Archon CLI v0.13.12026-10-29
Workflow names must match exactly, provider plugins move to protocol 3, and a batch of fixes for Postgres boots, Codex runs, cleanup and cost accounting.
> Upgrading from 0.13.0 with providers installed: use v0.13.2. In 0.13.1, archon provider add refused to replace providers installed by 0.13.0 (#5272). 0.13.2 fixes it, and archon provider add claude codex pi works in one step.
Breaking
- The provider plugin protocol is now 3. After upgrading, run archon provider add again for each provider you use. Third-party provider plugins must rebuild against protocol 3. A provider installed by 0.13.0 is reported as unavailable, naming its protocol, this Archon's protocol and the command that fixes it, instead of failing workflow discovery with Invalid input. (#5252, #5270)
- Workflow names resolve by exact name everywhere: archon workflow run, the chat /workflow command, the run API, workflow: children, include:, composed fan-outs and resume. archon workflow run review no longer selects archon-review; the error suggests the closest names instead. A workflow: or include: target that is not an exact name fails to load, so adding a workflow can no longer change which child runs. Project-name matching is unchanged. (#5259)
- Output field references accept properties declared through allOf, anyOf and oneOf, including nested compositions and schemas without root properties. A reference to a field that a composition-only schema declares nowhere now fails at load instead of being read schemaless. (#5255)
Changed
- SDLC build queue: before it continues an item, the answer checks the whole change against the issue's contract and asks when a correction would add machinery the contract does not name, naming the subtractive option. Review proposes removing machinery a change added before extending it. A Ready, green pull request whose only remaining check no round can run is asked about once instead of relaunched. The answer declares its question's class (stop or scope), and the queue records that class instead of deriving it from the round count. (#5266)
Fixed
- PostgreSQL: a boot that finds the schema current skips convergence, and so takes no lock and runs no DDL. Deadlock and serialization victims (40P01, 40001) retry, and concurrent boots on a fresh database recover from schema deadlocks, so a status query or timer tick no longer kills a live run with "deadlock detected". (#4543, #5249)
- A provider turn that stays silent after its result now fails at the node's idle_timeout instead of hanging the run indefinitely. (#5252)
- A usage-limit wait across a fan-out or loop group no longer resumes before the reset a provider reported. (#5250)
- Codex: a "no usable login" credential check is rechecked once in a fresh process, so concurrent runs sharing one refreshing login are not refused. Schema conversion keeps conditional oneOf/anyOf/allOf output schemas satisfiable, and a shape it cannot express is refused before the run with the keyword and path. (#5245, #5253)
- Abandoning a run keeps the checkout of a sub-run whose owner may still be live instead of deleting it. A crash before a container write-back apply no longer resumes as applied and discards the overlay. (#5251, #5243)
- Cleanup of an unregistered checkout, and terminal writes in the file store, no longer hold a lock while they delete files or walk artifacts. (#5238, #5242)
- A refreshed OAuth credential keeps its label, and a lost token rotation is reported as needing a reconnect instead of healthy. (#5240)
- When the newest of several servers on one Archon home stops, server.json is handed to a server still listening instead of being removed. (#5237)
- On Windows, a registered repository is no longer refused as "a different clone" when its path differs only in slash direction or drive-letter case. (#5244)
- A run blocked because another worktree holds its branch names that worktree. (#5258)
- Resume keeps the setup error when the run cannot be marked failed. Workflow run reads reject an unsupported stored status instead of acting on it. (#5241, #5265)
- A dropped SSE reconnect during replay keeps its buffered events. (#5239)
- Loop and fan-out cost totals keep a partial cost marked partial across iterations, resume and retries. (#5262)
- /status keeps its worktree totals when the base branch cannot be resolved. (#5263)
- A forge plugin file that cannot be inspected is reported instead of aborting discovery or silently disappearing. (#5264)
- Dry runs keep an explicit null stub, and workflow run --resume help says it also selects stale running runs. (#5261, #5260)
Installation
Quick Install (Recommended)
macOS / Linux
``bash
curl -fsSL https://archon.diy/install | bash
`
Windows (PowerShell)
`powershell
irm https://archon.diy/install.ps1 | iex
`
Homebrew (macOS / Linux)
`bash
brew install coleam00/archon/archon
`
Docker
`bash
docker run --rm -v "$PWD:/workspace" ghcr.io/coleam00/archon:latest workflow list
`
Manual Installation
macOS (Apple Silicon)
`bash
curl -fsSL https://github.com/coleam00/Archon/releases/latest/download/archon-darwin-arm64 -o /usr/local/bin/archon
chmod +x /usr/local/bin/archon
`
macOS (Intel)
`bash
curl -fsSL https://github.com/coleam00/Archon/releases/latest/download/archon-darwin-x64 -o /usr/local/bin/archon
chmod +x /usr/local/bin/archon
`
Linux (x64)
`bash
curl -fsSL https://github.com/coleam00/Archon/releases/latest/download/archon-linux-x64 -o /usr/local/bin/archon
chmod +x /usr/local/bin/archon
`
Linux (ARM64)
`bash
curl -fsSL https://github.com/coleam00/Archon/releases/latest/download/archon-linux-arm64 -o /usr/local/bin/archon
chmod +x /usr/local/bin/archon
`
Windows (Manual)
Download archon-windows-x64.exe from the assets below, rename to archon.exe, and add to your PATH.
Release assets
| Asset | Purpose |
| --- | --- |
| archon-<os>-<arch>[.exe] | CLI |
| archon-server-<os>-<arch>[.exe] | Version-matched server executable fetched by archon serve |
| archon-forge-github-<os>-<arch>[.exe] | GitHub forge plugin |
| archon-chat-slack-<os>-<arch>[.exe] | Slack chat plugin |
| archon-provider-claude-<os>-<arch>[.exe] | Claude provider plugin, installed by archon provider add claude |
| archon-provider-codex-<os>-<arch>[.exe] | Codex provider plugin, installed by archon provider add codex |
| archon-provider-pi-<os>-<arch>[.exe] | Pi provider plugin, installed by archon provider add pi |
| archon-web.tar.gz | Web UI |
| checksums.txt | SHA-256 checksums for every asset |
Verify installation
`bash
archon version
``
Archon CLI v0.11.02026-09-25
Workflow packs install straight from GitHub, workflows can start from schedules and GitHub events, forge plugins carry the delivery pack's pull-request work, and all three provider SDKs move to their latest releases.
Breaking changes: read before upgrading
These change behaviour you may rely on. Check each one against your setup before you upgrade.
- A repository's .archon/.env can no longer set ARCHONHOME, HOME, USERPROFILE, ARCHONDOCKER, WORKSPACE_PATH or PATH. Archon refuses to start and names the file and the key. Move the key to your shell environment or to ~/.archon/.env. This keeps one Archon home per process tree, so detached runs, trigger execute, detached approve/resume and CLI calls from workflow scripts read plugins from the same home as their parent. (#3501, #3506)
- assistants.<provider> in .archon/config.yaml (global and repo) is now validated when the config loads, by the provider's own parser. A value the provider would have silently dropped, such as modelReasoningEffort: extreme, a misspelled key, or a settingSources entry other than project or user, stops the load with a message naming the file, provider, key and accepted values. archon doctor reports it as a failed Config files check, and the settings API answers 400 naming the refused key. A settings change that fixes an invalid value already in ~/.archon/config.yaml is still saved. Two OpenCode keys are now refused: assistants.opencode.agent, which nothing read, and assistants.opencode.baseUrl, which every OpenCode run already rejected. (#2582, #3461)
- nonefailedminonesuccess now blocks a dependency that was skipped because something upstream failed, even when another dependency succeeded, including across chains, includes and resumed runs. A join that used to run after a failure now skips and keeps the original failed node as its skip cause. Condition skips and on_timeout: skip skips still count as admissible. (#3156, #3249)
- thinking: is removed. effort: is the only reasoning-depth setting in workflows, tiers, aliases and config, and any thinking: fails validation with a message pointing to effort:. The effort ladder gains a persistent rung. (#3094)
- Abandon and cancel now mean different things. A failed run is stopped and can be resumed; a cancelled run is discarded. archon workflow abandon, chat /workflow abandon, the Slack cancel button, the manage_run cancel action and POST /api/workflows/runs/{id}/abandon now stop a reachable detached owner before recording the run cancelled, so its worktree and resource slot are no longer released while the old process is still writing. If no owner answers, abandon still cancels and shows the recorded host, pid and last activity. If an owner answers but cannot be stopped, abandon fails and leaves the run unchanged (HTTP 409). A run executing inside a live Archon server is refused; cancel it from that server. (#2325, #3446)
- A when: condition that compares an object or array, from $node.output.field or $INPUTS.<name>, now fails the node with an actionable error instead of comparing JSON text and quietly skipping. (#2995, #3037)
- A plain loop.until or loop_group.until signal must now sit alone on the final line of the output. A mention inside prose, such as "not COMPLETE yet", no longer ends the loop. XML-wrapped signals work as before. (#2994)
- A workflow: node can no longer declare its own output_format. The child's returns: node owns the result contract, and a caller-side schema is a load error naming that node. (#3148)
Most likely to affect you: if a repository's .archon/.env sets ARCHONHOME, HOME, USERPROFILE, ARCHONDOCKER, WORKSPACE_PATH or PATH, Archon will now refuse to start in that repository. Move the key to your shell environment or to ~/.archon/.env. If you use thinking: in workflows or config, switch to effort:.
What's new
Added
- archon plugin install owner/repo[/path][@tag] installs a workflow pack from GitHub as one complete tree at one commit. archon plugin update, remove and list manage installed packs, and archon plugin copy <id> makes an editable copy under .archon/workflows/. (#2767, #3485)
- Installed packs run from the workflow catalog as owner/pack:entrypoint from the CLI, chat, the router, run management and the API. Only a pack's declared entrypoints can be dispatched. A run freezes the pack at the commit it started with, so update or remove affects only later runs. (#2835, #3486)
- Forge plugins: Archon can talk to a code forge through an installable executable plugin. Each release publishes the GitHub plugin, installed with archon plugin install coleam00/Archon/plugins/forge-github[@tag], including inside Docker. The delivery pack reads pull-request checks, including external commit statuses, through it, and with ARCHONSDLCFORGE=forge also creates, edits, comments on and flips pull requests through it, verifying each write by reading it back. (#3417, #3447, #3457)
- Event-triggered workflows: trigger bindings start ordinary governed runs from native macOS schedules and verified GitHub issue, pull-request, label, check and status webhooks. A binding names a shared resource and chooses skip or FIFO queue when a run already holds it, and queued work survives restarts. The server hosts triggers when ARCHONTRIGGERHOST is set; the CLI adds archon trigger fire, drain and whoami. See the workflow triggers guide. (#3415)
- Provider SDKs updated to latest: Claude Agent SDK 0.3.282, Codex SDK 0.157.0, Pi 0.87.1. Claude cost stays per query after a resume or fork, and Pi usage now includes cache-warming and compaction calls. (#3076, #3504)
- concurrency.providers.<provider-id>: N in ~/.archon/config.yaml caps simultaneous attempts per provider across every process sharing the database. There are no default caps; archon ai capacity release frees a slot held by an owner on another host. (#3460)
- archon workflow logs <run-id> [--follow] prints or follows a run's transcript, and workflow get shows its path. The transcript now marks resumes with a workflowresume row and records every gate approve, reject or response as a gatedecision row. (#3096, #3482)
- An interrupted run records why it stopped. After Ctrl-C, workflow get shows Stopped: interrupted by the operator (SIGINT) and the console labels the run Interrupted; it stays resumable. (#3502)
- Runs record the checkout they started from (checkout_baseline in workflow get and the API), and every node attempt records its starting checkout. (#3443, #3463)
- A workflow's authored outcome is shown beside its execution status in the CLI, console, chat, Slack and dry runs. Every terminal run keeps a structured record (status, outcome, node states, skip causes, artifact inventory) that workflow get and the API can read even when reporting nodes never ran. (#3082, #3110, #3254)
- A bash: or script: node can declare outputformat and certify its own JSON result, and a result may point at an artifact file with { "type": "archonartifact", "runid", "path" }. An outputformat schema that does not compile is a load error. (#3148, #3158)
- bash: and script: nodes accept on_timeout: skip, so optional work can time out without failing the run. (#3174)
- Workflows can declare an action-required durable wait. Delivery runs pause on inherited or environment CI failures and resume after an operator acts, and a concluded GitHub check wakes the waiting run right away. Waits report owner_lost when a run's owning process disappears. (#3115, #3172, #3175)
- Packaged Bun and Python scripts can import pack-local .shared modules, in source checkouts, global workflows and binary installs. (#3252)
- Script and bash nodes receive $TYPEDARTIFACTSFILE, a listing of typed artifacts grouped by outputtype with read errors reported, and ADOPTEDRUN_DIR for adopted runs