels can't silently fall behind a release. Every downstream publisher (AUR/Homebrew/scoop/winget) is gated behind its build leg, and a manifest-reconciliation job re-pins hashes from a release's SHA256SUMS automatically after every run and weekly.
• Homebrew is a clean two-line install: brew tap 1ay1/tap && brew install agentty.

Get it

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

See the full changelog and Providers & Models for the 1M-context and picker details.