marfrit/aish
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
cb37fa861a6ed58048e931877df172d9c41e2951
aish/config.lua
T
marfrit cb37fa861a phase10: config.lua example for cfg.norris.{preplanner,executor} 
Documents the new Phase 10 / #89 surface for users: when, why, and
how to set cfg.norris.preplanner + .executor + .tasks_max. Notes the
graceful single-model fall-back when preplanner is unset OR fails,
and the design choice that preplan does NOT route via call_broker
(retry would silently swap planning models).

C5: config-only commit; no behavior change.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-17 08:22:22 +00:00

412 lines
20 KiB
Lua
Raw Blame History

-- 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.
--
-- Phase 9 (docs/PHASE9.md): a `.aish.lua` in/above your cwd (walking up
-- to $HOME) overlays this user config. First encounter prompts to trust;
-- sha256-pinned in ~/.aish/trusted-projects. Use it for repo-specific
-- model presets, permissions, hooks, etc.
--
-- IMPORTANT: shallow merge. If your `.aish.lua` sets a top-level block
-- (models, permissions, cost, shell, ...), it REPLACES the user's
-- entire block — list every entry you want available OR omit the block
-- to keep the user's. Inspect the merge via `:config show` at runtime.
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,
-- 2026-05-13: qwen3-30b not loaded on hossenfelder right now;
-- using deepseek-coder-v2-lite (16B MoE, ~2.4B active) for the
-- time being. Restore qwen3-30b when the slot is back up.
model = "deepseek-coder-v2-lite",
timeout_ms = 300000, -- 5 min; MoE inference is faster than dense 30B
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
-- Issue #10: prompt template. When set, replaces the default
-- "[aish:<model>]> " prompt. Variables (substituted via {name}):
-- {model} {ctx_used} {ctx_max} {turn}
-- {cwd} {cwd_short} (cwd with $HOME -> ~)
-- {last_status} (last exec exit code, empty if none yet)
-- {mode} (norris / plan / normal)
-- prompt = "[{model} {ctx_used}/{ctx_max}t T{turn} {mode}] {cwd_short} > ",
},
context = {
max_turns = 40,
token_budget = 4096,
},
history = {
dir = (os.getenv("HOME") or ".") .. "/.local/share/aish",
},
-- Issue #3: pre/post CMD hooks. Optional shell scripts triggered around
-- every CMD: execution. Each hook receives the command on stdin and
-- AISH_CMD / AISH_TURN / AISH_CWD as env vars. Non-zero exit on pre_cmd
-- aborts execution; post_cmd's exit code is ignored but its stdout is
-- logged. Default off (no hooks). Uncomment to enable.
-- hooks = {
-- pre_cmd = (os.getenv("HOME") or ".") .. "/.aish/hooks/pre-cmd",
-- post_cmd = (os.getenv("HOME") or ".") .. "/.aish/hooks/post-cmd",
-- },
-- Issue #13: secret redaction. Vault is a separate file at ~/.aish/
-- secrets.lua (mode 0600 enforced). When set, outbound broker messages
-- are scrubbed: vault literals + autodetect heuristics (OpenAI sk-,
-- OpenRouter sk-or-v1-, GitHub ghp_/gho_/ghs_, AWS AKIA, JWT eyJ...,
-- SSH/GPG PRIVATE KEY headers) become $AISH_SECRET_NNN placeholders.
-- The streamed reply is rehydrated before display so the user sees
-- real values. Per-broker override via models[*].redact:
-- "off" -- no scrubbing (trusted local)
-- "vault" -- vault literals only
-- "vault+autodetect" -- + heuristics (default when vault loaded)
-- "stealth" -- + heuristics, opaque decoys, no rehydrate
-- Default per-broker is the global config.secrets.default, falling
-- back to "vault+autodetect" when vault loaded, else "off".
-- secrets = {
-- vault = "~/.aish/secrets.lua",
-- default = "vault+autodetect", -- applies when models[*].redact is nil
-- },
-- Issue #8: background CMD (CMD&: marker). Requires history.dir set
-- (logs land at <history.dir>/bg/<id>.log + .status sidecar). The
-- feature is always-on once history.dir exists — no config flag — but
-- only fires when the model emits "CMD&: " or the user runs :bg-spawn.
-- Issue #9: permission policy DSL for AI-suggested CMD: lines. When set,
-- supersedes shell.confirm_cmd. Patterns are Lua patterns (NOT regex)
-- per substrate invariant §3 (no compiled extensions). Priority order:
-- deny > confirm > allow; first match in the chosen category wins.
-- Unmatched commands default to "confirm". Probe with :perms check <cmd>.
-- permissions = {
-- allow = { "^ls%s", "^cat%s", "^git status", "^git diff" },
-- confirm = { "^rm%s", "^git push", "^docker%s", "^sudo%s" },
-- deny = { "^ssh%s+root@", "^curl%s+http[^s]" },
-- },
-- 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,
-- },
-- Phase 4 (docs/PHASE4.md): cross-session memory.jsonl + startup
-- injection + :memory management surface. The block is OFF by
-- default (no startup injection); uncomment to tune. Note that
-- :remember / :memory list / :memory forget / :memory summarize
-- all work without this block — they store to <history.dir>/
-- memory.jsonl regardless. The block only configures the
-- injection-into-system-prompt behavior at startup.
--
-- memory = {
-- -- Cap on total characters injected at startup. ~2000 chars ≈
-- -- 500 tokens. LRU-by-ts selection if your memory.jsonl has
-- -- more recent items than fit. Older items remain in the
-- -- file; only injection is bounded. Suppressed entirely in
-- -- Norris mode (R-C1).
-- inject_max_chars = 2000,
--
-- -- Which configured model to use for :memory summarize.
-- -- Defaults to the active model when nil. Use "fast" for
-- -- speed; "deep" or "cloud" for better extraction quality
-- -- (cloud may have variable cost per session).
-- summarizer_model = "fast",
-- },
-- Phase 5 (docs/PHASE5.md): multi-model routing + cloud fallback +
-- summarize-on-evict. OFF by default — auto-routing can spend money
-- silently on the cloud preset; require explicit opt-in.
--
-- routing = {
-- -- Enable auto-routing per request. When true, router.classify_model
-- -- inspects each prompt and may switch the model for THAT request
-- -- only (the :model selection is preserved across requests).
-- -- Default false. Toggle at runtime with :route on / :route off.
-- auto = true,
--
-- -- Class → model mapping. nil = "keep current" (heuristic fires
-- -- but no override). Ships with reasoning = nil because mapping
-- -- "explain ..." prompts to a paid cloud model would spend money
-- -- silently — opt in by uncommenting the reasoning line below.
-- classes = {
-- code = "deep", -- code-like prompts to local deep
-- -- reasoning = "cloud", -- OPT-IN: "explain"/"why"/"how does" → paid
-- -- default = nil, -- keep active model
-- },
--
-- -- Single-hop retry on transport failure (HTTP 5xx, 408,
-- -- 404 model_not_found, DNS, connection refused, timeouts).
-- -- Retries against fallback_model once. Skipped if any text
-- -- has already streamed (no partial-output duplication).
-- -- Toggle at runtime with :fallback on / :fallback off.
-- fallback = false, -- default off (cost-safety)
-- fallback_model = "cloud",
--
-- -- Issue #86: per-class system_prompt override. When the
-- -- classified request falls into a class with an entry here,
-- -- the BASE system_prompt is REPLACED for that one request
-- -- (dynamic blocks — [background], [project], [earlier
-- -- summary], NORRIS suffix — still compose on top). Mostly
-- -- useful for tightening small local models' instruction
-- -- adherence. Default {} (no override).
-- system_prompts = {
-- code = [[You are a code assistant. Rules:
-- 1. Output ONLY the requested code or command.
-- 2. No prose explanation unless explicitly asked.
-- 3. Wrap shell commands in CMD: prefix.
-- 4. Max response: 200 tokens.]],
-- default = [[You are a shell assistant.
-- Output shell commands as: CMD: <command>
-- Output answers as single short sentences.
-- Do not ask clarifying questions.]],
-- -- reasoning routes to cloud; no override usually needed
-- },
--
-- -- Issue #88: per-class GBNF grammar passthrough. llama.cpp
-- -- constrains the sampler to ONLY emit tokens matching the
-- -- grammar — eliminates format drift on small models. Cloud
-- -- (Anthropic/Bedrock) silently ignores the field, so default
-- -- passthrough is safe; no per-model opt-out needed. Misformed
-- -- grammar surfaces as a broker error at request time.
-- grammars = {
-- code = [[root ::= "CMD: " [^\n]+ "\n"]],
-- default = [[root ::= ("CMD: " [^\n]+ "\n") | [^\n]+ "\n"]],
-- },
-- },
--
-- Issue #88 (continued): for the safety LLM probe (YES/NO
-- destructive classification), set safety.probe_grammar to force
-- the probe model to emit exactly YES or NO. Eliminates the
-- regex-match fallback for unparseable verdicts; small models
-- become reliable enough to use as the probe.
--
-- safety = {
-- llm_second_opinion = true,
-- llm_model = "fast",
-- probe_grammar = [[root ::= ("YES" | "NO")]],
-- },
-- ── Issue #87 (route-aware context compression).
-- When a routed model preset has `local_compress = true`, each
-- broker call against THAT preset gets a compressed view of
-- ctx.turns: only the last `keep_turns` turns; any turn whose
-- content exceeds `max_turn_chars` is tail-truncated. The full
-- context lives on (visible via :history); compression is purely
-- per-request for small models that effectively use a fraction
-- of their advertised context window.
--
-- Set the per-model opt-in on models[<name>]:
-- models.fast = { ..., local_compress = true }
-- Defaults live under context.compress:
-- context = {
-- ...
-- compress = { keep_turns = 2, max_turn_chars = 800 },
-- }
--
-- Trade-off documented in the FR: tool turns lose information
-- when tail-truncated. Acceptable for shell-output blocks (the
-- tail is usually the relevant bit); known limitation for
-- structured tool results. Disable per-model if it bites.
-- ── Issue #89 / Phase 10: cloud preplanner → local executor split.
-- When cfg.norris.preplanner names a model preset, :norris launch
-- fires ONE broker.chat against that preset asking for a sequence
-- of TASK: <imperative> lines. Parsed list (capped at tasks_max)
-- becomes ctx.norris_tasks; the executor model (cfg.norris.executor,
-- defaulting to the active :model selection) runs each task with
-- the current task shown in the per-step header.
--
-- Goal: small fast local models are cheap per step but easily
-- distracted on multi-step plans; cloud is capable at planning
-- but expensive per step. Use cloud ONCE for the plan, local for
-- every step. Falls back to single-model Norris (existing
-- behavior) when preplanner unset / fails / produces no TASKs.
--
-- norris = {
-- preplanner = "anthropic", -- model name in cfg.models;
-- -- this preset is called ONCE per
-- -- :norris launch. Omit to run
-- -- single-model (Phase 6 behavior).
-- executor = "fast", -- model that runs each step.
-- -- Omit to use the active :model.
-- tasks_max = 16, -- cap on preplan list size.
-- -- preplan_system = "...", -- override the built-in prompt
-- },
--
-- :cost detail separates norris-preplan and norris rows so you
-- can see cloud planning cost vs local execution cost. The
-- preplan call does NOT retry via fallback_model (a different
-- model = a different decomposition; clean hard-fail to single-
-- model is safer).
-- ── Phase 5 context summarization on sliding-window eviction.
-- Set INSIDE the context = { ... } block above to enable:
-- context = {
-- max_turns = 40,
-- token_budget = 4096,
-- summarize_on_evict = true,
-- summarizer_model = "fast", -- model name in models{}
-- max_summary_chars = 2000,
-- },
-- When summarize_on_evict is true, evicted turn pairs are fed to
-- summarizer_model and the result lives on ctx.summary, appended to
-- the system prompt as [earlier conversation summary]. Suppressed
-- in Norris mode (R-C4 — planner stays on its goal). If broker
-- fails, falls back to Phase 0 silent eviction (no crash).
-- Phase 6 (docs/PHASE6.md): project file-tree context + :diff /
-- :tree / :highlight metas. The :diff and :tree metas work without
-- any config. The `project` block below only controls the
-- AUTO-injection-at-startup behavior; manual `:tree` always works
-- regardless. Uncomment to enable startup auto-inject.
--
-- project = {
-- auto_tree = true, -- run `:tree` once at startup
-- tree_depth = 3, -- depth filter for the scan (find fallback only;
-- -- git ls-files emits full repo-relative paths)
-- tree_max_chars = 4096, -- truncate the injected block above this
-- },
--
-- :highlight has no config flag in v1 — toggled at runtime only.
-- Requires the external `tree-sitter` CLI plus configured parser-
-- directories with cloned + built `tree-sitter-<lang>` grammars
-- (see `:highlight on` for the install hints).
-- Phase 7 (docs/PHASE7.md): cost / usage observability. broker.lua
-- captures `usage` (+ `cost` for cloud) from every chat/chat_stream
-- call and routes via ctx:add_usage to a per-session accumulator.
-- `:cost` / `:cost detail` / `:cost reset` surface the totals.
-- The `cost` block below configures OPTIONAL warn thresholds —
-- a single status line fires the first time the cumulative
-- crosses each threshold. Default off. Useful when paid cloud
-- presets are in play so runaway-cost sessions get a nudge.
--
-- cost = {
-- warn_at_dollars = 0.50, -- one-shot warn when cumulative cost crosses
-- warn_at_tokens = 100000, -- one-shot warn when cumulative tokens crosses
-- },
--
-- Both flags are independent (R4 — first-to-fire doesn't suppress
-- the other); `:cost reset` re-arms both. Per-turn usage is also
-- written to session/*.jsonl (assistant-turn `usage` field) for
-- after-the-fact scripting; cross-session aggregation deferred
-- to a future phase (Q-C2).
-- Phase 8 (docs/PHASE8.md): accurate tokenization via the broker's
-- /tokenize endpoint, replacing the Phase 0 §8 char/4 heuristic.
-- Two consequences when use_endpoint=true:
-- (1) Context:estimate_tokens hits <endpoint>/tokenize once per
-- new turn (cached on the turn dict thereafter). Network
-- cost is one round-trip (~30ms) per fresh turn; subsequent
-- calls reuse the cache.
-- (2) Context:enforce_budget actually ENFORCES token_budget now
-- (previously only max_turns was checked). Sessions that
-- fit under char/4 may evict earlier — raise token_budget
-- to match your model's real context window if needed.
-- Cloud endpoints (OpenRouter) don't expose /tokenize; capability
-- cached as unsupported on first probe -> silent char/4 fallback.
--
-- tokenize = {
-- use_endpoint = true,
-- },
}
Reference in New Issue View Git Blame Copy Permalink