Cy Orquestrate Issue

3. Create and bind the executor

Refresh the remote base without changing the root checkout. Use the repository's target branch, normally origin/main. Keep a compact mapping of issue URL, base SHA, worktree ID/path/branch, session ID, prompt idempotency key, and eventual PR URL in the orchestration context. Preserve that mapping in a handoff when needed; daemon state remains authoritative.

Perform these dependent operations in order:

1. compozy__worktree_create: workspace, base ref, unique name such as issue-<number>-<slug>, branch such as fix/issue-<number>-<slug>.
2. compozy__worktree_inspect: confirm creation reached ready, and retain the returned path/ID.
3. compozy__session_create: selected agent, descriptive issue name, exact workspace/worktree ID.
4. Load assets/executor-prompt.md, fill every field, and send the full briefing with compozy__session_prompt. Set runtime explicitly on this first prompt: {"provider":"codex","model":"gpt-6-astra","reasoning_effort":"high"} unless overridden. Use unique stable message_id and idempotency_key; use wait:false for background dispatch.
5. compozy__session_status: confirm the prompt is accepted, runtime ready with the requested effective selection, and the session bound to the intended worktree. The executor verifies cwd and branch before editing. Inspect compozy__goal_control status to confirm actual activation; distinguish requested, activated, and executing successfully.

If creation or prompt admission has an ambiguous result, reconcile the worktree/session catalog against the recorded mapping before retrying. Replay the same prompt idempotency key where supported; do not blindly create duplicates. Respect session limits; never change the session type to bypass a denial. This flow uses logical worktree-bound sessions, not the spec-task spawn/lease workflow.

If a required operation fails, complete independent work and report the precise blocker. Do not claim a pending worktree, unbound runtime, or merely accepted prompt is executing successfully.

4. Follow-ups and delivery

Send additional evidence only to the same issue's recorded executor. Inspect its current turn and busy-input capabilities: queue nonurgent additions; steer only when timely correction is needed, using the expected active turn and checking whether the runtime would interrupt. Do not assume a slash-shaped string in an agent prompt activates a skill or Goal; require the native operation.

The executor prompt owns implementation, validation, PR description, and the complete review cycle. Do not duplicate its checks or routinely interrupt it for status. If supervising, investigate finished CI failures while other jobs run and inspect new review evidence after relevant pushes. An idle turn or active Goal flag alone is neither completion nor a blocker. If Goal supervision fails while ordinary work continues, preserve the evidence, avoid repeated activation, and continue authorized work; do not repair the host or clear runtime records as a shortcut.

Report the issue links, worktree/branch and session mapping, verified runtime/Goal state, and actual launch or delivery status concisely. Do not stop the executor or remove its worktree after dispatch. At delivery, verify the PR head, required checks, both reviewers' completed coverage, all finding dispositions, and current description. Missing/red/pending CI or missing reviewer evidence means delivery is still in progress or externally blocked. No merge unless separately requested.