summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--internal/llm/anthropic.go4
-rw-r--r--internal/llm/ollama.go4
-rw-r--r--internal/llm/openai.go4
-rw-r--r--internal/llm/openrouter.go4
-rw-r--r--internal/llm/policy/policy.go68
-rw-r--r--internal/llm/policy/policy_test.go67
-rw-r--r--internal/llm/provider.go2
-rw-r--r--internal/llm/resilience.go18
-rw-r--r--internal/llm/yousearch.go6
9 files changed, 164 insertions, 13 deletions
diff --git a/internal/llm/anthropic.go b/internal/llm/anthropic.go
index 3f7a616..00af17a 100644
--- a/internal/llm/anthropic.go
+++ b/internal/llm/anthropic.go
@@ -11,6 +11,7 @@ import (
"strings"
"time"
+ "codeberg.org/snonux/hexai/internal/llm/policy"
"codeberg.org/snonux/hexai/internal/logging"
)
@@ -116,7 +117,8 @@ func newAnthropicWithTimeout(baseURL, model, apiKey string, defaultTemp *float64
model = "claude-3-5-sonnet-20240620"
}
if timeoutSec <= 0 {
- timeoutSec = 30
+ // Fall back to the shared default chat timeout from the policy package.
+ timeoutSec = policy.DefaultRequestTimeoutSeconds
}
return anthropicClient{
httpClient: &http.Client{Timeout: time.Duration(timeoutSec) * time.Second},
diff --git a/internal/llm/ollama.go b/internal/llm/ollama.go
index d8f911f..83823bc 100644
--- a/internal/llm/ollama.go
+++ b/internal/llm/ollama.go
@@ -13,6 +13,7 @@ import (
"strings"
"time"
+ "codeberg.org/snonux/hexai/internal/llm/policy"
"codeberg.org/snonux/hexai/internal/logging"
)
@@ -79,7 +80,8 @@ func newOllamaWithTimeout(baseURL, model, apiKey string, defaultTemp *float64, t
model = "gemma4:31b-cloud"
}
if timeoutSec <= 0 {
- timeoutSec = 30
+ // Fall back to the shared default chat timeout from the policy package.
+ timeoutSec = policy.DefaultRequestTimeoutSeconds
}
return ollamaClient{
httpClient: &http.Client{Timeout: time.Duration(timeoutSec) * time.Second},
diff --git a/internal/llm/openai.go b/internal/llm/openai.go
index d475de0..983723a 100644
--- a/internal/llm/openai.go
+++ b/internal/llm/openai.go
@@ -11,6 +11,7 @@ import (
"strings"
"time"
+ "codeberg.org/snonux/hexai/internal/llm/policy"
"codeberg.org/snonux/hexai/internal/logging"
)
@@ -125,7 +126,8 @@ func newOpenAIWithTimeout(baseURL, model, apiKey string, defaultTemp *float64, t
model = "gpt-4.1"
}
if timeoutSec <= 0 {
- timeoutSec = 30
+ // Fall back to the shared default chat timeout from the policy package.
+ timeoutSec = policy.DefaultRequestTimeoutSeconds
}
return openAIClient{
httpClient: &http.Client{Timeout: time.Duration(timeoutSec) * time.Second},
diff --git a/internal/llm/openrouter.go b/internal/llm/openrouter.go
index d728275..970ec29 100644
--- a/internal/llm/openrouter.go
+++ b/internal/llm/openrouter.go
@@ -9,6 +9,7 @@ import (
"strings"
"time"
+ "codeberg.org/snonux/hexai/internal/llm/policy"
"codeberg.org/snonux/hexai/internal/logging"
)
@@ -56,7 +57,8 @@ func newOpenRouterWithTimeout(baseURL, model, apiKey string, defaultTemp *float6
model = "openrouter/auto"
}
if timeoutSec <= 0 {
- timeoutSec = 30
+ // Fall back to the shared default chat timeout from the policy package.
+ timeoutSec = policy.DefaultRequestTimeoutSeconds
}
return openRouterClient{
httpClient: &http.Client{Timeout: time.Duration(timeoutSec) * time.Second},
diff --git a/internal/llm/policy/policy.go b/internal/llm/policy/policy.go
new file mode 100644
index 0000000..e133e27
--- /dev/null
+++ b/internal/llm/policy/policy.go
@@ -0,0 +1,68 @@
+// Package policy centralizes the timeout, retry, and circuit-breaker tuning
+// values used throughout the LLM HTTP call path. Previously these durations and
+// counts were scattered as bare literals across each provider constructor and
+// across resilience.go / circuitbreaker.go, which made the operational policy
+// hard to discover and easy to drift out of sync.
+//
+// Consolidating them here gives a single, documented source of truth: anyone
+// adjusting how aggressively Hexai times out, retries, or sheds load against an
+// upstream LLM API only has to look in one place, and every provider inherits
+// the same defaults automatically.
+package policy
+
+import "time"
+
+// Request timeouts.
+//
+// DefaultRequestTimeout is the per-request HTTP client timeout applied to chat
+// providers (Anthropic, OpenAI, OpenRouter, Ollama) when the user has not
+// configured an explicit RequestTimeout. Thirty seconds is generous enough for
+// normal completions yet short enough that a hung connection fails fast rather
+// than blocking interactive use indefinitely.
+//
+// ResearchRequestTimeout applies to the You.com Research provider, whose
+// multi-step research pipeline routinely runs much longer than a single chat
+// completion. It therefore gets a substantially larger default so legitimate
+// long-running research is not cut off prematurely.
+const (
+ DefaultRequestTimeout = 30 * time.Second
+ ResearchRequestTimeout = 120 * time.Second
+)
+
+// DefaultRequestTimeoutSeconds and ResearchRequestTimeoutSeconds expose the
+// timeouts as whole seconds, matching the int-seconds shape that provider
+// constructors and the user-facing Config.RequestTimeout field use.
+const (
+ DefaultRequestTimeoutSeconds = int(DefaultRequestTimeout / time.Second)
+ ResearchRequestTimeoutSeconds = int(ResearchRequestTimeout / time.Second)
+)
+
+// Retry policy.
+//
+// These govern how transient LLM HTTP failures (network resets, 5xx, 429 rate
+// limits) are retried with exponential backoff plus jitter. The defaults are
+// intentionally conservative: a handful of fast attempts that smooth over brief
+// upstream blips without hammering the API or stalling interactive use.
+//
+// - RetryMaxAttempts: total tries including the first; <=1 disables retries.
+// - RetryBaseDelay: delay before the first retry; doubles each subsequent attempt.
+// - RetryMaxDelay: upper bound on any single backoff so growth stays bounded.
+// - RetryJitterFraction: +/- fraction of random jitter to avoid thundering-herd retries.
+const (
+ RetryMaxAttempts = 3
+ RetryBaseDelay = 200 * time.Millisecond
+ RetryMaxDelay = 2 * time.Second
+ RetryJitterFraction = 0.2
+)
+
+// Circuit breaker.
+//
+// The shared circuit breaker guards the whole LLM HTTP path. After
+// CircuitFailureThreshold consecutive transient failures it trips open and
+// rejects requests for CircuitCooldown, giving an unhealthy upstream time to
+// recover instead of forcing every caller to wait out full retry+timeout
+// cycles.
+const (
+ CircuitFailureThreshold = 5
+ CircuitCooldown = 30 * time.Second
+)
diff --git a/internal/llm/policy/policy_test.go b/internal/llm/policy/policy_test.go
new file mode 100644
index 0000000..976b41e
--- /dev/null
+++ b/internal/llm/policy/policy_test.go
@@ -0,0 +1,67 @@
+package policy
+
+import (
+ "testing"
+ "time"
+)
+
+// TestRequestTimeouts verifies the documented default and research timeouts and
+// that their whole-second mirrors stay in sync with the duration constants.
+func TestRequestTimeouts(t *testing.T) {
+ if DefaultRequestTimeout != 30*time.Second {
+ t.Fatalf("DefaultRequestTimeout = %v, want 30s", DefaultRequestTimeout)
+ }
+ if ResearchRequestTimeout != 120*time.Second {
+ t.Fatalf("ResearchRequestTimeout = %v, want 120s", ResearchRequestTimeout)
+ }
+ if DefaultRequestTimeoutSeconds != 30 {
+ t.Fatalf("DefaultRequestTimeoutSeconds = %d, want 30", DefaultRequestTimeoutSeconds)
+ }
+ if ResearchRequestTimeoutSeconds != 120 {
+ t.Fatalf("ResearchRequestTimeoutSeconds = %d, want 120", ResearchRequestTimeoutSeconds)
+ }
+ // The integer mirrors must equal the duration constants converted to seconds.
+ if got := int(DefaultRequestTimeout / time.Second); got != DefaultRequestTimeoutSeconds {
+ t.Fatalf("DefaultRequestTimeoutSeconds out of sync: %d vs %d", DefaultRequestTimeoutSeconds, got)
+ }
+ if got := int(ResearchRequestTimeout / time.Second); got != ResearchRequestTimeoutSeconds {
+ t.Fatalf("ResearchRequestTimeoutSeconds out of sync: %d vs %d", ResearchRequestTimeoutSeconds, got)
+ }
+ // Research must allow a longer window than a normal chat completion.
+ if ResearchRequestTimeout <= DefaultRequestTimeout {
+ t.Fatalf("ResearchRequestTimeout (%v) must exceed DefaultRequestTimeout (%v)", ResearchRequestTimeout, DefaultRequestTimeout)
+ }
+}
+
+// TestRetryPolicyConstants verifies the retry tuning values stay within sane,
+// documented bounds.
+func TestRetryPolicyConstants(t *testing.T) {
+ if RetryMaxAttempts != 3 {
+ t.Fatalf("RetryMaxAttempts = %d, want 3", RetryMaxAttempts)
+ }
+ if RetryBaseDelay != 200*time.Millisecond {
+ t.Fatalf("RetryBaseDelay = %v, want 200ms", RetryBaseDelay)
+ }
+ if RetryMaxDelay != 2*time.Second {
+ t.Fatalf("RetryMaxDelay = %v, want 2s", RetryMaxDelay)
+ }
+ if RetryBaseDelay > RetryMaxDelay {
+ t.Fatalf("RetryBaseDelay (%v) must not exceed RetryMaxDelay (%v)", RetryBaseDelay, RetryMaxDelay)
+ }
+ if RetryJitterFraction <= 0 || RetryJitterFraction >= 1 {
+ t.Fatalf("RetryJitterFraction = %v, want in (0,1)", RetryJitterFraction)
+ }
+}
+
+// TestCircuitBreakerConstants verifies the circuit-breaker tuning values.
+func TestCircuitBreakerConstants(t *testing.T) {
+ if CircuitFailureThreshold != 5 {
+ t.Fatalf("CircuitFailureThreshold = %d, want 5", CircuitFailureThreshold)
+ }
+ if CircuitCooldown != 30*time.Second {
+ t.Fatalf("CircuitCooldown = %v, want 30s", CircuitCooldown)
+ }
+ if CircuitFailureThreshold <= 0 {
+ t.Fatalf("CircuitFailureThreshold must be positive to trip the breaker")
+ }
+}
diff --git a/internal/llm/provider.go b/internal/llm/provider.go
index dbfa603..27d8cc2 100644
--- a/internal/llm/provider.go
+++ b/internal/llm/provider.go
@@ -75,7 +75,7 @@ func WithStop(stop ...string) RequestOption {
// Config defines provider configuration read from the Hexai config file.
type Config struct {
Provider string
- RequestTimeout int // seconds; 0 means use default (30s)
+ RequestTimeout int // seconds; 0 means use the provider default (see internal/llm/policy)
// OpenAI options
OpenAIBaseURL string
OpenAIModel string
diff --git a/internal/llm/resilience.go b/internal/llm/resilience.go
index 4d82153..ea6a32b 100644
--- a/internal/llm/resilience.go
+++ b/internal/llm/resilience.go
@@ -9,6 +9,7 @@ import (
"net/http"
"time"
+ "codeberg.org/snonux/hexai/internal/llm/policy"
"codeberg.org/snonux/hexai/internal/logging"
)
@@ -37,15 +38,17 @@ type retryPolicy struct {
randFloat func() float64
}
-// defaultRetryPolicy returns the policy used for all LLM HTTP calls. Three
-// total attempts (two retries) with 200ms base backoff keeps recovery fast for
+// defaultRetryPolicy returns the policy used for all LLM HTTP calls. The
+// concrete tuning values (attempts, backoff, jitter) live in the policy package
+// so the whole resilience policy has a single source of truth. Three total
+// attempts (two retries) with a 200ms base backoff keeps recovery fast for
// transient errors while staying well within typical request timeouts.
func defaultRetryPolicy() retryPolicy {
return retryPolicy{
- maxAttempts: 3,
- baseDelay: 200 * time.Millisecond,
- maxDelay: 2 * time.Second,
- jitterFraction: 0.2,
+ maxAttempts: policy.RetryMaxAttempts,
+ baseDelay: policy.RetryBaseDelay,
+ maxDelay: policy.RetryMaxDelay,
+ jitterFraction: policy.RetryJitterFraction,
sleep: sleepWithContext,
randFloat: rand.Float64,
}
@@ -54,7 +57,8 @@ func defaultRetryPolicy() retryPolicy {
// sharedBreaker guards the whole LLM HTTP path. It trips after several
// consecutive transient failures and stays open briefly so a single unhealthy
// upstream does not cause every caller to wait out full retry+timeout cycles.
-var sharedBreaker = newCircuitBreaker(5, 30*time.Second)
+// Its threshold and cooldown come from the policy package.
+var sharedBreaker = newCircuitBreaker(policy.CircuitFailureThreshold, policy.CircuitCooldown)
// sleepWithContext sleeps for d unless ctx is cancelled first, in which case it
// returns ctx.Err(). This keeps backoff waits responsive to cancellation and
diff --git a/internal/llm/yousearch.go b/internal/llm/yousearch.go
index 7539e8d..66cfb35 100644
--- a/internal/llm/yousearch.go
+++ b/internal/llm/yousearch.go
@@ -12,6 +12,7 @@ import (
"strings"
"time"
+ "codeberg.org/snonux/hexai/internal/llm/policy"
"codeberg.org/snonux/hexai/internal/logging"
)
@@ -52,7 +53,10 @@ func youSearchProviderFactory(cfg Config, keys ProviderKeys) (Client, error) {
}
timeoutSec := cfg.RequestTimeout
if timeoutSec <= 0 {
- timeoutSec = 120
+ // The Research API runs a long multi-step pipeline, so it uses the
+ // larger research timeout from the policy package rather than the
+ // default chat timeout.
+ timeoutSec = policy.ResearchRequestTimeoutSeconds
}
return youSearchClient{
httpClient: &http.Client{Timeout: time.Duration(timeoutSec) * time.Second},
generated by cgit v1.2.3 (git 2.25.1) at 2026-08-07 04:28:51 +0000