EPIC: v0.9.0 WhaleFlow branch/leaf workflow mode

[Hmbown]

Goal

Build CodeWhale's own WhaleFlow workflow module/mode: a typed branch-and-leaf workflow runtime that can run background workflow pods, fan out bounded agents/leaves, replay traces deterministically, and promote only validated lessons into a cached-main overlay.

This is similar in product scope to Claude's workflow mode as observed in the transcript: a user can ask for a workflow, the system starts a background workflow, shows done/total progress, lets the user inspect it via a workflow monitor, and returns with a diagnosis/plan before writing code. CodeWhale should implement the concept in a CodeWhale-native way: typed IR, Rust executor, Starlark authoring, ARMH/RLM memoization, TraceStore, and teacher/student validation.

Mental model

• rooted trunk: canonical repo/history/user-visible main state
• working branch: isolated speculative strategy or implementation path
• leaf: bounded cognition/action inside a branch
• fruit: useful artifact produced by a branch/leaf, including losing branches
• cached-main overlay: validated reusable knowledge/workflow/test layer for future runs
• teacher: privileged trace reviewer proposing reusable lessons
• student: constrained model proving those lessons are useful
• janitor: stale invalidation, cache cleanup, candidate demotion

Scope

The epic should decompose into small PRs: IR, executor, authoring, ARMH/RLM leaves, provider abstraction, replay, teacher review, student promotion, cached-main overlay, janitor, and CLI/TUI workflow-mode surface.

Non-goals

• No model-weight RL in v0.9.0.
• No arbitrary JS/Python scripts as source of truth.
• No script-level async/await.
• No provider-specific workflow logic.
• No uncontrolled self-modifying agent.

Source: /Users/huntermbown/Downloads/CodeWhale_v0_9_0_BranchLeafTeacher_Workflows_UPDATED.md and the Claude workflow-mode observation from June 3, 2026.

[Hmbown]

WhaleFlow v0.9.0 issue stack created/mapped:

• WhaleFlow: typed workflow IR and TraceStore state migrations #2668 WhaleFlow: typed workflow IR and TraceStore state migrations
• WhaleFlow: Rust executor skeleton for branch/leaf workflows #2669 WhaleFlow: Rust executor skeleton for branch/leaf workflows
• WhaleFlow: Starlark authoring layer, repair loop, and compile gate #2670 WhaleFlow: Starlark authoring layer, repair loop, and compile gate
• WhaleFlow: ARMH/RLM leaves with shared branch memoization and telemetry #2671 WhaleFlow: ARMH/RLM leaves with shared branch memoization and telemetry
• WhaleFlow: model-provider registry and role-based model policy #2672 WhaleFlow: model-provider registry and role-based model policy
• WhaleFlow: deterministic replay from recorded leaf outputs #2673 WhaleFlow: deterministic replay from recorded leaf outputs
• WhaleFlow: TeacherReview GEPA candidate generation from traces #2674 WhaleFlow: TeacherReview GEPA candidate generation from traces
• WhaleFlow: StudentReplay and PromotionGate for validated lessons #2675 WhaleFlow: StudentReplay and PromotionGate for validated lessons
• WhaleFlow: cached-main overlay and janitor for promoted artifacts #2676 WhaleFlow: cached-main overlay and janitor for promoted artifacts
• WhaleFlow: evaluate Aleph-style external memory as a default context substrate #2677 WhaleFlow: evaluate Aleph-style external memory as a default context substrate
• WhaleFlow: CLI/TUI workflow mode and pod monitor surface #2678 WhaleFlow: CLI/TUI workflow mode and pod monitor surface
• WhaleFlow: ship rlm_cache_change.star as the first dogfood workflow #2679 WhaleFlow: ship rlm_cache_change.star as the first dogfood workflow

Existing draft PRs to inspect/harvest into this stack:

• feat: add WhaleFlow — declarative multi-agent workflow orchestration #2482 feat: add WhaleFlow — declarative multi-agent workflow orchestration
• Feat/whaleflow cost tracking #2486 Feat/whaleflow cost tracking

Implementation policy: use one PR per logically related slice, target the v0.9.0 integration branch, and keep scripts pure: Starlark builds IR, Rust executes IR.

[Hmbown]

Added a related v0.9.0 investigation issue for the retrieval gap surfaced from the semantic/sparse embedding discussion: #2680 WhaleFlow: hybrid semantic code search and retrieval substrate.

This should stay distinct from #2671 ARMH/RLM memoization and #2677 Aleph-style external memory: #2680 is specifically about model-facing concept/code retrieval over the repo using auditable hybrid search signals.