feat(agent): Kimi model profiles + compaction side-call timeout scaling

docs/PROVIDERS.md

@@ -38,6 +38,7 @@ odek ships with built-in **model profiles** that automatically apply sensible de
 | `deepseek-chat` | DeepSeek (legacy) | (provider default) | 120s | 128K | General |
 | `deepseek-v4-flash` | DeepSeek v4 Flash | — (faster/cheaper) | 90s | 128K | Quick tasks, coding |
 | `deepseek-v4-pro` | DeepSeek v4 Pro | `enabled` | 180s | **1M** | Deep reasoning |
+| `kimi-…`, `k3…` (e.g. `kimi-for-coding`, `k3-256k`) | Kimi | (provider default) | 300s | 256K | Agentic coding |
 | *(any other)* | Generic | (profile default) | 120s | (no limit) | Custom models |
 
 ### How profiles work

internal/loop/loop.go

@@ -269,6 +269,11 @@ type Engine struct {
 	// groups (Config.Compaction). compactDigest holds the current summary.
 	compaction    bool
 	compactDigest string
+
+	// sideCallTimeout bounds the compaction and progress-summary side calls.
+	// Zero means use the default (30s). Callers scale it off the resolved
+	// client timeout so slow providers don't silently lose the digest.
+	sideCallTimeout time.Duration
 }
 
 // New creates a new loop Engine.
@@ -371,6 +376,22 @@ func (e *Engine) SetDangerousConfig(cfg *danger.DangerousConfig) { e.dangerousCf
 // so it is wrapped with the engine's untrusted-content wrapper when set.
 func (e *Engine) SetCompaction(enabled bool) { e.compaction = enabled }
 
+// SetSideCallTimeout sets the bound for the compaction digest and
+// progress-summary side calls. 0 or negative restores the default (30s).
+func (e *Engine) SetSideCallTimeout(d time.Duration) { e.sideCallTimeout = d }
+
+// SideCallTimeout returns the effective bound for the compaction digest and
+// progress-summary side calls (default 30s).
+func (e *Engine) SideCallTimeout() time.Duration { return e.sideTimeout() }
+
+// sideTimeout returns the effective side-call timeout.
+func (e *Engine) sideTimeout() time.Duration {
+	if e.sideCallTimeout > 0 {
+		return e.sideCallTimeout
+	}
+	return 30 * time.Second
+}
+
 // ── Token Estimation ─────────────────────────────────────────────────
 //
 // Zero-dependency heuristic: 1 token ≈ 4 chars for English text.
@@ -931,9 +952,9 @@ func (e *Engine) refreshDigest(ctx context.Context, messages []llm.Message, drop
 }
 
 // summarizeDropped builds the summarizer input from the dropped messages and
-// the previous digest, then calls the LLM with a bounded timeout. Returns an
-// empty string on any failure — compaction is best-effort and must never
-// break the agent loop.
+// the previous digest, then calls the LLM with a bounded timeout (sideTimeout).
+// Returns an empty string on any failure — compaction is best-effort and must
+// never break the agent loop.
 func (e *Engine) summarizeDropped(ctx context.Context, dropped []llm.Message) string {
 	if e.client == nil {
 		return ""
@@ -968,7 +989,7 @@ func (e *Engine) summarizeDropped(ctx context.Context, dropped []llm.Message) st
 		return ""
 	}
 
-	callCtx, cancel := context.WithTimeout(ctx, 30*time.Second)
+	callCtx, cancel := context.WithTimeout(ctx, e.sideTimeout())
 	defer cancel()
 	res, err := e.client.Call(callCtx, []llm.Message{
 		{Role: "system", Content: compactionSystemPrompt},
@@ -981,8 +1002,9 @@ func (e *Engine) summarizeDropped(ctx context.Context, dropped []llm.Message) st
 }
 
 // summarizeProgress renders the tail of the conversation into a bounded
-// summarizer input and makes one final tool-less LLM call (30s bound, same
-// pattern as the compaction side-call) asking for a progress summary.
+// summarizer input and makes one final tool-less LLM call (bounded by
+// sideTimeout, same pattern as the compaction side-call) asking for a
+// progress summary.
 // Returns an empty string on any failure — including a non-compliant
 // response that still requests tool calls — so the caller can fall back to
 // the plain budget-exhaustion error.
@@ -1016,7 +1038,7 @@ func (e *Engine) summarizeProgress(ctx context.Context, messages []llm.Message)
 		return ""
 	}
 
-	callCtx, cancel := context.WithTimeout(ctx, 30*time.Second)
+	callCtx, cancel := context.WithTimeout(ctx, e.sideTimeout())
 	defer cancel()
 	res, err := e.client.Call(callCtx, []llm.Message{
 		{Role: "system", Content: budgetSummarySystemPrompt},

internal/loop/loop_trim_test.go

@@ -8,6 +8,7 @@ import (
 	"net/http/httptest"
 	"strings"
 	"testing"
+	"time"
 
 	"github.com/BackendStack21/odek/internal/llm"
 	"github.com/BackendStack21/odek/internal/tool"
@@ -616,6 +617,21 @@ func TestTrimToSurvival_IncludesPrecedingSystemMessages(t *testing.T) {
 
 // ── Coverage: summarizeDropped branches ────────────────────────────────
 
+func TestSideTimeout_DefaultAndOverride(t *testing.T) {
+	engine := &Engine{}
+	if got := engine.sideTimeout(); got != 30*time.Second {
+		t.Errorf("default sideTimeout = %v, want 30s", got)
+	}
+	engine.SetSideCallTimeout(120 * time.Second)
+	if got := engine.sideTimeout(); got != 120*time.Second {
+		t.Errorf("sideTimeout = %v, want 120s", got)
+	}
+	engine.SetSideCallTimeout(0)
+	if got := engine.sideTimeout(); got != 30*time.Second {
+		t.Errorf("zero SetSideCallTimeout must restore 30s default, got %v", got)
+	}
+}
+
 func TestSummarizeDropped_NilClient(t *testing.T) {
 	engine := &Engine{}
 	if got := engine.summarizeDropped(context.Background(), []llm.Message{{Role: "assistant", Content: "x"}}); got != "" {

odek.go

@@ -284,6 +284,24 @@ var KnownProfiles = []struct {
 	Prefix  string
 	Profile ModelProfile
 }{
+	{
+		Prefix: "kimi-",
+		Profile: ModelProfile{
+			Label:      "Kimi",
+			Timeout:    300,     // reasoning models can be slow to first byte
+			MaxContext: 262_144, // 256K safe default; /models discovery takes priority
+		},
+	},
+	{
+		// Kimi Code also ships models under the "k3" family name (k3, k3-256k),
+		// which the "kimi-" prefix does not match.
+		Prefix: "k3",
+		Profile: ModelProfile{
+			Label:      "Kimi",
+			Timeout:    300,
+			MaxContext: 262_144,
+		},
+	},
 	{
 		Prefix: "deepseek-v4-pro",
 		Profile: ModelProfile{
@@ -589,6 +607,14 @@ func New(cfg Config) (*Agent, error) {
 	engine := loop.New(client, registry, cfg.MaxIterations, cfg.SystemMessage, cfg.Renderer, maxContext)
 	engine.PromptCaching = cfg.PromptCaching
 	engine.SetCompaction(cfg.Compaction)
+	// Side calls (compaction digest, progress summary) use the same client and
+	// model, so scale their bound off the resolved request timeout — a slow
+	// provider would otherwise blow the 30s default and silently drop the digest.
+	sideTimeout := time.Duration(timeout) * time.Second
+	if sideTimeout > 120*time.Second {
+		sideTimeout = 120 * time.Second
+	}
+	engine.SetSideCallTimeout(sideTimeout)
 	engine.SetUntrustedWrapper(cfg.UntrustedWrapper)
 	if cfg.MaxToolParallel > 0 {
 		engine.SetMaxToolParallel(cfg.MaxToolParallel)

odek_test.go

@@ -528,6 +528,33 @@ func TestLookupProfile_NoMatch(t *testing.T) {
 	}
 }
 
+func TestLookupProfile_KimiMatch(t *testing.T) {
+	// Any kimi-* model (e.g. kimi-for-coding) matches the "kimi-" profile:
+	// a longer timeout for slow reasoning responses, 256K context fallback.
+	p := LookupProfile("kimi-for-coding")
+	if p == nil {
+		t.Fatal("LookupProfile(\"kimi-for-coding\") returned nil")
+	}
+	if p.Timeout != 300 {
+		t.Errorf("Timeout = %d, want 300", p.Timeout)
+	}
+	if p.MaxContext != 262_144 {
+		t.Errorf("MaxContext = %d, want 262144", p.MaxContext)
+	}
+
+	// The k3 family (k3, k3-256k) is the same Kimi Code line under a
+	// different prefix — it must get the same profile.
+	for _, model := range []string{"k3", "k3-256k"} {
+		p := LookupProfile(model)
+		if p == nil {
+			t.Fatalf("LookupProfile(%q) returned nil", model)
+		}
+		if p.Timeout != 300 || p.MaxContext != 262_144 {
+			t.Errorf("LookupProfile(%q) = %+v, want Timeout=300 MaxContext=262144", model, p)
+		}
+	}
+}
+
 func TestProfileLabel_Known(t *testing.T) {
 	if label := ProfileLabel("deepseek-v4-pro"); label != "DeepSeek v4 Pro" {
 		t.Errorf("ProfileLabel = %q, want %q", label, "DeepSeek v4 Pro")
@@ -600,6 +627,30 @@ func TestNew_ProfileTimeout_Pro(t *testing.T) {
 	}
 }
 
+func TestNew_SideCallTimeoutScaling(t *testing.T) {
+	// The compaction/progress-summary side-call bound scales off the resolved
+	// client timeout, capped at 120s.
+	cases := []struct {
+		model string
+		want  time.Duration
+	}{
+		{"kimi-for-coding", 120 * time.Second}, // 300s client timeout → capped at 120s
+		{"k3-256k", 120 * time.Second},         // same profile via the k3 prefix
+		{"deepseek-v4-pro", 120 * time.Second}, // 180s → capped at 120s
+		{"deepseek-v4-flash", 90 * time.Second},
+		{"gpt-4o", 120 * time.Second}, // unknown model → 120s default
+	}
+	for _, tc := range cases {
+		agent, err := New(Config{APIKey: "sk-test", Model: tc.model})
+		if err != nil {
+			t.Fatalf("New(%q): %v", tc.model, err)
+		}
+		if got := agent.engine.SideCallTimeout(); got != tc.want {
+			t.Errorf("New(%q) side-call timeout = %v, want %v", tc.model, got, tc.want)
+		}
+	}
+}
+
 func TestNew_DefaultModelNoProfile(t *testing.T) {
 	// deepseek-chat is not in KnownProfiles — no profile defaults applied
 	cfg := Config{