gabeujin/Harness Dashboard State Manager

Harness Dashboard State Manager

Use this skill when the generated dashboard surfaces need truthful updates that stakeholders and AI Agents can scan quickly.

Core Rules

1. Treat events/harness-events.jsonl as canonical truth; projection JSON and HTML are rebuildable.
2. Keep dashboard state in JSON files only. Do not introduce a database.
3. Prefer short stakeholder-friendly labels over engineering shorthand.
4. Distinguish observed, declared, planned, inferred, and stale facts clearly.
5. Keep work status explicit: waiting, in-progress, completed, blocked, failed, needs-user, and real-world.
6. Do not store secret values; references only.

Required Sections

• projectWorldModel: mission, stakeholders, products, services, environments, repos, data assets, dependencies, risks, decisions, cadence, metrics, constraints
• stakeholderBrief: current goal, what changed, why it matters, current risk, required decision, owner, due date, next milestone, evidence links
• agentResumeBrief: current project goal, active session, last safe checkpoint, next safest action, blockers, validation commands, authoritative files, confidence
• governanceEvidenceBrief: evidence coverage, missing claims, unlinked VCS changes, unresolved decisions, stale projections, failed validations
• taskQueues, decisionContracts, agile, operationsTimeline, releaseReadiness, incidentReadiness, sloSli, databaseReadiness
• dictionary, ontology, embeddingProjection, versionControl, vcsChangeRecords, and listener

Evidence Guidance

• Treat healthy, released, operational, connected, approved, or complete states as evidence claims.
• Record event sequence, source, timestamp, freshness, confidence, owner, and evidence links.
• Link Git commits or SVN revisions to sessions, tasks, decisions, releases, or warnings.
• Never show placeholder release, incident, SLO, database, or service-health evidence as current.

World Model Maturity Guidance

• Treat governanceActionabilityScore.score as the current governance setup/readiness-to-govern score, not as production readiness.
• Treat governanceActionabilityScore.operationalEvidenceScore as the separate operational proof score.
• Improve both scores gradually by adding evidence, resolving decisions, linking VCS, clarifying owners, and converting missing claims into owned work.
• Never inflate a score by hiding warnings, deleting missing evidence, or treating declared/tacit context as observed fact.
• Use worldModelImprovementContract as the Agent contract for maturity loops, score improvement policy, tab ownership, and non-duplication rules.
• Ask the user for tacit real-world context when repository evidence cannot explain operational constraints, stakeholder intent, release windows, support expectations, or external account state. Store it as declared with owner, timestamp, confidence, and reversal condition.

Tab Ownership And Non-Duplication

• Overview answers only: what changed, why it matters, current trust, current decision, and score meaning.
• Work owns open tasks, progress, Kanban/Gantt, blockers, remaining effort, and KPI timing inputs.
• Evidence owns claim-to-proof, missing evidence, VCS linkage, validation, and falsification status.
• Governance owns decisions, gates, values, cadence, reviews, retros, and improvement actions.
• Operations owns services, environments, dependencies, releases, incidents, SLO/SLI, database, and listener health.
• Tech Stack owns architecture, runtime stack, infrastructure, deployment path, data stores, and harness integration.
• Do not duplicate the same table across tabs. If a fact must appear in multiple tabs, each tab must answer a different stakeholder question about that fact.

Dashboard Collaboration Protocol

• When the user references a dashboard tab, task, claim, decision, score, or blocker, respond as if both user and Agent are looking at the same dashboard projection.
• Name the relevant tab, item id or label, evidence path, current score impact, and safest next action.
• Prefer generated harness skills and get_harness_dashboard_context over ad hoc reading of unrelated files.

Work Guidance

• Map legacy TODO, FIXME, OPEN, and unchecked boxes to waiting.
• Use needs-user for decisions, approvals, or confirmations.
• Use real-world for work agents cannot perform, such as purchases, external account setup, physical device access, or service procurement.
• Use failed only when there is clear evidence of failure.

Safety Guidance

• Do not introduce a database dependency.
• Do not turn the read-only local listener into a hidden backend app.
• Do not store secrets, raw tokens, private key material, environment variable values, or private deployment details in dashboard state, preview, or logs.
• Do not embed restricted or secret material unless an explicit governance event permits it.

Output Standard

• A stakeholder can open the single HTML dashboard and understand project state in seconds.
• A fresh AI Agent can read dashboard-index.json or call get_harness_dashboard_context and resume work without chat history.