revert family→model/variant rename; add (family, version, variant) tags

config + :model + -m use `model` (the API string, possibly shortened
to portion after last slash). Internally Profile carries `family`,
`version`, `variant` from a hardcoded known-good lookup. KnownGoodCombos
is now a 5-tuple. No :variant command. commonModelPrefix handles
single-model providers.

bump 0.2.3

Author: capocasa

config.example

@@ -12,35 +12,35 @@ current = "baseten.glm5"
 name = "baseten"
 url = "https://inference.baseten.co/v1"
 key = "..."
-models = "glm5, qwen3c"
+models = "glm5 qwen3c"
 
 [provider]
 name = "openai"
 url = "https://api.openai.com/v1"
 key = "sk-..."
-models = "gpt-4o-mini, gpt-4o"
+models = "gpt-4o-mini gpt-4o"
 
 [provider]
 name = "groq"
 url = "https://api.groq.com/openai/v1"
 key = "gsk_..."
-models = "llama-3.3-70b-versatile, llama-3.1-8b-instant"
+models = "llama-3.3-70b-versatile llama-3.1-8b-instant"
 
 [provider]
 name = "together"
 url = "https://api.together.xyz/v1"
 key = "..."
-models = "meta-llama/Llama-3.3-70B-Instruct-Turbo"
+model_prefix = "meta-llama/"
+models = "Llama-3.3-70B-Instruct-Turbo"
 
-# Optional per-provider `mode` selects the agent protocol:
-#   tools (default) — OpenAI tool_calls; harness sends a tool schema
-#   text            — model emits fenced ```bash / path-then-fence / SEARCH/
-#                     REPLACE blocks; harness parses them. No schema sent.
-# The `:mode [text|tools]` REPL command toggles per-session for experiments.
+# `family` is an experimental-only override (only honored under
+# `--experimental`) that picks the system-prompt branch by family
+# name, when the (provider, model) combo isn't on the known-good
+# list. Known-good combos ignore it.
 ;
 ; [provider]
 ; name = "deepseek"
 ; url = "https://api.deepseek.com/v1"
 ; key = "..."
-; mode = "text"
+; family = "deepseek"
 ; models = "deepseek-chat"

src/threecode.nim

@@ -23,7 +23,7 @@ proc runTurns*(p: Profile, messages: var JsonNode, session: var Session) =
       stdout.styledWriteLine styleDim, "  · interrupted", resetStyle
       interrupted = false
       return
-    let window = contextWindowFor(p.variant)
+    let window = contextWindowFor(p.model)
     var summarized = 0
     case decideContextAction(usage.promptTokens, window, messages.len)
     of caSummarize:
@@ -86,7 +86,7 @@ proc runTurns*(p: Profile, messages: var JsonNode, session: var Session) =
             stderr.writeLine "3code: tool_call " & name &
               " has malformed arguments JSON (" & e.msg & "): " & argsStr
             newJObject()
-        let act = toolCallToAction(p.model, name, args)
+        let act = toolCallToAction(p.family, name, args)
         let idx = session.toolLog.len + 1
         let silent = isSkillRead(act)
         # Pre-exec: dim bullet + dim banner text — "in flight" signal. After
@@ -332,7 +332,7 @@ proc main() =
       &"({messages.len} msg" & (if messages.len == 1: "" else: "s") & ")",
       resetStyle
     stdout.write "\n"
-    replaySessionTail(messages, session.toolLog, prof.model)
+    replaySessionTail(messages, session.toolLog, prof.family)
     stdout.write "\n"
     if prompt != "":
       messages.add %*{"role": "user", "content": buildUserMessage(messages, prompt)}

src/threecode/actions.nim

@@ -14,9 +14,9 @@ import types, util, shell
 # runAction returns a clean error rather than crashing.
 # ---------------------------------------------------------------------------
 
-proc unknownTool(model, name: string): Action =
+proc unknownTool(family, name: string): Action =
   Action(kind: akBash,
-         body: "echo 'tool not offered to model " & model & ": " & name &
+         body: "echo 'tool not offered to family " & family & ": " & name &
                "' >&2; exit 1")
 
 proc dispatchGlmOrQwen(name: string, args: JsonNode): Action =
@@ -58,14 +58,14 @@ proc dispatchGptOss(rawName: string, args: JsonNode): Action =
   else:
     unknownTool("gpt-oss", rawName)
 
-proc toolCallToAction*(model, name: string, args: JsonNode): Action =
-  ## Routes a tool_call to the dispatcher for the active model. The model
-  ## label comes from `Profile.model` ("glm" / "qwen" / "gpt-oss"); the
+proc toolCallToAction*(family, name: string, args: JsonNode): Action =
+  ## Routes a tool_call to the dispatcher for the active family. The family
+  ## label comes from `Profile.family` ("glm" / "qwen" / "gpt-oss"); the
   ## case statement below mirrors `setup` in `prompts.nim`.
-  case model
+  case family
   of "glm", "qwen", "deepseek": dispatchGlmOrQwen(name, args)
   of "gpt-oss": dispatchGptOss(name, args)
-  else: die "unknown model in tool dispatch: '" & model & "'"
+  else: die "unknown family in tool dispatch: '" & family & "'"
 
 proc nearestLineHint(content, search: string): string =
   ## When a patch search block didn't match, point the model at the most

src/threecode/api.nim

@@ -504,7 +504,7 @@ template hint(args: varargs[untyped]) =
 proc callModel*(p: Profile, messages: JsonNode, usage: var Usage, lastPromptTokens: int): JsonNode =
   ensureReasoningField(messages)
   var body = %*{
-    "model": p.variantPrefix & p.variant,
+    "model": p.modelPrefix & p.model,
     "messages": messages,
     "stream": true,
     "stream_options": {"include_usage": true}
@@ -515,7 +515,7 @@ proc callModel*(p: Profile, messages: JsonNode, usage: var Usage, lastPromptToke
   let t0 = epochTime()
   decayLevel(serverRetryLevel, serverLastTs, t0)
   decayLevel(rateRetryLevel, rateLastTs, t0)
-  let window = contextWindowFor(p.variant)
+  let window = contextWindowFor(p.model)
   let baseLabel = contextLabel(lastPromptTokens, window)
   # Blank line above the upcoming spinner / bullet so the assistant
   # output isn't flush against the prior content (user prompt or last
@@ -619,7 +619,7 @@ proc verifyProfile*(p: Profile): (bool, string) =
     "Content-Type": "application/json"
   })
   let body = $(%*{
-    "model": p.variantPrefix & p.variant,
+    "model": p.modelPrefix & p.model,
     "messages": [%*{"role": "user", "content": "ping"}],
     "max_tokens": 1,
     "stream": false
@@ -636,9 +636,9 @@ proc verifyProfile*(p: Profile): (bool, string) =
   except CatchableError as e:
     return (false, e.msg)
 
-proc fetchVariants*(url, key: string): seq[string] =
+proc fetchModels*(url, key: string): seq[string] =
   ## GET /models on the provider — that endpoint name is OpenAI's; what it
-  ## returns is the list of variant ids this provider exposes.
+  ## returns is the list of model ids this provider exposes.
   let client = newHttpClient(timeout = 20000)
   defer: client.close()
   client.headers = newHttpHeaders({"Authorization": "Bearer " & key})

src/threecode/compact.nim

@@ -7,11 +7,11 @@ const
   CompactedMarker* = "[compacted — tool output elided; use :show to view]"
   SupersededMarker* = "[superseded — later action on same path elided this]"
 
-proc contextWindowFor*(variant: string): int =
-  ## Heuristic: substring match on well-known variant slugs. Cheap, and no
+proc contextWindowFor*(model: string): int =
+  ## Heuristic: substring match on well-known model slugs. Cheap, and no
   ## known collisions exist in practice. Update if a provider ships a
-  ## colliding variant name.
-  let m = variant.toLowerAscii
+  ## colliding model name.
+  let m = model.toLowerAscii
   if "kimi-k2" in m: 128_000
   elif "qwen3-coder" in m or "qwen3_coder" in m: 262_144
   elif "qwen" in m: 128_000
@@ -184,7 +184,7 @@ proc applySummary*(messages: JsonNode, summary: string,
 proc callSummarizer(p: Profile, messages: JsonNode): string =
   ## Fires a single meta-call to the model with a dedicated summarizer
   ## system prompt and no tools. Returns "" on any failure.
-  if p.name == "" or p.url == "" or p.key == "" or p.variant == "": return ""
+  if p.name == "" or p.url == "" or p.key == "" or p.model == "": return ""
   # Build a trimmed payload: the summarizer prompt + every non-system
   # message from the live conversation. Tool_call messages are allowed —
   # most OpenAI-compatible providers accept them in chat completions even
@@ -205,7 +205,7 @@ proc callSummarizer(p: Profile, messages: JsonNode): string =
     "Content-Type": "application/json"
   })
   let body = %*{
-    "model": p.variantPrefix & p.variant,
+    "model": p.modelPrefix & p.model,
     "messages": payload,
     "max_tokens": SummarizeMaxTokens,
     "stream": false