plat 305

← Pulse platform index

PLAT-305 — Separate QA, Architecture and Strategy and close the improvement loop

2026-09-10 update: PLAT-306 supersedes mandatory Markdown checkpoint/report maintenance with optional SQLite review notes on the existing result tool. Typed lifecycle records remain authoritative; historical design discussion below is retained. No reporting-only turn is required.

Coordination	Value
Assigned agent	Codex
Ticket state	implemented role/UI/cadence follow-through locally; deployment and live workflow acceptance pending
Last synchronized	2026-09-17

User problem

Technical Review mixes bug repair with architectural improvement. Recurring platform defects occupy review attention while opportunities in learning, KB, reports, databases, scripts and orchestration receive little independent focus. Strategy needs internal/external investigation and tracked outcomes rather than only reading the existing plan/dashboard. Approval, application and impact must be distinguishable.

Implemented

• Canonical Architecture module, focus history, finding/decision ownership, reports and research guidance; historical identities remain intact.
• Independent due decisions and sequential module stages with their own receipts and recovery. Protected Architecture/Strategy dates persist across Gate skips; explicit dated deferrals remain possible. See PLAT-303.
• Typed research scope on background dispatch: inherited authorized MCP/browser setup, research and Pulse custom tools, scoped checkpoint writes, no workflow edit/production execution custom tools. External connection grants are preserved; there is no new universal external read-only proxy.
• Architecture improvement records reuse the existing impact ledger. Decision answers and application receipts atomically advance linked Architecture and Strategy proposals. Architecture application/adoption require appropriate application/outcome evidence; applied records cannot reset to proposed.
• Health / Architecture / Strategy UI with separate Drift Check and compact improvement-to-outcome history. Full reports remain readable.

Acceptance

Automated tests cover all four modules due, independent completion after QA failure, protected boundary persistence, research tool filtering, Architecture coverage and finding ownership, and proposal/approval/application/assessment/ adoption transitions. Frontend tests cover areas, filters, history, and honest outcome presentation. Full affected Go packages, focused frontend tests and production frontend build pass; see the implementation document for exact scope.

Remaining / not claimed complete

Deploy and observe an actual autonomous workflow improvement. Finish immutable partial-run evidence under PLAT-047/089 and builder contract/learning prevention under PLAT-257/298. An automated lifecycle fixture is not evidence of business impact. No legacy workflow schema or learning file is rewritten by this change.

Implementation and operational contract

2026-09-17 role, control and outcome-focus follow-through

The first release still presented Health, Architecture and Strategy as peer work areas, allowed workflow-specific playbook focus to leak into technical and architecture review, and had no direct one-off reviewer controls. The role contracts and Pulse workspace now make the intended hierarchy explicit:

• Plan Drift is the exclusive compatibility prerequisite after plan changes. Technical, Architecture and Strategy wait for a clean current plan instead of reviewing stale structure or evidence.
• Technical owns concrete correctness failures and regressions. Architecture owns evidence-backed improvements to the technical construction of an otherwise working plan. Strategy owns goals, outcomes, measurement, assumptions, useful alternatives and self-improvement.
• Installed playbook custom focus is Strategy-only. Technical and Architecture retain their canonical platform-owned scopes rather than receiving arbitrary workflow-specific focus lists.
• The Pulse UI leads with goals, metrics and Strategy. Technical, Architecture and Plan Drift are grouped as platform health and stability. Strategy opens by default without filtering the complete issue backlog.
• Goal progress keeps the primary outcomes scannable: supporting/secondary metric cards are collapsed by default and appear only when the user opens their linked primary metric. Opening one primary does not reveal another primary's supporting metrics; unassigned supporting metrics produce only a compact setup notice until they are linked.
• The first review area now keeps Strategic Review and its proposal lifecycle together. strategy_experiment interventions appear as Strategic proposals directly beneath the Strategy card, before platform maintenance or the general decision queue. The former mixed Improvements block is now a lower Platform improvements section limited to Technical fix bundles and Architecture improvements, so those records no longer displace or duplicate Strategy's proposals.
• Technical, Architecture and Strategy each expose a durable Run automatically toggle and an independent Run now action. A disabled automatic reviewer can still run once by hand. Plan Drift remains mandatory, cannot be disabled, and has its own manual drift-check action. A due or unavailable Plan Drift status disables the three downstream manual actions.
• Manual actions reuse the canonical guided review contracts in the workflow's interactive Builder chat; they do not invent a second execution path.

Scheduling now has an explicit research-horizon asymmetry without adding new cron jobs or a universal fixed interval. Architecture normally waits across several comparable producing runs for stable structural evidence. Strategy is reconsidered at the next meaningful goal, outcome, feedback, experiment, decision or measurement checkpoint, so it normally receives the shorter research horizon. Material evidence and reached checkpoints can still override either wait.

Strategy also has a symptom-loop guard. A recurring operational label such as booking-heavy is Technical context, not a sufficient strategic agenda. The reviewer deduplicates or hands off the concrete defect, assumes it is fixed, then asks what would still limit the primary goal. Every completed Strategic Review must contain a goal-derived conclusion (including an honest no-change conclusion); a technical handoff alone is not completion.

Local verification

• 73 Pulse-focused frontend tests pass.
• The focused Goal Progress, metric-grouping and Pulse workspace suites pass, including primary-specific supporting-metric disclosure.
• Frontend TypeScript build passes.
• Complete cmd/server/guidance Go test suite passes, including regression coverage for distinct Architecture/Strategy horizons and the recurring symptom guard.
• git diff --check passes.

Deployment and a live RTS Pulse acceptance pass remain open. Live acceptance must confirm that Architecture is selected less often than Strategy over several comparable schedules, Strategy produces a goal-derived conclusion, and an existing booking-heavy Technical issue is not recreated or used as the entire Strategic Review.