Role Directives and Formula Overlays

Operator-customizable agent behavior without modifying the Go binary.

Reference examples: docs/contrib-harnesses/ contains copy-and-adapt directives and overlays that contributors can drop into their own rig. See for example polecat-pr-flow/ for a rig that gates work on GitHub PR review rather than the canonical Refinery merge queue.

Problem

The MEOW stack embeds formulas and role templates in the binary — intentionally centralized for consistency, but leaving no override path. Operators cannot customize agent behavior at the rig or town level.

Concrete failure: Multiple crew members autonomously posted gh pr review comments on GitHub during PR review tasks. The formula says "post to GitHub," and there was no way for the operator to say "actually, in this rig, report back instead."

Design: Two Levels

Level 1: Role Directives

Per-role behavioral boundaries injected at prime time. Operator-authored Markdown that modifies how agents of a given role behave, regardless of which formula they are running.

File layout:

~/gt/directives/<role>.md              # Town-level (all rigs)
~/gt/<rig>/directives/<role>.md        # Rig-level (wins by appearing last)

Injection point: After the role template, before context files and handoff content. Directives carry an authority marker: "Rig Policy — overrides formula instructions where they conflict."

Precedence: Town and rig directives concatenate. If both exist, the combined output is <town content>\n<rig content>. The rig directive gets the last word, so it effectively overrides the town directive on conflicting instructions.

Implementation:

• Loader: internal/config/directives.go → LoadRoleDirective(role, townRoot, rigName) string
• Integration: internal/cmd/prime_output.go → outputRoleDirectives(ctx RoleContext)
• Called in the gt prime pipeline after outputPrimeContext()

Level 2: Formula Overlays

Per-formula, per-step overrides at rig or town scope. CSS-like step modifications applied post-parse before rendering at prime time.

File layout:

~/gt/formula-overlays/<formula>.toml        # Town-level
~/gt/<rig>/formula-overlays/<formula>.toml  # Rig-level (full precedence)

Precedence: Rig-level overlays fully replace town-level overlays (not merged). If a rig overlay exists, the town overlay is completely ignored. This prevents conflicting step modifications from merging unpredictably.

Implementation:

• Loader: internal/formula/overlay.go → LoadFormulaOverlay(formulaName, townRoot, rigName) (*FormulaOverlay, error)
• Applier: internal/formula/overlay.go → ApplyOverlays(f *Formula, overlay *FormulaOverlay) []string
• Integration: internal/cmd/prime_molecule.go → applyFormulaOverlays() called in showFormulaStepsFull()

TOML Format (Overlays)

Formula overlays use TOML with a [[step-overrides]] array:

[[step-overrides]]
step_id = "submit-review"
mode = "replace"
description = """
Report your review findings back to the conversation instead of posting
to GitHub. Format as a structured summary with grade and findings."""

[[step-overrides]]
step_id = "build"
mode = "append"
description = """
Also run integration tests: npm run test:integration"""

[[step-overrides]]
step_id = "deprecated-step"
mode = "skip"

Override Modes

Skip Mode Dependency Handling

When a step is skipped, steps that depended on it inherit its needs (dependencies). This preserves the formula DAG integrity. For example, if step B depends on step A, and step A is skipped, then step B inherits whatever step A depended on.

Validation Rules

• step_id is required on every override
• mode must be one of: replace, append, skip
• Malformed TOML returns an error during load
• Step IDs that don't match any formula step generate warnings (stale overrides)

Directive Format (Markdown)

Role directives are plain Markdown files. There is no special syntax — the content is injected verbatim into the agent's prime output with an authority header.

## PR Review Policy

Do NOT post review comments directly to GitHub via `gh pr review`.
Instead, report your findings back in the conversation as a structured summary.

## Code Style

Always run `npm run lint --fix` before committing.
Follow existing patterns in the codebase.

CLI Commands

Note: CLI commands are being added in gt-3kg.5. The interface below reflects the planned design.

Directive Commands

gt directive show <role> [--rig <rig>]    # Show active directive with source
gt directive edit <role> [--rig <rig>]    # Open in editor (creates file if needed)
gt directive list                         # List all directive files

Overlay Commands

gt formula overlay show <formula> [--rig <rig>]   # Show active overlay with source
gt formula overlay edit <formula> [--rig <rig>]   # Open in editor (creates file if needed)
gt formula overlay list                           # List all overlay files

The edit commands create the directory and file if they don't exist (following the gt hooks override precedent). The show commands display the resolved content with source annotation (town vs rig).

gt doctor Integration

The overlay-health doctor check validates formula overlays:

gt doctor                    # Runs all checks including overlay health

What it checks:

• Scans all town-level and rig-level overlay TOML files
• Parses each overlay and loads the corresponding embedded formula
• Validates every step_id exists in the current formula version
• Reports stale step IDs (formula was updated, overlay wasn't)

Results:

• OK: "N overlay(s) healthy" or "no overlay files found"
• Warning: Stale step IDs found (auto-fixable)
• Error: Malformed TOML (requires manual fix)

Auto-fix:

gt doctor --fix              # Removes stale step-override entries

The fix removes step overrides that reference non-existent step IDs. If all overrides in a file are stale, the entire file is removed. Malformed TOML is left untouched.

Implementation: internal/doctor/overlay_health_check.go

Worked Example: The PR Review Override

This is the motivating use case that drove the feature.

The Problem

The mol-polecat-work formula has a step called submit-review that tells polecats to post review results to GitHub using gh pr review --comment. In the gastown rig, the operator wants polecats to report findings back in conversation instead.

The Solution

Step 1: Create a rig-level formula overlay.

mkdir -p ~/gt/gastown/formula-overlays

Create ~/gt/gastown/formula-overlays/mol-polecat-work.toml:

[[step-overrides]]
step_id = "submit-review"
mode = "replace"
description = """
Report your review findings back to the conversation. Format as:

## Review: <file or component>
**Grade:** A-F
**Findings:**
- CRITICAL: ...
- MAJOR: ...
- MINOR: ...

Do NOT post comments to GitHub via gh pr review."""

Step 2: Verify with gt doctor.

gt doctor
# ✓ overlay-health: 1 overlay(s) healthy

Step 3: Test with gt prime.

gt prime --explain
# Shows: "Formula overlay: applying 1 override(s) for mol-polecat-work (rig=gastown)"

Now any polecat in the gastown rig running mol-polecat-work will see the replacement step instead of the original "post to GitHub" instruction.

What If the Formula Changes?

If a future gt release renames submit-review to post-results, the overlay's step_id becomes stale. On next gt doctor run:

⚠ overlay-health: stale step IDs in gastown/formula-overlays/mol-polecat-work.toml:
  - step_id "submit-review" not found in formula mol-polecat-work

Running gt doctor --fix removes the stale override. The operator then creates a new override targeting post-results.

Design Rationale

Why Two Levels, Not One?

Directives and overlays solve different problems at different granularities:

A role directive saying "never post to GitHub" applies everywhere — any formula, any step. An overlay targeting submit-review in mol-polecat-work applies only to that specific step in that specific formula.

Both are needed: directives for broad guardrails, overlays for surgical fixes.

Why Not Modify Formulas Directly?

Formulas are embedded in the Go binary. Modifying them requires rebuilding and redeploying. Directives and overlays are external config files that take effect immediately on the next gt prime.

Architectural Harmony

• Fits gt prime pipeline: Role template → directives → context → handoff → formula
• Follows hooks override precedent: ~/.gt/hooks-overrides/<target>.json
• Extends property layers: Rig > town > system precedence
• ZFC-compliant: Go transports the content, agents interpret the instructions
• Only touches gt: bd doesn't render formulas, so overlays are gt-only

Dissonance to Manage

• Conflicting instructions: Directive says "don't X", formula says "do X" → mitigated with clear authority framing at injection ("Rig Policy — overrides formula instructions where they conflict")
• Unstable step IDs: Formula steps are not a stable API; step IDs can change across versions → gt doctor warns about stale overlays
• Discoverability: gt prime --explain shows active directives/overlays with source annotations

File Reference