50666d092f
Phase 3 commit #6 (final) per docs/PHASE3.md §12. Documentation-only; commented-out example showing the safety schema: - llm_second_opinion (bool, default true) - llm_model (string, default deep→default_model fallback) - max_norris_steps (int, default 8) The block notes the model-selection trade-off (R-B2): cloud is the independent-class fast option (costs money), deep is the local-but-slow option, fast is self-policing and NOT recommended. No behavior change to existing configs — safety defaults kick in when the block is absent. Phase 3 implementation complete: #1bd59ce7safety static patterns (34 rules) + 87-case test corpus #22abd5daLLM second-opinion + session cache + opts.max_tokens #3d2a53d2renderer Norris frames #411b1f56safety.norris_step planner (single iteration) #5a404b2arepl driver + \C-n real binding + :norris/:safety meta + readline rl_insert_text/rl_redisplay #6 (this) config.lua safety example block Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
127 lines
5.4 KiB
Lua
127 lines
5.4 KiB
Lua
-- config.lua — model registry, routing rules, user preferences.
|
|
-- Loaded with dofile() at startup; returns a plain Lua table.
|
|
-- See docs/PHASE0.md §10 for resolution order and full schema.
|
|
--
|
|
-- Per issue #12: hossenfelder is the canonical single-URL broker. It does
|
|
-- model-aware routing server-side (local models on boltzmann; cloud routes
|
|
-- through OpenRouter using its own bearer auth — no client-side key here).
|
|
-- Discovery: GET http://hossenfelder.fritz.box:8082/v1/models.
|
|
|
|
local HOSSENFELDER = "http://hossenfelder.fritz.box:8082"
|
|
|
|
return {
|
|
default_model = "fast",
|
|
|
|
models = {
|
|
fast = {
|
|
endpoint = HOSSENFELDER,
|
|
model = "qwen2.5-coder-1.5b-q4_k_m.gguf",
|
|
temperature = 0.2,
|
|
},
|
|
deep = {
|
|
endpoint = HOSSENFELDER,
|
|
model = "qwen3-30b-a3b-instruct",
|
|
timeout_ms = 1800000, -- 10 min; Nemo on RK3588 is patient work
|
|
temperature = 0.1,
|
|
},
|
|
cloud = {
|
|
endpoint = HOSSENFELDER,
|
|
model = "anthropic/claude-haiku-4.5",
|
|
temperature = 0.2,
|
|
},
|
|
},
|
|
|
|
shell = {
|
|
known_commands = {
|
|
"ls", "cat", "cd", "grep", "find", "cp", "mv", "rm",
|
|
"mkdir", "rmdir", "git", "make", "cmake", "gcc", "clang",
|
|
"python3", "luajit", "ssh", "scp", "curl", "wget",
|
|
},
|
|
capture_output = true, -- inject exec output into context
|
|
confirm_cmd = true, -- prompt before executing CMD: suggestions
|
|
},
|
|
|
|
context = {
|
|
max_turns = 40,
|
|
token_budget = 4096,
|
|
},
|
|
|
|
history = {
|
|
dir = (os.getenv("HOME") or ".") .. "/.local/share/aish",
|
|
},
|
|
|
|
-- Phase 2 (docs/PHASE2.md): MCP server registry + tool-call policy.
|
|
-- The block is OFF by default — connect-at-startup happens only when
|
|
-- `servers` is non-empty. Uncomment + adjust per your fleet.
|
|
--
|
|
-- mcp = {
|
|
-- servers = {
|
|
-- -- Each entry: alias = { url = "...", auth_token = "..." | auth_env = "..." }
|
|
-- -- auth_token literal > auth_env env-var indirection > nil (no auth).
|
|
-- -- Aliases become the namespace prefix on tool names sent to the model
|
|
-- -- ("<alias>__<tool>" — e.g. "boltzmann__list_dir"). The separator is
|
|
-- -- "__" (two underscores) because Anthropic via Bedrock validates tool
|
|
-- -- names against ^[a-zA-Z0-9_-]{1,128}$ — dots are rejected.
|
|
-- -- Aliases themselves must not contain "__".
|
|
-- boltzmann = {
|
|
-- url = "http://boltzmann.fritz.box:8080/mcp",
|
|
-- auth_env = "BOLTZMANN_MCP_TOKEN",
|
|
-- },
|
|
-- hertz = {
|
|
-- url = "http://hertz.fritz.box:8080/mcp",
|
|
-- auth_env = "HERTZ_MCP_TOKEN",
|
|
-- },
|
|
-- broglie = {
|
|
-- url = "http://broglie.fritz.box:8080/mcp", -- LAN-only, no auth
|
|
-- },
|
|
-- },
|
|
--
|
|
-- -- Per-call confirm gate auto-approve policy.
|
|
-- -- Key forms:
|
|
-- -- "<alias>__<tool>" — auto-approve one specific tool
|
|
-- -- "<alias>__*" — auto-approve every tool on that server
|
|
-- -- Anything not matched falls back to the [y/N] prompt.
|
|
-- auto_approve = {
|
|
-- ["boltzmann__read_file"] = true,
|
|
-- ["boltzmann__list_dir"] = true,
|
|
-- ["boltzmann__search_files"] = true,
|
|
-- ["hertz__*"] = true, -- trust the hub fully
|
|
-- },
|
|
--
|
|
-- -- Tool-call sub-loop budget per ask_ai turn. Hitting the cap surfaces
|
|
-- -- a status and breaks; default 8 if absent.
|
|
-- max_tool_depth = 8,
|
|
-- },
|
|
|
|
-- Phase 3 (docs/PHASE3.md): Chuck Norris autonomous mode + destructive-op
|
|
-- heuristic. The block is OFF by default (sane defaults kick in when
|
|
-- absent); uncomment to tune.
|
|
--
|
|
-- safety = {
|
|
-- -- LLM second-opinion on commands the static patterns don't flag.
|
|
-- -- Default true. Set false for static-only operation (faster, but
|
|
-- -- misses novel destructive patterns the static list doesn't know
|
|
-- -- about — bash -c content, custom destructive idioms, etc.).
|
|
-- llm_second_opinion = true,
|
|
--
|
|
-- -- Which configured model to use for the YES/NO destructive probe.
|
|
-- -- Precedence: this field → models.deep → models[default_model].
|
|
-- -- R-B2: prefer an INDEPENDENT model class from the action-emitting
|
|
-- -- model (avoids self-policing). Recommended values:
|
|
-- -- "cloud" — anthropic/claude-haiku-4.5 via openrouter. Fast and
|
|
-- -- reliable. Costs money per probe (typical Norris
|
|
-- -- session = 16 probes max, often cached).
|
|
-- -- "deep" — local large model (qwen3-30b on this fleet). Free
|
|
-- -- but slow on RK3588 hardware (~1-3s per probe).
|
|
-- -- Falls back here automatically if not set.
|
|
-- -- "fast" — same model as the action-emitter. NOT RECOMMENDED
|
|
-- -- (circular trust); use only when no other option.
|
|
-- llm_model = "cloud",
|
|
--
|
|
-- -- Norris planning-loop budget. Iterations of safety.norris_step.
|
|
-- -- Each iteration is one broker round-trip + dispatch of actions.
|
|
-- -- Default 8. Bump for long-running goals; cap low for testing.
|
|
-- max_norris_steps = 8,
|
|
-- },
|
|
}
|