feat(dashboard-api): add per-service resource metrics endpoint

[yasinBursali]

What

• Add GET /api/services/resources returning per-service CPU, RAM, and disk metrics
• Create new routers/resources.py with parallel data fetching
• Register resources router in main.py

Why

• Dashboard needs per-service resource breakdown for monitoring and settings pages
• Container stats (CPU/RAM) available from host agent but no dashboard-api consumer exists
• Disk usage per service helps users understand storage consumption

How

• Parallel fetch: asyncio.to_thread for both host agent HTTP call (container stats) and local disk scan simultaneously
• Split cache TTLs: Container stats cached 20s (live data), disk scan cached 60s (slow I/O)
• Container name mapping: Builds reverse map from SERVICES dict container_name → service_id, correctly handling mismatched names (dream-webui → open-webui)
• Disk mapping: _DATA_DIR_MAP maps data directory names to service IDs (e.g., models → llama-server)
• Graceful degradation: When host agent unavailable, returns disk-only data (container stats null)

Three Pillars Impact

• Install Reliability: No installer changes
• Broad Compatibility: asyncio.to_thread (Python 3.9+, container is 3.11). Negligible resource usage
• Extension Coherence: Reports orphaned data directories not belonging to known services

New Files

• dashboard-api/routers/resources.py (137 lines)

Modified Files

• dashboard-api/main.py — router import + registration (2 lines)

Testing

Automated

• Python compile: PASS
• Existing tests: 173 pass

Manual

• GET /api/services/resources returns services with CPU/RAM/disk data
• Host agent down → returns disk data only, container fields null
• Verify cache: second request within 20s returns cached container stats
• No auth → 401

Review

• CG: ⚠️ APPROVED WITH WARNINGS (recommend adding test file — non-blocking)
• Security: PASS (auth enforced, read-only, no file mutations)

Platform Impact

• All platforms supported. WSL2 returns disk-only data (no host agent)

Sequence

PR 6 of 7 (Phase 3). Depends on PR 3 (#804 — host agent stats + dir_size_gb) and PR 5 (#808 — container_name in SERVICES)

[Lightheartdevs]

Audit: APPROVE

Good architecture — parallel fetch with asyncio.to_thread and split cache TTLs (20s for container stats, 60s for disk) is well-designed. The container name reverse-mapping handles naming mismatches correctly.

Auth enforced, read-only. Clean.

Minor notes (non-blocking):

• _DATA_DIR_MAP is hardcoded — new services require manual updates. Consider deriving from the SERVICES dict or manifests.
• 10s timeout on host agent fetch could cause the entire endpoint to hang when the agent is down. The 20s cache mitigates for subsequent requests.

Depends on #804. Should merge after it.

[Lightheartdevs]

Revised Audit: REQUEST CHANGES — fatal import crash

Withdrawing my earlier approval after deeper review.

BLOCKER: from helpers import dir_size_gb crashes without PR #804
dir_size_gb does not exist in helpers.py on main — it's a nested function inside main.py:_compute_storage(). PR #804 extracts it. If #810 merges first, the module-level import fails and the entire dashboard-api refuses to start — all endpoints go down, not just this one.

Soft dependency on #808: Without it, SERVICES[sid] lacks container_name. The fallback f"dream-{sid}" works for most services but breaks for open-webui (container is dream-webui, fallback generates dream-open-webui) and langfuse (dream-langfuse-web vs dream-langfuse). These services show null container stats even when running.

Other findings (non-blocking):

• _DATA_DIR_MAP is hardcoded and incomplete (~10 services missing) — unmapped services fall back to dir name convention
• Cache thundering herd on simultaneous requests with expired TTL — two fetches dispatched, second write wins. Acceptable for dashboard.
• opencode has container_name: "" creating an empty-string key in the reverse map (cosmetic)

What's good: Auth enforced, no SSRF (URL from server config), no path traversal (child.name only), no info leaks (relative paths), bounded resource usage (always 2 threads max regardless of service count), cache design with split TTLs is well-thought-out.

Required merge order: #804 → #808 → #810

[yasinBursali]

Addressing review feedback

Re: from helpers import dir_size_gb BLOCKER — Acknowledged. This PR requires #804 to merge first. The function is extracted to module-level in #804.

Re: container_name soft dependency — Acknowledged. Without #808, open-webui and langfuse show null container stats due to non-standard naming. #808 adds container_name to the SERVICES dict which fixes this.

Required merge order: #804 → #808 → #810

No code changes needed in this PR — all issues are addressed by the dependency chain.

[Lightheartdevs]

Note: The Rust dashboard-api rewrite (#821) merged and deleted the Python files this PR modifies.

This PR added a Python endpoint and router that no longer exist. The per-service resource metrics feature needs to be reimplemented as a Rust endpoint in the dashboard-api crate.

Please rebase or rewrite against the current main branch.