Merge remote-tracking branch 'origin/master' into v2.3

docs/plans/v2.3/04-ctx-error-docs.md

@@ -18,8 +18,9 @@ shape, so the client handles crossbar's refusal like the server's:
 
 ## Files
 
-- Copy: the **replacement** `internal/proxy/ctxguard_test.go` (was v2's; only the 400-body
-  assertions changed; the v2.3 copy is now protected)
+- Copy: the **replacements** `internal/proxy/ctxguard_test.go` (was v2's) and
+  `internal/proxy/ctxguard_router_test.go` (was v2.1's); only their 400-body assertions changed;
+  the v2.3 copies are now protected
 - Modify: `internal/proxy/ctxguard.go` (`refuseCtx`), `README.md`, `docs/implementer-log.md`
 
 ## Rules

docs/plans/v2.3/README.md

@@ -27,7 +27,7 @@ context refusal that is not llama-server's `exceed_context_size_error`.
   `tools/smoke.sh` (v2: adds check 6, the dedicated listener).
 - **04-ctx-error-docs** — the context refusal in llama-server's shape
   `{"error":{"code":400,"type":"exceed_context_size_error","message":"prompt too large","n_prompt_tokens":N,"n_ctx":M}}`;
-  README. Given: replaces `proxy/ctxguard_test.go` (v2).
+  README. Given: replaces `proxy/ctxguard_test.go` (v2) and `proxy/ctxguard_router_test.go` (v2.1).
 
 **Order matters:** 02's affinity test uses `/slots` (01); 03's listener test uses route affinity
 (02). Each task is green on its own given tests plus all earlier ones.
@@ -49,8 +49,8 @@ the client's chat goes anyway, so this is the load the client asked for.
 
 - Everything in `AGENTS.md`. Branch `v2.3` from `master`. One task, one fresh OpenCode session,
   one commit. Given files are copied and never edited; earlier plans' given files stay protected,
-  except the four this plan replaces (`proxy/proxy_test.go`, `proxy/ctxguard_test.go`,
-  `example.toml`, `tools/smoke.sh`), whose v2.3 copies are then the protected ones.
+  except the five this plan replaces (`proxy/proxy_test.go`, `proxy/ctxguard_test.go`,
+  `proxy/ctxguard_router_test.go`, `example.toml`, `tools/smoke.sh`), whose v2.3 copies are then the protected ones.
 
 ## Changes during the run
 
@@ -65,3 +65,10 @@ the client's chat goes anyway, so this is the load the client asked for.
   chunk, asserts the slot is still held; fails on the hack, passes without it). Owner removed the
   `onFlush` hook and the `release` parameter from `forward`. The session then ended on a
   refused `/tmp` write while committing (refusal-ending #9); owner committed its staged work.
+- Task 03: first session emitted a stray `</tool_call>` after reading files and ended with no
+  change (model); restarted unchanged, done in 16 min (`fa1c398`).
+- Task 04: **owner fault, correct stop.** The v2.1 given `ctxguard_router_test.go` also asserts
+  the refusal body (`e["max"]`); I grepped only for the `"prompt too large"` string when
+  writing the replacement list. Ornith implemented the new shape, saw the two protected tests
+  demand incompatible bodies, committed only its `stopped` row, and reported — exactly the
+  AGENTS.md rule. Replacement `ctxguard_router_test.go` (reads `error.n_ctx`) added.

docs/plans/v2.3/_files/internal/proxy/ctxguard_router_test.go

@@ -0,0 +1,110 @@
+package proxy_test
+
+import (
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"git.wntrmute.dev/kyle/crossbar/internal/proxy"
+)
+
+// routerUpstream is a fake shaped like llama-server's router mode: the plain /props carries no
+// context (role router, n_ctx 0), /v1/models lists models with a status, and /props?model=X
+// answers for one loaded model. The guard must work from the per-model figures.
+func routerUpstream(t *testing.T, name string, models map[string][2]int, unloaded ...string) *upstream {
+	u := &upstream{name: name}
+	mux := http.NewServeMux()
+	mux.HandleFunc("/health", func(w http.ResponseWriter, r *http.Request) { fmt.Fprint(w, `{"status":"ok"}`) })
+	mux.HandleFunc("/v1/models", func(w http.ResponseWriter, r *http.Request) {
+		fmt.Fprint(w, `{"object":"list","data":[`)
+		first := true
+		for id := range models {
+			if !first {
+				fmt.Fprint(w, ",")
+			}
+			first = false
+			fmt.Fprintf(w, `{"id":%q,"status":{"value":"loaded"}}`, id)
+		}
+		for _, id := range unloaded {
+			fmt.Fprintf(w, `,{"id":%q,"status":{"value":"unloaded"}}`, id)
+		}
+		fmt.Fprint(w, `]}`)
+	})
+	mux.HandleFunc("/props", func(w http.ResponseWriter, r *http.Request) {
+		model := r.URL.Query().Get("model")
+		if model == "" {
+			fmt.Fprint(w, `{"role":"router","default_generation_settings":{"n_ctx":0}}`)
+			return
+		}
+		m, ok := models[model]
+		if !ok {
+			w.WriteHeader(http.StatusInternalServerError)
+			fmt.Fprint(w, `{"error":"asked for a model that is not loaded"}`)
+			return
+		}
+		fmt.Fprintf(w, `{"default_generation_settings":{"n_ctx":%d},"total_slots":%d}`, m[0], m[1])
+	})
+	mux.HandleFunc("/", func(w http.ResponseWriter, r *http.Request) {
+		u.hits.Add(1)
+		w.Header().Set("Content-Type", "application/json")
+		fmt.Fprint(w, `{"choices":[{"message":{"role":"assistant","content":"ok"}}],"usage":{"prompt_tokens":1,"completion_tokens":1}}`)
+	})
+	u.srv = httptest.NewServer(mux)
+	t.Cleanup(u.srv.Close)
+	return u
+}
+
+// On routers the guard reads the per-model per-slot context: `small` serves "shared" from a
+// 8192-context child with two slots (4096 per slot), `big` from a 131072-context child with one.
+// The plain /props of both says nothing, so a v2 guard that only knew host-level figures would
+// stay inert and let the oversized prompt overflow `small`.
+func TestRouterGuardUsesPerModelContext(t *testing.T) {
+	small := routerUpstream(t, "small", map[string][2]int{"shared": {8192, 2}})
+	big := routerUpstream(t, "big", map[string][2]int{"shared": {131072, 1}})
+	r := newRig(t, ctxHosts, small, big)
+
+	resp := r.post("/r/v1/chat/completions", bodyOfTokens(100))
+	drain(resp)
+	if got := resp.Header.Get(proxy.HostHeader); got != "small" {
+		t.Fatalf("small prompt went to %q, want small (weight 10)", got)
+	}
+	resp = r.post("/r/v1/chat/completions", bodyOfTokens(6000))
+	drain(resp)
+	if resp.StatusCode != 200 || resp.Header.Get(proxy.HostHeader) != "big" {
+		t.Fatalf("6000-token prompt: status %d host %q, want 200 on big", resp.StatusCode, resp.Header.Get(proxy.HostHeader))
+	}
+	if got := resp.Header.Get("X-Crossbar-Ctx"); got != "moved:small>big" {
+		t.Errorf("X-Crossbar-Ctx = %q, want moved:small>big", got)
+	}
+}
+
+// A model the router lists as unloaded is not resident there: the guard must not move a prompt
+// to that host, and the refusal names the largest per-slot context among hosts that do serve it.
+func TestRouterUnloadedModelIsNotACandidate(t *testing.T) {
+	small := routerUpstream(t, "small", map[string][2]int{"shared": {8192, 2}})
+	big := routerUpstream(t, "big", map[string][2]int{"other": {131072, 1}}, "shared") // shared unloaded on big
+	r := newRig(t, ctxHosts, small, big)
+
+	resp := r.post("/r/v1/chat/completions", bodyOfTokens(6000))
+	body := drain(resp)
+	if resp.StatusCode != 400 {
+		t.Fatalf("status %d body %s, want 400: shared is loaded only on small, where it does not fit", resp.StatusCode, body)
+	}
+	if big.hits.Load() != 0 {
+		t.Errorf("big served %d requests for a model it does not have loaded", big.hits.Load())
+	}
+	// v2.3: the refusal is llama-server's exceed_context_size_error shape; n_ctx is what "max" was.
+	var e struct {
+		Error struct {
+			NCtx float64 `json:"n_ctx"`
+		} `json:"error"`
+	}
+	if err := json.Unmarshal([]byte(body), &e); err != nil {
+		t.Fatalf("body %q is not JSON: %v", body, err)
+	}
+	if e.Error.NCtx != 4096 {
+		t.Errorf("error.n_ctx = %v, want 4096: the largest per-slot context among hosts that have shared loaded", e.Error.NCtx)
+	}
+}