Skip to main content
Status: accepted · ADR-40 · Filed 2026-05-01

​
Persona role reshape — PO / Architect / Engineering split

​
Status

accepted — trial period (Frank greenlight 2026-05-01)

​
Context

Pre-2026-05-01 persona roles concentrated design authority in Lola (Claude Desktop) with Donna (Claude Code) executing. As the agent fleet has grown to ~7 active personas, the design-vs-build seam has become a bottleneck:
  • Specs authored in terminal don’t render well; Claude Code is suboptimal for long-form spec reading. Markdown native to Desktop is the right surface for product validation.
  • All design-side personas were running on the Anthropic SDK (claude_code / claude_desktop) — same model family, shared cognitive priors. Cross-surface diversity (Codex vs Anthropic) is missing from the architectural review path.
  • “Ask Frank” is the default routing for cross-persona uncertainty. Peer-direct routing is harder than it should be (see signal vocab gap memory — 5 instances of StatusRequest/Question shoehorning).

​
Decision

Personas are assigned the following authority scopes:
PersonaSurfaceRoleAuthority
Lolaclaude_desktopProduct Owner”What to build and why.” Authors product requirement specs. Validates completed work against original requirements. PO sign-off gates spec sealing.
TexicodexSystem Architect”How it should be designed.” Architectural review/veto authority across all surfaces. Authors architectural specs. Cross-surface (Codex ≠ Anthropic) — divergent thinking enforcer.
Donnaclaude_codeEngineering / Code Execution”Make it work.” Implements specs after architectural sign-off. May draft implementation specs (subject to Texi review).
Porscheclaude_codeTelemetry / Observability laneDashboard / OTel / metrics surface. Specs reviewed by Texi.
Desireeclaude_codeDocumentation laneOwns docs/* tree.
Lafondaclaude_codeInstall / deployment laneOwns install + LAN deploy.
Candiclaude_code (Win)Cross-platform / Windows tester

​
Spec consensus flow (trial)

  1. Need identified → Lola drafts product requirement (in Desktop’s rendered markdown).
  2. Lola signals Texi: ReviewRequested with the requirement.
  3. Texi authors architectural spec; signals Lola for product sign-off.
  4. Lola signals back: approve / revisions_requested. Loop until approved.
  5. Texi signals Donna (or relevant lane owner): TaskAssigned for implementation.
  6. Donna ships; signals Texi (architectural conformance check) AND Lola (requirements satisfied check).
  7. Both ack → spec sealed via prism_seal.
Cross-surface review-pairing default: ANY claude_code-authored spec (Donna, Porsche, Lafonda, Desiree, Candi) routes to Texi (codex) for architectural review by default. The cross-surface guarantee IS the divergent-thinking enforcer.

​
Rationale

  • Surface-fit. Desktop renders specs natively; terminal is read-by-cat. Lola in Desktop is the right surface for spec reading and product validation.
  • Cognitive diversity. Codex (Texi) vs Anthropic (Lola/Donna/Porsche) is the only cross-family seam available. Routing architectural review through it ensures specs are stress-tested by a different model family before commit.
  • Authority discipline. Per feedback_engineering_authority memory — when Frank designates a system engineer, peers defer. Texi as system architect formalizes that pattern across the fleet.

​
Alternatives considered

  • Status quo (Lola=designer, Donna=executor). Rejected: collapses architecture into product design; no cross-surface diversity.
  • Texi as architect, Lola unchanged. Rejected: leaves the spec-reading surface mismatch unresolved.
  • Multiple architects (Texi + Lola). Rejected: defeats the divergent-thinking goal; consensus loop requires distinct authority scopes per role.

​
Consequences

  • Donna routes all spec authorship for architectural review through Texi by default.
  • Lola in Desktop becomes the product-validation gate before any spec seals.
  • The spec lifecycle gains state machine semantics (draft → in_review → revisions_requested → approved → accepted). Implementation of the state machine is its own SPEC, deferred.
  • The signal vocab gap (Question / Answer / Veto types) now blocks the consensus loop’s clean expression. Vocab extension SPEC priority bumps to immediate.
  • Texi’s memory scope tightens to architecture / contracts / boundaries. Memory namespace work is a separate track (research branch territory).

​
Trial period — failure modes to watch

  • Lola not online enough in Desktop to gate specs (requires Frank’s Desktop usage cadence). Mitigation: Lola can pre-author batches of product specs; consensus loop is async.
  • Texi (Codex) shim instability (per project_codex_shim_no_respawn memory) causes review backlogs.
  • Cross-surface review delay > engineering throughput, becomes drag rather than diversity-enforcer.
If any of these emerge, revisit and either reassign or build tooling to mitigate.
Last modified on May 18, 2026