You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Pipeline block ① (Generate: prompt → workflow IR). Part of the compile-to-weights epic.
Context. There is no surface that turns a natural-language workflow description into a stargraph graph. authoring.DurableGraph (authoring.py:218-438) synthesizes an IRDocument from caller-supplied functions + an edge map, but at compile time only — no LLM generation, no prompt input. The compile-to-weights story starts with "given a prompt for a workflow, generate the workflow."
Proposal. New GraphAuthorSkill that emits a validatedIRDocument from an NL prompt.
Backbone: the existing DSPy/LiteLLM seam (adapters/dspy.pybind(), nodes/dspy.py), default Claude — schema-constrained generation.
Validate: IRDocument.model_validate(...) → construct Graph → structural_hash (graph/hash.py:6) → a verify pass (node-kind resolvability via cli/run.py:_resolve_node_factory:260, rule/goto target integrity). Reject + repair-loop on invalid IR.
Human-review gate before anything downstream compiles (generated graph is reviewed, then frozen by hash).
Invalid generations trigger a bounded repair loop, not a crash.
Generated graph is content-hash-frozen and surfaced for human review before compile.
Caveat. Open-ended generation can emit unresolvable node kinds / dangling goto targets — the verify pass is mandatory, not optional. Relates to #43 (KG-augmented authoring), #42 (discovery registry).
Source: arXiv:2605.22502v1 — "Compiling Agentic Workflows into LLM Weights" (subterranean agents). Part of the end-to-end compile-to-weights pipeline epic; file refs verified against current main. Determinism: per-graph knob (fully-compiled | hybrid | rules-only) defaults to honor #117 (no learned core router).
Pipeline block ① (Generate: prompt → workflow IR). Part of the compile-to-weights epic.
Context. There is no surface that turns a natural-language workflow description into a stargraph graph.
authoring.DurableGraph(authoring.py:218-438) synthesizes anIRDocumentfrom caller-supplied functions + an edge map, but at compile time only — no LLM generation, no prompt input. The compile-to-weights story starts with "given a prompt for a workflow, generate the workflow."Proposal. New
GraphAuthorSkillthat emits a validatedIRDocumentfrom an NL prompt.adapters/dspy.pybind(),nodes/dspy.py), default Claude — schema-constrained generation.IRDocument.model_validate(...)→ constructGraph→structural_hash(graph/hash.py:6) → a verify pass (node-kind resolvability viacli/run.py:_resolve_node_factory:260, rule/goto target integrity). Reject + repair-loop on invalid IR.Where.
src/stargraph/skills/graph_author/_skill.py(model onskills/extract/orskills/triage/); register viastargraph.skillsentry point (relates to [P2] EPIC: Rules + Skills + Plugins discovery registry #42).authoring.DurableGraph._build_graph(authoring.py:288-313) IR-synthesis path.Acceptance criteria.
IRDocumentthat passesmodel_validate+ node-kind resolution + structural hashing.Caveat. Open-ended generation can emit unresolvable node kinds / dangling goto targets — the verify pass is mandatory, not optional. Relates to #43 (KG-augmented authoring), #42 (discovery registry).
Source: arXiv:2605.22502v1 — "Compiling Agentic Workflows into LLM Weights" (subterranean agents). Part of the end-to-end compile-to-weights pipeline epic; file refs verified against current
main. Determinism: per-graph knob (fully-compiled | hybrid | rules-only) defaults to honor #117 (no learned core router).