Status: accepted v0.2 — body sourced from TriGraph spec entity. Docs-mirror file added by docs lane to surface ratified specs that lacked an on-disk mirror.
SPEC-078: Consensus-First Parallelism and Method Fragments
SPEC-078 — Consensus-First Parallelism and Method Fragments
Status
Accepted v0.2 — ratified by Donna under delegated PO authority 2026-05-04 (Plan #10 v0.3 §1; Texi clean signoff via signal 17ff83d3).
Author: Candi
Architecture reviewer: Texi
PO / ratifier: Donna
Acceptance panel: Porsche
Docs sweep: Desiree
Plan: Plan #10 Wave 2 Blocker #4
v0.2 Changelog
v0.2 folds Texi architectural review a6a0242b: 10 binding amendments, 6 nits, 4 cross-blocker references, resolution of 5 v0.1 open questions.
Summary
Prism parallelism is consensus-first, not brute-force-first. This SPEC defines how Prism agents decide when to work alone, when to use tools, when to request peer review, when Texi must gate architecture, and when a task requires a multi-agent convergence loop before implementation. It also defines the first governed method-fragment schema so reusable methodology can be routed to agents and skills without creating prompt sprawl.
The standard exists because faster implementation is not the highest-order goal. As the project grows, the damaging failure mode is agents hacking to a passing result while taking architectural shortcuts that create downstream bugs, rework, and operator frustration.
Shared Definitions
These terms consume ADR #48 v0.2 and must not redefine its field model.
| Term | Definition |
|---|
identity | Stable human-addressable agent/persona name used for registration and routing. |
surface | Runtime host capability context (Codex, Claude Code, Claude Desktop, Cursor, VSCode). |
specialization | Durable expertise lane used for routing, memory locality, skill matching, method-fragment ownership. NOT governance authority. |
assignment | Current task responsibility for an identity in a session or work unit. |
style | Optional communication preference. Never authority. |
method_fragment | Governed, triggerable unit of methodology with use/skip criteria, owner specialization, required context, outputs, review gates. |
tool | Deterministic/repeatable operation producing same kind of output for same input without judgment-heavy tradeoff analysis. |
agent | Judgment-bearing actor applying context, tradeoff analysis, review, synthesis, conflict resolution. |
consensus_loop | Bounded multi-agent review workflow that converges on a decision before implementation. |
lane_owner | Agent identity responsible for a specialization lane or assigned work stack. |
implementation_authority | Donna, or Donna-delegated lane owner, authorizing implementation after review gates pass. |
review_gate | Required review step for a risk tier, method fragment, or assignment. |
risk_tier_assignment | Structured record of proposed/final R0-R3 tier, who set it, and any override. |
Problem
Three recurring risks: shortcut risk (agents solve symptom while violating architecture), single-model blind spot (one model family misses what another would catch), coordination debt (more agents without ownership/convergence rules increases conflicts/duplication/unclear handoff). Parallelism without consensus worsens all three.
Goals
Define risk tiers determining review depth and agent participation. Define when a task should use tool vs single agent vs peer review vs consensus loop. Define ownership contracts for parallel work. Preserve Texi’s architectural gate. Preserve Donna’s execution authority. Route prompt/method conflicts to Candi. Convert reusable methodology into method fragments with governed schema. Provide Porsche-visible acceptance signals. Keep all automated routing behind a default-off feature flag.
Non-goals
No automatic swarm spawning in initial implementation. No specialization-based automatic routing in initial implementation. No Tri-Graph production recall dependency in initial implementation. No replacement of Donna/Texi/Candi human-readable signal workflows. No change to ADR #48 field semantics. No runtime enforcement while feature flag is off.
Feature Flag Contract
| Field | Value |
|---|
| Name | PRISM_CONSENSUS_PARALLELISM_ENABLED |
| Default | false |
| Owner | Candi (methodology semantics); Donna (implementation authorization) |
| Operator-visible state | Future boot/report surfaces show enabled/disabled only when implementation exists. |
| Shadow mode | No. Passive collection while flag off requires separate SPEC/amendment. |
| Rollback | Set flag to false; no migration rollback required for advisory metadata. |
false: no automatic consensus gate enforced, no automatic specialization-based routing, no automatic method-fragment injection, no task blocked by automation, manual signals/reviews/Donna-directed assignments work exactly as before.
When true in future implementation: Prism may recommend or require review gates by risk tier, surface applicable method fragments, emit structured route/review recommendations, report missing required gates before task completion.
No shadow mode. Default-off invariant test: start session with flag unset, assign R2/R3 task, verify Prism does not block/auto-route/inject fragments/require gates, verify manual ReviewRequested/TaskAssigned/StatusUpdate/checkpoint flows unchanged.
Risk Tiers
Risk-tier authority: task owner proposes; Donna sets/overrides for Plan #10; Texi may escalate architectural risk; Candi may escalate governance/methodology/memory/prompt/conflict risk; future advisory tooling recommends only.
R0 — Routine / Single-Agent: one-file/small-scope changes, local bug fixes with established patterns, doc wording updates, read-only inspection, deterministic validation. Single agent may proceed; checkpoint outcome.
R1 — Bounded / Peer-Reviewed: small multi-file changes with low architectural impact, implementation with clear precedent, test additions or refactors not altering contracts, low-risk launcher/docs/install updates. Owner drafts; one peer review; Donna or lane owner authorizes; checkpoint with decisions and tests.
R2 — Architectural / Consensus: new architecture, cross-module contracts, governance rules, memory semantics, routing behavior, install/session boot behavior, surface capability semantics, data model changes, anything that creates durable project debt if wrong. Workflow: Proposal → Candi methodology/governance review (when applicable) → Texi architectural review → Donna engineering feasibility review → revision → PO/Donna tie-break (when project goals decide) → implementation authorization.
R3 — Program / Multi-Stack: multi-wave plans, cross-agent programs, release trains, governance foundation work. Donna-led plan, explicit ownership map, milestone checkpoints, ratification gates, RCA for all issues, final Donna summary.
Use a tool when operation is deterministic, same input produces same class of output, no tradeoff analysis needed, validation can be automated, failure is structured and bounded. Examples: syntax parse, test run, lint, schema validation, file inventory, static graph query, status-card render check.
Use an agent when tradeoff analysis required, architecture can be affected, prompt/governance conflict exists, requirements are ambiguous, synthesis across sources needed, disagreement must be resolved, quality depends on judgment. Examples: architectural review, method conflict resolution, spec design, cross-surface critique, implementation strategy, memory domain taxonomy, risk-tier assignment.
Standing Agents (7-row)
| Identity | Specialization | Primary role |
|---|
| Donna | engineering | Plan owner, implementation authorization, final summary. |
| Texi | architecture | Architectural gate; veto/revision authority for R2/R3 work. |
| Candi | governance | Prompt conflict resolution, method fragments, consensus standard, CI loop. |
| Lafonda | install | Install-lane execution; launcher/install feedback. |
| Samantha | rte | Runtime validation; cross-surface smoke support. |
| Desiree | docs | Docs sweep; generated/manual surface clarity. |
| Porsche | dashboard | Acceptance panels; probes; measured performance truth. |
Ownership Contract
Every parallel assignment must include: identity, surface, specialization, assignment, risk tier, owned files/modules/artifacts, expected output, decision authority, review gates, test/verification expectation, handoff target, status recipient (usually Donna for Plan #10).
Parallel work allowed only when ownership boundaries are explicit. Two agents must NOT independently edit the same file/artifact unless one is clearly reviewer and one is clearly implementer, OR Donna explicitly authorizes the shared edit plan.
Consensus Workflow
Low risk: single agent proceeds; checkpoint outcome.
Medium risk: owner drafts plan/patch; one peer review; Donna or lane owner authorizes.
High architectural risk: proposal → Candi methodology review (when applicable) → Texi architecture review → Donna engineering feasibility review → revision → PO/Donna tie-break (if project goals decide) → implementation authorization.
Dual-model review mandatory for R2/R3 only when: cross-surface behavior affected, model-family disagreement risk material, tool capability/prompt behavior affected, memory/routing semantics affected, OR Donna/Texi/Candi marks required. NOT mandatory for every R2 by default.
Conflict routing
| Conflict type | Resolver |
|---|
| Prompt/method/rule/governance | Candi |
| Architecture | Texi |
| Engineering feasibility/sequencing | Donna |
| Product/project-goal tie | PO / Donna under Plan #10 |
| Metrics/performance truth | Porsche |
| Documentation SOR | Desiree (with Candi/Donna as needed) |
ReviewCompleted signal type with explicit verdict semantics (revisions_required, veto, accept_with_amendments, accepted). No new signal type required.
Method Fragments
Reusable governed methodology units, smaller than SPEC, more structured than prose. Answer: when should this trigger, who owns it, what context required, what output expected, what review gates apply, what memory should be read/written.
Schema (yaml)
schema_version: 1
version: 1
id: method.<slug>
title: <short title>
status: draft|accepted|deprecated|superseded
source_of_record: prism_typed_artifact
generated_mirror_paths: [docs/method-fragments/<slug>.md]
authority_layer: global|tenant|org|project|user_session_runtime
owning_specialization: <specialization>
owner_identity: <identity> # optional
applies_to_specializations: [<specialization>]
applies_to_surfaces: [codex, claude_code]
use_when: [<trigger>]
skip_when: [<non-trigger>]
risk_tier_default: R0|R1|R2|R3
risk_tier_override_allowed: true|false
required_context: [<memory domain or artifact>]
expected_outputs: [<artifact or result>]
required_prism_verbs: [<verb name>]
review_gates: [<identity or specialization>]
review_required: true|false
memory_writes: [<memory domain>]
conflict_policy: <routing rule>
telemetry_key: method.<slug>
supersedes: [<method id>]
superseded_by: [<method id>]
source_refs: [<path or artifact id>]
Minimum acceptance requirements
A method fragment is not accepted until it exists as a Prism typed artifact in the database (SOR), use_when/skip_when both present, owner specialization named, applies-to specializations named, expected outputs named, risk tier named, risk-tier override behavior explicit, review gates explicit (or explicitly empty), review-required boolean explicit, memory writes explicit (or explicitly empty), telemetry key present, source refs present.
Initial candidate fragments
| Method fragment | Owner | Purpose |
|---|
method.consensus.risk-tier-assignment | Candi | Classify work as R0-R3. |
method.consensus.architecture-gate | Texi | Trigger arch gate for R2/R3 work. |
method.conflict.prompt-governance | Candi | Route prompt/method conflicts to Candi. |
method.parallel.ownership-contract | Donna/Candi | Require ownership fields before parallel work. |
method.review.dual-model | Texi/Candi | Require Claude/Codex cross-model review for selected R2 work. |
method.completion.done-definition | Donna | Enforce design+built+tested+committed unless deferral approved. |
Completion Semantics
For Plan #10, the operating contract defines “complete” as: designed, built, tested, committed, status reported to Donna, relevant checkpoint recorded, RCA committed to memory for any issue or dead end. Deferrals require Donna’s knowledge and a good reason. Milestones are checkpoints, not rest stops.
This is operating-contract authority. SPEC-078 automation must NOT auto-enforce this definition while PRISM_CONSENSUS_PARALLELISM_ENABLED=false.
Telemetry and Acceptance Panel
Porsche acceptance panels measure whether this standard improves engineering quality without creating hidden drag.
Minimum signals: risk-tier assignment events (proposed tier, final tier, setter, override reason); consensus loop lifecycle events (opened, review_requested, amended, authorized, closed); review turnaround time; review gate status events (missing, requested, satisfied, waived); conflicts routed by type; Candi conflict resolutions; Texi architecture amendments; Donna implementation authorizations; override/deferral events (who authorized, why, expiry/follow-up); RCA count for consensus failures; method fragments surfaced; method fragments invoked; tasks blocked by missing required gates when flag enabled.
No performance claims may enter prompts or governance rules until measured.
Storage / Memory Hooks
This SPEC does not define final memory storage. Blocker #5 (SPEC-079 Memory Domain Contracts) owns that. SPEC-079 must consume these terms: risk tier, risk_tier_assignment, method fragment, ownership contract, consensus loop, conflict resolution, review gate, implementation authorization, deferral, RCA.
Blocker #6 (Tri-Graph Governance Recall) should consume these graph entities: MethodFragment, RiskTier, ReviewGate, ConsensusLoop, ConflictResolution, ImplementationAuthorization. Minimum edges: APPLIES_TO_SPECIALIZATION, REQUIRES_GATE, SATISFIES_METHOD, SUPERSEDES.
Implementation Phases
Phase 0 — Ratification only: SPEC review by Texi, Donna ratification, no behavior-changing code.
Phase 1 — Manual practice and docs: Desiree docs sweep; add accepted method-fragment templates; Candi maintains CI loop tracking; no automatic routing.
Phase 2 — Default-off advisory support (requires PRISM_CONSENSUS_PARALLELISM_ENABLED=true): surface applicable method fragments in prism_start or dedicated recall verb; emit advisory risk-tier recommendations; emit missing ownership-contract warnings. Phase 2 is advisory only — recommend/surface/warn, NOT block completion.
Phase 3 — Enforcement gates (requires separate implementation authorization): block completion when required gates missing; require explicit Donna override for deferrals; integrate with Tri-Graph Governance Recall once Blocker #6 ratifies.
Acceptance Criteria
- SPEC reviewed by Texi and ratified by Donna. ✓
- Feature flag named and default-off invariant specified. ✓
- Risk tiers R0-R3 defined with required workflow. ✓
- Agent/tool boundary defined. ✓
- Ownership contract defined. ✓
- Conflict routing table defined. ✓
- Method-fragment schema defined. ✓
- Initial candidate fragments listed. ✓
- SPEC-079 Memory Domain Contracts can consume this vocabulary without redefining terms. ✓
- Telemetry events/fields listed. ✓
- With feature flag unset/false, no live behavior changes. ✓
- Method fragments specified as Prism typed artifacts in database SOR with docs mirrors generated from that source. ✓
Review Decisions Folded in v0.2
- Method fragments live as Prism typed artifacts in DB SOR with generated docs mirrors.
- Dual-model review mandatory only for cross-surface/model/tool/prompt/memory/routing-sensitive R2/R3 or Donna/Texi/Candi-marked.
- Risk tier proposed by owner, set/overridden by Donna for Plan #10, escalated by Texi/Candi for arch/gov risk; advisory recommends only.
- Architecture veto/revision uses
ReviewCompleted verdict semantics; no new signal type.
prism_start exposes compact counts/ids only; full retrieval via dedicated recall verb or future Tri-Graph.
References
- Plan #10 v0.3 (Donna ownership map)
- ADR #47 v0.2 (Governance Layer Precedence + Override Semantics)
- SPEC-077 v0.2 (Governance Resolution + Boot Report)
- ADR #48 v0.2 (Identity / Specialization / Assignment / Style Model)
- SPEC-063 (Postmortem + Retrospective Discipline)
- Texi architectural review
a6a0242b
- Texi clean signoff
17ff83d3
