Adds a temporary local HTTP server in `arc install` that handles the
GitHub App manifest registration flow, enabling standalone users to
register a GitHub App without running the arc-web server.
Also gates server-only install steps (pre-flight checks, server.toml,
secrets/certs, start servers) behind #[cfg(feature = "server")] and
removes the `gh` CLI from system dependency checks. Reloads .env before
running arc doctor so it picks up freshly written credentials.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Self-hosted runner lacks credentials for HTTPS checkout.
Switch both compile (Linux) and release jobs to ubuntu-latest.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Live service probes now run by default instead of requiring --live.
Added --dry-run flag to skip probes when needed. Install command now
runs doctor with live probes to validate setup.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Remove rust-toolchain/rust-cache setup (pre-installed on self-hosted),
disable checkout clean, enable sccache and incremental compilation.
Split macOS tests into a separate job that only runs on workflow_dispatch
since release.yml already gates on macOS tests.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Extend conclusion.json with per-stage operational data (cost, duration,
retries) so the Arc Details section of PR descriptions no longer depends
on retro.json. The retro section remains optional and only appears when
the retro agent runs.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add release.yml: tag-triggered (v*) workflow that cross-compiles
for aarch64-apple-darwin and x86_64-unknown-linux-gnu, packages
archives with SHA256 checksums, and creates a GitHub Release
- Add workflow_dispatch, concurrency groups, and cache-on-failure
to rust.yml and typescript.yml
- Run clippy and test on both ubuntu-latest and macos-15
- Add release profile with thin LTO and strip to Cargo.toml
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Project-level arc.toml can now set [pull_request] with enabled/draft
fields, slotting between cli.toml defaults and workflow.toml overrides.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add app_id(), slug(), and git_author() methods to CliConfig to
replace repeated .git.as_ref().and_then(...) chains across 5 call
sites in main.rs and doctor.rs.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Move GitHub App credentials (app_id, slug) and run defaults (llm,
sandbox, pull_request) into cli.toml so standalone commands (run, pr
create, llm) no longer require server.toml. The server.toml loading is
now limited to `arc serve` and server-mode doctor checks.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The one-time environment setup command (LLMs, certs, GitHub) is now
`arc install`, freeing up naming space for a future `arc init` command
for per-project initialization.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Surfaces existing final.patch and per-node diff.patch files via CLI.
Supports live diffs from in-progress runs by reconnecting to the sandbox.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
ubuntu-latest-16-cores requires a GitHub Team/Enterprise org plan and
was causing all jobs to stay queued indefinitely.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When no explicit provider is set on a request, resolve_provider now
consults the model catalog (e.g. gpt-5.4 → openai) before falling back
to the client's default provider. This fixes PR description generation
sending cross-provider model IDs to the wrong provider.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Standalone subcommand that creates SSH credentials for a run's Daytona
sandbox and connects directly (or prints the command with --print).
Also parameterizes create_ssh_access TTL instead of hardcoding 60 min.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
In verbose mode, completed tool call bars accumulated in indicatif's
MultiProgress without limit. After ~80+ bars exceeded the terminal
height, indicatif's cursor repositioning broke and all rendering stopped.
Now in verbose mode, completed tool calls are printed as permanent output
via multi.println() and their bars are removed from MultiProgress. Only
running tool calls remain as live bars with spinner + elapsed time. This
keeps MultiProgress small regardless of how many tool calls execute.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Provides CLI access to preview URLs for ports exposed by Daytona sandboxes.
Supports standard (URL + token + curl), signed (self-contained URL), and
--open (browser launch) modes.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Remove --fail-with-body from curl so the response body (including error
JSON) is always captured. Previously, set -e killed the script at the
curl line on non-2xx responses before the error-checking code could run,
leaving the agent with an opaque exit code 22 and no actionable message.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace Gemini API with OpenAI /v1/images/generations endpoint in the
imagegen script. Remove unsupported response_format parameter (gpt-image-1
returns b64_json by default). Add standalone imagegen-openai workflow for
testing image generation independently.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Per-run data now lives in ~/.arc/runs/ while daily CLI log files
stay in ~/.arc/logs/. Renames: logs_root → run_dir (RunConfig field,
Handler trait param, all handlers), logs_dir → run_dir (CLI arg,
local variables), default_logs_base → default_runs_base (path fn).
Adds DB migration 002 to rename workflow_runs.logs_dir → run_dir.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds a switchable tracing layer that writes to ~/.arc/logs/{run-dir}/cli.log,
so all diagnostic info for a run is co-located with its other artifacts instead
of buried in the shared daily log file.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The Playwright MCP server provides a browser_install tool, so we can
drop the separate command node and let the agent install the browser
via MCP instead.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Display collected assets in an "=== Assets ===" section after the run
result, so users can see which files were captured without digging
through the logs directory.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Use npx directly instead of npm install, which fails due to the repo's
workspace:* protocol in package.json.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The GitHub App check was gated behind #[cfg(feature = "server")] so it
never ran in CLI-only builds. It also only checked whether env vars were
set, not whether the key could actually sign JWTs.
- Move GithubAppStatus and check_github_app out of the server feature gate
- Try sign_app_jwt to validate the private key, catching corrupt keys
- Show app_id and slug in doctor output
- Keep server-only fields (client_id, client_secret, webhook_secret)
behind #[cfg(feature = "server")]
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Port the OpenAI Codex develop-web-game skill (Playwright-based game dev
loop) into skills/develop-web-game/ with paths adapted for Arc's
workspace-relative skill discovery. Add a web-game-demo workflow that
verifies the skill and its asset files (playwright client script,
action payloads JSON) are properly discovered and accessible in a
Daytona sandbox.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Arc API, Arc Web, GitHub App, and cryptographic key checks are only
relevant when the server feature is enabled. Without this gate, `arc
doctor` showed misleading connectivity errors for services that cannot
even be started.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Return worktree_path from the match expression instead of mutating
a pre-initialized variable.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When the sandbox does not support preview URLs (e.g., local sandbox),
fall back to http://localhost:{port} instead of erroring. This makes
the sandbox MCP transport work transparently on both local and remote
(Daytona) sandboxes.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- docs/agents/mcp.mdx: Remove pre-release warning. Add sandbox transport
docs alongside stdio and HTTP. Restructure configuration section to
cover both cli.toml and run config TOML. Replace filesystem example
with Playwright browser automation example.
- docs/execution/run-configuration.mdx: Add [mcp_servers] section with
field reference for all three transport types. Add to full example.
- docs/reference/cli-configuration.mdx: Add sandbox transport subsection.
Cross-reference workflow-level MCP config.
- docs/integrations/daytona.mdx: Add MCP sandbox transport to feature table.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Wire MCP server configs from workflow TOML into agent sessions so
MCP tools are available to agent stages during workflow runs.
Changes:
- run_config.rs: add mcp_servers HashMap to WorkflowRunConfig
- backend.rs: AgentApiBackend.with_mcp_servers() passes configs to SessionConfig
- run.rs: extract mcp_servers from RunConfig and pass to backend
- demo/mod.rs: add default mcp_servers field to struct literals
New workflow: arc/workflows/playwright-demo/
- Runs on Daytona with daytona-medium snapshot
- Configures Playwright MCP as a sandbox transport (port 3100)
- Agent browses Hacker News, takes screenshots at each step
- asset_globs captures screenshots/** into logs directory
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Adds McpTransport::Sandbox variant that starts an MCP server inside the
sandbox, waits for it to listen, gets a Daytona preview URL, and rewrites
to HTTP transport for the MCP client connection.
Key changes:
- Sandbox trait: add get_preview_url(port) for authenticated port access
- DaytonaSandbox: implement get_preview_url via Daytona SDK preview links
- McpTransport::Sandbox: new variant with command, port, env fields
- Session: resolve_sandbox_mcp_servers() starts servers and rewrites to HTTP
- Integration test: end-to-end Playwright MCP in Daytona - navigates to
example.com and verifies "Example Domain" in accessibility snapshot
Config example:
[mcp_servers.playwright]
type = "sandbox"
command = ["npx", "@playwright/mcp@latest", "--port", "3100", "--headless"]
port = 3100
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
