run directory semantics, status interpretation, and practical troubleshooting steps

Run Artifacts and Debugging

Every task execution creates a run directory:

.downcity/task/<taskId>/<timestamp>/

Core artifacts

File	Purpose
`input.md`	definition snapshot for this run
`output.md`	output body (`agent`: final task output text; `script`: script output)
`result.md`	human-readable summary
`run.json`	final status payload
`run-progress.json`	live progress snapshot
`dialogue.md`	multi-round summary (`agent`)
`dialogue.json`	structured round records (`agent`, including query / output / raw assistantMessage snapshots / feedback)
`messages.jsonl`	execution context message stream (including executor query / assistant reply)
`user-simulator/messages.jsonl`	user-simulator context message stream (only when `review=true`, including simulator query / assistant reply)
`error.md`	failure summary

resultStatus=invalid
- usually output too short
- for review=true agent tasks, inspect dialogue.md / dialogue.json first for each round's:
  - executor query
  - executor output
  - user simulator query
  - user simulator output
  - raw assistantMessage snapshot
  - feedbackForNextRound
executionStatus=failure
- usually script execution failure, agent process failure, failed chat_send delivery, or an agent run that only performed tool calls without producing any user-visible output
- for ACP agents such as Claude ACP, task now normalizes known startup / hook / transport internal errors into a stable failure message first; if the same task keeps failing, inspect agent-server logs and hook configuration
Task skipped (already running)
- overlap trigger skipped by per-task serial guard
one-shot when=time:... no longer auto-runs
- expected behavior after execution: paused + @manual