Skip to content

Observability — workflows

Target

This page describes intended behaviour or structure. No implementation is claimed unless a section is explicitly marked Current implementation or Evidence.

Rewritten 2026-10-09 (design plan P8)

The June-2026 page is archived under catalogs 2026-06.

W1 — A run becomes a timeline

sequenceDiagram
    participant RC as RunControlService
    participant AR as AgentRuntimeService
    participant MR as ModelRouterService
    participant TA as ToolAdapterService
    participant CU as CostUsageService
    participant D as DashboardService
    participant BFF as Studio BFF
    RC-)D: RunAccepted (runId, snapshot hash, traceId)
    RC-)D: AttemptDispatched
    AR-)D: AttemptPhaseChanged (context, think, act, validate)
    MR-)D: ModelInvoked (receipt: profile, deployment, tokens, cost)
    TA-)D: ToolInvoked (receipt: tool, class, outcome)
    CU-)D: UsageRecorded (estimated → reserved → observed → reconciled)
    RC-)D: RunCompleted
    D-)BFF: RunProjectionUpdated (stream)
    BFF->>D: GetRunTimeline → the §51 tree as rows with receipt references
Hold "Alt" / "Option" to enable pan & zoom

Rules: the projection waits a short window for an out-of-sequence event, then rebuilds the run from the journal; content never enters the timeline, only hashes and references; usage amounts keep their status and uncertain amounts are never summed.

W2 — An objective breaches and someone is told

sequenceDiagram
    participant D as DashboardService
    participant S as SloService
    participant A as AlertRuleService
    participant N as StudioNotificationService
    participant C as ConnectorRegistryService
    D-)S: OperationsDigestProduced (window 1h)
    S->>S: evaluate objectives, burn rates, error budgets
    S-)A: SloBreached (task success below 0.9 for project X)
    A->>A: rule match, dedup key "slo:{sloId}", severity Warning
    A-)N: AlertFired → Studio
    A->>C: deliver to the tenant's chat connector (external class, receipt)
    Note over S: later window: value back above target
    S-)A: SloRecovered → AlertResolved
Hold "Alt" / "Option" to enable pan & zoom

W3 — A release is evaluated and gated

sequenceDiagram
    participant M as Asset maintainer
    participant AR as AgentRegistryService
    participant Q as QualityScoreService
    participant F as FeedbackService
    participant RC as RunControlService
    participant AS as ArtifactService
    M->>AR: submit release 1.4.0 of agent "backend-generator"
    AR-)Q: AssetReleaseSubmitted
    Q->>F: GetDatasetSnapshot (suite agent-core-v3 → dataset backend-tasks v7)
    Q->>RC: submit tasks as runs (evaluation project, sandbox class) for 1.4.0 and baseline 1.3.2
    RC-)Q: RunCompleted × n (receipts)
    Q->>Q: score, thresholds hold, baseline rule: task success 0.93 ≥ 0.91, tool errors 0.02 ≤ 0.03
    Q->>AS: evidence record (immutable)
    Q-)AR: EvaluationCompleted (Passed, evidenceRef)
    AR->>Q: GetVerdict at Published gate → Passed
    Note over Q,AR: a Failed verdict keeps the candidate InReview, a waiver needs an approval of class publication bound to the evaluation id
Hold "Alt" / "Option" to enable pan & zoom

W4 — Production drifts below the evaluated score

flowchart LR
    O[AttemptOutcomeRecorded · ModelInvoked · ToolInvoked per release] --> S[production outcome statistics]
    S --> DM[DriftMonitor every 6 h]
    DM -- observed < evaluated − tolerance --> DD[DriftDetected]
    DD --> AL[AlertRuleService → alert]
    DD --> FB[FeedbackService → re-evaluation proposal]
    DD --> MR[ModelRouterService capability evidence refreshed]
Hold "Alt" / "Option" to enable pan & zoom

W5 — Feedback becomes a reviewed proposal

sequenceDiagram
    participant R as Reviewer
    participant F as FeedbackService
    participant MP as MemoryPromotionService
    participant M as Asset maintainer
    participant AR as AgentRegistryService
    R->>F: RecordFeedback (correction on artifact, corrected artifact attached)
    MP-)F: SkillCandidateProposed (reinforced procedure)
    Note over F: nightly ProposalGenerator clusters evidence per asset (≥ 3 items), drafts a proposal with the structured profile
    F-)M: LearningProposalRaised (Studio notification)
    M->>F: DecideLearningProposal (accept)
    F->>AR: open Draft release of the skill with the proposed diff
    F-)AR: LearningProposalDecided (Accepted, resultingReleaseId)
    Note over AR: the release goes through W3 like any change
Hold "Alt" / "Option" to enable pan & zoom

W6 — The projection store is restored

The store is last in the D30 restore order. After a restore, ProjectionRebuilder replays Run Control's journal per run and the producers' events from the inbox checkpoints; objectives are re-evaluated from the rebuilt digests; alert instances that were open are re-derived from current objective and health states; evaluation evidence is unaffected because it lives in artifact-blob and is referenced, never copied.