feat(ai): proactive model prewarming and vram eviction governance (#127)

[dan-petty]

Closes #127. Implements devops ai prewarm and ai_prewarm_models FastMCP tool.

[Copilot]

🟡 Changes recommended

Unresolved issues affect request semantics, node selection, eviction behavior, and endpoint redaction.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

Adds Ollama model prewarming and VRAM eviction through CLI and FastMCP interfaces.

Changes:

• Adds configurable model, node, keep-alive, and eviction controls.
• Adds devops ai prewarm and ai_prewarm_models.
• Adds tests, help text, documentation, and task tracking.

File summaries

File	Summary
tests/test_ai_prewarm.py	Tests client, CLI, and MCP behavior.
src/devops_cli/lang/en/help.py	Adds localized command help.
src/devops_cli/config/defaults.py	Adds prewarm defaults.
src/devops_cli/commands/ai.py	Adds CLI command and result rendering.
src/devops_cli/ai/mcp/server.py	Adds FastMCP prewarm tooling.
src/devops_cli/ai/client/ollama.py	Implements prewarming and eviction requests.
README.md	Updates the command listing.
docs/MCP_TOOLS.md	Documents the MCP tool.
docs/commands/ai.md	Documents CLI usage.
docs/CLI_REFERENCE.md	Updates the CLI reference.
docs/agent/tasks/task-127-ai-model-prewarming-vram-eviction.md	Records task objectives and verification.

Review details

Suppressed comments (3)

src/devops_cli/ai/mcp/server.py:1038

• When an MCP caller passes all_nodes=False, this branch emits no replacement flag. The subprocess then uses the CLI's default all_nodes=True, so the tool still prewarms every configured node instead of honoring the caller's request; pass an explicit single-node/negative flag supported by the CLI.

    if all_nodes:
        cmd.append("--all-nodes")

src/devops_cli/commands/ai.py:749

• The help and docs allow --keep-alive 0 as an eviction form, but the rendered action is derived only from --evict. Thus prewarm --keep-alive 0 sends an unload request and reports prewarming/prewarmed into VRAM instead of eviction. Derive the display action from the effective keep-alive value or reject 0 unless --evict is set.

        json_output=json_output,

src/devops_cli/commands/ai.py:669

• These new console lines interpolate the raw configured node URL. print_success and print_error do not apply the output secret sanitizer, so a URL containing userinfo or other credential material can be copied into terminal or MCP output. Redact the endpoint before formatting it and keep the raw URL only for the request.

    for endpoint, ok in results.items():
        if ok:
            print_success(f"  ✓ {endpoint}: confirmed HTTP 200 (keep_alive={keep_alive})")
        else:
            print_error(f"  ✗ {endpoint}: request failed or timed out", prefix=False)

• Files reviewed: 11/11 changed files
• Comments generated: 5
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.