fix(resume): replay original event log into dashboard on --web (#167)

[jrob5756]

Summary

Fixes #167. conductor resume … --web / --web-bg previously opened
an empty dashboard because the original run's events lived in a
different process and were never carried forward across the checkpoint
boundary. Now the dashboard's history is fully seeded before the
server starts accepting clients, so prior agents stay visible alongside
the live resumed run.

Changes

Checkpoint schema (engine/checkpoint.py, engine/workflow.py)

• Added run_id: str = "" and event_log_path: str = "" to
  CheckpointData (defaults preserve backward-compat for older
  checkpoints, no version bump needed).
• WorkflowEngine._save_checkpoint_on_failure forwards
  self._run_context.run_id and self._run_context.log_file.

Dashboard replay (web/server.py)

• WebDashboard.prepend_workflow_started(data) — inserts a fresh
  workflow_started (built from the current YAML via
  engine.build_workflow_started_data()) at position 0 of
  _event_history so historical events apply to the correct topology.
• WebDashboard.replay_events_from_jsonl(path) — line-by-line replay
  of the original JSONL log, tolerating malformed/partial lines via
  the existing web/replay.py::_load_events parser. Skips root-level
  workflow_started / workflow_completed / workflow_failed /
  checkpoint_saved events (detected by absent data.subworkflow_path),
  preserving subworkflow lifecycle events so per-context wfDepth
  stays balanced.
• WebDashboard.replay_synthetic_from_context(ctx, config, ts) —
  fallback that synthesises type-aware *_started / *_completed
  pairs from WorkflowContext.execution_history. Parallel /
  for-each shapes include success_count / failure_count /
  elapsed so the frontend doesn't mis-render groups as failed.
• All three methods append directly to _event_history (no
  _queue.put_nowait); they emit logger.warning if invoked after
  dashboard.start().

Engine (engine/workflow.py)

• Extracted WorkflowEngine.build_workflow_started_data() from the
  inline _execute_loop emit so both the engine and the CLI can build
  the same event shape.
• WorkflowEngine.suppress_workflow_started_emit() sets a flag that
  _execute_loop checks; suppresses the engine's own emit on resume
  so the dashboard sees exactly one root start (avoids wfDepth
  double-counting).

Event log continuity (engine/event_log.py)

• EventLogSubscriber.__init__ accepts optional existing_path and
  existing_run_id kwargs; when both are provided and the file is a
  writable regular file, opens it in append mode and reuses the
  run_id. Falls back to a fresh log on OSError so resume never fails
  due to a locked / readonly log file.

CLI wiring (cli/run.py)

• resume_workflow_async reordering: load config and checkpoint → open
  registry → build EventLogSubscriber (append-mode when available) →
  build engine with the resolved run_id / log_file →
  dashboard.prepend_workflow_started(...) →
  dashboard.replay_events_from_jsonl(...) (or synthetic fallback) →
  engine.suppress_workflow_started_emit() →
  dashboard.start() + print URL → subscribe console + event log →
  engine.resume().

Docs

• Updated AGENTS.md "Run / Resume Parity" note.
• Added Unreleased CHANGELOG entry.

Why it's correct

Two rubber-duck reviews caught blind spots that informed the final
design:

1. Single root workflow_started — The frontend distinguishes root
   vs subworkflow workflow_started by wfDepth (not by
   data.subworkflow_path). Without suppression, the resumed engine
   would emit a second root workflow_started, which the frontend
   would treat as a child workflow and routes the resumed run's
   events into a phantom child context. Now exactly one root event
   exists.

2. Topology before historical events — Frontend
   agent_started / parallel_agent_completed etc. assume topology
   (groups, routes) has already been initialised by workflow_started.
   Replay now emits the topology event first, so historical events
   land on the right nodes.

3. No mid-replay races — Replay populates _event_history
   completely before dashboard.start(). The very first
   GET /api/state and the first WebSocket connect both see the
   full history.

Tests

• 2669 passing (15 skipped), up from baseline 2646.
• New / updated tests:
  • test_engine/test_checkpoint.py — run_id + event_log_path
    round-trip; back-compat defaults to "".
  • test_engine/test_event_log.py — append-mode reuse of
    existing_path / existing_run_id; missing-path + missing-run_id
    fallbacks.
  • test_engine/test_resume.py — engine forwards RunContext values
    into checkpoint; build_workflow_started_data() shape;
    suppress_workflow_started_emit() skips engine emit on resume.
  • test_web/test_server.py — replay populates /api/state; skips
    root lifecycle events; preserves subworkflow lifecycle events;
    does not enqueue on _queue; handles missing/corrupt/partial
    files; synthetic replay emits per node type with correct fields;
    prepend_workflow_started inserts at position 0.
  • test_cli/test_resume_command.py — resume_workflow_async seeds
    the dashboard with JSONL replay when the log exists, falls back to
    synthetic events when it doesn't.

Out of scope (deferred)

• Dashboard banner visually marking the resume boundary in the UI.

How to verify locally

make check    # lint + typecheck
make test     # full suite (2669 passed, 15 skipped on this branch)

Reproducer from the issue:

uv run conductor run examples/some-workflow.yaml --web --input ...
# Trigger a failure to save a checkpoint, then:
uv run conductor resume examples/some-workflow.yaml --web
# Open the dashboard — prior agents are now visible.

Closes #167.

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com