flowmind-author

Authors FlowMind .flow.yaml graphs with validate, fix, promote, and measure steps. Use when user writes or audits flows, fixes violations, or advances lifecycle quality.

SKILL.md

GitHub

/flowmind-author — Write, Validate, Fix, Promote, Measure

Intent Router

Intent Route What happens
Write Authoring guide Decision matrix + pattern catalog + golden rule
Validate Audit flow Deterministic probes + domain lenses + typed findings
Fix Meta-prompt KB + constraint-iterative Structured reasoning chain per violation
Promote Promotion ladder Typed PromotionDecision with gate evidence
Measure Telemetry analysis FlowTelemetry: tokens, cost, outcome, wall time

Write a Flow

Load the authoring guide first:

# Canonical reference
cat ~/lev/docs/design/design-flowmind-authoring-guide.md

The Golden Rule

If you can express the check as bash -c "..." && echo pass || echo fail, it is a lev.validate node. Full stop.

The Two Primitives

Primitive Op AI involved?
Validate lev.validate No — shell command, exit 0/1
Exec lev.exec Yes — prompt dispatched to adapter

Everything else is topology over these two primitives.

Required Fields

name: my-flow           # unique identifier
entry: first_node       # where execution starts

policy:
  determinism: true     # enforce reproducibility

knobs:                  # parameterize the flow
  my_param:
    type: string        # MUST have type declaration
    default: value

nodes:
  first_node:
    type: action
    description: |      # instruction-shaped, not vague
      What this node does and why.
    op: lev.validate    # or lev.exec
    ...

Flow Patterns (from authoring guide)

Pattern When Shape
Linear gate chain Sequential deterministic checks A→B→C→done
Implement-validate loop Iterative constraint engineering impl→gate→review→impl (loop)
Fan-out analysis Independent domain lenses load→{lens1,lens2,lens3}→synthesize
Adversarial review Red-team a constraint verify→probe→validate→generate→report

Validate a Flow

Run the audit flow against any .flow.yaml:

# Audit all flows in the repo
lev exec "audit all flows" \
  --flow ~/.lev/flows/lev-flowmind-audit.flow.yaml

# Audit a specific directory
lev exec "audit sdlc flows" \
  --flow ~/.lev/flows/lev-flowmind-audit.flow.yaml \
  --set scan_path=plugins/sdlc/flows/

# Audit a single flow
lev exec "audit this flow" \
  --flow ~/.lev/flows/lev-flowmind-audit.flow.yaml \
  --set scan_path=path/to/my.flow.yaml

What the Audit Checks

  1. Syntax: Valid YAML, required fields, reachable nodes, terminal nodes
  2. Spec parity: op: is lev.validate or lev.exec, branch keys valid, knobs typed
  3. Patterns: Golden rule enforced, iteration bounded, prompts exist
  4. Telemetry: Terminal nodes present, descriptions are instructional
  5. Parity: Duplicate flows across plugin directories flagged

Fix a Flow

When the audit finds violations, use the remediation meta-prompt pattern:

VIOLATION: {description}
FLOW: {flow_file}
CANON RULE: {gate_id}

REASONING CHAIN:
1. Read {flow_file}
2. Identify the specific violation
3. Determine fix strategy
4. Apply fix (touch ONLY the flow file)
5. Verify with deterministic command

For project-specific constraint flows (like NAAC), the meta-prompt KB lives at: .lev/flows/prompts/naac-auditr-revival/kb-remediation-metaprompts.md

General FlowMind fixes:

  • Missing entry: → add entry field pointing to first node
  • Orphaned node → wire it into a branch path or remove it
  • lev.exec for deterministic check → change to lev.validate
  • Untyped knob → add type: declaration
  • Missing prompt file → create at path referenced by prompt: field

Promote a Flow

Flows follow the same promotion ladder as other Lev artifacts:

workshop (experimental)
  → poc (proven in one project)
  → plugin (proven across projects)
  → core (canonical)

Promotion requires a typed PromotionDecision:

  • gate_pass_ratio == 1.0 for all authoritative gates
  • blocking_findings == []
  • Evidence artifacts exist
  • Transition is on an allowed edge

Measure a Flow

After execution, check FlowTelemetry:

  • total_tokens — cost signal
  • wall_time_ms — performance signal
  • outcome — green/red/partial
  • gate_pass_ratio — quality signal

Over N runs, flows build a track record. Successful patterns get promoted; failing patterns get mutated or retired.

Related

Skill Relationship
/work Lifecycle router — calls flowmind-author for flow-related work
/levloop Loop scheduler — flows are the execution unit per tick
/dump → /capture Capture pipeline — proposed flows should pass audit before filing
/lev-builder POC→core placement — uses promotion ladder
/naac NAAC operator console — naac-auditr-revival is the reference implementation

Reference Documents

Doc Location What
Authoring guide ~/lev/docs/design/design-flowmind-authoring-guide.md Decision matrix + patterns
FlowMind spec ~/lev/docs/specs/spec-flowmind.md Behavioral contract
Lifecycle design ~/lev/docs/design/design-flowmind-lifecycle.md Observability events
Evolutionary learning ~/lev/docs/_inbox/flowmind-evolutionary-learning.md Self-improving flow thesis
Meta-substrate ~/lev/docs/design/design-meta-substrate-declaration-compiler-ir-runtime.md 4-layer doctrine
NAAC reference impl naac/.lev/flows/naac-auditr-revival.flow.yaml Working constraint audit
NAAC meta-prompt KB naac/.lev/flows/prompts/naac-auditr-revival/kb-remediation-metaprompts.md Remediation templates