GraphQL client with 6 public functions: execute_graphql, fetch_viewer_id,
create_comment, update_issue_state, fetch_candidate_issues, and
fetch_issues_by_ids. Includes Issue/BlockerRef domain types with manual
normalization from GraphQL JSON responses.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The inherent method shadowed the identical Sandbox trait method,
creating a risk of subtle divergence if one were changed without
the other.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds a generic SSH sandbox that connects to any user-provided host via
openssh, without VM lifecycle management. Supports git clone with GitHub
App credentials, configurable working directory, and arc cp reconnection.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Extract duplicated compaction check+trigger blocks in the session loop into
Session::compact_if_needed(). Replace the hardcoded 20_000 token budget with
COMPACTION_USER_MESSAGE_TOKEN_BUDGET constant. Change extract_recent_user_messages
to take ownership of discarded turns instead of borrowing and cloning.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Reframe the summarization prompt as creating a handoff for a different assistant
(rather than self-summarization) to produce more thorough, explicit summaries.
Add a pre-turn compaction check so context is trimmed before building the LLM
request, not just after. Extract recent user messages from discarded turns during
compaction so the model retains the user's actual words alongside tool call/result
pairs.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
A workflow run hung for 33 minutes in a Daytona shell command despite
a 5-minute client-side timeout in tokio::select!. The root cause is
unknown — tests confirm the select pattern works correctly. Adding
info-level tracing at method entry (with command and timeout_ms),
after process service acquisition, and in each select arm so the
next occurrence reveals exactly where execution stalls.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Enable serde_json's preserve_order feature so Map uses insertion-order
IndexMap instead of sorted BTreeMap. Extract build_event_envelope()
from the on_event closure to deduplicate the envelope construction.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Mirrors `arc validate` workflow resolution. Shells out to `dot` (graphviz)
with friendly error if not installed. Supports --format svg|png and -o output path.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Wire ProgressUI and EventEmitter into the retro agent so it shows
stage spinners with nested tool calls, matching the look of agent
nodes in the main workflow.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Previously, AgentError::Aborted was a bare variant that displayed as
just "Aborted", making it impossible to tell whether a session was
killed by wall clock timeout, user cancellation, or something else.
Now it carries an AbortReason (WallClockTimeout or Cancelled) so
terminal output like the retro agent failure message includes the cause.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The command takes several seconds (especially with live probes) but
previously showed no output until fully complete. Now a braille spinner
animates while checks run.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Embeds the arc-create-workflow skill files via include_str! and installs them
to user (~/.claude/skills/ or ~/.agents/skills/) or project-level directories.
Supports --dir (claude|agents), --for (user|project), and --force flags.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Regenerate 2026-03-10 changelog with full commit set (installer, devcontainer
support, arc pr subcommands, --no-dotenv removal). Add 2026-03-11 changelog
for shadow branch run records and arc diff improvements. Update docs: CLI
reference (arc pr list/view/merge/close, --shortstat), devcontainers page
(replace placeholder with full content), run-configuration (devcontainer
field), checkpoints (per-node files on metadata branch), models catalog
(gemini-3.1-pro-preview-customtools). Improve preflight to show per-model
LLM checks instead of a single merged provider line.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Remote sandboxes (Daytona, Exe) clone into their own environment, so
creating a local git worktree is unnecessary. The worktree decision
previously only checked worktree_mode and git_clean, ignoring the
sandbox provider entirely.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When workflow prompts tell agents to write results to a file with
embedded JSON status fields, the engine now falls back to reading the
last file the agent wrote if neither the response text nor status.json
contain extractable status fields.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Store execution trace data (prompts, responses, status, diffs, etc.) on
the shadow branch alongside checkpoint/manifest data. This makes the
shadow branch a complete, self-contained record of each run.
- scan_node_files() scans nodes/ for allowlisted files (<512KB)
- Checkpoint commits now include per-node files
- init_run includes sandbox.json
- Finalize commit after workflow writes retro.json + final node files
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Use std::path::Path::strip_prefix instead of manual string manipulation,
and simplify redundant format! wrapper.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Change parallel branch entries from tool-call style (6-space indent,
spinner) to subagent info style (8-space indent, ▸ prefix, dim/static
while running). Keeps ✓/✗ glyphs on completion.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Drop `HEAD` from the live diff command so `git diff {base_sha}` compares
against the working tree instead of only committed changes. Extracts
`build_live_diff_cmd` for testability.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Tool output previously showed absolute sandbox paths (e.g.
write_file(/home/daytona/workspace/output/js/physics.js)) because
shorten_path() only checked the host CWD. Now ProgressUI stores the
sandbox working directory and strips it from displayed paths, producing
cleaner output like write_file(output/js/physics.js).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When `arc run` uses a Daytona sandbox, `initialize()` re-detects the
branch from cwd — but by that point the worktree has switched HEAD to
`arc/run/{run_id}`, which was never pushed, causing the clone to fail
with HTTP 400. Pass the already-known base branch (e.g. `main`) from
the caller instead of re-detecting it inside the sandbox.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Remove `devcontainer_env()` and `resolve_devcontainer()` trivial wrappers,
inline at call sites
- Fix initializeCommand Args quoting: use shlex::try_quote instead of naive join
- Fix initializeCommand silently dropping Parallel commands: now executes them
- Convert three repetitive lifecycle calls to a data-driven loop
- Use crate::millis_u64() in setup command loop for consistency
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When `devcontainer = true` in TOML sandbox config, Arc now resolves
devcontainer.json from the repo, uses its Dockerfile for the Daytona
snapshot, runs lifecycle hooks (on_create, post_create, post_start)
in the sandbox, and merges devcontainer env vars under TOML env.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Model list tests (6): exact stdout snapshots in .trycmd files, replacing
insta .snap files. Error validation tests (5): TOML with stderr-to-stdout
checking specific error messages. Doctor tests (3): TOML with env.inherit
= false for env_clear(). Validate tests (7): TOML with fs.cwd/sandbox for
file path access. Dry-run tests (6): TOML with fs.cwd/sandbox. Exec help
and no-prompt tests (2): TOML with env config.
Remaining in cli.rs: negative-assertion tests, file-system side-effect
test, tmpdir+env test, and 11 ignored API-key-dependent tests.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Remote sandboxes clone from origin, so "clean" should mean no uncommitted
changes AND the local branch is pushed. Replaces the hardcoded `false` with
a new `ensure_clean_and_pushed` that reuses `ensure_clean` + `branch_needs_push`.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Activate the dormant Segment-based telemetry infrastructure by wiring
lifecycle events in main.rs. Each CLI invocation now emits a "Command Run"
or "Command Error" event with sanitized args, duration, hashed repo
identifier, and CI detection.
- Add sanitize.rs: redacts sensitive CLI args (paths, values) while
keeping flags, booleans, and numeric patterns
- Add git.rs: MD5-hashes normalized origin remote URL so HTTPS/SSH
variants produce the same identifier
- Extend command_name to include nested subcommands (e.g. "pr create",
"llm prompt", "system prune", "model list")
- Restructure main/main_inner to capture timing and return command name
alongside result for telemetry
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Covers run, validate, cp, preview, ssh, model, doctor, init, install,
pr, and system. Each test verifies --help exits 0 and produces output.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Daytona's Rust SDK doesn't support uploading build context files, so
COPY/ADD instructions referencing local files fail at build time with
confusing errors. Detect this early in resolve() and return a clear
UnsupportedCopyAdd error. Multi-stage COPY --from= and ADD http(s)://
are still allowed.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace assert_eq!/contains() chains with insta::assert_snapshot! in
dockerfile unit tests (11 inline snapshots) and CLI model/help tests
(7 file-based snapshots), making test output easier to read and update.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The file was only deleted after successful deserialization + send.
Wrap the read/deserialize/send in an async block so remove_file
runs regardless of where the error occurs.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace tokio::spawn fire-and-forget with a detached subprocess pattern:
send() serializes Track to ~/.arc/tmp/, spawns `arc __send_analytics <path>`
with ARC_TELEMETRY=off, and the child sends to Segment then cleans up the file.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Rename PullRequestDetail fields from user_wrapper/head_wrapper/
base_wrapper to user/head/base (serde renames no longer needed)
- Make GITHUB_API_BASE_URL public, replace hardcoded strings in pr.rs
- Replace opaque 5-tuple with PrRow struct in pr_list_from
- Parallelize GitHub API calls in pr_list using futures::join_all
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add PR lifecycle commands that operate on workflow runs via their
pull_request.json records. Each loads the PR record from a run directory
and calls the GitHub API: list scans all runs and fetches live state,
view shows details, merge supports squash/merge/rebase, close sets
state to closed.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Signoffs are stamps of approval for (control, repository, commit SHA)
tuples, enabling external tools and humans to mark a control as
pass/fail/pending for a specific commit. Adds OpenAPI spec, demo
handlers with 5 fixtures, route registration, and regenerated TS client.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The local .env loading was unnecessary since arc install writes all
config to ~/.arc/.env. Simplifies dotenv behavior to unconditionally
load from the home directory only.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When compiled without the server feature, these args are useless since
there is no server to delegate to. Wraps the CLI fields, parser,
resolve_mode, build_server_client, and related tests in cfg(feature =
"server"), with standalone-only fallback paths in each command handler.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Fix /tmp/arc-commit-msg race: use unique path per run_id+node_id for concurrent local runs
- Add shell_quote for author.name/email interpolated into shell commands
- Unify git_push_meta_host and git_push_run_host into a single git_push_host helper
- Remove unused git_head_sha_remote function
- Extract repeated !sandbox.is_remote() && git_checkpoint_enabled into local_git_checkpoint
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
All git checkpoint, diff, push, worktree, and merge operations now go
through sandbox.exec_command() instead of maintaining parallel Host
(spawn_blocking) and Remote code paths. The enum is replaced by
git_checkpoint_enabled + host_repo_path on RunConfig, and
sandbox.is_remote() is used where behavior still differs.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Split preflight into Repository (setup commands, git clean) and Workflow
(nodes, edges, goal, model, provider) sections. Merge LLM providers and
Provider checks into a single LLM check. Add terminal-width-aware
truncation for long detail lines. Render "Git clean: false" in red.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
