Files
aish/broker.lua
T
marfrit 2abd5da3a6 safety: LLM second-opinion + session cache (Phase 3 commit #2)
Phase 3 commit #2 per docs/PHASE3.md §12. Adds the LLM-probe gate on
top of commit #1's static patterns. Together they form is_destructive.

broker.lua extension:
  - opts.max_tokens (A2) — passed through to the request body. Phase 3
    probes cap at 4 tokens for YES/NO replies.
  - opts.timeout_ms — overrides model_cfg.timeout_ms per-call. Probe
    uses 15000ms cap regardless of the model's normal timeout
    (the user's deep model has 1800000ms for long generations; the
    probe must stay snappy).
  - M.chat now accepts an opts table (same shape as chat_stream's).
    Backwards compatible — existing callers passing (cfg, msgs)
    unaffected.

safety.lua additions:
  - llm_probe(cfg, system, cmd): single broker.chat call returning
    "YES"/"NO"/"YES_FAILSAFE"/"YES_UNPARSEABLE" — fail-safe defaults.
  - llm_second_opinion(cmd, cfg): two-probe protocol per R-B2.
    Probe 1: "Is this destructive?" — YES → flag.
    Probe 2 (only if probe 1 said NO): "Is this safe?" inverted
    question — NO → flag (disagreement = HALT).
    Both NO → safe.
  - Session-scoped cache _llm_cache keyed by normalized command
    (lowercased + whitespace-collapsed). Mitigates Q23 latency for
    repeated commands within a Norris run.
  - Model-selection precedence: cfg.safety.llm_model (explicit)
    → cfg.models.deep (independent local class) → cfg.models[default].
    Fail-safe YES if none configured.
  - is_destructive(cmd, cfg): runs static patterns first (always),
    then LLM if cfg present + not explicitly opted-out. cfg=nil
    yields static-only mode (handy for tests).

End-to-end verified against hossenfelder using qwen-coder-7b-32k as
the deep probe (qwen3-30b-a3b-instruct in repo's config.lua isn't
currently loaded on the local backend):
  cat /etc/hostname              → hit=false (LLM: NO, NO inverted = safe)
  rm /tmp/x.log                  → hit=true  (LLM flagged; static missed
                                              because no -r/-f flags)
  cp /etc/passwd /tmp/passwd.bak → hit=false (safe copy)
  cache: second probe on same cmd → 0s wall time
  static-only (cfg=nil): rm -rf /tmp/x → static hit, no LLM call
  opt-out (llm_second_opinion=false): cp x y → hit=false, no probe

Test corpus (test_safety.lua, 87 cases) still all pass — cfg=nil
preserves the static-only behavior.

Note: production config.lua currently has `deep = qwen3-30b-a3b-instruct`
which isn't loaded on the proxy backend right now; Norris users will
hit the fail-safe (everything flagged destructive) until either the
deep model is brought up OR cfg.safety.llm_model = "cloud" is set
to route the probe through anthropic/claude-haiku-4.5. Update the
config or model deployment for production use — covered by Phase 3
verify test case.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-12 23:36:06 +00:00

174 lines
7.4 KiB
Lua

-- broker.lua — llama.cpp HTTP client.
-- Phase 0: blocking POST via ffi/curl + vendored dkjson.
-- Phase 1: streaming POST via ffi/curl.post_sse with an OpenAI-shape decoder
-- on top. M.chat becomes a thin buffering wrapper around M.chat_stream so the
-- one streaming path covers both incremental and sync callers.
-- Phase 2: optional opts.tools array passed through to the request body
-- (omitted entirely when nil/empty per §12 risk row 1). The chat_stream
-- on_delta callback widens to (kind, payload) where kind is "text" or
-- "tool_call"; tool_call deltas are accumulated by `index` (default 0 if
-- absent per C2) and emitted as complete records on finish_reason "tool_calls".
-- broker.lua does NOT depend on mcp.lua — the caller assembles opts.tools
-- and passes it in. See docs/PHASE0.md §6, PHASE1.md §3, PHASE2.md §3 / §5.
local curl = require("ffi.curl")
local json = require("dkjson")
local M = {}
local function build_headers(model_cfg)
local h = { "Content-Type: application/json" }
if model_cfg.key_env then
local key = os.getenv(model_cfg.key_env)
if key and key ~= "" then
h[#h + 1] = "Authorization: Bearer " .. key
end
end
return h
end
local function build_request(model_cfg, messages, stream, tools, max_tokens)
if not (model_cfg and model_cfg.endpoint and model_cfg.model) then
return nil, "broker: model_cfg.endpoint and .model are required"
end
local url = model_cfg.endpoint:gsub("/+$", "") .. "/v1/chat/completions"
local req = {
model = model_cfg.model,
messages = messages,
stream = stream and true or false,
temperature = model_cfg.temperature or 0.2,
}
-- Per PHASE2.md §12 risk row "Empty tools array": some servers reject
-- "tools": []. Only set the field when the list has entries.
if tools and #tools > 0 then req.tools = tools end
-- Phase 3 (A2): max_tokens passthrough — used by safety.is_destructive
-- to cap YES/NO probes at ~4 tokens. Omitted when nil (Phase 1/2
-- callers unaffected — model defaults still apply).
if max_tokens then req.max_tokens = max_tokens end
return url, json.encode(req), build_headers(model_cfg),
(model_cfg.timeout_ms or 60000)
end
-- Streaming /v1/chat/completions.
-- Signature widens vs Phase 1: opts is optional and may carry .tools.
-- on_delta is called as on_delta(kind, payload):
-- on_delta("text", content_string) - per text chunk
-- on_delta("tool_call", { id, name, arguments }) - once per completed
-- tool call (on finish_reason "tool_calls").
-- arguments is the raw JSON-string accumulated across
-- deltas; caller does the dkjson.decode.
-- Returns:
-- true stream ended cleanly
-- nil, errmsg transport / API failure
function M.chat_stream(model_cfg, messages, on_delta, opts)
opts = opts or {}
local url, body, headers, timeout_ms =
build_request(model_cfg, messages, true, opts.tools, opts.max_tokens)
if not url then return nil, body end -- url slot carries err on bad cfg
-- Phase 3: opts.timeout_ms overrides the model's default. Used by
-- safety.is_destructive's LLM probe to cap YES/NO checks at ~15s even
-- when the model's normal timeout is much higher (e.g. user's deep
-- model has 1800000ms for long generations).
if opts.timeout_ms then timeout_ms = opts.timeout_ms end
local done = false
local api_err
-- Tool-call accumulator keyed by index. Each slot is filled across
-- many deltas: id+name come on the opener, arguments arrives as
-- character-fragment JSON-string chunks (PHASE2-baseline.md §4).
local tc_by_index = {}
local tc_index_order = {} -- preserve emission order
local index_absent_warned = false
local function on_event(data)
if done then return end
if data == "[DONE]" then done = true; return end
local doc = json.decode(data)
if not doc then return end -- ignore unparseable events
-- Some servers emit an SSE-framed error envelope at the start of the
-- stream — surface it.
if doc.error then
local m = (type(doc.error) == "table" and doc.error.message)
or tostring(doc.error)
api_err = m
done = true
return
end
local choice = doc.choices and doc.choices[1]
local delta = choice and choice.delta
-- Text path (unchanged from Phase 1 semantics; kind widened).
local content = delta and delta.content
if type(content) == "string" and #content > 0 then
on_delta("text", content)
end
-- Tool-call accumulation (Phase 2).
local tcs = delta and delta.tool_calls
if type(tcs) == "table" then
for _, tc in ipairs(tcs) do
local idx = tc.index
if idx == nil then
idx = 0
if not index_absent_warned then
index_absent_warned = true
-- One-shot debug status per stream; printed to
-- stderr so it doesn't interleave with renderer
-- stdout output.
io.stderr:write(
"[aish] broker: tool_calls[].index absent; assuming 0\n")
end
end
local slot = tc_by_index[idx]
if not slot then
slot = { id = nil, name = nil, arguments = "" }
tc_by_index[idx] = slot
tc_index_order[#tc_index_order + 1] = idx
end
if tc.id then slot.id = tc.id end
if tc["function"] then
local fn = tc["function"]
if fn.name then slot.name = fn.name end
if fn.arguments then
slot.arguments = slot.arguments .. fn.arguments
end
end
end
end
-- On finish_reason "tool_calls", emit all accumulated calls.
if choice and choice.finish_reason == "tool_calls" then
for _, idx in ipairs(tc_index_order) do
on_delta("tool_call", tc_by_index[idx])
end
tc_by_index = {}
tc_index_order = {}
end
end
local ok, err = curl.post_sse(url, body, headers, on_event, timeout_ms)
if api_err then return nil, "api: " .. api_err end
if not ok then return nil, "transport: " .. tostring(err) end
return true
end
-- Send a /v1/chat/completions request and return the full assistant text.
-- Thin buffering wrapper over M.chat_stream — same path as the streaming
-- consumer, so the broker keeps one HTTP shape (stream:true always).
-- M.chat's external contract is unchanged in Phase 2 (C1 in review):
-- it returns the text string and silently ignores any tool_call kinds
-- the model might emit (no caller of M.chat passes opts.tools).
-- Returns:
-- assistant_content_string on success
-- nil, errmsg on transport / decode / API failure
function M.chat(model_cfg, messages, opts)
local parts = {}
local ok, err = M.chat_stream(model_cfg, messages, function(kind, payload)
if kind == "text" then parts[#parts + 1] = payload end
end, opts)
if not ok then return nil, err end
return table.concat(parts)
end
return M