K8s fleet roster/observe UI — console can't drill into an existing k8s fleet

[brettchien]

Problem

Clicking a k8s-runtime fleet in the Fleets panel doesn't drill into its member roster. selectFleet (console/src/main.ts:448-463) explicitly declines and shows an info toast — this is a known, intentional stopgap (see PR #145's description), not a regression. Before #145 (2026-09-06) k8s fleets weren't even visible in the Fleets list, so this is newly reachable but was never wired.

Root cause (confirmed by reading current main, 2026-09-07)

The entire roster/observe read path is ECS-only, several layers deep:

• oab-mcp's target() (crates/oab-mcp/src/lib.rs:449-459) hard-errors for a k8s-runtime fleet: "fleet {name} is a k8s-runtime fleet; this tool only supports ecs fleets".
• t_runtime_context (crates/oab-mcp/src/lib.rs:850) unconditionally calls scp::observe_identity(&aws) (AWS/STS) — never branches to the already-existing observe_k8s_identity (crates/studio-cp/src/lib.rs:281, built in Phase 1 / studio#63 slice 3e) even though RuntimeContext's shape is already unified across both drivers.
• There is no k8s-side roster tool at all — deploy_list/t_list has no equivalent of "list Deployments in this namespace, mapped to the console's Deployment shape" the way ECS's roster_over_mcp (src-tauri/src/lib.rs:261) does.
• console/src/main.ts's tick(), and source.ts's listDeployments/runtimeContext, are cluster-keyed throughout — no context/namespace equivalent path exists on the console side either.

ProvisionDriver (crates/oabctl/src/driver.rs) only has apply/scale/delete — no list/observe method on the trait, so there's no existing seam to hang a k8s roster read off of.

Scope (rough slices, same shape as studio#63/#104's phased delivery)

1. studio-cp: a k8s roster read — list Deployments in a (context, namespace), scoped to a fleet's members (mirroring crates/oabctl/src/k8s_driver.rs's existing kube client setup), mapped to a shape the console can render alongside ECS's Deployment type (replicas/ready vs ECS's desired/running count, image, etc.).
2. oab-mcp: fix t_runtime_context to dispatch on binding.runtime (observe_k8s_identity for k8s, existing path for ecs); add the new roster tool from (1); relax target()'s hard error only for the calls that now support both runtimes (identity/roster), not for the ones that genuinely stay ecs-only (apply/scale/delete stay behind the existing k8s dispatch added in studio#104/feat(studio-cp,oabctl): k8s deploy_provision dispatch (studio#104, resumed after #111) #114 instead, not this trait).
3. src-tauri: extend deploy_list/runtime_context bridge commands (or add k8s-specific ones) to pass context+namespace instead of cluster when the target fleet is k8s. New bridge commands need the usual "no local signal, bundle-macos CI only" caveat ([[studio-tauri-bridge-layer-gotcha]]).
4. console: main.ts's selectFleet/tick branch on fleet.runtime instead of declining; source.ts interface + Deployment/RuntimeContext types grow the k8s-shaped fields; fixtures/tests updated.

Not in scope

Scaling/starting/stopping k8s deployments from the console — that's a separate action surface (ProvisionDriver::scale/delete already exist for k8s via studio#104/#114's provider dispatch on the create path; wiring the existing-fleet scale/delete buttons to it is a follow-on, not blocking the read-only roster view this issue is about).

🤖 Filed by Orca

[brettchien]

More detail (2026-09-07, after reading agent-lifecycle/k8s_driver.rs more closely)

Good news: the design work for this is already done, only the code is missing. Corrects/sharpens the slice-1 scope from the issue body.

The canonical 6-state model already has a k8s row. docs/adr/agent-lifecycle.md §6's projection table (line ~140) maps k8s Pod signals onto the same Discriminator axes ECS uses: deletionTimestamp != null ⟺ ECS desiredStatus STOPPED (Stopping), Pod phase + readinessProbe ⟺ Starting/Running, Unknown (node lost) → Unhealthy(fenced), etc. This was drafted alongside ECS's row, not an afterthought.

Only the ECS side got implemented. crates/agent-lifecycle/src/ecs.rs implements RuntimeDriver for ECS (EcsDriver::project(&EcsTask, verified_before) -> Discriminator, ~230 lines incl. tests) — there is no k8s.rs sibling. crates/oabctl/src/k8s_driver.rs's own module doc says this explicitly (written during studio#63 slice 3b, i.e. this was known and deliberately deferred, not missed): "Observing k8s state into the canonical 6-state ... is also out of scope here — it's substantial enough on its own ... to land as its own follow-up." This issue is that follow-up.

The console side needs less than slice-4 originally said. console/src/types.ts's Deployment (desired/current/ready + instances: InstancePhase[], each InstancePhase.state: AgentState) is already runtime-agnostic — it's expressed in the canonical 6-state, not ECS-specific fields. So a k8s roster tool that fills the same JSON shape needs no console type changes for the roster data itself. What still needs console changes is purely the read-path plumbing: main.ts's selectFleet/tick pass an ECS cluster string today; they need to pass k8s context/namespace instead when fleet.runtime === "k8s".

Revised slice 1 (studio-cp / agent-lifecycle): add crates/agent-lifecycle/src/k8s.rs — a K8sPod native-observation struct (phase, conditions/readiness, deletionTimestamp, a CP/director cordon signal for accepting_work) + K8sDriver::project() implementing RuntimeDriver, mirroring ecs.rs's shape and directly off the ADR table (no new design decisions needed, a fairly mechanical port). Then a kube::Api<Pod> (or Deployment+Pod) list scoped to (context, namespace) + fleet members, feeding each observation through project(), shaped into the existing Deployment/InstancePhase JSON.

Slices 2–4 (oab-mcp dispatch fix + new roster tool, src-tauri bridge, console read-path branch) are unchanged from the issue body.