API and configuration

All presenter and internal evidence endpoints require Authorization: Bearer TOKEN. The static SPA shell is public. The OpenAI API key remains in server configuration.

Presenter API

Method	Endpoint	Purpose
GET	/api/state	Execution mode, release availability, latest evaluation, and recent activity
GET	/api/progress	Read live stage, analysis check budget, per-cohort counts, and investigation readiness without starting agents
GET	/api/cluster	Rollout and Argo CD Application status
POST	/api/reset	Clear faults and current presentation; restore v1 through Git when needed, preserving gates
POST	/api/release/{version}	Commit v1/v2/v3 to GitHub, request immediate Argo CD refresh; return version, SHA, commit URL, release identity, and refresh status
POST	/api/evaluate	Start an investigation
GET	/api/evaluations/{id}	Inspect one investigation
POST	/api/evaluations/{id}/approve	Approve an eligible pending production decision
GET	/api/logs/{source}	Read a bounded Kubernetes log tail
POST	/api/fault/{seconds}	Inject or clear the shared inventory slowdown
GET	/api/analysis?revision=v2	Return the effective verdict for Argo analysis

Reset and release actions reject concurrent investigations. A reset on an already Healthy stable v1 does not create a new rollout. Offline reset does not touch Git or Kubernetes. /api/state includes resetEnabled and releaseIntent; /api/cluster includes actual stableVersion and candidateVersion labels. The empty initial decision is not a GO: /api/analysis remains HOLD until eligible fresh evidence exists.

Dashboard release responses include refreshRequested as the string "true" or "false". The server requests a hard refresh of the named canary-checkout Application only after Git accepts the commit. A failed refresh does not roll back or invalidate that commit; the response reports the automatic-sync fallback. The control service needs its existing Application read permission plus the named-Application patch Role installed by scripts/configure-release-refresh.sh. This presenter action is not an agent tool, does not patch the Rollout, and does not bypass analysis or approval gates.

Cluster evidence includes canaryPodHash, captured from Rollout status independently of trace observations. Withholding traces preserves this metadata and records the missing prerequisite; it does not turn a current HOLD into an unrelated historical investigation. The SPA, progress guidance, and analysis endpoint still reject mismatched rollout stages, pod-template identities, and expired authorization. Required-evidence failures identify the actual gap separately from insufficient request counts.

Start an investigation

The endpoint accepts query parameters and does not require a JSON body:

curl -X POST \
  -H "Authorization: Bearer $DEMO_API_TOKEN" \
  'http://localhost:8095/api/evaluate?scene=retry&profile=development&revision=v3'

Valid scenes are healthy, slowdown, retry, and missing. Valid profiles are development and production. Candidate versions are v1, v2, and v3.

In cluster mode, a scene selection does not fabricate normal telemetry. The collector reads actual running stable and canary pods. The Missing evidence scene deliberately withholds trace observations and labels that injected outage.

Log sources

agents is a UI view of investigation events. Kubernetes log endpoints accept:

• checkout
• inventory
• quarkus
• rollouts
• argocd

The SPA polls every four seconds, with up to three matching pods and bounded line counts. This is a live console for the demo, not a centralized logging backend.

Environment

Variable	Default or requirement	Effect
OPENAI_API_KEY	Required for live model calls	Server-side GPT-4o mini credential
DEMO_AI_MODE	live	Use fixture for explicit deterministic rehearsal
DEMO_API_TOKEN	Required in packaged deployment	Authenticate presenter and internal requests
DEMO_PORT	8095 in the local launcher	Local development HTTP port
DEMO_CLUSTER_ENABLED	false	Enable cluster evidence collection
DEMO_NAMESPACE	canary-demo	Workload namespace; installer currently supports this fixed name
DEMO_GITOPS_NAMESPACE	openshift-gitops	Runtime Argo CD namespace
DEMO_ROLE	control	Run the image as control, checkout, inventory, or traffic
DEMO_RELEASE_ENABLED	false (cluster control manifest enables it)	Allow presenter Git release actions when a credential is configured and mode is live
DEMO_GITHUB_TOKEN	Optional server-side Secret	Repository contents write credential; never sent to the SPA
DEMO_RELEASE_REPOSITORY	danieloh30/intelligent-canary-delivery	GitHub repository; must match Argo CD source
DEMO_RELEASE_BRANCH	main	Release branch; must match Argo CD target revision
DEMO_RELEASE	v1	Select the workload behavior
OTEL_ENDPOINT	http://jaeger:4317	OTLP gRPC collector endpoint
REPO_URL	This GitHub repository	Installer's Argo CD source repository
GITOPS_CHANNEL	Operator catalog default	Select a GitOps channel when installing a new operator

Local development defaults the presenter token to demo-local. Production manifests inject a generated token from a Secret.

Outcome contract

The application distinguishes three things:

1. Recommendation: the model's proposed verdict and explanation.
2. Effective decision: the result after Decision Policy checks.
3. Rollout outcome: the controller's actual status after consuming the analysis result.

The decision API rechecks release and stage identity. A pending, expired, fixture, or mismatched result returns HOLD. Never infer that a rollout advanced merely because an agent said GO; inspect the controller state.

Progress and investigation timing

GET /api/progress returns state, ready, revision, podHash, stepIndex, targetWeight, stages, stableRequests, canaryRequests, stableP95Ms, canaryP95Ms, samplesAvailable, capturedAt, missing, and AnalysisRun details (analysisName, analysisPhase, checksCompleted, checksTotal, interval). It is authenticated, briefly cached, and read-only. Its counts precede investigation and never replace the frozen evidence used for a decision. The browser withholds readiness for stale snapshots, mismatched stages/candidates, and unavailable telemetry. The investigation endpoint collects fresh evidence again when explicitly invoked.

The configured 15-second pause occurs before the inline analysis step. During analysis, Argo polls the saved verdict up to 180 times at five-second intervals. This is roughly a 15-minute analysis budget, not a 15-second click window or 180 LLM calls. A completed Inconclusive run cannot be restarted by polling progress or starting another investigation; request a fresh Git release/reset.