v0.14.5

Release v0.14.5

Installation

One-liner (Recommended)

curl -fsSL https://asimi.dev/installer | bash

Homebrew (macOS/Linux)

brew install afittestide/tap/asimi

Binary Download

Download the appropriate binary for your platform below.

Added

• Private-network provider egress opt-in — a new llm.allow_private_network config option lets the provider SDK reach RFC1918 (10.x/172.16.x/192.168.x) and ULA endpoints for self-hosted/on-prem models. Off by default because enabling it relaxes the SDK's SSRF egress guard for every provider; link-local/metadata addresses (169.254.0.0/16, fe80::) remain blocked regardless (e912)

Fixed

• Daemon disconnect recovery — the TUI reconnects automatically when the daemon connection drops: the court client is wrapped in a reconnecting client that respawns a dead daemon and re-runs the SetContext handshake on every new connection, drop detection re-arms after each drop, and if reconnect does not complete within the poll window the user is told "Daemon unreachable — still trying to reconnect…" instead of being stranded on "Reconnecting…" (e911)
• Judge evidence files no longer clutter the project root — the Judge no longer writes judge_evidence_*.json files into the working tree; its treason clause now requires a verdict to include its evidence rather than render it to a JSON file

---

Full Changelog: v0.14.4...v0.14.5

v0.14.1

Release v0.14.1

Installation

One-liner (Recommended)

curl -fsSL https://asimi.dev/installer | bash

Homebrew (macOS/Linux)

brew install afittestide/tap/asimi

Binary Download

Download the appropriate binary for your platform below.

Added

• Live model picker — :models (alias :model) now enumerates models live from every configured provider's /v1/models and caches the catalog for the session (reopening is instant and makes no network calls); :models refresh forces a refetch. Rows carry per-model metadata (context window, price), list the current provider's models first followed by the others alphabetically, show provider icons, and render a dim unavailable: <provider> line when a provider cannot be reached (e903, e906, e910)
• Direct model selection — :model <provider>/<name>, :model <provider> <name>, and :model <name> (current provider) select a model without opening the picker; the picker also ends with a "Type a model name…" row for provider/model input (e903, e906, e910)

Changed

• Context window reporting — the context window is now provider-reported when available, falling back to a local registry estimate rendered with a leading ~ (e.g. ~128.0k) so an estimate is never shown as exact (e908)
• Model rows drop the modality token — model picker rows no longer show a text->text-style modality token (e910)

Fixed

• Automatic context overflow recovery — a request rejected for exceeding the model's context window no longer halts the session: Asimi learns the true window from the provider's error, compacts the conversation to fit, and re-sends once; if it still cannot fit, the session stays alive with an actionable hint (context window N exceeded — use :compact or :new) (e908)
• Malformed tool-call arguments — a tool call whose function.arguments is not valid JSON is normalized to {} in place and reported back to the model as an error, instead of poisoning history and wedging the session on every later request (e854)
• Unknown rituals — enact_ritual now rejects an unknown ritual name with a clear error listing the available rituals, instead of emitting an event that can never run (e841)
• Precedent verification on reforged manifests — check_precedent_approved now evaluates only current manifests, so a rejected precedent left on a superseded manifest no longer blocks an edict from ascending (sibling of the verdict-gate bug)
• Fork race condition — removed a data race in fork/join ritual execution (e900)

---

Full Changelog: v0.14.0...v0.14.1

v0.13.7

Release v0.13.7

Installation

One-liner (Recommended)

curl -fsSL https://asimi.dev/installer | bash

Homebrew (macOS/Linux)

brew install afittestide/tap/asimi

Binary Download

Download the appropriate binary for your platform below.

Chaged

• vendored package where updated based on security requirments

---

Full Changelog: v0.13.5...v0.13.7

v0.13.6

Release v0.13.6

Installation

One-liner (Recommended)

curl -fsSL https://asimi.dev/installer | bash

Homebrew (macOS/Linux)

brew install afittestide/tap/asimi

Binary Download

Download the appropriate binary for your platform below.

Added

• Gitless ground as a first-class mode — RepoInfo now carries HasVCS as the single authoritative VCS-detection flag, so Asimi runs cleanly in a plain directory with no git repository: the project slug falls back to the directory basename, branch slugs stay empty, earth expressions and heaven snapshots report the condition explicitly, borderland manifests are built from a file inventory, and the honest "no VCS detected" log line replaces the previous false "git repository detected" (e888)

Changed

• Host runner timeout discipline — commands run outside the sandbox (--isolated-host / host runner) now run under a command deadline (default 10m) and bounded approval waits (default 2m), matching the Podman runner. Both are configurable via [run_shell_command].timeout_minutes and the sandbox approval_timeout, and are plumbed through SessionConfig into every ephemeral HostRunner construction path (e886, e889)
• Manifest permissions merge — create_manifest is dual-registered under intent Write so the forge can record its own manifests; registry.Register now merges permissions instead of panicking on re-registration (e888)
• asimisql uses the embedded driver — court queries now go through the embedded driver (ToolContext DB) instead of shelling out to a sqlite3 CLI binary (e888)

Fixed

• ask_ruler inside a ritual — answers submitted on a ritual tab are now delivered as the zhengming answer instead of being treated as a ritual interjection (which paused or aborted the step); the TUI tracks pending zhengming per routing channel and the Court refuses to mint interactive sessions for ritual channels (e894)
• Headless log noise — headless mode (-p) now consumes tool-scheduling and stream-lifecycle events at debug level, so a headless run no longer floods its log with unhandled-event warnings (598d32e)
• bash step commands that are not found — a given step failing with exit 127 (command not found) now produces a note for the minister instead of failing the ritual step, leaving the verdict gate as the single retry mechanism and ending the harbor timeout spiral (e888)

---

Full Changelog: v0.13.5...v0.13.6

v0.13.5

Release v0.13.5

Installation

One-liner (Recommended)

curl -fsSL https://asimi.dev/installer | bash

Homebrew (macOS/Linux)

brew install afittestide/tap/asimi

Binary Download

Download the appropriate binary for your platform below.

Added

• Ling execution engine (built, default off) — lings (task orders) can now execute over a shared DAG engine that honors declared dependencies; lings carry an owning minister and batches of insert_ling calls with satisfied dependencies can be ignited automatically. Ignition is disabled by default (ling_ignition_enabled = false) until per-ling worktree isolation lands (e531) — concurrent lings on one working tree are unsafe without it (e847)
• Kimi K3 — added Moonshot's Kimi K3 to the model roster/config defaults

Changed

• Provider error reporting — provider errors now carry provider name and HTTP status context, improving operator diagnosis (e879)

Fixed

• Release tarball wrapping — release tarballs no longer double-wrap the asimi binary under an ./asimi directory, fixing the checksum/installer layout (e864)
• OPENAI_BASE_URL compatibility — openai-compatible providers honor custom base URLs again (e863)
• ask_ruler routing — ruler questions are routed to the correct tab (e866)
• --isolated-host log output — headless-mode log output restored under --isolated-host (e881)
• Precedent tools — record_precedent / query_precedents fixes (e874, e875, e876)
• Duplicate seal crash — crash on duplicate seal grant; seal granting is now idempotent (e882)

---

Full Changelog: v0.13.4...v0.13.5

v0.13.4

Release v0.13.4

Installation

One-liner (Recommended)

curl -fsSL https://asimi.dev/installer | bash

Homebrew (macOS/Linux)

brew install afittestide/tap/asimi

Binary Download

Download the appropriate binary for your platform below.

Fixed

• Heaven's ledger
• support for clit prompts starting with - as in -p "- <prompt>
• ask_ruler chat option send the users prompt as the reply

---

Full Changelog: v0.13.3...v0.13.4

v0.13.3

Release v0.13.3

Installation

One-liner (Recommended)

curl -fsSL https://asimi.dev/installer | bash

Homebrew (macOS/Linux)

brew install afittestide/tap/asimi

Binary Download

Download the appropriate binary for your platform below.

Added

• Minister communication over the Tian ledger — ministers can now communicate through the Tian ledger

Fixed

• Harbor integration — honor ASIMI_HOME for the court DB path and degrade gracefully when $HOME is unwritable, fixing the read-only-$HOME startup abort under --isolated-host; initLogger no longer panics when its log directory cannot be created (e836)
• Init-project ritual TODO comments — the ritual now uses TODO comments correctly
• suggest_edict description — improved the tool description for smoother activation

---

Full Changelog: list...v0.13.3

v0.13.2

Release v0.13.2

Installation

One-liner (Recommended)

curl -fsSL https://asimi.dev/installer | bash

Homebrew (macOS/Linux)

brew install afittestide/tap/asimi

Binary Download

Download the appropriate binary for your platform below.

Fixed

• Aborted and failed ritual no longer blocks the rituals queue

Changed

• Default context size set to 256K for unknown models

---

Full Changelog: v0.13.1...v0.13.2

v0.13.1

Release v0.13.1

Installation

One-liner (Recommended)

curl -fsSL https://asimi.dev/installer | bash

Homebrew (macOS/Linux)

brew install afittestide/tap/asimi

Binary Download

Download the appropriate binary for your platform below.

Added

• --reasoning-effort flag — new --reasoning-effort CLI flag (plus ASIMI_REASONING_EFFORT env, subcommand, and [llm] config plumbing) for setting the reasoning effort / model thinking level on supported providers, useful for harness-driven and headless runs (e798)

Changed

• Repo rename completed — docs, scripts/install.sh, .goreleaser.yml, and version strings updated to finish the repository name change
• TUI update speed (e780) — chat content re-renders are now debounced while streaming via a new [ui] tick_interval config (default 100ms), letting the TUI batch streaming updates for smoother, faster rendering
• Edict selector seals use emojis — the judge and chancellor seal glyphs in the :edict list are now rendered as emojis (👩‍⚖️ / ☯️) instead of kanji (刑 / 門)
• Git diff TTL caching — repo status/diff statistics (internal/repo) are now cached and refreshed with a TTL instead of being recomputed on every call, and tool-call messages are batched as dirty content updates instead of forcing full re-renders; both reduce unnecessary TUI work

Fixed

• **ATIF trajectory completeness ** — fixes to the --atif recorder so recorded trajectories are well-formed
• Slowdown - getting context size from the provider in a loop led to sluginsh off the ui as Asimi was busy waiting for the model's context size

---

Full Changelog: v0.13.0...v0.13.1

v0.13.0

Release v0.13.0

Installation

One-liner (Recommended)

curl -fsSL https://asimi.dev/installer | bash

Homebrew (macOS/Linux)

brew install afittestide/tap/asimi

Binary Download

Download the appropriate binary for your platform below.

Added

• --model & --provider CLI flags — new --model and --provider flags plus
  ASIMI_MODEL / ASIMI_PROVIDER environment-variable overrides, so the model can be
  selected from the command line or environment for harness-driven runs (terminal-bench).
  Precedence: CLI flag > env var > config [llm] section > defaults (e785)
• Vertex AI provider — new vertex provider that talks to Google Vertex AI via gcloud
  application-default credentials (ADC), with all gcloud-specific values read from
  environment variables (GOOGLE_CLOUD_PROJECT, GOOGLE_CLOUD_REGION, optional
  GOOGLE_APPLICATION_CREDENTIALS); no VERTEX_API_KEY needed (e786)

---

Full Changelog: v0.12.0...v0.13.0