Reference — the discovery sidecar, plan-tree, and roster

Note

Diátaxis: reference. The slots, the plan-tree template, and the skill/agent roster of the discovery loop, for look-up. For the why, see the explanation The discovery loop; to run it, the how-to Run a discovery. The authoritative, versioned definition is carried in the discovery-loop skill’s references/sidecar-schema.md; this page summarizes it for adopters.

The sidecar — two tiers, one forest

The discovery loop keeps its working state in a typed sidecar. There is one plan-tree per initiative (a forest, not one master tree); initiatives cross-link by stable id.

• Tier 1 — _state/ (the working store, durable while active): blackboard, open-questions, traceability, decision-log, meta, plan-tree.
• Tier 2 — committed artifacts (durable beyond the run): domain-framing, scope-boundary, journey-map, service-blueprint, screens/, decision-brief, backlog.

Layout (paths resolved config → default → discover-by-marker; dirs created lazily): <discovery-root>/<initiative-slug>/_state/… with the committed artifacts beside it.

Every instance carries a schema_version stamp (0.1 today). Downstream consumers — the traceability lint, work-loop, the release loop — read instances by convention + the stamp, never importing the definition.

The five working slots

Slot	Shape (key fields)
blackboard	typed artifact slots: {id, type, status, version, produced_by, lens, path?, round_last_touched, parent_id?} + a meta block
open-questions	{id, raised_by, target_discipline, question, status, resolution, round} — the queue lenses answer each other through (never chat)
traceability	the lint-authoritative JSON: {root, leaf_kind, nodes:[{id, kind, backed_by}], edges:[{from, to}]} — an orphan is a defect
decision-log	{ts, gate, decision, ratified_by, reversibility_class, rationale, prev_hash, hash} — append-only, hash-chained, actor-attested
meta	{round, round_cap, cost_budget, cost_spent, gate, status, saturation} — counters the controller increments (no runtime)

Two extension slots: a validation-plan (assumption → kill-condition → activity → validation-status) and the decision-brief’s required success_metrics slot (a brief cannot reach G3 without it).

Three status namespaces

Namespace	Values
slot status	draft | proposed | ratified | stale | rejected
node lifecycle	draft → diverging → converging → ratified | stale, plus parked / abandoned
node validation status	hypothesis → validating → validated | refuted
meta gate-state	awaiting-human / paused-at-bound / stalled-at-cap

The plan-tree template

A node = an intent slot + parent_id + the four per-node namespaces above + a sub-idea index (open / parked / done) + the divergence shapes (a candidate set + a selection, not-chosen retained) + a validation hook (kill-condition + real-world activity). The controller copies the template per initiative; the traceability lint walks it. No planner engine — the template is the mechanism.

The verdict set

approve / approve-with-constraint / redirect (steer) / explore-alternatives / abandon / park (defer) / extend (override). Every blackboard-changing verdict reuses one cascade mechanism (walk traceability out-edges → mark stale → re-run only the affected lenses), bound by two guards: impact-before-blast and no-jumping-ahead.

The roster

Role	Agent / skill	Required?
Controller	discovery-lead (agent)	the loop
Divergence	explore-options (skill)	the loop
Validation	plan-validation (skill)	the loop
Threat / compliance lens	discovery-threat-reviewer (agent)	required at G2
Reliability / operability lens	discovery-reliability-reviewer (agent)	required at G2
UX/design lens	experience pack (+ experience-reviewer)	optional, detect-and-degrade
Architecture lens	architect pack (+ design-reviewer)	optional, detect-and-degrade
Research lens	research pack	optional, detect-and-degrade

The two discovery reviewers are distinct agents from work-loop’s code security-reviewer / quality-engineer and degrade only in depth, never to nothing. The CHARTER’s “three reviewers is the ceiling” stays a work-loop/code-review cap; the discovery roster is a tracked exception.

The gates

Gate	Phase	Human acts?
G0	Intake	consent — ratify the value seed
G1	Strategy	auto unless a risk trigger fires
G1.5	Domain & MVP (after divergence)	consent — ratify the altitude/MVP bet
G2	Convergence (after self-coverage + reviewers)	consent — ratify the “what”; adjudicate value conflict
G3	Handoff	hand off to work-loop

Bounds (round / cost / concentration ~40% / depth / breadth) are pause-and-confirm gates, never auto-terminals: hitting one sets paused-at-bound and surfaces the verdict set.