Finance Research Dashboard

New Task

Title Details Plan and rank useful data before gathering

Difficulty Task Type

Codex Claude

Launch selected agents now

Tasks

How Dashboard Research Works

The dashboard does not automatically start work when a task is created. It first stores task state, creates a workspace, and waits for a launch request. Logs only appear after Codex or Claude is actually started for that task.

create only launch required for logs review when process exits cleanly

Step By Step

Create task in the UI. The browser sends the title, details, task type, difficulty, workflow flag, selected engines, and the launch checkbox.
Server creates task storage. It makes data/tasks/<task-id>/, then writes task.json, workspace/, and logs/.
Optional research workflow is seeded. If the workflow checkbox is on, the task gets planning status, default subtasks, and an empty ranked source list.
If launch is off, the task stays idle. This is the state your flowchart task was left in, so its log stayed empty.
Launch starts a run. The UI calls POST /api/tasks/<id>/launch when you use Launch Codex, Launch Claude, or create with launch enabled.
The server queues a run record. It stores attempt metadata, target engine, log path, and status in task.json.
A background thread builds the prompt. The prompt includes the task description, workspace path, finance research instructions, and workflow update instructions.
The CLI process is spawned. Codex runs through codex --search exec. Claude runs through claude -p.
The dashboard opens the engine log. It writes a run header with timestamp, cwd, and command to logs/<engine>.log.
Process output streams into the log. Stdout and stderr are redirected into that log file for the whole run.
The agent works inside the task workspace. It writes report.md and any supporting files under workspace/.
The agent can self-update the workflow. It can call POST /api/tasks/<id>/workflow to update subtasks, source status, ranks, raw links, synthesis notes, and events.
The UI keeps polling. The browser reloads task state every 5 seconds and refreshes the selected engine log, so the page reflects run progress.
Workspace files become visible. Anything written under the task workspace shows up in the file list for direct review.
Exit updates final run state. Exit code 0 becomes ready_for_human. Nonzero exit codes become error.
The task can continue later. The continue form launches a new run with a continuation prompt for the same task.

Flow Chart

Create task in UI
        |
        v
POST /api/tasks
        |
        v
Create task folder + task.json + workspace/ + logs/
        |
        v
Workflow enabled?
   yes / \ no
      /   \
     v     v
Seed subtasks   Skip workflow scaffold
        \       /
         v     v
      Launch requested?
        |
   no   |   yes
   |    |    |
   v    |    v
Task stays    POST /api/tasks/<id>/launch
pending       |
              v
        Queue run in task.json
              |
              v
       Build prompt for engine
              |
              v
     Spawn Codex or Claude CLI
              |
              v
     Open logs/<engine>.log and stream output
              |
              v
   Agent researches in workspace and writes report.md
              |
              v
 Agent POSTs workflow updates as it gathers evidence
              |
              v
     UI polls tasks/logs and updates the page
              |
              v
        Process exits with code
          /            \
         v              v
exit 0 -> ready     nonzero -> error

Empty log usually means one of two things: the task was never launched, or the selected engine has no run record yet.