After each checkpoint commit, the metadata branch (checkpoint.json, manifest,
graph DOT, artifacts) is now pushed from the host process to the GitHub remote
using a GitHub App installation token. The local custom ref (refs/arc/{run_id})
is mapped to refs/heads/arc/meta/{run_id} on the remote since GitHub rejects
branch names starting with "refs/".
Changes:
- Move ssh_url_to_https to github_app.rs as pub fn for reuse
- Add push_ref() to git.rs for pushing a ref to an explicit URL
- Add github_app field to RunConfig to thread credentials into the engine
- Add git_push_meta_host() async wrapper in engine.rs
- Call git_push_meta_host after each remote checkpoint
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Introduces a [checkpoint] config table with exclude_globs in both run.toml
(per-run) and server.toml (defaults). Globs are merged (union + dedup) when
both are present. Non-empty excludes use git pathspec :(glob,exclude) syntax
to prevent staging matching files during checkpoint commits.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replaces flat ModelInfo fields with nested sub-structs (ModelLimits,
ModelFeatures, ModelCosts) and adds family, training, and
cache_input_cost_per_mtok fields to enrich the model catalog.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Filter engine-internal keys (internal.*, graph.*, thread.*, current*)
from the context diff returned by SubWorkflowHandler, preventing child
run state from overwriting parent values. Pass the parent's preamble
into the child context so child workflows have awareness of what the
parent already accomplished.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The extract_status_fields function recognized "outcome" as a field for
detecting status JSON objects but never read its value — LLM responses
like {"outcome": "fail", "failure_reason": "tests failed"} were silently
ignored and the outcome was always Success.
Now extract_status_fields reads the outcome field to set the node status
and failure_reason to populate the failure detail. Also adds a fallback:
if no routing directives are found in the response text, the handler
reads status.json from the sandbox CWD (written by agents that prefer
file output over inline JSON). Response text always takes priority.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
New `arc-exe` crate that runs agent tool operations inside ephemeral
exe.dev VMs via SSH. Uses two SSH connections: a management plane
(`ssh exe.dev`) for VM lifecycle and a data plane (`ssh vmname.exe.xyz`)
for command execution and file I/O.
Includes SshRunner trait with MockSshRunner for unit tests and
OpensshRunner for real SSH, with raw_mode for the exe.dev management
plane's custom command handler.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The active_stages map was only populated inside a TTY renderer guard,
so Plain mode never tracked counts. Move counters to a separate
stage_counts map that is always populated regardless of renderer.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
resolve_clone_credentials was short-circuiting for public repos,
returning no token. This broke git push from the sandbox since push
requires authentication regardless of repo visibility. Always generate
an installation access token.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Replace apt-get/NodeSource install (requires root) with direct Node.js
binary download to ~/.local (works as non-root daytona user)
- Run node install + npm install in single shell so PATH persists
- Add ~/.local/bin to PATH in env file and version check
- Fall back to stdout for error details when stderr is empty (Daytona
always returns empty stderr)
- Add e2e assertion that cli_stdout.log is written during poll
- Verified on Daytona with haiku: ensure_cli installs in 2s, full
workflow succeeds
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
AgentCliBackend now detects missing CLIs at runtime and installs them
on-demand (including Node.js via NodeSource if needed), removing the
need for custom Dockerfiles that pre-install CLI tools.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Remove hardcoded default_model_for_provider() from arc-workflows and
default_model() from arc-agent, delegating both to the catalog via
arc_llm::catalog::default_model_for_provider(). Add claude-sonnet-4-6
to catalog and move "sonnet"/"claude-sonnet" aliases to it.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Sync cli_stdout.log and cli_stderr.log to stage_dir each poll iteration
so there is visibility into what the CLI agent is doing before it finishes.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Daytona's POST /process/execute blocks until all descendant processes
exit. The backgrounded claude process kept the API hanging, causing a
60-second HTTP timeout. Using setsid creates a new session so the child
is fully detached — the API now returns in ~200ms.
Also touch the event emitter during the poll loop to prevent the stall
watchdog from killing the stage while waiting for claude to finish.
Falls back gracefully on macOS where setsid isn't available (not needed
since the local exec implementation doesn't wait for grandchildren).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Daytona's POST /process/execute is synchronous and blocks until the
command finishes. Long-running CLI agent sessions (claude, codex, gemini)
cause HTTP proxy timeouts. Replace the single blocking exec_command with
a background launch + poll pattern:
- Generate UUID-based temp file paths to avoid collisions between
concurrent CLI nodes
- Disable sandbox auto-stop before launching (new Sandbox trait method)
- Launch command in background, capture PID
- Poll every 5s for exit code file
- Read stdout/stderr from temp files after completion
- Cleanup temp files
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When a workflow runs on a Daytona sandbox, commits on the run branch
are lost when the sandbox is deleted. This pushes the run branch to
origin after each checkpoint so the full commit history is preserved.
- Upgrade GitHub App token permission from contents:read to contents:write
- Add refresh_push_credentials to Sandbox trait (default no-op)
- Store origin URL on DaytonaSandbox and configure push credentials after clone
- Implement refresh_push_credentials on DaytonaSandbox to rotate expiring tokens
- Add git_push_remote helper, called after each remote GitCheckpoint
- Add e2e test verifying the branch appears on GitHub after push
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Write provider API key to /tmp/arc_cli_env.sh and source it before
running the CLI tool (Daytona exec API doesn't support env vars)
- Add --verbose flag to claude command (required with -p + stream-json)
- Log cli_stdout.log and cli_stderr.log on failure for debugging
- Fall back to stdout in error message when stderr is empty
Known limitation: CLI backend on Daytona times out for long-running
claude sessions due to Daytona proxy execute API timeout.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Collect provider API key env vars and pass them to sandbox exec_command
so CLI tools can authenticate in the sandbox environment
- Wire up previously-ignored env_vars parameter in DaytonaSandbox, explicitly
set all ExecuteCommandOptions fields (catches new fields at compile time)
- Write stdout/stderr logs on CLI failure, fall back to stdout in error detail
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Skip start/exit nodes from all preamble modes (compact, summary high/medium/low)
- Exclude blank-string context values (e.g. failure_class: "") from context sections
- Emit section headers lazily so they don't appear when all content is filtered
- Exclude meta nodes from summary:high pipeline progress counts
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add !patch.is_empty() guards so per-node diff.patch and top-level
final.patch are only written when there are actual code changes.
Eliminates noise from empty patch files.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- CLI backend: show last 500 chars of stderr (not first), and include
the command itself when stderr is empty (e.g. exit code 127)
- Test runner: add PARALLEL env var for concurrent execution
- backend-demo.dot: use API backend since claude CLI isn't in sandbox
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The is_repo_public function called GET /repos/{owner}/{repo} with
the App JWT, but GitHub returns 401 for App JWTs on the repos
endpoint (they need an installation token). Previously this 401
was treated as an auth error, failing sandbox init.
Now 401 and 403 are treated like 404: assume private and proceed
to create an installation access token, which has the right perms.
Also add preflight phase to the DOT test runner.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Addresses agreed items from the openapi-hitl-and-run-outputs review:
rename retrieveRunDiff operationId, add 409s to steer/preview, bound
expires_in_secs, add selected_option_keys for multi-select end-to-end,
document skip/na semantics, add slug and require type on
RunVerificationControl, require file on CodeLocation, and remove the
checkpoint "all" sentinel in favor of omitting the parameter.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace raw TOML / UI-oriented SettingGroup responses with structured
JSON that mirrors the Rust config types (WorkflowRunConfig, ServerConfig).
- Add Serialize derives to all config types in arc-workflows and arc-api
- Add manual Serialize impl for DaytonaNetwork enum
- Update OpenAPI spec: /runs/{id}/configuration returns RunConfiguration,
/settings returns ServerConfiguration, WorkflowDetail.config is now
RunConfiguration object instead of TOML string
- Remove SettingGroup/SettingField/SettingFieldType schemas
- Update demo handlers to construct and serialize real config structs
- Regenerate TypeScript client
- Simplify settings page to JSON display, update run-configuration and
workflow-definition pages to show JSON
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Extract millis_u64 and save_json/load_json to shared crate root, replacing
identical copies in engine, parallel, human, conclusion, manifest,
checkpoint, and retro modules
- Remove exists() pre-checks before Manifest::load and Conclusion::load
in CLI runs scanner (TOCTOU anti-pattern)
- Merge duplicate cancel_run match arms for Queued/Starting/Running
- Change Conclusion.status from String to StageStatus enum
- Extract RunFilterArgs shared struct from RunsListArgs/RunsPruneArgs
- Replace hand-rolled formatDuration with formatDurationMs wrapper over
existing formatDurationSecs
- Extract duplicated ToolRow/ToolBlock components to shared tool-use.tsx
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Introduces a strongly typed Fidelity enum (Full, Truncate, Compact,
SummaryLow, SummaryMedium, SummaryHigh) with Display/FromStr for
string roundtripping and a degraded() method for checkpoint resume.
Updates all consumers: resolve_fidelity, build_preamble, PreambleTransform,
FidelityValidRule, and Context::fidelity() accessor now use the enum.
Invalid fidelity strings in DOT attributes are rejected at parse boundaries
rather than silently falling through to a default at usage sites.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Introduces context/keys.rs with 24 static constants, 4 prefix constants, and
4 helper functions for dynamic keys. Adds 6 typed accessor methods on Context
(run_id, fidelity, preamble, thread_id, node_visit_count, current_node_id).
Replaces all bare string literals across 13 files with constants/accessors.
Fixes bug in manager_loop.rs where "internal.node_visit" was read instead of
"internal.node_visit_count".
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace ad-hoc serde_json::Value construction/parsing with typed Manifest
and RunFinal structs, matching the pattern used by Checkpoint and Retro.
This gives compile-time guarantees for field access and eliminates
stringly-typed indexing.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add three end-to-end tests for the GitHub App Installation Access Token clone
flow (private repo clone, public repo optimization, not-installed error). Also
return an error from is_repo_public on non-404 HTTP failures instead of
attempting to parse the response as JSON.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Share a single reqwest::Client across GitHub API calls in resolve_clone_credentials
to reuse the TLS connection pool
- Accept raw PEM (not just base64-encoded) in build_github_app_credentials, matching
the existing decode_pem_env/decode_pem_value convention
- Move github_app into DaytonaSandbox::new() instead of cloning
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace `gh auth token` with GitHub App IATs scoped to `contents: read` for
Daytona sandbox git cloning. Public repos are auto-detected and cloned without
credentials. Private repos get short-lived, repo-scoped tokens. Clear error
messages for each failure mode (app not installed, suspended, no repo access,
auth failure). Falls back gracefully when no GitHub App is configured.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Per-node `max_visits` now overrides the graph-level `max_node_visits`
(and the dry-run default of 10) for individual nodes, giving tighter
control over specific loops like fix-and-verify cycles.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Allows literal $ signs in prompts and DOT files by writing $$. Also
refactors VariableExpansionTransform to use expand_vars instead of
string replace, fixing a substring-matching bug with $goal.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add DaytonaNetwork enum (block/allow_all/allow_list) with custom serde
Deserialize for TOML string-or-table syntax
- Wire network config through run_config defaults merging and base_params
- Document network access in sandboxing, environments, and run-configuration
- Fill in execution docs: checkpoints, environments, failures, interviews,
run configuration, observability, retros
- Rename compounding.mdx → retros.mdx, insights.mdx → observability.mdx
- Use DaytonaConfig::default() in tests to reduce boilerplate
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Aligns context key names with the handler name (CommandHandler), which was
previously inconsistent.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Extract shared strip+parse+trailing-check into parser::parse_ast()
so both parser::parse() and parse_command reuse it
- Accept impl Write in parse_command so tests verify actual JSON output
instead of re-parsing independently
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Parses a DOT file and outputs its AST as pretty-printed JSON, useful for
debugging and tooling. Adds Serialize/Deserialize to all AST types.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Replace useless format!() with .to_string() in parse_decision
- Derive Default for HookDecision instead of manual impl
- Use contains_key() instead of get().is_none() in semantic parser
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace freeform LLM text generation with schema-constrained
generate_object() for prompt hooks, eliminating the need for
JSON formatting instructions in the system prompt and the
code-fence stripping workaround.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace flat Vec<Clause> parser with AST-based recursive descent parser
supporting full operator precedence (&& binds tighter than ||, ! is prefix).
New operators: >, <, >=, <= (numeric), contains (substring/array membership),
matches (regex, validated at parse time). Simplify ConditionSyntaxRule to
delegate entirely to parse_condition(). Public API unchanged.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The pattern Some("agent") | Some("agent_loop") | Some("prompt") |
Some("one_shot") was duplicated across preamble.rs (3x) and
validation/rules.rs (1x). Centralizes into a single function in
graph/types.rs. Also fixes stale doc comment on default_registry.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Keep legacy aliases (agent_loop, one_shot) in the handler registry
and validation rules for backwards compatibility. Add codergen_mode
attribute support in the DOT parser, translating legacy values to
the new type names.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Rename handler type strings: codergen → agent_loop, wait.human → human,
script → command, wait.timer → wait
- Rename handler modules/structs to match: AgentHandler, HumanHandler,
CommandHandler, WaitHandler
- Split one_shot into PromptHandler (handler/prompt.rs) with shape=tab mapping
- Remove CodergenMode enum and codergen_mode attribute — one_shot is now its
own handler type, not a mode flag on the agent loop handler
- Update all demo DOT files: codergen_mode="one_shot" → shape=tab
- Update spec, README, validation rules, preamble, and hook tests
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Extract shared prompt/agent hook setup (model resolution, system prompt,
user message, timeout wrapper) into reusable helpers
- Fix blocking I/O: std::process::Command → tokio::process::Command for
host-mode hook execution
- Cache reqwest::Client per TLS mode via OnceLock instead of rebuilding
per HTTP hook call
- Return Cow from resolved_hook_type() to avoid cloning HookType on
every call
- Rename command_executor → executor in HookRunner
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Change HookExecutor trait to take Arc<dyn Sandbox> so agent hooks can
share the sandbox with the ToolRegistry via ToolContext
- Replace hand-rolled 2-tool dispatch with register_core_tools() giving
agent hooks the full tool set (read_file, write_file, shell, grep, glob)
- Add strip_code_fences() to handle LLMs wrapping JSON in markdown
- Set max_tokens(1024) on prompt hooks to avoid exceeding model limits
- Add e2e tests: TOML parsing, prompt proceed/block, agent proceed,
agent with tool use (reads a file via read_file tool)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Prompt hooks make a single-turn LLM call returning {"ok": true/false}.
Agent hooks run a multi-turn LLM tool loop with sandbox access (exec_command, read_file).
Both fail-open on errors/timeouts. Prompt hooks default to 30s timeout,
agent hooks to 60s with max 50 tool rounds. Default model is "haiku".
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
