4.13.0: plans are short intents, the lead chooses where agents run

[cdowin]

Closes #317
Closes #316
Closes #310
Closes #311
Closes #299
Closes #318
Closes #319
Closes #320
Closes #321
Closes #322
Closes #323
Closes #324
Closes #325
Closes #326
Closes #327

What changes

• Plans are a short ordered list of intents (Plans are a short ordered list of intents; the worker's code is the detail #327). plan takes the lead's intent list as the graph with 0 agents, or drafts one with 1 agent. Wave and flow give a task with no brief a self-brief: the worker writes a few lines, then the code, and returns its brief in the report. heavy: true keeps the old path (brief-writers, specs, critic) for a one-way door.
• The lead chooses where agents run (flow and wave: the lead chooses where each agent runs; workflows name work, not places #316, contract: a runtime may dispatch its agents elsewhere; slot cap and load are limits of where agents run #320, wave, flow and split: with a remote runtime, prompts name work, not places #325). runtime.remote marks a runtime whose agents run elsewhere; prompts then name the work (branch, base, brief, oracle, gate, report) and never a local path; the slot cap comes from the runtime, and the host load check does not apply.
• Prompts state the task is the lead's request (wave: worker and sub-lead prompts state that the task is the lead's request (agents saw a chat message and refused the task) #310, wave: worker and sub-lead prompts state that the task is the lead's request (agents saw a chat message and refused the task) #324).
• One source for the helpers wave and flow share (flow: shared helpers copied from wave.js drift silently; and 4 review minors #311, flow: the helpers it shares with wave come from 1 source and cannot drift #319), with a drift check in tests/workflows.js.
• A null brief retries once instead of crashing plan (plan: a brief-writer that returns null crashes the run at the Briefs phase #318); an escalated plan keeps its specs (plan: a joint spec escalation returns the specs of the green chains #321).
• Flow fix landings go through the landing lane; a group ships only after its reviews (flow: a holds fix lands through the landing lane, and ship waits for a batched review in flight #326). Short-id stale checks (wave: a short task id in prose does not mark a task stale #323). Behaviour tests for prune and reach (tests: wave-prune and wave-reach check behaviour, not only prompt text #322).

Proof

• sh tests/run.sh: 318 passed, 0 failed. node tests/workflows.js: 1285 passed, 0 failed.
• A cold review ran every workflow's prompts from main and from this branch under the same test scenarios (1917 prompt pairs, local runtime) and diffed them: every difference traces to wave: worker and sub-lead prompts state that the task is the lead's request (agents saw a chat message and refused the task) #324, flow: a holds fix lands through the landing lane, and ship waits for a batched review in flight #326 or Plans are a short ordered list of intents; the worker's code is the detail #327; with heavy: true the plan prompts are byte-identical to main. With runtime.remote, no prompt names a worktree root or git worktree add.

Not verified

• No live run of the light plan path or the self-brief yet; Plans are a short ordered list of intents; the worker's code is the detail #327's measure (under 3 min from agreed work to first code) is measured on the next real run.
• The brief-retry prompt (plan: a brief-writer that returns null crashes the run at the Briefs phase #318) was not prompt-diffed (its test has its own stub).

🤖 Generated with Claude Code