router: classify_model heuristic + 31-case corpus (Phase 5 commit #1)

Phase 5 commit #1 per docs/PHASE5.md §11. Pure-Lua per-request model
routing — no IO, no LLM probe in v1.

router.classify_model(text, cfg) -> (model_name | nil, class_label):
  1. classify_class(text) walks heuristics in priority order:
       code class:
         - triple-backtick fence anywhere
         - "traceback" / "stacktrace" / "stack trace" (ci)
         - "error:" / "exception:" in first 60 chars (ci)
         - path-with-code-extension token (.py/.lua/.c/.js/.go/.rs/.cpp/.h/.ts)
         - 5+ lines with indented content (looks like a paste)
       reasoning class (requires text >= 15 chars to skip bare keywords):
         - "explain" / "why " / "how does" / "compare" (ci)
         - "?" + length > 100 chars
       default class: everything else
  2. Map class via cfg.routing.classes[class] → model name (or nil = keep current).
  3. Return (model_name_or_nil, class_label).

ALWAYS evaluates regardless of cfg.routing.auto — caller (repl.ask_ai
in commit #3) gates on the flag. This separation lets `:route check`
introspect the heuristic even when routing is off (N1).

M._classify_class exposed for testing.

Test corpus (test_router_model.lua, 31 cases):
  - 13 code-class positives (fence, traceback, paths, multi-line paste)
  - 6 reasoning-class positives (explain/why/how does/compare/?+length)
  - 8 default-class (short queries, bare keywords below 15-char threshold,
    non-code paths like .md/.txt)
  - 3 model-mapping cases (code→"deep", reasoning→"cloud", default→nil)
  - 1 R-N2 default test: classes.reasoning=nil → reasoning text yields
    nil model override (heuristic still fires, no swap)
  - All 31 pass; 15-char threshold catches "how does ASLR work?" without
    false-positive on bare "explain".

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-05-13 11:17:22 +00:00
parent 2e389c1475
commit 3e57824684
2 changed files with 189 additions and 8 deletions
+81 -8
View File
@@ -1,12 +1,9 @@
-- router.lua — task classifier: meta / shell / AI.
-- See docs/PHASE0.md §5.
-- router.lua — task classifier: meta / shell / AI / model-routing.
-- See docs/PHASE0.md §5 and docs/PHASE5.md §4 for Phase 5 additions.
--
-- Pure function. Takes (line, config). Returns (kind, payload).
-- kind : "meta" | "shell" | "ai"
-- payload: the (possibly stripped) line that the dispatcher should act on.
--
-- Empty / whitespace-only lines are returned as ("ai", ""); the repl loop
-- skips them before dispatching.
-- M.classify(line, config) (kind, payload) for input dispatch (Phase 0).
-- M.classify_model(text, cfg) → name | nil for per-request model routing
-- (Phase 5; pure-Lua heuristics, no IO).
local M = {}
@@ -63,4 +60,80 @@ function M.classify(line, config)
return "ai", line
end
-- ---------------------------------------------------------------- classify_model
-- Phase 5: per-request model routing heuristic. Pure-Lua, no IO.
-- Returns the NAME of a model preset (string) to switch to for this
-- request, or nil to keep the active model unchanged.
--
-- The mapping from class to model name lives in `cfg.routing.classes`.
-- A class with value `nil` means "keep current" — even though the
-- heuristic fires, no override happens (used by default for the
-- `reasoning` class per R-N2 cost-safety policy).
--
-- This function ALWAYS evaluates the heuristic regardless of
-- `cfg.routing.auto` — the caller (repl.ask_ai) gates on the flag.
-- This separation lets `:route check <text>` introspect the heuristic
-- even when routing is disabled (N1).
local function classify_class(text)
if not text or text == "" then return "default" end
-- ── Code class — looks like a paste or contains code markers
if text:find("```", 1, true) then return "code" end
local lower = text:lower()
if lower:find("traceback", 1, true)
or lower:find("stacktrace", 1, true)
or lower:find("stack trace", 1, true) then
return "code"
end
-- exception/error markers near beginning (first 60 chars)
if lower:sub(1, 60):find("error:", 1, true)
or lower:sub(1, 60):find("exception:", 1, true) then
return "code"
end
-- path with code-extension token
if text:match("[%./~][%w%-_/.]+%.([%w]+)") then
local ext = text:match("[%./~][%w%-_/.]+%.([%w]+)")
if ext == "py" or ext == "lua" or ext == "c"
or ext == "js" or ext == "go" or ext == "rs"
or ext == "cpp" or ext == "h" or ext == "ts" then
return "code"
end
end
-- multi-line + indented (looks like a code paste)
local nlines = 0
for _ in (text .. "\n"):gmatch("[^\n]*\n") do nlines = nlines + 1 end
if nlines > 4 and text:find("\n%s+%S") then return "code" end
-- ── Reasoning class
-- Min length 15 — catches "how does X work" but excludes bare "why" / "explain"
if #text >= 15 then
if lower:find("explain", 1, true)
or lower:find("why ", 1, true) -- trailing space (not "whyever")
or lower:find("how does", 1, true)
or lower:find("compare", 1, true) then
return "reasoning"
end
end
if text:find("?", 1, true) and #text > 100 then
return "reasoning"
end
return "default"
end
-- Public API.
function M.classify_model(text, cfg)
local class = classify_class(text)
local classes = (cfg and cfg.routing and cfg.routing.classes) or {}
local target = classes[class]
-- nil target = keep current (this is the R-N2 default for "reasoning")
if target == nil then return nil, class end
-- Caller may want the class label for the status line; return both.
return target, class
end
-- Exposed for `:route check` introspection (N1).
M._classify_class = classify_class
return M