local config support on top of #185

[i386]

Refs #184.
Depends on #185.

Users can now put startup models in ~/.mesh-llm/config.toml and run mesh-llm serve without repeating model flags in every launch path.

What changed

• mesh-llm serve now loads startup models from ~/.mesh-llm/config.toml
• the same local config file now covers both [[models]] and [[plugin]]
• startup models can be repeated and each model can set its own ctx_size
• GGUF multimodal entries can carry an explicit mmproj
• explicit CLI --model / --gguf still win over configured [[models]]
• explicit CLI --context-size still wins over configured per-model ctx_size
• service installs now run mesh-llm serve directly and expect startup models in config

Examples

Default config path:

version = 1

[gpu]
assignment = "auto"

[[models]]
model = "Qwen3-8B-Q4_K_M"

[[models]]
model = "bartowski/Qwen2.5-VL-7B-Instruct-GGUF/qwen2.5-vl-7b-instruct-q4_k_m.gguf"
mmproj = "bartowski/Qwen2.5-VL-7B-Instruct-GGUF/mmproj-f16.gguf"
ctx_size = 8192

[[plugin]]
name = "blackboard"
enabled = true

Start from config:

mesh-llm serve
mesh-llm serve --config /path/to/config.toml

Override config explicitly:

mesh-llm serve --model Qwen3-8B-Q4_K_M
mesh-llm serve --gguf ~/models/qwen2.5-vl.gguf --mmproj ~/models/mmproj.gguf
mesh-llm serve --model Qwen3-8B-Q4_K_M --context-size 4096

Bare mesh-llm serve with no configured [[models]] now warns, shows help, and exits cleanly:

⚠️ `mesh-llm serve` needs at least one startup model.
  Add `[[models]]` to ~/.mesh-llm/config.toml, or pass `--model` / `--gguf` explicitly.

Services

Background service installs now use mesh-llm serve as the service command:

• macOS launchd loads ~/.config/mesh-llm/service.env and executes mesh-llm serve
• Linux systemd --user writes mesh-llm serve into ExecStart=
• startup models now come from ~/.mesh-llm/config.toml, not service.args or custom service argv

The service restart policy now matches the new clean-exit path so a missing config does not loop forever:

• systemd: Restart=on-failure
• launchd: restart only on unsuccessful exit

Scope

Included here:

• unified local config.toml parsing for startup models plus plugins
• config-backed startup model resolution in mesh-llm serve
• per-model ctx_size
• optional mmproj
• service installer updates and docs

Still not in scope:

• pinned gpu_id assignment
• mesh-wide reconciliation/editor work

Validation

• cargo test -p mesh-llm
• bash -n install.sh

[Copilot]

Pull request overview

Adds phase-2 local node startup configuration so mesh-llm serve can load startup models (and their per-model settings) from the unified local config file, and updates service installers/docs to rely on that config instead of custom service argv.

Changes:

• Load and resolve startup models from ~/.mesh-llm/config.toml (with CLI models/context size overriding config).
• Extend unified config parsing/validation to include version, [gpu], and [[models]] entries alongside existing [[plugin]].
• Update install/service templates and documentation so background services run mesh-llm serve directly and don’t restart-loop on clean exit.

Reviewed changes

Copilot reviewed 12 out of 14 changed files in this pull request and generated 8 comments.

Show a summary per file

File	Description
README.md	Documents startup config, precedence rules, and service behavior changes.
mesh-llm/src/runtime/mod.rs	Implements startup-model planning/resolution from config + CLI precedence and “no models” clean-exit path.
mesh-llm/src/runtime/local.rs	Adds optional mmproj override support when starting local llama-server.
mesh-llm/src/runtime/discovery.rs	Adjusts “start new mesh” behavior to respect presence of configured startup models.
mesh-llm/src/plugin/mod.rs	Re-exports new config types (MeshConfig/ModelConfigEntry) for use in runtime.
mesh-llm/src/plugin/config.rs	Adds unified config schema for version/gpu/models and validates unsupported settings.
mesh-llm/src/cli/mod.rs	Makes --config user-visible and updates CLI help text to mention config-backed serve.
mesh-llm/README.md	Updates developer docs to describe unified local config (models + plugins).
mesh-llm/docs/TESTING.md	Adds a startup-config smoke checklist for manual testing.
mesh-llm/docs/DESIGN.md	Documents local-node config scope and CLI precedence rules.
install.sh	Switches services to run mesh-llm serve, removes service-args workflow, and updates messaging.
docs/USAGE.md	Documents config-backed serve and updated service install behavior.
dist/mesh-llm.service	Changes restart policy to on-failure to avoid restart loops on clean exit.
dist/com.mesh-llm.mesh-llm.plist	Restarts launchd job only on unsuccessful exit.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Design doc references --context-size, but the CLI flag is --ctx-size and there is no --context-size alias. Align the docs with the current CLI surface, or add an alias so the documented command works.

Suggested change

	- explicit `--context-size` overrides configured `ctx_size`
	- explicit `--ctx-size` overrides configured `ctx_size`

[i386]

@copilot apply changes based on the comments in this thread

[Copilot]

@copilot apply changes based on the comments in this thread

Applied all changes from the review thread in 100a649:

• runtime/mod.rs: Warning now uses plugin::config_path() to show the actual effective config path (--config override, MESH_LLM_CONFIG env var, or default) instead of the hardcoded ~/.mesh-llm/config.toml
• Docs (README.md, docs/USAGE.md, mesh-llm/README.md, mesh-llm/docs/DESIGN.md, mesh-llm/docs/TESTING.md): All references to --context-size updated to --ctx-size to match the actual CLI flag
• install.sh: Removed dead service-args plumbing — write_service_args_file, serialize_shell_args, parse_service_args, read_existing_systemd_args, and resolve_systemd_service_args (including the SYSTEMD_ARGS_COMMENT_PREFIX reference)