fix up command

[khaliqgant]

Summary

Collection of CLI and broker fixes addressing entrypoint symlink resolution, agent management improvements, and deduplication correctness.

Changes

CLI Entrypoint

• Fix symlink resolution in CLI entrypoint guard — uses fs.realpathSync to resolve symlinks before comparing paths, with path.resolve fallback (src/cli/index.ts)
• Update cliScript fallback from bootstrap.js to index.js to match the single-entrypoint model (src/cli/commands/core.ts)
• New test suite validating entrypoint behavior and package.json bin mapping (src/cli/entrypoint.test.ts)

Agent Management

• Forward release reason in HTTP agent management client — sends { reason } in DELETE body (src/cli/commands/agent-management.ts)
• Refactor agent management to use HTTP-based client with broker auto-discovery, port scanning, and auto-start capability
• Always resolve API port in broker lifecycle, even without dashboard (src/cli/lib/broker-lifecycle.ts)

Broker (Rust)

• Fix dedup key priority for spawn requests — prefers event_id before falling back to agent name, preventing false-positive deduplication when multiple spawns share a name (src/main.rs)
• Auto-format Rust code with cargo fmt

CI/CD

• Add standalone macOS smoke test to package-validation workflow
• Add ci-standalone-smoke.sh script for standalone binary validation
• Update publish workflow

SDK

• Various test fixes and updates for lifecycle, workflow runner, and integration tests
• Python SDK client improvements

Known limitation

The release reason is now sent from the TS HTTP client, but the Rust listen_api.rs DELETE handler does not yet parse the request body — so the reason is silently dropped for HTTP-based releases. The WebSocket-based release path (release_agent in main.rs) correctly handles the reason field. A follow-up PR should add body parsing to the Rust HTTP endpoint.

Commits

• d1c770e7 fix up command
• ca63089e test fixes and updates for lifecycle
• 4fde8ed5 style: auto-format Rust code with cargo fmt
• 738cb7c7 fix: resolve symlinks in CLI entrypoint guard
• 89ca732a fix: forward release reason in HTTP agent management client
• 2d2bf7f4 fix: dedup key priority — event_id before agent name for spawn requests

Breaking changes

None.

Testing done

• All 584 TypeScript tests pass
• All 233 Rust tests pass
• CLI --version and --help work correctly
• Symlink resolution verified manually (ln -sf ... && node /tmp/test-relay-symlink --version)
• who command works against running broker
• release --help confirms command structure

🤖 Generated with Claude Code

[khaliqgant]

Bug report from the latest full integration rerun:

The original agent-relay up / installer / stale-broker fixes are in, but the full broker integration suite is still not green because workflow "local mode" is incomplete.

Root cause:

• WorkflowRunner sets AGENT_RELAY_WORKFLOW_DISABLE_RELAYCAST=1
• but it still constructs AgentRelay with channels: [channel]
• AgentRelayClient always launches the broker with init --channels ...
• the broker then still attempts Relaycast registration / workspace calls
• so tests that should be purely local still fail against https://api.relaycast.dev/...

Observed failure pattern:

• after moving some workflow suites to useRelaycast: false, they still fail with broker exit / Relaycast workspace errors
• this shows the disable flag currently stops JS-side channel setup, but does not stop broker-side Relaycast usage

Representative failing suites after the rerun:

• workflow-dag
• workflow-lifecycle
• workflow-verification
• parts of workflow-runner
• parts of workflow-patterns

There are also a couple of likely real behavior mismatches behind the infrastructure noise:

• workflow-dag: detects dependency cycle expects a rejected execution path, but the runner currently returns a failed run instead
• workflow-lifecycle: abort cancels a running workflow currently completes instead of cancelling

Repro:

RELAY_INTEGRATION_CLI_LIFECYCLE=1 node --test tests/integration/broker/dist/*.test.js

Recommended follow-up:

1. Make workflow local mode truly local end-to-end: when Relaycast is disabled, do not start the broker with Relaycast channels / workspace registration.
2. Rerun the workflow integration slice in local mode.
3. Fix the remaining semantic regressions (dependency cycle, abort cancels) once the Relaycast dependency is removed from those tests.

[khaliqgant]

Fixed in 79b1616. Here's what was wrong and what changed:

Root cause: When AGENT_RELAY_WORKFLOW_DISABLE_RELAYCAST=1 was set, the runner still passed channels: [channel] to the AgentRelay constructor, which caused the broker to run init --channels <channel> and attempt Relaycast registration.

Fixes applied (all in one commit on fix-up):

1. Don't pass channels when Relaycast is disabled — runner.ts now passes channels: [] when relaycastDisabled is true
2. Skip --channels arg when empty — client.ts no longer emits --channels '' which confused the broker
3. Validate config in execute() — DAG cycle detection (detectCycles) was only called via validateConfig during YAML parsing, not in execute(). Added this.validateConfig(resolved) so assert.rejects() works correctly
4. Initialize abortController early — it was set in runWorkflowCore but abort() could be called before that method ran (race condition). Moved to execute()/resume() so abort works immediately
5. Mark pending steps as failed on cancel — when abort fires, pending/running steps now emit step:failed events
6. Fix continue error strategy — allCompleted check now treats failed steps as done when using continue/skip strategy, so the run completes instead of failing
7. Fix verification test false positive — changed task text to avoid triggering the anti-injection check (taskHasToken path)

All 28 tests across workflow-dag, workflow-lifecycle, workflow-runner, workflow-verification, and workflow-patterns pass.