Render Dispatcher

Central entry point + menu for the canvas/state render fleet. Routes intent to specialists or surfaces what's available. Cross-cutting views are research-gated and NOT emitted by default. Completes the v0.40.0-announced render fleet foundation.

When to use

• "Show me the OST" / "render the diamond state" / "visualize the cycles" → dispatcher recommends + routes to the specialist (recommend, do NOT auto-invoke).
• "Visualize how opportunity X became cycle Y" → cross-cutting traceability intent. Dispatcher surfaces that the view is research-gated; does not emit a default rendering.
• Discovery: "what renderable views does Mycelium have?" → dispatcher lists specialists + their states.

When NOT to use

• If the user names a specialist directly (/mycelium:diamond-render, /mycelium:ost-render, /mycelium:cycle-render), the specialist is the one-hop path. Do NOT re-route through the dispatcher.
• For cross-cutting views that already have dedicated skills (e.g., /mycelium:framework-health is its own audit; not a render specialist).

Identifier exposure

Declared: MIXED

Scope (per-canvas table — REQUIRED for MIXED)

Canvas	Exposure	Identifier-bearing fields	Activated when
.claude/canvas/opportunities.yml	YES	evidence_sources, notes	--view traceability (deferred), --view ost-summary
.claude/canvas/cycle-history.yml	YES	learnings.process prose, related_corrections	--view traceability (deferred), --view cycle-summary
.claude/canvas/archived-solutions.yml	YES	discard_notes prose	--view traceability (deferred), --view discards
.claude/canvas/landscape.yml	YES (low)	notes, evidence_sources	--view competitive-landscape (deferred)
.claude/diamonds/active.yml	NONE	none	any view including the diamond

Rationale

Dispatcher behavior depends on which cross-cutting view is invoked. MIXED with per-canvas scoping is the only honest declaration: YES would force registry consultation on diamond-only views (unnecessary cost); NONE would silently leak identifiers on traceability views (incident). Each cross-cutting view's implementation MUST consult the registry per engine/render-conventions.md#hard-rule-consent--privacy-gate for the YES-exposure canvases it activates and skip the consultation only for NONE-exposure canvases.

For v0.40.3 ship: the dispatcher ships WITHOUT cross-cutting views. All MIXED-relevant identifier surfaces are inactive. The MIXED declaration is forward-looking — it names the canvases that WILL activate when cross-cutting views land in subsequent patches per the research-first methodology.

Anon-label convention

Per engine/render-conventions.md#anon-label-convention. Numbering shared across a single dispatcher invocation regardless of canvas count (cohort-tester-2 in cycle-003 IS cohort-tester-2 in opp-004 evidence).

Consent value semantics

Per engine/render-conventions.md#consent-value-semantics.

Worked examples (forward-looking)

--view traceability (deferred): activates opportunities.yml + cycle-history.yml + archived-solutions.yml. Registry consultation runs. Multi-canvas anon-label numbering shared.

--view diamond-only (always available — alias for routing to diamond-render): touches only active.yml. Registry consultation SKIPPED per the per-canvas exposure table. No anon labels emitted.

Fixture pointer

• tests/bash/fixtures/render/dispatcher-recommend-not-invoke.yml (routes to specialist)
• tests/bash/fixtures/render/dispatcher-list-options.yml (discovery mode)
• tests/bash/fixtures/render/dispatcher-view-traceability-mixed-consent.yml (deferred — fires once traceability view ships)
• tests/bash/fixtures/render/dispatcher-view-diamond-only-no-redaction.yml (alias-to-specialist behavior)
• tests/bash/fixtures/render/dispatcher-view-traceability-no-registry-entry.yml (deferred — fail-loud on missing identifier)

Arguments

Arg	Default	Values	Effect
--list	false	bool	Print the dispatch table + deferred cross-cutting views, do not route
--view	null	traceability | confidence-trajectory | cluster-graduation-flow | canvas-dependency-graph	Cross-cutting view selector (all currently deferred per research-first methodology)
--format	mermaid	mermaid | ascii | json	Format hint for routed specialists (overridable per specialist)
--theme	base	base | dark	Theme hint for routed specialists (dark opt-in for WCAG-by-construction per engine/render-conventions.md#wcag-aa-theme-convention)

Workflow

Step 1: Parse intent

Identify whether the request maps to:

• A single canvas → route to specialist (Step 2)
• A cross-cutting view → surface the deferral (Step 3)
• Discovery / --list → enumerate options (Step 4)

Step 2: Route to specialist

Map canvas to specialist. Recommend, do NOT auto-invoke (architecture decision Q1):

Canvas / intent	Specialist	Beta risk
opportunities.yml	/mycelium:ost-render	none
diamonds/active.yml	/mycelium:diamond-render	none
cycle-history.yml	/mycelium:cycle-render	none
landscape.yml	/mycelium:landscape-render (deferred)	wardley-beta
bvssh-health.yml	/mycelium:bvssh-render (deferred)	radar beta
dora-metrics.yml	/mycelium:dora-render (deferred)	xychart-beta
scenarios.yml	/mycelium:scenarios-render (deferred)	none

Output shape for routing:

> Recommend: `/mycelium:ost-render` (--format <inherited>, --theme <inherited>)
> The canvas remains source of truth; this render is a snapshot.

Specialist not yet shipped → surface deferral + promotion-trigger.

Step 3: Cross-cutting view (currently all deferred)

Each --view value names a candidate view; all are research-gated:

View	Spans canvases	Status	Promotion trigger
traceability	opportunities + cycle-history + archived-solutions	DEFERRED	Research-first methodology converges on a candidate shape; founder visual eval ≥4.0
confidence-trajectory	diamonds/active.yml history + decision-log	DEFERRED	Cohort tester asks for confidence-over-time visualization
cluster-graduation-flow	corrections.md + cluster-instances.md	DEFERRED	External receipts-case need for graduation-flow visualization
canvas-dependency-graph	every canvas's cross_ref + dependencies	DEFERRED	Canvas-health drift incident traces to cross-canvas dependency staleness

Output shape for deferred-view requests:

> ⓘ View `traceability` is deferred.
> The shape requires research-first methodology + founder visual eval
> before a default rendering can ship. The framework intentionally does
> NOT emit a "best guess" view to avoid teaching users to think with a
> shape that wasn't validated.
>
> Promotion trigger: <named in table>

Step 4: List options (discovery mode)

Print the Step 2 dispatch table + Step 3 deferred view list. Each entry names: skill or view, what it renders, supported formats, beta risk, current status (shipped / deferred-with-trigger).

Rules

1. Read engine/render-conventions.md before any emit.
2. Never modify canvas state. Renders are read-only by class.
3. Apply canonical disclaimer + lossy-on-export template per the engine doc (when routing, the user emits the render via the specialist; when listing options, no disclaimer needed).
4. Routing: NEVER silently sub-invoke. Name the recommended specialist + arg-set explicitly. User retains one-hop control.
5. Deferred views: NEVER emit a "best guess" rendering. The trap is exactly "ship a looks-reasonable-to-me view"; the dispatcher's job is to gate against it.
6. Format/theme hints route through: when recommending a specialist, pass --format <inherited> and --theme <inherited> so the user's invocation matches dispatcher intent.

Counter-Argument Check

Before emitting:

1. "Is this render conveying current state, or a stale snapshot the canvas no longer matches?" Each routed specialist runs its own staleness check (per engine/render-conventions.md#staleness-check-distinction); dispatcher does not pre-emptively check (the specialist owns it).
2. "Am I about to silently auto-invoke a specialist instead of recommending?" The architecture decision is load-bearing — silent sub-invocation defeats the dispatcher's reason to exist beyond menu-discovery. If routing logic ever becomes "if intent matches X, just emit X's output," surface the regression and stop.
3. "Am I about to emit a 'best guess' cross-cutting view?" If --view <name> is set and the named view's status is deferred, the only correct output is the deferral message. NEVER fall through to a default rendering.

What this skill does NOT do

• Does NOT emit single-canvas renders directly. That's the specialists.
• Does NOT modify canvas state. Read-only by class.
• Does NOT auto-invoke specialists. Recommends + names the user-facing invocation.
• Does NOT emit cross-cutting views until the research-first methodology converges + manual visual eval lands. The deferral is the feature.

Theory citations

• Hick's Law (single recommended specialist; deferred-view fail-loud avoids decision-tax)
• Argyris triple-loop (the dispatcher routes between loops; doesn't claim to BE a loop)
• Counter-Goodhart (never emit a "best guess" cross-cutting view; the deferral IS the discipline)