[dominno]

Summary

Distills the actionable parts of a "graph engineering" writeup (Slate / Random Labs) into our existing Workflow runtime — which already is a graph runtime (nodes = agents, edges = data hand-offs, parallel/pipeline, scoped failure). Three real gaps were closed, then the change was run through our own /improve-skills gate, which caught a security regression that is fixed here too.

What changed

Model tiering (all 6 workflows)

• High-volume fan-out nodes (critics, per-story evidence, per-dimension gaps, per-category E2E cases, meta-critics) → model: 'sonnet'.
• High-stakes gates/judgment (adversarial verifiers, status verifier, synthesis) → model: 'opus'.
• Tiers are FANOUT_MODEL/JUDGE_MODEL constants, overridable via args.models = { fanout, judge }.
• Deliberate exception: the loop-iteration implementer writes real code, so it inherits the session model rather than being silently downgraded.

Budget as a hard stop (soft + hard control)

• Guarded on budget.total (null ⇒ no cap, so no accidental Infinity loop).
• improve-skills skips the verify fan-out below a token floor and routes findings to needsDesign (unverified ⇒ never auto-applied).
• loop-iteration escalates all items below its floor instead of failing mid-fix.
• Fail-safe direction preserved: a budget stop never approves unverified work.

Diagram-first

• New docs/workflow-graphs.md: a Mermaid node/edge graph for every workflow, annotated with the per-node model tier — the "see the graph before you run it" view, version-controlled next to the scripts.

Gate findings applied (from /improve-skills run on this diff — 1 blocker + 7 important)

• BLOCKER — loop-iteration denylist bypass: the deterministic gate trusted the implementer's self-reported changedFiles, but the caller applies the real diff. An under-reported file list touching a denylisted path (.claude/, .env) could slip past. Now parses paths from the actual diff (parseDiffPaths) and gates on the union of diff-paths + changedFiles. Unit-tested: a diff editing .claude/loop.md while changedFiles omits it now escalates.
• Ported the JSON-string args normalization to the other 5 workflows (was only in improve-skills), rewiring every args.X read to the normalized a so the args.models override no longer silently no-ops.
• workflow-dsl lens now also checks model-tiering + budget-guard fail-safety; consistency-dry lens and /dream lint now check docs/workflow-graphs.md for drift vs the scripts; added the diagram to improve-skills DEFAULT_TARGETS (the drift check was documented but not wired).
• improve-skills.md now documents the budgetStop return field and instructs disclosing a skipped-verification run.

Nice-to-haves

• Fan-out agent labels now include the map index (${key}-${i}) for consistency with the project's own DSL rule.
• Diagram note that uiInScope:false drops the 5 UI critics.
• skills-index.md documents the args.models override once.

Memory wiki

• Filed the tiering, soft/hard-budget, diagram-drift, and "a deterministic gate is only as trustworthy as the input it reads — derive checked paths from the applied artifact (the diff), not a self-report" conventions in topics/workflow.md + log.md.

Not adopted (out of scope)

The post's quant/hedge-fund framing and the standalone Slate runtime — we already have the graph runtime, and the trading claims overstate what orchestration solves (data quality, transaction costs, and overfitting remain the hard part).

Verification

• All six workflow scripts validated (node --check wrapped in the async-function harness).
• The denylist-union blocker fix has a unit test (attack case escalates; clean case passes).
• The change was dogfooded through /improve-skills on the new tiering itself (21 agents, Sonnet fan-out + Opus verify), and every confirmed finding is fixed in this branch.

🤖 Generated with Claude Code

---

Generated by Claude Code