Supervise native Ollama serve lifecycle in Orchestrator

Context

Parent #13852 tracks the full native-Ollama lifecycle gap, including live heavy-session evidence. The current Orchestrator implementation slice is narrower: it wires an orchestrator-owned local-dev ollama serve task and reuses the already-delivered native Ollama readiness helper (#13865) for model residency/context convergence. This child exists so the PR can close the delivered L2 wiring without auto-closing #13852 while its live-evidence AC remains open.

Release classification: boardless (close-target split for an already implemented L2 wiring slice; not a release-blocking scope expansion).

Live latest-open sweep: checked latest 20 open issues on 2026-06-22; #13852 is the parent, #13865 is the readiness-only sibling, #13862 is compose-only, and no existing narrow ollama serve Orchestrator task leaf exists. A2A recency sweep: latest 30 messages checked; only @neo-gpt's own #13852/#13865/#13863 claims overlap this scope.

The Problem

lint-pr-body requires a truthful Resolves #N target. Closing #13852 from the L2 wiring PR would overstate delivery because #13852 still asks for live evidence: with provider set to native Ollama, a heavy session digests after orchestrator boot. The code slice is valuable and testable, but it should close a narrower leaf while #13852 remains the live-validation parent.

The Architectural Reality

• ai/daemons/orchestrator/taskDefinitions.mjs owns pure child-process task descriptors. It must not read env vars or re-resolve AiConfig directly.
• ai/daemons/orchestrator/Orchestrator.mjs is the entrypoint that reads AiConfig and forwards concrete values into buildTaskDefinitions().
• ai/services/graph/providerReadinessHelper.mjs#buildOllamaReadinessConfig is the native-Ollama role-selector source of truth.
• ai/services/graph/providerReadinessHelper.mjs#ensureOllamaModelsReady owns native Ollama warm-up and loaded-context verification. This ticket consumes that helper; it does not fork readiness logic.
• ADR 0019 applies: AiConfig is read at the use-site entrypoint and passed as concrete values into pure helpers.

The Fix

Add the local-dev native Ollama lifecycle lane:

1. Add orchestrator.ollama.enabled to the tracked Tier-1 config template.
2. Have Orchestrator build native Ollama readiness config from AiConfig and forward concrete host, roles, keep-alive, and resident-model requirements into task definitions.
3. Add an ollama continuous task that starts ollama serve only when roles select provider ollama.
4. Derive OLLAMA_HOST, OLLAMA_KEEP_ALIVE, OLLAMA_CONTEXT_LENGTH, and OLLAMA_MAX_LOADED_MODELS for the spawned server from canonical config values.
5. Use the existing ensureOllamaModelsReady() helper in liveness/preload paths so an already-running Ollama server converges residency/context without spawn churn.
6. Cover config defaults, pure task construction, Orchestrator getters, and continuous-task scheduling with focused unit tests.

Contract Ledger Matrix

Decision Record impact

Aligned with ADR 0019 (AiConfig reactive Provider SSOT). No ADR amendment required.

Acceptance Criteria

• Tier-1 config exposes orchestrator.ollama.enabled with env override.
• Orchestrator forwards native Ollama readiness config into buildTaskDefinitions() without duplicating role-selection logic.
• buildTaskDefinitions() creates no ollama task unless the lane is enabled and at least one role targets native Ollama.
• tasks.ollama starts ollama serve with environment derived from canonical provider/local-model config.
• Existing healthy Ollama endpoints are treated as shared infrastructure: no duplicate spawn churn, and readiness/preload still converges model residency/context.
• Focused unit tests cover the config, task descriptor, getter, and scheduling contracts.

Out of Scope

• Live heavy-session digestion evidence for the parent #13852.
• Compose-local-model serving caps (#13862 / PR #13863).
• Native Ollama loaded-context enforcement helper internals (#13865 / PR #13866).
• Model-name / JSON-schema provider parity (#13853/#13854).

Related

Parent: #13852. Siblings: #13862, #13865, #13851. Epic/root context: #13624, #13750.

Handoff Retrieval Hints: search Memory Core/A2A for #13852 local-dev native Ollama serve lifecycle, buildOllamaReadinessConfig, ensureOllamaModelsReady, and commit d11cee1c7f.