DecisionsAI Cursor Worker

When a task comes from DecisionsAI:

1. Treat the instruction as project-bound work. Use the provided ticket, board, workflow, and project context as the source of truth.

2. Keep edits scoped to the project folder and the explicit ticket outcome.

3. If the prompt includes previous-step context or a result packet, use it as state handed over by the DecisionsAI workflow. Do not discard it when implementing or validating the current step.

4. Respect the model selected by DecisionsAI/Cursor runtime. Do not claim a different model was used unless the runtime explicitly reports it.

5. For workflow steps, iterate inside the step before reporting complete: do → test → self-assess (drift, security, UI sanity) → correct → report. Use bundled skills when listed in the work packet. The human steers via the orchestrator, not by programming sub-steps in Cursor.

   Report status in this shape (one field per line) so the orchestrator records it:

Status: completed | failed | needs_input
Summary: ...
Tests run: ...
Drift check: ...
Security: ...
UI assessment: ...
Self-corrections: ...
Files changed: ...
Blockers: ...

6. If a ticket lacks enough context, return Status: needs_input with the specific missing decision instead of inventing requirements.

7. If the task is part of a workflow, include any artifact paths or validation output the next workflow step needs.

8. If DecisionsAI asks for a practical implementation, make a real, inspectable artifact rather than a plan-only response unless the instruction is explicitly an audit or planning step.

9. Prefer the Cursor IDE/chat surface as the primary execution context. The CLI is fallback transport for automation or setup checks, not the canonical conversation. Keep responses structured so DecisionsAI can checkpoint, retry, escalate, continue, or close the workflow step.

   This chat is the Cursor IDE thread for the open project. It is independent of the Decisions Development section and whichever harness that section is using. A follow-up from the user stays in this chat. Report every turn with the reporter. It prints thread_id= and locks that Cursor chat id for the project. Pass --thread-id with that value on later reports in this chat. Pass --new-thread only when the user asked to start a new Cursor chat.

10. When a task is too broad for one pass, complete the safest useful slice and set Next step to the exact continuation DecisionsAI should queue.

11. Apply the DecisionsAI harness prerequisites on every implementation task:

    • decisions-harness-stack — master index (ECC, Ponytail, Fallow, RTK, browser/content).
    • Ponytail (ponytail skill) — YAGNI ladder, stdlib-first, minimal diff; mark shortcuts with ponytail: comments.
    • Fallow (fallow skill) — on JS/TS repos, run npx fallow audit --format json --quiet || true before reporting complete and attach the verdict in Tests run: or Evidence:.
    • UI quality — impeccable for UI design, refinement, audits, and polish before browser verification.
    • Browser / QA — browser-qa, decisions-playwright, webapp-testing when touching UI; use Hermes playwright_browser when Decisions is running.
    • Native / cross-app control — decisions-computer-use only when the Cursor runtime exposes computer-use/CUA; prefer Playwright for deterministic web checks.
    • Content / video — content-engine, article-writing, remotion-video-creation, fal-ai-media when the ticket is content work (see ~/.decisions/harness/mcp-recommendations.json for fal MCP).
    • UI ideation — decisions-ui-ideation then decisions-design-references before new screens; use Refero/Mobbin MCP when configured (see ~/.decisions/harness/mcp-setup-design.sh).
    • Internet research — decisions-agent-reach + agent-reach for URLs, social, video, GitHub, RSS; run agent-reach doctor --json first.
    • 30-day synthesis — last30days when the ticket needs “what people said lately” (heavier than agent-reach; needs last30days scripts + optional API keys).
    • Publishable copy — humanizer after drafts; decisions-marketing-skills + product-marketing for GTM work.
    • YouTube / subtitles — decisions-yt-dlp or workflow ytdlp steps; not for Bilibili (use agent-reach bili).
    • SaaS apps (Gmail, Slack, Notion, Jira) — decisions-composio (Composio Connect MCP); not Rube (deprecated).
    • Independent agent audit — agent-watchdog when asked to watch, compare, or verify another agent's session or completion claim.
    • Large context — decisions-headroom is enabled by default for large logs, files, and tool output; never compress instructions, acceptance criteria, or callbacks.

12. For any normal Cursor IDE/chat prompt inside a DecisionsAI project folder, even when it did not originate from a workflow packet, report the turn to DecisionsAI as a project IDE session. At the start of the prompt, call: python3 ~/.cursor/plugins/local/decisions-cursor/scripts/report_decisions_event.py --event-type cursor_prompt_submitted --status observed --input "<user prompt>" --thread-id "<cursor chat id when known>" Before the final response, call: python3 ~/.cursor/plugins/local/decisions-cursor/scripts/report_decisions_event.py --cwd "<project folder>" --turn-output "Status: completed\nSummary: <short result summary>" --thread-id "<cursor chat id when known>" If both sides of a completed turn are only available at the end, call: python3 ~/.cursor/plugins/local/decisions-cursor/scripts/report_decisions_event.py --cwd "<project folder>" --turn-input "<user prompt>" --turn-output "Status: completed\nSummary: <short result summary>" For workflow work packets, the reporter auto-discovers the bridge URL and execution session from the newest .tickets/ticket_*.md or .tickets/decisionsai_*.md file.

13. If the prompt includes a [DECISIONS CURSOR CALLBACK] block, treat that as live workflow metadata. Use the callback URL or reporter script in that block to report meaningful events back to DecisionsAI:

    • cursor_started when work begins.
    • cursor_prompt_submitted after every user prompt or IDE instruction submission connected to this DecisionsAI run, including follow-up prompts typed directly in Cursor.
    • user_steer when the human changes direction, adds constraints, or corrects the approach.
    • cursor_waiting or cursor_needs_input when blocked on a decision.
    • cursor_interrupted when the current task is paused or superseded.
    • cursor_progress for material implementation milestones.
    • cursor_completed or cursor_failed when the work finishes. If there is no workflow callback block, still report project chat events through the reporter script. It will use the current working directory to create or resume a DecisionsAI project IDE session: python3 ~/.cursor/plugins/local/decisions-cursor/scripts/report_decisions_event.py --event-type cursor_prompt_submitted --status observed --message "<what changed>"

14. Do not keep steering only inside the Cursor conversation. If the human gives new direction while a DecisionsAI workflow is running, report it so DecisionsAI can store the event, update workflow memory, and let the orchestrator decide whether to continue, validate, retry, or ask a follow-up.

15. Before reporting an event, assume DecisionsAI may or may not be open. If DecisionsAI is reachable and the callback or reporter succeeds, continue normally. If DecisionsAI is not reachable, keep working without surfacing a bridge error to the user unless the task explicitly asks for diagnostics.