v0.144.1

What's New

Fixes

• Homebrew update detection — agentico update no longer demands the Go toolchain for a Homebrew-installed binary. The unambiguous Homebrew Cellar check now runs before the go-install check, so brew installations are classified correctly even when their build-info version looks like a real module version.

v0.144.0

What's New

Features

• Model catalog discovery — Agentic now discovers each provider's available models at startup instead of relying on a hardcoded list. Claude models are resolved by probing each alias (so you get the real model and context window behind opus, sonnet, haiku, etc.), and Codex refreshes from codex debug models. Catalogs are cached per CLI version with an offline fallback.
• Selectable context-window variants — Each context window is now its own selectable model entry (e.g. opus[200K], opus[1M], gpt-5.4[272K], gpt-5.4[1M]). Bare names like opus remain as aliases for the default window. Codex passes the chosen window through as a context-window override.
• Claude Fable support — Added support for Anthropic's Claude Fable models.

Improvements

• Concurrent catalog discovery with live progress — Model catalogs are now discovered concurrently across providers, with live progress shown during startup so you're not left waiting on a blank screen.
• Centralized textarea styling & refined wizard navigation — Text input fields share a single consistent style across the TUI, and the setup wizard's navigation has been smoothed out.

v0.143.0

What's New

Features

• agentico update self-update command — Update your installation in place without re-running a manual install. The command detects how Agentic was installed (go install, Homebrew tarball, etc.) and takes the right update path automatically. Use agentico update --check to see whether a newer release is available without applying it.

Improvements

• Safer self-updates — The updater never downgrades: if your current build is equal to or newer than the latest published release, it reports that you're already up to date (and tells you when you're ahead). Release-asset matching is tolerant of naming variations, and file mode/ownership are preserved across the swap.
• Privileged-location handling — When Agentic lives in a privileged path, the updater runs a pre-flight check and re-signs the binary on macOS after replacement, avoiding broken installs.
• Dev-build guard — Updating a locally built (dev) binary is refused with a clear message instead of silently replacing it.

Fixes

• CI coverage — Added a CI workflow running go vet and tests on every change, and made the knowledge-base fallback test independent of the host PATH for reliable runs.

v0.142.0

What's New

Features

• Install via Homebrew — releases now publish a formula to the doordash-oss/homebrew-agentic-orchestrator tap, so you can install and upgrade agentico with brew instead of downloading archives by hand.

Improvements

• Clearer install docs — the README now covers installing from prebuilt binaries and configuring a single LLM provider, so first-time setup no longer assumes a full multi-provider build.

Fixes

• No more blank input gates — when an implement iteration pauses for your input, the questionnaire is now backfilled from the run's progress handoff. Previously a blank agent record could surface an empty, unanswerable gate even when the real summary and questions existed.

v0.141.0

What's New

Features

• Sub-agent delegation — Agents can now spawn focused sub-agents, parallelizing broad independent work and honoring SKILL instructions that explicitly request delegation instead of serializing it.
• Provider authentication checks at startup — Agentic now verifies that each detected provider CLI is actually authenticated (claude auth status, codex login status), not just present on PATH. Unready providers are filtered out of model routing, startup surfaces actionable notices, and you get a provider-specific remedy when none are ready.

Improvements

• Resilient final review — Features that hit a terminal failure are now surfaced and recoverable end-to-end, even when status had drifted to a success state. The TUI renders a failure box, a ✗ row, a "failed" status, and a [l] Logs action, and restart routes through the configured iteration budgets.
• Smoother streaming output — Streaming partials are coalesced across passive tool-progress events for cleaner live output.
• Comment-hygiene directive — The implement, final-fix, and tweak-session skills now steer agents toward cleaner, less noisy comments in generated code.

Fixes

• Clear error for repos with no commits — Creating a worktree in a repository with an unborn HEAD now returns an actionable message instead of an opaque "invalid reference" error, and the true git error is reported instead of being masked by the fallback.
• Stable Codex question phases — Stopped Codex agents from restarting grill-me phases when their question format drifted into open-ended prose on later turns; marker existence is now the authoritative completion signal.

v0.140.0

What's New

Features

• Redesigned repo + branch selection wizard — Step 2 is now a chip/pill picker: selected repos appear as chips at the top, the list shows only unselected repos, and an explicit Continue button replaces the old footer-only "Enter to advance" affordance. Enter is contextual (adds the focused repo, opens Browse/Create, or activates Continue), Backspace on an empty filter removes the last chip, and Ctrl+D advances from anywhere. The branch warning is now a dedicated per-repo screen so each off-default repo can independently choose whether to reuse its current branch or start fresh.

Improvements

• Rebase and review-comments cycles pause on NEED_USER_INPUT — both loops now reach parity with the refactor cycle: when an iteration surfaces a NEED_USER_INPUT gate, the staged repos transition to RepoCycleNeedUserInput with the gate path attached, instead of being written as failed. You can resolve the gate and resume rather than restarting the cycle.

v0.139.4

What's New

Fixes

• Fewer false-positive deferral warnings — Removed regex-based prose hedge detection in the progress.md parser, which was flagging legitimate prose referencing future phases (including agent self-reports about fixing prior feedback) as unstructured deferrals. Structured deferral checks (unclosed entries, missing reasons) still apply.

v0.139.3

What's New

Improvements

• Planning agents stay in their lane — Inquirer, Designer, Roadmap, and PhasePlan roles are now explicitly forbidden from writing code. They can only write to their designated output roots, even if a user asks them to edit source. This keeps planning phases focused on planning and prevents accidental code edits before review.

v0.139.2

What's New

Improvements

• Bounded protocol retries for single-shot phases — Inquire, Research, Brainstorm, per-repo Knowledge-Base, and refactor-plan phases now retry with structured feedback when their contract fails, instead of immediately failing the feature. Backed by a shared retry helper with per-phase bounds.
• Polished review-gate UX — The dashboard CTA now stays "Review" while a review gate is open instead of reverting to stale "Approve" or "Answer" prompts. Stale help/permission chips clear when entering review or rewinding, the live preview yields to the overview during needs-review, and you can restart a phase from the dashboard while a review is still active.
• Faster startup — Embedded skills and guidelines reconcile via a content-hash stamp, so unchanged content short-circuits the full filesystem walk on every launch. Provider version checks now run concurrently, and the reconcile is covered by a single launch spinner that only appears when there's actual work to do.

Fixes

• Ignore stale session events from prior incarnations — When a session was replaced under the same ID, late-arriving SDK or done events from the previous incarnation could trigger spurious retries against the replacement. Events are now stamped with the originating session's start time and dropped when they no longer match the live session.

v0.139.1

What's New

Fixes

• Apple Terminal color rendering — Agentic now falls back to ANSI 256-color output when running inside macOS Terminal.app, so colors render correctly instead of looking washed out or wrong.

Improvements

• README polish — minor cleanup to the project README.