Context
From the WhaleFlow vision-alignment assessment (docs/WHALEFLOW_VISION_ALIGNMENT.md). WhaleFlow's goal is ultracode-style orchestration with heterogeneous-model workers (DeepSeek / GLM / MiniMax / Moonshot-Kimi / OpenAI) plus a kimi-code–style swarm mode.
Problem
Every credible swarm (kimi-code, Claude agent-teams, Anthropic's orchestrator–worker) coordinates through shared state — a task list with claim-by-status, dependency ordering, idle-reassignment, and failed-retry — not free-form chat. WhaleFlow has none of that live today: the only real fan-out is the parent model voluntarily calling agent_open.
The kicker: crates/whaleflow already contains the IR for exactly this — WorkflowPlan models a phase DAG, cross-phase result-dependency ordering, parallel-write-scope conflict validation, and IsolationMode::Worktree. But it is orphaned: zero workspace consumers, mock execution only.
Proposal
Make the Fleet ledger the shared task substrate that a swarm/goal run reads and writes:
- claim-by-status (pending → in-progress → done / blocked)
- dependency ordering (consume the
WorkflowPlan phase DAG)
- idle-worker reassignment + failed-worker retry
- parallel-write-scope conflict checks (already modeled in the IR)
…and have it consume the whaleflow IR instead of leaving it dead.
Scope
- Minimal slice (could land in 0.8.61): Fleet ledger as a shared task list + wire the
whaleflow IR to one real consumer (no longer mock execution).
- Full scheduler (0.9.0): dependency scheduler, reassignment, retry, conflict gating.
Relationships
This is the single biggest architectural lever for honoring both the ultracode and kimi-code-swarm visions.
🤖 Filed from the WhaleFlow vision-alignment analysis.
Context
From the WhaleFlow vision-alignment assessment (
docs/WHALEFLOW_VISION_ALIGNMENT.md). WhaleFlow's goal is ultracode-style orchestration with heterogeneous-model workers (DeepSeek / GLM / MiniMax / Moonshot-Kimi / OpenAI) plus a kimi-code–style swarm mode.Problem
Every credible swarm (kimi-code, Claude agent-teams, Anthropic's orchestrator–worker) coordinates through shared state — a task list with claim-by-status, dependency ordering, idle-reassignment, and failed-retry — not free-form chat. WhaleFlow has none of that live today: the only real fan-out is the parent model voluntarily calling
agent_open.The kicker:
crates/whaleflowalready contains the IR for exactly this —WorkflowPlanmodels a phase DAG, cross-phase result-dependency ordering, parallel-write-scope conflict validation, andIsolationMode::Worktree. But it is orphaned: zero workspace consumers, mock execution only.Proposal
Make the Fleet ledger the shared task substrate that a swarm/goal run reads and writes:
WorkflowPlanphase DAG)…and have it consume the whaleflow IR instead of leaving it dead.
Scope
whaleflowIR to one real consumer (no longer mock execution).Relationships
crates/whaleflow.This is the single biggest architectural lever for honoring both the ultracode and kimi-code-swarm visions.
🤖 Filed from the WhaleFlow vision-alignment analysis.