Workflow absorb and task replay

Done

• Workflow DSL exists in JSON and YAML.
• Workflow engine supports triggers, agent nodes, transforms, branches, rules, actions, user tasks, subprocesses, signals, timers, checkpoints, end nodes, retries, and maxChildDepth.
• Workflow CLI supports authoring, validation, running, tracing, pausing, resuming, and canceling runs.
• Task traces are stored in SQLite with task, step, workflow, token, duration, and error fields.
• Workflow definitions now support generated-workflow metadata: status, tags, entity, intentPath, generatedFrom, sourceTaskIds, confidence, workflowVersion, ownerAgent, lastMatchedAt, and matchCount.
• Workflow drafts are stored under .agentx/workflows/_drafts/ and are never registered as live triggers.
• New CLI surfaces:
  • agentx workflow draft-from-task <taskId>
  • agentx workflow absorb
  • agentx workflow drafts
  • agentx workflow promote <draftId>
  • agentx workflow reject <draftId>
  • agentx workflow replay-task <taskId>
• A conservative workflow matcher seam exists behind workflows.matching.enabled; v1 logs suggestions and falls back to normal agent execution.

Pending

• Add a workflow index once active workflows exceed roughly 100 definitions.
• Persist replay comparison metrics: original vs replay duration, usage, model, status, and output summary.
• Add richer clustering signals: intent graph path, action sequence similarity, file/action targets, and outcome category.

Shipped recently

• LLM-assisted workflow architect (src/workflows/architect.ts) — when --model <id> is passed to agentx workflow draft-from-task or agentx workflow absorb, the trace + step list is sent through the Anthropic API with a forced emit_workflow tool_use. The model decomposes the procedure into action.run, transform, branch, etc. nodes; the result is validated against workflowSchema (with one retry on schema failure); deterministic single-node fallback runs on any LLM error. Requires ANTHROPIC_API_KEY set in the daemon's environment.
• Dashboard edit-then-replay loop — drafts are editable in the workflow detail panel; "Save" persists to disk, "Save & Replay" fires an ad-hoc replay against the first source trace's input.
• workflows.matching.mode: "auto" fires matched workflows directly via dispatcher.dispatchWorkflow; high-confidence matches return an empty AgentResponse with metadata.handledByWorkflow so the agent doesn't double-reply.

Recommended cron

Operators wire scheduled workflow mining via crons in agentx.json. The cron only generates drafts — promotion stays explicit:

crons:
  workflow-absorb-nightly:
    enabled: false
    schedule: "0 2 * * *"
    agent: workflow-architect
    timeout: 900
    prompt: >
      Run agentx workflow absorb --since 24h --min-cluster-size 3 --max 10 --commit.
      Report generated drafts, skipped clusters, and validation failures.

Flip enabled: true once an agent is configured for the role and the operator has reviewed the first hand-run drafts.

Decisions

• Workflow DSL is the canonical representation for reusable task execution.
• YAML is preferred for generated drafts because it is easier to review.
• JSON remains supported and editor-compatible.
• Generated workflows default to status: draft and state: disabled.
• Scheduled mining produces drafts only; no generated workflow is activated automatically.
• Workflow absorb mirrors the wiki absorb philosophy, but it uses workflow draft storage rather than wiki article storage.
• v1 matching is conservative and non-blocking; normal agent execution remains the fallback.

Acceptance Criteria

• Existing YAML and JSON workflows remain valid.
• Draft workflows validate but never auto-trigger.
• agentx workflow absorb --dry-run writes nothing and exits cleanly on empty trace sets.
• agentx workflow draft-from-task <taskId> --commit writes a valid draft with task provenance.
• agentx workflow promote <draftId> validates and activates the workflow explicitly.
• agentx workflow reject <draftId> archives the draft without deleting evidence.
• agentx workflow replay-task <taskId> can run an ad-hoc workflow from a task trace without promoting it.
• Workflow matching can be disabled globally and never blocks normal task execution in v1.