WhaleFlow: Starlark authoring layer, repair loop, and compile gate

[Hmbown]

Problem

Models should be able to author workflows, but raw arbitrary scripts are unsafe and unreliable. CodeWhale needs a small Starlark dialect that compiles to IR and fails closed.

Scope

• Add Starlark parser/bindings for branch, rlm, agent, test, review, search, shell, ctx.parallel, ctx.sequence, ctx.tournament, ctx.loop_until, ctx.when, ctx.expand, ctx.teacher.review.
• Add dialect spec prompt for model-authored workflows.
• Reject imports, classes, while loops, recursion, async/await, direct filesystem/network/provider calls.
• Add safe mechanical rewrite pass for common generated Pythonisms, starting with f-string rewrite where intent is obvious.
• Add one reflect-and-repair pass using parser/validator errors.
• Add compile gate: invalid workflows never run.

Acceptance criteria

• starlark_compiles_to_ir passes.
• starlark_repair_loop passes.
• invalid_workflow_rejected passes.
• Example workflows include rlm_cache_change.star and issue_fix_tournament.star.

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

[Hmbown]

v0.9 stewardship progress update: I opened #2824 as the first narrow Starlark authoring/compile-gate slice.

Scope in #2824:

• compile model-authored .star workflow files into WorkflowSpec
• reject unsupported constructs before evaluation (load, import, class, while, async, await, open)
• add one-pass repair for common ctx.* generated aliases
• add starlark_compiles_to_ir, starlark_repair_loop, and invalid_workflow_rejected tests
• add the named example files workflows/rlm_cache_change.star and workflows/issue_fix_tournament.star

Still intentionally out of scope: live workflow_run, runtime execution, provider calls, TraceStore writes, replay, and TUI workflow mode. The issue should stay open after this slice until those surfaces are proven safe.

Local verification before opening: cargo test -p codewhale-whaleflow --locked, cargo clippy -p codewhale-whaleflow --all-targets --locked -- -D warnings, cargo fmt --all -- --check, ./scripts/release/check-versions.sh, ./scripts/release/check-ohos-deps.sh, and git diff --check.

[Hmbown]

v0.9 stewardship update: #2837 landed as a narrow follow-up to the Starlark compile-gate slice.

What changed:

• Starlark-authored workflows now fail closed when teacher-review candidates reference unknown node ids.
• The typed IR validator also rejects unknown leaf depends_on_results, reducer inputs, and teacher-review candidates before mock execution or replay.
• Dynamic Expand generated nodes still use shape validation only, so we do not accidentally over-specify runtime/generated-source semantics before the workflow runtime is ready.

Verification before merge: cargo test -p codewhale-whaleflow --locked, cargo clippy -p codewhale-whaleflow --all-targets --locked -- -D warnings, cargo check --workspace --all-features --locked, cargo fmt --all -- --check, ./scripts/release/check-versions.sh, ./scripts/release/check-ohos-deps.sh, and git diff --check.

Leaving this issue open because runtime execution, provider calls, TraceStore writes, replay commands, and TUI workflow mode are still intentionally deferred. Thanks @AdityaVG13 for the broader WhaleFlow draft direction that shaped these safe crate-only slices.

[Hmbown]

Fixed by the v0.9 Starlark authoring/compile-gate slices now merged into codex/v0.9.0-stewardship: #2824 added the Starlark graph-builder compiler, one-pass repair loop, compile gate, and the named example workflows; #2837 tightened fail-closed reference validation for generated/teacher-review workflows.\n\nAcceptance evidence on current stewardship head efbcc681a:\n- workflows/rlm_cache_change.star exists and is compiled by starlark_compiles_to_ir.\n- workflows/issue_fix_tournament.star exists and is covered by issue_fix_tournament_example_compiles.\n- starlark_repair_loop passes.\n- invalid_workflow_rejected passes.\n- Focused local verification: cargo test -p codewhale-whaleflow starlark --locked passed, 7 tests / 0 failures.\n\nI’m closing this issue because its written acceptance criteria are now satisfied. The broader runtime work remains tracked separately in the WhaleFlow execution/replay/TUI issues, especially #2669, #2673, #2678, and #2679. Thanks @AdityaVG13 for the broader WhaleFlow draft direction that shaped these safe crate-only slices.