PR #TBD — One Dispatch Path: Queue Launcher Stops Reimplementing Spawn

The Broader Goal at the End of the Previous PR

Fable's external review (blog-adjacent strategy doc, PR #188) laid out a multi-year arc for synlynk as the OS for multi-agent development. One concrete finding was independent of the local-agent MLX work: the daemon/scheduler queue path and the interactive dispatch path had silently forked. Fleet scheduler v1 (synlynk schedule --execute) is designed to call _dispatch_ready_jobs as its execution primitive — so any isolation or flag gap there becomes a fleet-wide failure class, not a niche daemon bug.

Strategic Shifts in This PR (if any)

No goalpost move. The review agreed we should not delete _dispatch_ready_jobs (it owns priority ordering, dependency gating, and concurrency caps that dispatch_agent correctly does not absorb). The fix is collapse-to-one-execution-implementation: keep the scheduler brain, delete the second spawn body.

Cost-accounting hardcoding at the reconcile path remains out of scope (#189).

What This PR Shipped

dispatch_agent(..., job_id=None)

Optional job_id lets the queue path hand the pre-existing daemon_jobs primary key into the shared launcher. When the row already exists, the DB write is an UPDATE (preserves priority, depends_on, enqueued_at) instead of INSERT OR REPLACE defaults.

_dispatch_ready_jobs delegates spawn

The inline block that built cli + non_interactive_flags only, skipped _preflight_dispatch, and Popen'd sh -c with no cwd is gone. Per ready queue row it now calls:

dispatch_agent(agent, task, story_id=story_id, force_agent=True, job_id=job_id)

Scheduling logic (priority ORDER BY, failed/unmet depends_on, max_parallel slots) is unchanged. Preflight/worktree failures mark the queue row failed so the daemon does not spin forever.

Regression test

test_dispatch_ready_jobs_creates_worktree_and_applies_dispatch_flags asserts that a queued row launched via _dispatch_ready_jobs:

• creates the job worktree layout (worktrees/<job_id>/.synlynk/{logs,prompts})
• applies Claude dispatch_flags (--dangerously-skip-permissions)
• sets Popen(..., cwd=worktree_path)
• records worktree_path / worktree_branch on the jobs.json entry under the same job_id

Priority / dependency / concurrency tests continue to pass.

On Track Toward Full Autonomous Multi-Agent Dispatch

The scheduler and the interactive CLI no longer have two half-copies of "how to run an agent." When fleet dispatch launches concurrent cross-agent jobs, each one gets worktree isolation and the same permission/dispatch flag translation as a human synlynk dispatch. That is a prerequisite for safe parallel execution under synlynk schedule --execute.

New Goalpost

• Land #190; confirm daemon + schedule paths dogfood the unified launcher.
• #189 still open for cost-accounting rate-table bypasses (including the reconcile path next to this code).
• Local-agent / MLX driver work can assume queue isolation is real rather than re-litigating it.