session: rename exports for clean public API
- Core interface → Session interface - NewAgentCore → New - AgentCoreConfig → Config - Config (agent JSON) → AgentConfig - types.go → session.go (defines Session interface) - session.go → history.go (History type, persistence)
This commit is contained in:
Levi Neely
parent
6ee949bd7c
commit
916b31ee89
|
|
@ -58,7 +58,7 @@ func (p Prompt) MarshalJSON() ([]byte, error) {
|
|||
return json.Marshal(p.Value)
|
||||
}
|
||||
|
||||
type Config struct {
|
||||
type AgentConfig struct {
|
||||
Hooks map[string]HookCmds `json:"hooks,omitempty"`
|
||||
Prompt Prompt `json:"prompt,omitempty"`
|
||||
Backend string `json:"backend,omitempty"`
|
||||
|
|
@ -90,8 +90,8 @@ type Config struct {
|
|||
}
|
||||
|
||||
// Load parses a Config from r.
|
||||
func Load(r io.Reader) (*Config, error) {
|
||||
var cfg Config
|
||||
func Load(r io.Reader) (*AgentConfig, error) {
|
||||
var cfg AgentConfig
|
||||
if err := json.NewDecoder(r).Decode(&cfg); err != nil {
|
||||
return nil, err
|
||||
}
|
||||
|
|
@ -100,6 +100,6 @@ func Load(r io.Reader) (*Config, error) {
|
|||
|
||||
// ToolsEnabled reports whether tool use is enabled. Defaults to true when
|
||||
// the field is omitted from the config.
|
||||
func (c *Config) ToolsEnabled() bool {
|
||||
func (c *AgentConfig) ToolsEnabled() bool {
|
||||
return c.Tools == nil || *c.Tools
|
||||
}
|
||||
|
|
|
|||
|
|
@ -96,7 +96,7 @@ func newCore(t *testing.T, be backend.Backend, hooks Hooks) *harness {
|
|||
Hooks: hooks,
|
||||
Preamble: "test system prompt",
|
||||
}
|
||||
c := NewAgentCore(AgentCoreConfig{
|
||||
c := New(Config{
|
||||
Backend: be,
|
||||
AgentName: "test",
|
||||
AgentsDir: t.TempDir(),
|
||||
|
|
@ -111,7 +111,7 @@ func newCore(t *testing.T, be backend.Backend, hooks Hooks) *harness {
|
|||
}
|
||||
|
||||
// collectEvents runs Submit synchronously and returns all emitted events.
|
||||
func collectEvents(ctx context.Context, c Core, input string) []Event {
|
||||
func collectEvents(ctx context.Context, c Session, input string) []Event {
|
||||
var mu sync.Mutex
|
||||
var evs []Event
|
||||
sub := c.Bus().Subscribe("event", func(ev Event) {
|
||||
|
|
@ -138,7 +138,7 @@ func byRole(evs []Event, role string) []string {
|
|||
}
|
||||
|
||||
// waitState blocks until c.State() == want, failing after 2 s.
|
||||
func waitState(t *testing.T, c Core, want string) {
|
||||
func waitState(t *testing.T, c Session, want string) {
|
||||
t.Helper()
|
||||
ctx, cancel := context.WithTimeout(context.Background(), 2*time.Second)
|
||||
defer cancel()
|
||||
|
|
@ -1779,7 +1779,7 @@ func TestSetSessionID_UpdatesPreamble(t *testing.T) {
|
|||
env := &Runtime{
|
||||
Preamble: "session is " + oldID + " end",
|
||||
}
|
||||
c := NewAgentCore(AgentCoreConfig{
|
||||
c := New(Config{
|
||||
Backend: defaultBE(),
|
||||
AgentName: "test",
|
||||
AgentsDir: t.TempDir(),
|
||||
|
|
@ -2280,7 +2280,7 @@ func newCoreWithExecServer(t *testing.T, srv *mockEnvServer) *harness {
|
|||
Preamble: "test system prompt",
|
||||
Dispatcher: d,
|
||||
}
|
||||
c := NewAgentCore(AgentCoreConfig{
|
||||
c := New(Config{
|
||||
Backend: defaultBE(),
|
||||
AgentName: "test",
|
||||
AgentsDir: t.TempDir(),
|
||||
|
|
@ -2396,7 +2396,7 @@ func TestBuildRuntime_NilConfig(t *testing.T) {
|
|||
func TestBuildRuntime_PromptBecomesPreamble(t *testing.T) {
|
||||
setupCfgDir(t)
|
||||
d := tools.NewDispatcher()
|
||||
cfg := &Config{Prompt: Prompt{Value: []string{"the prompt"}}}
|
||||
cfg := &AgentConfig{Prompt: Prompt{Value: []string{"the prompt"}}}
|
||||
env := BuildRuntime(cfg, d, t.TempDir(), nil)
|
||||
|
||||
if env.Preamble != "the prompt" {
|
||||
|
|
@ -2408,7 +2408,7 @@ func TestBuildRuntime_HooksAndParams(t *testing.T) {
|
|||
setupCfgDir(t)
|
||||
d := tools.NewDispatcher()
|
||||
temp := 0.7
|
||||
cfg := &Config{
|
||||
cfg := &AgentConfig{
|
||||
Hooks: map[string]HookCmds{"preTurn": {"echo hi"}},
|
||||
MaxTokens: 512,
|
||||
Temperature: &temp,
|
||||
|
|
@ -2492,7 +2492,7 @@ func TestBuildRuntime_ToolsDisabled(t *testing.T) {
|
|||
tools: []tools.ToolInfo{{Server: "s1", Name: "mytool", Description: "desc", InputSchema: json.RawMessage(`{}`)}},
|
||||
}
|
||||
f := false
|
||||
cfg := &Config{Tools: &f}
|
||||
cfg := &AgentConfig{Tools: &f}
|
||||
env := BuildRuntime(cfg, d, t.TempDir(), nil)
|
||||
if len(env.Tools) != 0 {
|
||||
t.Errorf("expected no tools when disabled; got %+v", env.Tools)
|
||||
|
|
@ -2502,7 +2502,7 @@ func TestBuildRuntime_ToolsDisabled(t *testing.T) {
|
|||
func TestBuildRuntime_PromptOnly(t *testing.T) {
|
||||
setupCfgDir(t)
|
||||
d := tools.NewDispatcher()
|
||||
cfg := &Config{Prompt: Prompt{Value: []string{"only agent"}}}
|
||||
cfg := &AgentConfig{Prompt: Prompt{Value: []string{"only agent"}}}
|
||||
env := BuildRuntime(cfg, d, t.TempDir(), nil)
|
||||
if env.Preamble != "only agent" {
|
||||
t.Errorf("preamble = %q; want %q", env.Preamble, "only agent")
|
||||
|
|
@ -2512,7 +2512,7 @@ func TestBuildRuntime_PromptOnly(t *testing.T) {
|
|||
func TestBuildRuntime_ExecPrompt(t *testing.T) {
|
||||
setupCfgDir(t)
|
||||
d := tools.NewDispatcher()
|
||||
cfg := &Config{Prompt: Prompt{
|
||||
cfg := &AgentConfig{Prompt: Prompt{
|
||||
Value: []string{"echo hello", "echo 'You are a security auditor.'", "echo world"},
|
||||
IsExec: true,
|
||||
}}
|
||||
|
|
@ -2532,7 +2532,7 @@ func TestBuildRuntime_ExecPromptFileResolution(t *testing.T) {
|
|||
t.Setenv("OLLIE_PROMPTS_PATH", promptsDir)
|
||||
|
||||
d := tools.NewDispatcher()
|
||||
cfg := &Config{Prompt: Prompt{
|
||||
cfg := &AgentConfig{Prompt: Prompt{
|
||||
Value: []string{"test-prompt", "echo extra"},
|
||||
IsExec: true,
|
||||
}}
|
||||
|
|
|
|||
|
|
@ -35,7 +35,7 @@ type toolClassifier func(name string) bool
|
|||
// env provides additional environment variables injected into prompt resolution
|
||||
// subprocesses (e.g. OLLIE_SESSION_ID=xxx).
|
||||
// The caller is responsible for registering all servers on d before calling this.
|
||||
func BuildRuntime(cfg *Config, d tools.Dispatcher, cwd string, env []string, baseLayers ...string) *Runtime {
|
||||
func BuildRuntime(cfg *AgentConfig, d tools.Dispatcher, cwd string, env []string, baseLayers ...string) *Runtime {
|
||||
var messages []string
|
||||
|
||||
var allToolInfos []tools.ToolInfo
|
||||
|
|
@ -269,8 +269,8 @@ type actionHandle struct {
|
|||
cancel context.CancelCauseFunc
|
||||
}
|
||||
|
||||
// AgentCoreConfig is the configuration for creating an agent.
|
||||
type AgentCoreConfig struct {
|
||||
// Config is the configuration for creating an agent.
|
||||
type Config struct {
|
||||
Backend backend.Backend
|
||||
ModelName string // if non-empty, overrides backend's default model
|
||||
AgentName string
|
||||
|
|
@ -392,7 +392,7 @@ func (a *harness) pushLockDir() {
|
|||
|
||||
}
|
||||
|
||||
var _ Core = (*harness)(nil) // compile-time interface check
|
||||
var _ Session = (*harness)(nil) // compile-time interface check
|
||||
|
||||
var sweepTmpOnce sync.Once
|
||||
|
||||
|
|
@ -435,8 +435,8 @@ func sweepStaleTmpDirs() {
|
|||
})
|
||||
}
|
||||
|
||||
// NewAgentCore creates an agent from the given configuration.
|
||||
func NewAgentCore(cfg AgentCoreConfig) Core {
|
||||
// New creates an agent from the given configuration.
|
||||
func New(cfg Config) Session {
|
||||
sweepStaleTmpDirs()
|
||||
if cfg.ModelName != "" {
|
||||
cfg.Backend.SetModel(cfg.ModelName)
|
||||
|
|
@ -450,7 +450,7 @@ func NewAgentCore(cfg AgentCoreConfig) Core {
|
|||
}
|
||||
// Store the backend on the runtime so it's the single source of truth.
|
||||
rt.Backend = cfg.Backend
|
||||
// A non-zero MaxSteps in AgentCoreConfig takes precedence over the
|
||||
// A non-zero MaxSteps in Config takes precedence over the
|
||||
// value loaded from the agent JSON.
|
||||
if cfg.MaxSteps > 0 {
|
||||
rt.MaxSteps = cfg.MaxSteps
|
||||
|
|
|
|||
|
|
@ -0,0 +1,696 @@
|
|||
package session
|
||||
|
||||
import (
|
||||
"context"
|
||||
"encoding/json"
|
||||
"fmt"
|
||||
"os"
|
||||
"slices"
|
||||
"strings"
|
||||
"time"
|
||||
|
||||
"ollie/backend"
|
||||
)
|
||||
|
||||
const (
|
||||
compactionPrompt = `You are performing a CONTEXT CHECKPOINT COMPACTION. Produce a JSON object summarizing the current task state. Output ONLY valid JSON matching this schema:
|
||||
|
||||
{
|
||||
"objective": "the user's overall goal",
|
||||
"plan_step": "what step you are currently on",
|
||||
"constraints": ["any constraints or user preferences"],
|
||||
"last_action": "brief: last tool/action + outcome",
|
||||
"open_questions": ["unresolved questions"],
|
||||
"next_decision": "what to do next"
|
||||
}
|
||||
|
||||
Be concise. Capture what another LLM needs to seamlessly continue.`
|
||||
|
||||
// hotTailSize is the number of most recent messages kept verbatim (hot zone).
|
||||
hotTailSize = 8
|
||||
|
||||
// warmIndexSize is the number of messages preceding the hot zone
|
||||
// summarized as a brief decision index (warm zone).
|
||||
warmIndexSize = 10
|
||||
)
|
||||
|
||||
type Reaction struct {
|
||||
ID string `json:"id"`
|
||||
ResponseID string `json:"responseId"`
|
||||
Emoji string `json:"emoji"`
|
||||
Category string `json:"category"`
|
||||
CreatedAt time.Time `json:"createdAt"`
|
||||
}
|
||||
|
||||
// PersistedSession is the on-disk format for a saved session.
|
||||
type PersistedSession struct {
|
||||
ID string `json:"id"`
|
||||
Agent string `json:"agent,omitempty"`
|
||||
Backend string `json:"backend,omitempty"`
|
||||
Model string `json:"model,omitempty"`
|
||||
CWD string `json:"cwd,omitempty"`
|
||||
Remote string `json:"remote,omitempty"`
|
||||
Messages []backend.Message `json:"messages"`
|
||||
TaskState *TaskState `json:"taskState,omitempty"`
|
||||
// Usage and cost tracking, persisted across restarts.
|
||||
TotalInputTokens int `json:"totalInputTokens,omitempty"`
|
||||
TotalCachedInputTokens int `json:"totalCachedInputTokens,omitempty"`
|
||||
TotalCacheCreationTokens int `json:"totalCacheCreationTokens,omitempty"`
|
||||
TotalOutputTokens int `json:"totalOutputTokens,omitempty"`
|
||||
TotalRequests int `json:"totalRequests,omitempty"`
|
||||
Estimated bool `json:"estimated,omitempty"`
|
||||
LastTurnCostUSD float64 `json:"lastTurnCostUSD,omitempty"`
|
||||
SessionCostUSD float64 `json:"sessionCostUSD,omitempty"`
|
||||
// Reaction tracking.
|
||||
PositiveReactions int `json:"positiveReactions,omitempty"`
|
||||
NegativeReactions int `json:"negativeReactions,omitempty"`
|
||||
Reactions []Reaction `json:"reactions,omitempty"`
|
||||
}
|
||||
|
||||
// TaskState is a compact structured overlay that summarizes the agent's
|
||||
// current position in a task. Injected at the top of every turn so the
|
||||
// model doesn't rely on full history recall.
|
||||
type TaskState struct {
|
||||
Objective string `json:"objective"`
|
||||
PlanStep string `json:"plan_step"`
|
||||
Constraints []string `json:"constraints,omitempty"`
|
||||
LastAction string `json:"last_action"`
|
||||
OpenQuestions []string `json:"open_questions,omitempty"`
|
||||
NextDecision string `json:"next_decision"`
|
||||
}
|
||||
|
||||
// render produces a compact text representation for injection into the message history.
|
||||
func (ts *TaskState) render() string {
|
||||
if ts.Objective == "" {
|
||||
return ""
|
||||
}
|
||||
var sb strings.Builder
|
||||
sb.WriteString("[task state]\n")
|
||||
sb.WriteString("objective: " + ts.Objective + "\n")
|
||||
if ts.PlanStep != "" {
|
||||
sb.WriteString("plan_step: " + ts.PlanStep + "\n")
|
||||
}
|
||||
if len(ts.Constraints) > 0 {
|
||||
sb.WriteString("constraints: " + strings.Join(ts.Constraints, "; ") + "\n")
|
||||
}
|
||||
if ts.LastAction != "" {
|
||||
sb.WriteString("last_action: " + ts.LastAction + "\n")
|
||||
}
|
||||
if len(ts.OpenQuestions) > 0 {
|
||||
sb.WriteString("open_questions: " + strings.Join(ts.OpenQuestions, "; ") + "\n")
|
||||
}
|
||||
if ts.NextDecision != "" {
|
||||
sb.WriteString("next_decision: " + ts.NextDecision + "\n")
|
||||
}
|
||||
sb.WriteString("[/task state]")
|
||||
return sb.String()
|
||||
}
|
||||
|
||||
// saveTo writes the full message history to path as JSON.
|
||||
func (s *History) saveTo(path, id, agentName, cwd string) error {
|
||||
return s.saveToFull(path, id, agentName, "", "", cwd, "")
|
||||
}
|
||||
|
||||
// saveToFull writes session state including backend/model info.
|
||||
func sanitizeMessages(msgs []backend.Message) []backend.Message {
|
||||
out := make([]backend.Message, 0, len(msgs))
|
||||
for _, m := range msgs {
|
||||
for j, tc := range m.ToolCalls {
|
||||
if len(tc.Arguments) > 0 {
|
||||
var dummy json.RawMessage
|
||||
if err := json.Unmarshal(tc.Arguments, &dummy); err != nil {
|
||||
// Corrupted arguments — replace with empty object so save succeeds.
|
||||
m.ToolCalls[j].Arguments = json.RawMessage("{}")
|
||||
}
|
||||
}
|
||||
}
|
||||
out = append(out, m)
|
||||
}
|
||||
return out
|
||||
}
|
||||
|
||||
func (s *History) saveToFull(path, id, agentName, backendName, modelName, cwd, remote string) error {
|
||||
ps := PersistedSession{
|
||||
ID: id,
|
||||
Agent: agentName,
|
||||
Backend: backendName,
|
||||
Model: modelName,
|
||||
CWD: cwd,
|
||||
Remote: remote,
|
||||
Messages: sanitizeMessages(s.messages),
|
||||
TaskState: s.TaskState,
|
||||
TotalInputTokens: s.TotalInputTokens,
|
||||
TotalCachedInputTokens: s.TotalCachedInputTokens,
|
||||
TotalCacheCreationTokens: s.TotalCacheCreationTokens,
|
||||
TotalOutputTokens: s.TotalOutputTokens,
|
||||
TotalRequests: s.TotalRequests,
|
||||
Estimated: s.Estimated,
|
||||
LastTurnCostUSD: s.LastTurnCostUSD,
|
||||
SessionCostUSD: s.SessionCostUSD,
|
||||
PositiveReactions: s.PositiveReactions,
|
||||
NegativeReactions: s.NegativeReactions,
|
||||
Reactions: s.Reactions,
|
||||
}
|
||||
data, err := json.Marshal(ps)
|
||||
if err != nil {
|
||||
return fmt.Errorf("session save: %w", err)
|
||||
}
|
||||
// Atomic write: write to temp file, then rename to preserve last good snapshot.
|
||||
tmpPath := path + ".tmp"
|
||||
if err := os.WriteFile(tmpPath, data, 0600); err != nil {
|
||||
return fmt.Errorf("session save tmp: %w", err)
|
||||
}
|
||||
return os.Rename(tmpPath, path)
|
||||
}
|
||||
|
||||
// LoadPersistedSession reads a PersistedSession from a JSON file.
|
||||
func LoadPersistedSession(path string) (*PersistedSession, error) {
|
||||
data, err := os.ReadFile(path)
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
var ps PersistedSession
|
||||
if err := json.Unmarshal(data, &ps); err != nil {
|
||||
return nil, fmt.Errorf("parse session %s: %w", path, err)
|
||||
}
|
||||
return &ps, nil
|
||||
}
|
||||
|
||||
// RestoreHistory reconstructs a Session from a persisted message list.
|
||||
func RestoreHistory(ps *PersistedSession) *History {
|
||||
s := &History{
|
||||
messages: ps.Messages,
|
||||
TaskState: ps.TaskState,
|
||||
TotalInputTokens: ps.TotalInputTokens,
|
||||
TotalCachedInputTokens: ps.TotalCachedInputTokens,
|
||||
TotalCacheCreationTokens: ps.TotalCacheCreationTokens,
|
||||
TotalOutputTokens: ps.TotalOutputTokens,
|
||||
TotalRequests: ps.TotalRequests,
|
||||
Estimated: ps.Estimated,
|
||||
LastTurnCostUSD: ps.LastTurnCostUSD,
|
||||
SessionCostUSD: ps.SessionCostUSD,
|
||||
PositiveReactions: ps.PositiveReactions,
|
||||
NegativeReactions: ps.NegativeReactions,
|
||||
Reactions: ps.Reactions,
|
||||
}
|
||||
for i := range s.messages {
|
||||
if s.messages[i].Role == "assistant" && s.messages[i].ID == "" {
|
||||
s.messages[i].ID = NewResponseID()
|
||||
}
|
||||
}
|
||||
for _, m := range ps.Messages {
|
||||
if m.Role == "user" {
|
||||
s.goal = m.Content
|
||||
break
|
||||
}
|
||||
}
|
||||
return s
|
||||
}
|
||||
|
||||
// History is an ephemeral in-memory state backend.
|
||||
type History struct {
|
||||
goal string
|
||||
messages []backend.Message
|
||||
TaskState *TaskState
|
||||
// Cumulative usage tracking.
|
||||
TotalInputTokens int
|
||||
TotalCachedInputTokens int
|
||||
TotalCacheCreationTokens int
|
||||
TotalOutputTokens int
|
||||
TotalRequests int
|
||||
Estimated bool // true if any usage was estimated rather than reported by the backend
|
||||
LastTurnCostUSD float64 // cost of the most recently completed turn in USD
|
||||
SessionCostUSD float64 // cumulative cost of all turns in this session in USD
|
||||
// per-turn accumulators; reset at the start of each turn
|
||||
turnInputTokens int
|
||||
turnCachedTokens int
|
||||
turnCreationTokens int
|
||||
turnOutputTokens int
|
||||
turnCostUSD float64 // >0 when backend reported cost directly (e.g. OpenRouter)
|
||||
// Reaction tracking.
|
||||
PositiveReactions int
|
||||
NegativeReactions int
|
||||
Reactions []Reaction
|
||||
}
|
||||
|
||||
// newHistory creates a new empty Session. The caller is responsible for
|
||||
// appending the initial user message via appendUserMessage.
|
||||
func newHistory(goal string) *History {
|
||||
return &History{goal: goal}
|
||||
}
|
||||
|
||||
// Checkpoint forks the session: returns a new Session that inherits the given
|
||||
// TaskState but starts with a clean message history. This enables narrow-context
|
||||
// sub-agents that know what to do without inheriting all parent message noise.
|
||||
func (s *History) Checkpoint(ts TaskState) *History {
|
||||
child := &History{
|
||||
goal: ts.Objective,
|
||||
TaskState: &ts,
|
||||
}
|
||||
// Seed with a user message so the child has a valid initial turn.
|
||||
child.messages = []backend.Message{{
|
||||
Role: "user",
|
||||
Content: ts.render(),
|
||||
}}
|
||||
return child
|
||||
}
|
||||
|
||||
func (s *History) history() []backend.Message {
|
||||
if len(s.Reactions) == 0 {
|
||||
return s.messages
|
||||
}
|
||||
// Append feedback after the conversation so assistant tool-call messages stay
|
||||
// adjacent to their tool results, as required by provider APIs.
|
||||
out := make([]backend.Message, 0, len(s.messages)+len(s.Reactions))
|
||||
out = append(out, s.messages...)
|
||||
for _, r := range s.Reactions {
|
||||
_, desc, _ := classifyReaction(r.Emoji)
|
||||
out = append(out, backend.Message{Role: "user", Content: "[reaction to assistant response " + r.ResponseID + ": " + r.Category + " (" + r.Emoji + ")] " + desc})
|
||||
}
|
||||
return out
|
||||
}
|
||||
|
||||
func (s *History) taskState() *TaskState {
|
||||
return s.TaskState
|
||||
}
|
||||
|
||||
func (s *History) updateTaskState(ts TaskState) {
|
||||
s.TaskState = &ts
|
||||
}
|
||||
|
||||
func (s *History) addUsage(u backend.Usage, estimated bool) {
|
||||
s.TotalInputTokens += u.InputTokens
|
||||
s.TotalCachedInputTokens += u.CachedInputTokens
|
||||
s.TotalCacheCreationTokens += u.CacheCreationTokens
|
||||
s.TotalOutputTokens += u.OutputTokens
|
||||
s.TotalRequests++
|
||||
s.turnInputTokens += u.InputTokens
|
||||
s.turnCachedTokens += u.CachedInputTokens
|
||||
s.turnCreationTokens += u.CacheCreationTokens
|
||||
s.turnOutputTokens += u.OutputTokens
|
||||
s.turnCostUSD += u.CostUSD
|
||||
if estimated {
|
||||
s.Estimated = true
|
||||
}
|
||||
}
|
||||
|
||||
func (s *History) recomputeReactionCounts() {
|
||||
s.PositiveReactions = 0
|
||||
s.NegativeReactions = 0
|
||||
for _, r := range s.Reactions {
|
||||
switch r.Category {
|
||||
case "positive", "excellent":
|
||||
s.PositiveReactions++
|
||||
case "negative", "terrible":
|
||||
s.NegativeReactions++
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
func (s *History) resetTurnAccumulators() {
|
||||
s.turnInputTokens = 0
|
||||
s.turnCachedTokens = 0
|
||||
s.turnCreationTokens = 0
|
||||
s.turnOutputTokens = 0
|
||||
s.turnCostUSD = 0
|
||||
}
|
||||
|
||||
// recordTurnCost computes and stores the last turn's cost, adding it to the
|
||||
// session total. model is used for the pricing-table fallback when the backend
|
||||
// did not report cost directly.
|
||||
func (s *History) recordTurnCost(model string) {
|
||||
var cost float64
|
||||
if s.turnCostUSD > 0 {
|
||||
cost = s.turnCostUSD
|
||||
} else {
|
||||
cost = computeCostUSD(model, backend.Usage{
|
||||
InputTokens: s.turnInputTokens,
|
||||
CachedInputTokens: s.turnCachedTokens,
|
||||
CacheCreationTokens: s.turnCreationTokens,
|
||||
OutputTokens: s.turnOutputTokens,
|
||||
})
|
||||
}
|
||||
s.LastTurnCostUSD = cost
|
||||
s.SessionCostUSD += cost
|
||||
}
|
||||
|
||||
func (s *History) update(assistant backend.Message, results []toolResult) {
|
||||
s.messages = append(s.messages, assistant)
|
||||
for _, r := range results {
|
||||
s.messages = append(s.messages, backend.Message{
|
||||
Role: "tool",
|
||||
Content: r.Content,
|
||||
ContentBlocks: r.ContentBlocks,
|
||||
ToolCallID: r.ToolCallID,
|
||||
})
|
||||
}
|
||||
}
|
||||
|
||||
// llmSummarizeToolResult uses the LLM to produce a concise summary of a tool result.
|
||||
func llmSummarizeToolResult(ctx context.Context, b backend.Backend, name, content string) string {
|
||||
prompt := fmt.Sprintf("Summarize this tool result from %q in 1-3 sentences, preserving key facts, file paths, values, and findings:\n\n%s", name, content)
|
||||
ch, err := b.ChatStream(ctx, []backend.Message{{Role: "user", Content: prompt}}, nil, backend.GenerationParams{})
|
||||
if err != nil {
|
||||
return content
|
||||
}
|
||||
var sb strings.Builder
|
||||
for ev := range ch {
|
||||
if ev.Content != "" {
|
||||
sb.WriteString(ev.Content)
|
||||
}
|
||||
if ev.Done {
|
||||
break
|
||||
}
|
||||
}
|
||||
if s := strings.TrimSpace(sb.String()); s != "" {
|
||||
return fmt.Sprintf("[summary of %s result]: %s", name, s)
|
||||
}
|
||||
return content
|
||||
}
|
||||
|
||||
// removeCancelledToolResults filters out tool results that were cancelled due
|
||||
// to interrupt, keeping completed work. Also removes the corresponding tool
|
||||
// calls from assistant messages to maintain a valid message sequence.
|
||||
func (s *History) removeCancelledToolResults() {
|
||||
// First pass: collect cancelled tool call IDs
|
||||
cancelled := make(map[string]bool)
|
||||
for _, m := range s.messages {
|
||||
if m.Role == "tool" && isCancelledToolResult(m.Content) {
|
||||
cancelled[m.ToolCallID] = true
|
||||
}
|
||||
}
|
||||
if len(cancelled) == 0 {
|
||||
return
|
||||
}
|
||||
|
||||
// Second pass: filter messages and prune tool calls from assistant messages
|
||||
filtered := s.messages[:0]
|
||||
for _, m := range s.messages {
|
||||
if m.Role == "tool" && cancelled[m.ToolCallID] {
|
||||
continue
|
||||
}
|
||||
if m.Role == "assistant" && len(m.ToolCalls) > 0 {
|
||||
// Remove cancelled tool calls from this assistant message
|
||||
kept := m.ToolCalls[:0]
|
||||
for _, tc := range m.ToolCalls {
|
||||
if !cancelled[tc.ID] {
|
||||
kept = append(kept, tc)
|
||||
}
|
||||
}
|
||||
if len(kept) == 0 && m.Content == "" {
|
||||
// All tool calls cancelled and no text content - skip message
|
||||
continue
|
||||
}
|
||||
m.ToolCalls = kept
|
||||
}
|
||||
filtered = append(filtered, m)
|
||||
}
|
||||
s.messages = filtered
|
||||
}
|
||||
|
||||
func isCancelledToolResult(content string) bool {
|
||||
return strings.Contains(content, `"status":"cancelled"`) ||
|
||||
strings.Contains(content, "tool execution interrupted by user")
|
||||
}
|
||||
|
||||
// PreCompactionSnapshot returns a copy of the current messages for persistence
|
||||
// before compaction. Call this before compact().
|
||||
func (s *History) PreCompactionSnapshot() []backend.Message {
|
||||
return slices.Clone(s.messages)
|
||||
}
|
||||
|
||||
// Compact summarizes the conversation via an LLM call, replacing the history
|
||||
// with system messages + preserved user messages + a structured summary.
|
||||
// Returns (n compacted, summary text, error); n==0 means nothing to compact.
|
||||
func (s *History) compact(ctx context.Context, b backend.Backend) (int, string, error) {
|
||||
if len(s.messages) <= hotTailSize+warmIndexSize {
|
||||
return 0, "", nil
|
||||
}
|
||||
|
||||
// Flatten tool calls into plain text so the compaction request
|
||||
// doesn't need tool schemas.
|
||||
flattened := flattenToolMessages(s.messages)
|
||||
flattened = append(flattened, backend.Message{
|
||||
Role: "user",
|
||||
Content: compactionPrompt,
|
||||
})
|
||||
|
||||
ch, err := b.ChatStream(ctx, flattened, nil, backend.GenerationParams{})
|
||||
if err != nil {
|
||||
return 0, "", fmt.Errorf("compact: %w", err)
|
||||
}
|
||||
|
||||
var summary strings.Builder
|
||||
for ev := range ch {
|
||||
if ev.Content != "" {
|
||||
summary.WriteString(ev.Content)
|
||||
}
|
||||
if ev.Done {
|
||||
break
|
||||
}
|
||||
}
|
||||
|
||||
summaryText := strings.TrimSpace(summary.String())
|
||||
if summaryText == "" {
|
||||
return 0, "", fmt.Errorf("compact: empty summary")
|
||||
}
|
||||
|
||||
// Parse structured task state from the model's JSON response.
|
||||
var ts TaskState
|
||||
raw := extractJSON(summaryText)
|
||||
if err := json.Unmarshal([]byte(raw), &ts); err != nil {
|
||||
// Fallback: treat as unstructured summary if JSON parse fails.
|
||||
ts = TaskState{Objective: summaryText}
|
||||
}
|
||||
s.TaskState = &ts
|
||||
|
||||
beforeCount := len(s.messages)
|
||||
s.messages = buildCompactedHistory(ts, s.messages)
|
||||
return beforeCount - len(s.messages), summaryText, nil
|
||||
}
|
||||
|
||||
// buildCompactedHistory constructs a three-zone post-compaction message list:
|
||||
// - Cold: structured task state summary (everything older than warm+hot)
|
||||
// - Warm: brief index of recent decisions/actions (warmIndexSize messages)
|
||||
// - Hot: last hotTailSize messages verbatim
|
||||
func buildCompactedHistory(ts TaskState, allMessages []backend.Message) []backend.Message {
|
||||
total := len(allMessages)
|
||||
|
||||
// Determine zone boundaries.
|
||||
hotStart := total - hotTailSize
|
||||
if hotStart < 0 {
|
||||
hotStart = 0
|
||||
}
|
||||
|
||||
// If hotStart lands on a tool message, walk backward to include the
|
||||
// preceding assistant message with tool_calls. Without this, the hot
|
||||
// zone starts with an orphaned tool result which violates the OpenAI
|
||||
// API constraint that tool messages must follow an assistant message
|
||||
// containing the corresponding tool_calls.
|
||||
for hotStart > 0 && allMessages[hotStart].Role == "tool" {
|
||||
hotStart--
|
||||
}
|
||||
|
||||
warmStart := hotStart - warmIndexSize
|
||||
if warmStart < 0 {
|
||||
warmStart = 0
|
||||
}
|
||||
|
||||
// Cold zone: structured task state.
|
||||
stateJSON, _ := json.Marshal(ts)
|
||||
cold := backend.Message{
|
||||
Role: "user",
|
||||
Content: "[compacted context — cold zone]\n```json\n" + string(stateJSON) + "\n```",
|
||||
}
|
||||
|
||||
var out []backend.Message
|
||||
out = append(out, cold)
|
||||
|
||||
// Warm zone: one-line summaries of decisions in the warm window.
|
||||
warmSlice := allMessages[warmStart:hotStart]
|
||||
if len(warmSlice) > 0 {
|
||||
var sb strings.Builder
|
||||
sb.WriteString("[warm zone — recent decisions]\n")
|
||||
for _, m := range warmSlice {
|
||||
line := summarizeMessage(m)
|
||||
if line != "" {
|
||||
sb.WriteString("- ")
|
||||
sb.WriteString(line)
|
||||
sb.WriteByte('\n')
|
||||
}
|
||||
}
|
||||
out = append(out, backend.Message{Role: "user", Content: sb.String()})
|
||||
}
|
||||
|
||||
// Hot zone: verbatim recent messages.
|
||||
out = append(out, allMessages[hotStart:]...)
|
||||
return out
|
||||
}
|
||||
|
||||
// summarizeMessage produces a one-line summary of a message for the warm index.
|
||||
func summarizeMessage(m backend.Message) string {
|
||||
switch m.Role {
|
||||
case "assistant":
|
||||
if len(m.ToolCalls) > 0 {
|
||||
names := make([]string, len(m.ToolCalls))
|
||||
for i, tc := range m.ToolCalls {
|
||||
names[i] = tc.Name
|
||||
}
|
||||
return "called: " + strings.Join(names, ", ")
|
||||
}
|
||||
text := m.Content
|
||||
if len(text) > 120 {
|
||||
text = text[:120] + "…"
|
||||
}
|
||||
return "said: " + text
|
||||
case "user":
|
||||
text := m.Content
|
||||
if len(text) > 120 {
|
||||
text = text[:120] + "…"
|
||||
}
|
||||
return "user: " + text
|
||||
case "tool":
|
||||
text := m.Content
|
||||
if len(text) > 80 {
|
||||
text = text[:80] + "…"
|
||||
}
|
||||
return "result(" + m.ToolCallID + "): " + text
|
||||
}
|
||||
return ""
|
||||
}
|
||||
|
||||
// extractJSON finds the first JSON object in s (handling markdown fences).
|
||||
func extractJSON(s string) string {
|
||||
// Strip markdown code fence if present.
|
||||
if i := strings.Index(s, "```json"); i >= 0 {
|
||||
s = s[i+7:]
|
||||
if j := strings.Index(s, "```"); j >= 0 {
|
||||
s = s[:j]
|
||||
}
|
||||
} else if i := strings.Index(s, "```"); i >= 0 {
|
||||
s = s[i+3:]
|
||||
if j := strings.Index(s, "```"); j >= 0 {
|
||||
s = s[:j]
|
||||
}
|
||||
}
|
||||
// Find first { ... }
|
||||
start := strings.Index(s, "{")
|
||||
if start < 0 {
|
||||
return s
|
||||
}
|
||||
end := strings.LastIndex(s, "}")
|
||||
if end < start {
|
||||
return s
|
||||
}
|
||||
return s[start : end+1]
|
||||
}
|
||||
|
||||
// flattenToolMessages converts tool call/result sequences into plain text
|
||||
// so the compaction request doesn't include tool-specific structures that
|
||||
// the API may reject when no tools are defined.
|
||||
func flattenToolMessages(messages []backend.Message) []backend.Message {
|
||||
out := make([]backend.Message, 0, len(messages))
|
||||
for _, m := range messages {
|
||||
switch {
|
||||
case m.Role == "assistant" && len(m.ToolCalls) > 0:
|
||||
var sb strings.Builder
|
||||
if m.Content != "" {
|
||||
sb.WriteString(m.Content)
|
||||
sb.WriteString("\n\n")
|
||||
}
|
||||
for _, tc := range m.ToolCalls {
|
||||
fmt.Fprintf(&sb, "[Tool call: %s(%s)]\n", tc.Name, string(tc.Arguments))
|
||||
}
|
||||
out = append(out, backend.Message{Role: "assistant", Content: sb.String()})
|
||||
case m.Role == "tool":
|
||||
text := m.Content
|
||||
if len(text) > 4000 {
|
||||
text = text[:4000] + "..."
|
||||
}
|
||||
out = append(out, backend.Message{
|
||||
Role: "user",
|
||||
Content: fmt.Sprintf("[Tool result for %s]:\n%s", m.ToolCallID, text),
|
||||
})
|
||||
default:
|
||||
out = append(out, m)
|
||||
}
|
||||
}
|
||||
return out
|
||||
}
|
||||
|
||||
func (s *History) appendUserMessage(content string) {
|
||||
s.messages = append(s.messages, backend.Message{Role: "user", Content: content})
|
||||
}
|
||||
|
||||
// cloneMessages returns a deep copy of the message slice.
|
||||
func cloneMessages(msgs []backend.Message) []backend.Message {
|
||||
out := make([]backend.Message, len(msgs))
|
||||
for i, m := range msgs {
|
||||
out[i] = backend.Message{
|
||||
Role: m.Role,
|
||||
Content: m.Content,
|
||||
ToolCallID: m.ToolCallID,
|
||||
}
|
||||
if len(m.ToolCalls) > 0 {
|
||||
out[i].ToolCalls = make([]backend.ToolCall, len(m.ToolCalls))
|
||||
for j, tc := range m.ToolCalls {
|
||||
out[i].ToolCalls[j] = backend.ToolCall{
|
||||
ID: tc.ID,
|
||||
Name: tc.Name,
|
||||
Arguments: append(json.RawMessage(nil), tc.Arguments...),
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
return out
|
||||
}
|
||||
|
||||
// estimateTokens returns a rough token count (~4 chars per token).
|
||||
func (s *History) estimateTokens() int {
|
||||
chars := 0
|
||||
for _, m := range s.messages {
|
||||
chars += len(m.Content)
|
||||
for _, tc := range m.ToolCalls {
|
||||
chars += len(tc.Name) + len(tc.Arguments)
|
||||
}
|
||||
}
|
||||
return chars / 4
|
||||
}
|
||||
|
||||
// stripColdResults summarizes large tool-result messages outside the hot tail
|
||||
// using an LLM call. Messages in the last hotTailSize slots are left verbatim.
|
||||
func (s *History) stripColdResults(ctx context.Context, b backend.Backend) {
|
||||
hot := len(s.messages) - hotTailSize
|
||||
for i := range s.messages {
|
||||
if i >= hot {
|
||||
break
|
||||
}
|
||||
m := &s.messages[i]
|
||||
if m.Role == "tool" && len(m.Content) > 200 {
|
||||
m.Content = llmSummarizeToolResult(ctx, b, m.ToolCallID, m.Content)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// stripCold implements the state interface.
|
||||
func (s *History) stripCold(ctx context.Context, b backend.Backend) {
|
||||
s.stripColdResults(ctx, b)
|
||||
}
|
||||
|
||||
// contextDebug returns a multi-line breakdown of the history.
|
||||
func (s *History) contextDebug() string {
|
||||
var sb strings.Builder
|
||||
sb.WriteString(fmt.Sprintf("=== %d messages ===\n", len(s.messages)))
|
||||
for i, m := range s.messages {
|
||||
preview := m.Content
|
||||
if len(preview) > 80 {
|
||||
preview = preview[:80] + "..."
|
||||
}
|
||||
chars := len(m.Content)
|
||||
for _, tc := range m.ToolCalls {
|
||||
chars += len(tc.Name) + len(tc.Arguments)
|
||||
}
|
||||
sb.WriteString(fmt.Sprintf(" [%d] role=%-10s chars=%-6d %q\n", i, m.Role, chars, preview))
|
||||
}
|
||||
return sb.String()
|
||||
}
|
||||
|
|
@ -2,695 +2,186 @@ package session
|
|||
|
||||
import (
|
||||
"context"
|
||||
"encoding/json"
|
||||
"fmt"
|
||||
"os"
|
||||
"slices"
|
||||
"strings"
|
||||
"time"
|
||||
"errors"
|
||||
|
||||
"github.com/simonfxr/pubsub"
|
||||
"ollie/backend"
|
||||
)
|
||||
|
||||
// WatchField names supported by Core.WaitChange.
|
||||
const (
|
||||
compactionPrompt = `You are performing a CONTEXT CHECKPOINT COMPACTION. Produce a JSON object summarizing the current task state. Output ONLY valid JSON matching this schema:
|
||||
|
||||
{
|
||||
"objective": "the user's overall goal",
|
||||
"plan_step": "what step you are currently on",
|
||||
"constraints": ["any constraints or user preferences"],
|
||||
"last_action": "brief: last tool/action + outcome",
|
||||
"open_questions": ["unresolved questions"],
|
||||
"next_decision": "what to do next"
|
||||
}
|
||||
|
||||
Be concise. Capture what another LLM needs to seamlessly continue.`
|
||||
|
||||
// hotTailSize is the number of most recent messages kept verbatim (hot zone).
|
||||
hotTailSize = 8
|
||||
|
||||
// warmIndexSize is the number of messages preceding the hot zone
|
||||
// summarized as a brief decision index (warm zone).
|
||||
warmIndexSize = 10
|
||||
WatchState = "state"
|
||||
WatchUsage = "usage"
|
||||
WatchCtxSz = "ctxsz"
|
||||
WatchCWD = "cwd"
|
||||
WatchAgent = "agent"
|
||||
)
|
||||
|
||||
type Reaction struct {
|
||||
ID string `json:"id"`
|
||||
ResponseID string `json:"responseId"`
|
||||
Emoji string `json:"emoji"`
|
||||
Category string `json:"category"`
|
||||
CreatedAt time.Time `json:"createdAt"`
|
||||
// ErrInterrupted is returned when the user cancels an agent turn (Ctrl-C).
|
||||
var ErrInterrupted = errors.New("interrupted")
|
||||
|
||||
// Event is a typed output event emitted during an agent turn or in response
|
||||
// to a command.
|
||||
type Event struct {
|
||||
Role string
|
||||
Name string
|
||||
Content string
|
||||
ResponseID string
|
||||
}
|
||||
|
||||
// PersistedSession is the on-disk format for a saved session.
|
||||
type PersistedSession struct {
|
||||
ID string `json:"id"`
|
||||
Agent string `json:"agent,omitempty"`
|
||||
Backend string `json:"backend,omitempty"`
|
||||
Model string `json:"model,omitempty"`
|
||||
CWD string `json:"cwd,omitempty"`
|
||||
Remote string `json:"remote,omitempty"`
|
||||
Messages []backend.Message `json:"messages"`
|
||||
TaskState *TaskState `json:"taskState,omitempty"`
|
||||
// Usage and cost tracking, persisted across restarts.
|
||||
TotalInputTokens int `json:"totalInputTokens,omitempty"`
|
||||
TotalCachedInputTokens int `json:"totalCachedInputTokens,omitempty"`
|
||||
TotalCacheCreationTokens int `json:"totalCacheCreationTokens,omitempty"`
|
||||
TotalOutputTokens int `json:"totalOutputTokens,omitempty"`
|
||||
TotalRequests int `json:"totalRequests,omitempty"`
|
||||
Estimated bool `json:"estimated,omitempty"`
|
||||
LastTurnCostUSD float64 `json:"lastTurnCostUSD,omitempty"`
|
||||
SessionCostUSD float64 `json:"sessionCostUSD,omitempty"`
|
||||
// Reaction tracking.
|
||||
PositiveReactions int `json:"positiveReactions,omitempty"`
|
||||
NegativeReactions int `json:"negativeReactions,omitempty"`
|
||||
Reactions []Reaction `json:"reactions,omitempty"`
|
||||
// EventHandler receives events from the agent.
|
||||
type EventHandler func(Event)
|
||||
|
||||
// Session is the interface between a frontend (TUI, HTTP handler, etc.) and the
|
||||
// agent engine. All output from the agent is delivered via the event bus.
|
||||
type Session interface {
|
||||
// Submit processes one line of user input. Slash commands and shell
|
||||
// shortcuts are dispatched synchronously; any other input starts an agent
|
||||
// turn that publishes events to the bus until the turn is complete.
|
||||
// After the turn, any queued prompts are drained sequentially.
|
||||
Submit(ctx context.Context, input string)
|
||||
|
||||
// Interrupt cancels the current in-progress agent turn.
|
||||
// Returns true if an action was running and was cancelled.
|
||||
Interrupt(cause error) bool
|
||||
|
||||
// Inject sends a message that will be appended to the next tool result
|
||||
// as a user interruption. If no turn is running, it is silently dropped.
|
||||
Inject(prompt string)
|
||||
|
||||
// Queue pushes a prompt onto the FIFO for execution after the current
|
||||
// turn completes.
|
||||
Queue(prompt string)
|
||||
|
||||
// Bus returns the session event bus.
|
||||
Bus() *pubsub.Bus
|
||||
|
||||
// PopQueue removes and returns the next queued prompt.
|
||||
// Returns ("", false) if the queue is empty.
|
||||
PopQueue() (string, bool)
|
||||
|
||||
// IsRunning returns true if an agent turn is currently in progress.
|
||||
IsRunning() bool
|
||||
|
||||
// State returns the current agent state: "idle", "thinking", or "calling: <tool>".
|
||||
State() string
|
||||
|
||||
// Reply returns the assistant text from the most recently completed turn.
|
||||
// Cleared when a new prompt is submitted.
|
||||
Reply() string
|
||||
|
||||
// AgentName returns the name of the active agent.
|
||||
AgentName() string
|
||||
|
||||
// BackendName returns the name of the active backend (e.g. "anthropic", "ollama").
|
||||
BackendName() string
|
||||
|
||||
// ModelName returns the name of the active model.
|
||||
ModelName() string
|
||||
|
||||
// CtxSz returns the estimated context size as a one-line summary.
|
||||
CtxSz() string
|
||||
|
||||
// Usage returns billed token counts as a one-line summary.
|
||||
Usage() string
|
||||
|
||||
// Cost returns per-turn and session cost as a two-line key=value summary.
|
||||
Cost() string
|
||||
|
||||
// ListModels returns available model names, one per line.
|
||||
ListModels() string
|
||||
|
||||
// CWD returns the current working directory used for tool execution.
|
||||
CWD() string
|
||||
|
||||
// SetCWD changes the working directory for tool execution and
|
||||
// updates the system prompt. Returns an error if the path does not exist.
|
||||
SetCWD(dir string) error
|
||||
|
||||
// SetSessionID renames the session: updates the in-memory ID, renames
|
||||
// persisted files on disk, and propagates to the execute server env.
|
||||
SetSessionID(newID string) error
|
||||
|
||||
// Context returns the current message history as it would be sent to the
|
||||
// backend: system prompt prepended, stale reads pruned. Does not include
|
||||
// tool definitions (see Tools) or generation params (see GenerationParams).
|
||||
Context() []backend.Message
|
||||
|
||||
// SystemPrompt returns the fully rendered system prompt for this session.
|
||||
SystemPrompt() string
|
||||
|
||||
// GenerationParams returns the current sampling parameters.
|
||||
GenerationParams() backend.GenerationParams
|
||||
|
||||
// SetGenerationParams replaces the current sampling parameters.
|
||||
// Returns an error if the agent is currently running.
|
||||
SetGenerationParams(params backend.GenerationParams) error
|
||||
|
||||
// CompactionModel returns the model override used for context compaction.
|
||||
CompactionModel() string
|
||||
|
||||
// SetCompactionModel changes the model used for context compaction.
|
||||
SetCompactionModel(model string)
|
||||
|
||||
// SetEnv injects a session-scoped environment variable into the shell
|
||||
// subprocesses. Does not affect the daemon process environment.
|
||||
SetEnv(key, value string)
|
||||
|
||||
// WaitChange blocks until the named field changes from current, then returns
|
||||
// the new value. Returns ("", false) if ctx is cancelled before a change.
|
||||
// Supported fields: WatchState, WatchUsage, WatchCtxSz, WatchCWD.
|
||||
WaitChange(ctx context.Context, field, current string) (string, bool)
|
||||
|
||||
// ToolCallCount returns the total number of tool calls executed in this
|
||||
// session since the agent was created. The counter is monotonically
|
||||
// increasing and never resets. Blocked calls (pre-tool hook exit 2) are
|
||||
// not counted. Timed-out or erroring calls are counted because execution
|
||||
// was attempted. Use modulo arithmetic in hooks to fire every N calls:
|
||||
// [ $(($(cat tcct) % 10)) -eq 0 ] && ...
|
||||
ToolCallCount() int64
|
||||
|
||||
// SaveSession writes the current session state to the given path.
|
||||
// The file includes all messages, agent/backend/model metadata, and task state.
|
||||
SaveSession(path string) error
|
||||
|
||||
// Close releases resources associated with the session, including its
|
||||
// temporary directory under /tmp/ollie/.
|
||||
Close()
|
||||
|
||||
// Detach detaches the currently running process from the agent.
|
||||
// The process continues running; returns false if nothing is executing.
|
||||
Detach() bool
|
||||
|
||||
// ListDetached returns info about all detached processes.
|
||||
ListDetached() []DetachedInfo
|
||||
|
||||
// SignalDetached sends a signal to a detached process by PID.
|
||||
SignalDetached(pid, signal int) error
|
||||
|
||||
// GetDetachedOutput returns the ring buffer output for a detached process.
|
||||
GetDetachedOutput(pid int) (string, error)
|
||||
|
||||
// DismissDetached removes an exited process from the list.
|
||||
DismissDetached(pid int) bool
|
||||
|
||||
// InjectSystemEvent appends a system-originated message to the session
|
||||
// context and emits it on the event bus. The agent sees it on its next
|
||||
// turn; it appears in the chat log immediately.
|
||||
InjectSystemEvent(content string)
|
||||
|
||||
// React records an emoji reaction to the most recent assistant response.
|
||||
React(emoji string)
|
||||
// ReactTo records an emoji reaction to a specific assistant response.
|
||||
ReactTo(responseID, emoji string) error
|
||||
// Reactions returns the current response ID to emoji mapping.
|
||||
Reactions() map[string]string
|
||||
}
|
||||
|
||||
// TaskState is a compact structured overlay that summarizes the agent's
|
||||
// current position in a task. Injected at the top of every turn so the
|
||||
// model doesn't rely on full history recall.
|
||||
type TaskState struct {
|
||||
Objective string `json:"objective"`
|
||||
PlanStep string `json:"plan_step"`
|
||||
Constraints []string `json:"constraints,omitempty"`
|
||||
LastAction string `json:"last_action"`
|
||||
OpenQuestions []string `json:"open_questions,omitempty"`
|
||||
NextDecision string `json:"next_decision"`
|
||||
}
|
||||
|
||||
// render produces a compact text representation for injection into the message history.
|
||||
func (ts *TaskState) render() string {
|
||||
if ts.Objective == "" {
|
||||
return ""
|
||||
}
|
||||
var sb strings.Builder
|
||||
sb.WriteString("[task state]\n")
|
||||
sb.WriteString("objective: " + ts.Objective + "\n")
|
||||
if ts.PlanStep != "" {
|
||||
sb.WriteString("plan_step: " + ts.PlanStep + "\n")
|
||||
}
|
||||
if len(ts.Constraints) > 0 {
|
||||
sb.WriteString("constraints: " + strings.Join(ts.Constraints, "; ") + "\n")
|
||||
}
|
||||
if ts.LastAction != "" {
|
||||
sb.WriteString("last_action: " + ts.LastAction + "\n")
|
||||
}
|
||||
if len(ts.OpenQuestions) > 0 {
|
||||
sb.WriteString("open_questions: " + strings.Join(ts.OpenQuestions, "; ") + "\n")
|
||||
}
|
||||
if ts.NextDecision != "" {
|
||||
sb.WriteString("next_decision: " + ts.NextDecision + "\n")
|
||||
}
|
||||
sb.WriteString("[/task state]")
|
||||
return sb.String()
|
||||
}
|
||||
|
||||
// saveTo writes the full message history to path as JSON.
|
||||
func (s *History) saveTo(path, id, agentName, cwd string) error {
|
||||
return s.saveToFull(path, id, agentName, "", "", cwd, "")
|
||||
}
|
||||
|
||||
// saveToFull writes session state including backend/model info.
|
||||
func sanitizeMessages(msgs []backend.Message) []backend.Message {
|
||||
out := make([]backend.Message, 0, len(msgs))
|
||||
for _, m := range msgs {
|
||||
for j, tc := range m.ToolCalls {
|
||||
if len(tc.Arguments) > 0 {
|
||||
var dummy json.RawMessage
|
||||
if err := json.Unmarshal(tc.Arguments, &dummy); err != nil {
|
||||
// Corrupted arguments — replace with empty object so save succeeds.
|
||||
m.ToolCalls[j].Arguments = json.RawMessage("{}")
|
||||
}
|
||||
}
|
||||
}
|
||||
out = append(out, m)
|
||||
}
|
||||
return out
|
||||
}
|
||||
|
||||
func (s *History) saveToFull(path, id, agentName, backendName, modelName, cwd, remote string) error {
|
||||
ps := PersistedSession{
|
||||
ID: id,
|
||||
Agent: agentName,
|
||||
Backend: backendName,
|
||||
Model: modelName,
|
||||
CWD: cwd,
|
||||
Remote: remote,
|
||||
Messages: sanitizeMessages(s.messages),
|
||||
TaskState: s.TaskState,
|
||||
TotalInputTokens: s.TotalInputTokens,
|
||||
TotalCachedInputTokens: s.TotalCachedInputTokens,
|
||||
TotalCacheCreationTokens: s.TotalCacheCreationTokens,
|
||||
TotalOutputTokens: s.TotalOutputTokens,
|
||||
TotalRequests: s.TotalRequests,
|
||||
Estimated: s.Estimated,
|
||||
LastTurnCostUSD: s.LastTurnCostUSD,
|
||||
SessionCostUSD: s.SessionCostUSD,
|
||||
PositiveReactions: s.PositiveReactions,
|
||||
NegativeReactions: s.NegativeReactions,
|
||||
Reactions: s.Reactions,
|
||||
}
|
||||
data, err := json.Marshal(ps)
|
||||
if err != nil {
|
||||
return fmt.Errorf("session save: %w", err)
|
||||
}
|
||||
// Atomic write: write to temp file, then rename to preserve last good snapshot.
|
||||
tmpPath := path + ".tmp"
|
||||
if err := os.WriteFile(tmpPath, data, 0600); err != nil {
|
||||
return fmt.Errorf("session save tmp: %w", err)
|
||||
}
|
||||
return os.Rename(tmpPath, path)
|
||||
}
|
||||
|
||||
// LoadPersistedSession reads a PersistedSession from a JSON file.
|
||||
func LoadPersistedSession(path string) (*PersistedSession, error) {
|
||||
data, err := os.ReadFile(path)
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
var ps PersistedSession
|
||||
if err := json.Unmarshal(data, &ps); err != nil {
|
||||
return nil, fmt.Errorf("parse session %s: %w", path, err)
|
||||
}
|
||||
return &ps, nil
|
||||
}
|
||||
|
||||
// RestoreHistory reconstructs a Session from a persisted message list.
|
||||
func RestoreHistory(ps *PersistedSession) *History {
|
||||
s := &History{
|
||||
messages: ps.Messages,
|
||||
TaskState: ps.TaskState,
|
||||
TotalInputTokens: ps.TotalInputTokens,
|
||||
TotalCachedInputTokens: ps.TotalCachedInputTokens,
|
||||
TotalCacheCreationTokens: ps.TotalCacheCreationTokens,
|
||||
TotalOutputTokens: ps.TotalOutputTokens,
|
||||
TotalRequests: ps.TotalRequests,
|
||||
Estimated: ps.Estimated,
|
||||
LastTurnCostUSD: ps.LastTurnCostUSD,
|
||||
SessionCostUSD: ps.SessionCostUSD,
|
||||
PositiveReactions: ps.PositiveReactions,
|
||||
NegativeReactions: ps.NegativeReactions,
|
||||
Reactions: ps.Reactions,
|
||||
}
|
||||
for i := range s.messages {
|
||||
if s.messages[i].Role == "assistant" && s.messages[i].ID == "" {
|
||||
s.messages[i].ID = NewResponseID()
|
||||
}
|
||||
}
|
||||
for _, m := range ps.Messages {
|
||||
if m.Role == "user" {
|
||||
s.goal = m.Content
|
||||
break
|
||||
}
|
||||
}
|
||||
return s
|
||||
}
|
||||
|
||||
// History is an ephemeral in-memory state backend.
|
||||
type History struct {
|
||||
goal string
|
||||
messages []backend.Message
|
||||
TaskState *TaskState
|
||||
// Cumulative usage tracking.
|
||||
TotalInputTokens int
|
||||
TotalCachedInputTokens int
|
||||
TotalCacheCreationTokens int
|
||||
TotalOutputTokens int
|
||||
TotalRequests int
|
||||
Estimated bool // true if any usage was estimated rather than reported by the backend
|
||||
LastTurnCostUSD float64 // cost of the most recently completed turn in USD
|
||||
SessionCostUSD float64 // cumulative cost of all turns in this session in USD
|
||||
// per-turn accumulators; reset at the start of each turn
|
||||
turnInputTokens int
|
||||
turnCachedTokens int
|
||||
turnCreationTokens int
|
||||
turnOutputTokens int
|
||||
turnCostUSD float64 // >0 when backend reported cost directly (e.g. OpenRouter)
|
||||
// Reaction tracking.
|
||||
PositiveReactions int
|
||||
NegativeReactions int
|
||||
Reactions []Reaction
|
||||
}
|
||||
|
||||
// newHistory creates a new empty Session. The caller is responsible for
|
||||
// appending the initial user message via appendUserMessage.
|
||||
func newHistory(goal string) *History {
|
||||
return &History{goal: goal}
|
||||
}
|
||||
|
||||
// Checkpoint forks the session: returns a new Session that inherits the given
|
||||
// TaskState but starts with a clean message history. This enables narrow-context
|
||||
// sub-agents that know what to do without inheriting all parent message noise.
|
||||
func (s *History) Checkpoint(ts TaskState) *History {
|
||||
child := &History{
|
||||
goal: ts.Objective,
|
||||
TaskState: &ts,
|
||||
}
|
||||
// Seed with a user message so the child has a valid initial turn.
|
||||
child.messages = []backend.Message{{
|
||||
Role: "user",
|
||||
Content: ts.render(),
|
||||
}}
|
||||
return child
|
||||
}
|
||||
|
||||
func (s *History) history() []backend.Message {
|
||||
if len(s.Reactions) == 0 {
|
||||
return s.messages
|
||||
}
|
||||
// Append feedback after the conversation so assistant tool-call messages stay
|
||||
// adjacent to their tool results, as required by provider APIs.
|
||||
out := make([]backend.Message, 0, len(s.messages)+len(s.Reactions))
|
||||
out = append(out, s.messages...)
|
||||
for _, r := range s.Reactions {
|
||||
_, desc, _ := classifyReaction(r.Emoji)
|
||||
out = append(out, backend.Message{Role: "user", Content: "[reaction to assistant response " + r.ResponseID + ": " + r.Category + " (" + r.Emoji + ")] " + desc})
|
||||
}
|
||||
return out
|
||||
}
|
||||
|
||||
func (s *History) taskState() *TaskState {
|
||||
return s.TaskState
|
||||
}
|
||||
|
||||
func (s *History) updateTaskState(ts TaskState) {
|
||||
s.TaskState = &ts
|
||||
}
|
||||
|
||||
func (s *History) addUsage(u backend.Usage, estimated bool) {
|
||||
s.TotalInputTokens += u.InputTokens
|
||||
s.TotalCachedInputTokens += u.CachedInputTokens
|
||||
s.TotalCacheCreationTokens += u.CacheCreationTokens
|
||||
s.TotalOutputTokens += u.OutputTokens
|
||||
s.TotalRequests++
|
||||
s.turnInputTokens += u.InputTokens
|
||||
s.turnCachedTokens += u.CachedInputTokens
|
||||
s.turnCreationTokens += u.CacheCreationTokens
|
||||
s.turnOutputTokens += u.OutputTokens
|
||||
s.turnCostUSD += u.CostUSD
|
||||
if estimated {
|
||||
s.Estimated = true
|
||||
}
|
||||
}
|
||||
|
||||
func (s *History) recomputeReactionCounts() {
|
||||
s.PositiveReactions = 0
|
||||
s.NegativeReactions = 0
|
||||
for _, r := range s.Reactions {
|
||||
switch r.Category {
|
||||
case "positive", "excellent":
|
||||
s.PositiveReactions++
|
||||
case "negative", "terrible":
|
||||
s.NegativeReactions++
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
func (s *History) resetTurnAccumulators() {
|
||||
s.turnInputTokens = 0
|
||||
s.turnCachedTokens = 0
|
||||
s.turnCreationTokens = 0
|
||||
s.turnOutputTokens = 0
|
||||
s.turnCostUSD = 0
|
||||
}
|
||||
|
||||
// recordTurnCost computes and stores the last turn's cost, adding it to the
|
||||
// session total. model is used for the pricing-table fallback when the backend
|
||||
// did not report cost directly.
|
||||
func (s *History) recordTurnCost(model string) {
|
||||
var cost float64
|
||||
if s.turnCostUSD > 0 {
|
||||
cost = s.turnCostUSD
|
||||
} else {
|
||||
cost = computeCostUSD(model, backend.Usage{
|
||||
InputTokens: s.turnInputTokens,
|
||||
CachedInputTokens: s.turnCachedTokens,
|
||||
CacheCreationTokens: s.turnCreationTokens,
|
||||
OutputTokens: s.turnOutputTokens,
|
||||
})
|
||||
}
|
||||
s.LastTurnCostUSD = cost
|
||||
s.SessionCostUSD += cost
|
||||
}
|
||||
|
||||
func (s *History) update(assistant backend.Message, results []toolResult) {
|
||||
s.messages = append(s.messages, assistant)
|
||||
for _, r := range results {
|
||||
s.messages = append(s.messages, backend.Message{
|
||||
Role: "tool",
|
||||
Content: r.Content,
|
||||
ContentBlocks: r.ContentBlocks,
|
||||
ToolCallID: r.ToolCallID,
|
||||
})
|
||||
}
|
||||
}
|
||||
|
||||
// llmSummarizeToolResult uses the LLM to produce a concise summary of a tool result.
|
||||
func llmSummarizeToolResult(ctx context.Context, b backend.Backend, name, content string) string {
|
||||
prompt := fmt.Sprintf("Summarize this tool result from %q in 1-3 sentences, preserving key facts, file paths, values, and findings:\n\n%s", name, content)
|
||||
ch, err := b.ChatStream(ctx, []backend.Message{{Role: "user", Content: prompt}}, nil, backend.GenerationParams{})
|
||||
if err != nil {
|
||||
return content
|
||||
}
|
||||
var sb strings.Builder
|
||||
for ev := range ch {
|
||||
if ev.Content != "" {
|
||||
sb.WriteString(ev.Content)
|
||||
}
|
||||
if ev.Done {
|
||||
break
|
||||
}
|
||||
}
|
||||
if s := strings.TrimSpace(sb.String()); s != "" {
|
||||
return fmt.Sprintf("[summary of %s result]: %s", name, s)
|
||||
}
|
||||
return content
|
||||
}
|
||||
|
||||
// removeCancelledToolResults filters out tool results that were cancelled due
|
||||
// to interrupt, keeping completed work. Also removes the corresponding tool
|
||||
// calls from assistant messages to maintain a valid message sequence.
|
||||
func (s *History) removeCancelledToolResults() {
|
||||
// First pass: collect cancelled tool call IDs
|
||||
cancelled := make(map[string]bool)
|
||||
for _, m := range s.messages {
|
||||
if m.Role == "tool" && isCancelledToolResult(m.Content) {
|
||||
cancelled[m.ToolCallID] = true
|
||||
}
|
||||
}
|
||||
if len(cancelled) == 0 {
|
||||
return
|
||||
}
|
||||
|
||||
// Second pass: filter messages and prune tool calls from assistant messages
|
||||
filtered := s.messages[:0]
|
||||
for _, m := range s.messages {
|
||||
if m.Role == "tool" && cancelled[m.ToolCallID] {
|
||||
continue
|
||||
}
|
||||
if m.Role == "assistant" && len(m.ToolCalls) > 0 {
|
||||
// Remove cancelled tool calls from this assistant message
|
||||
kept := m.ToolCalls[:0]
|
||||
for _, tc := range m.ToolCalls {
|
||||
if !cancelled[tc.ID] {
|
||||
kept = append(kept, tc)
|
||||
}
|
||||
}
|
||||
if len(kept) == 0 && m.Content == "" {
|
||||
// All tool calls cancelled and no text content - skip message
|
||||
continue
|
||||
}
|
||||
m.ToolCalls = kept
|
||||
}
|
||||
filtered = append(filtered, m)
|
||||
}
|
||||
s.messages = filtered
|
||||
}
|
||||
|
||||
func isCancelledToolResult(content string) bool {
|
||||
return strings.Contains(content, `"status":"cancelled"`) ||
|
||||
strings.Contains(content, "tool execution interrupted by user")
|
||||
}
|
||||
|
||||
// PreCompactionSnapshot returns a copy of the current messages for persistence
|
||||
// before compaction. Call this before compact().
|
||||
func (s *History) PreCompactionSnapshot() []backend.Message {
|
||||
return slices.Clone(s.messages)
|
||||
}
|
||||
|
||||
// Compact summarizes the conversation via an LLM call, replacing the history
|
||||
// with system messages + preserved user messages + a structured summary.
|
||||
// Returns (n compacted, summary text, error); n==0 means nothing to compact.
|
||||
func (s *History) compact(ctx context.Context, b backend.Backend) (int, string, error) {
|
||||
if len(s.messages) <= hotTailSize+warmIndexSize {
|
||||
return 0, "", nil
|
||||
}
|
||||
|
||||
// Flatten tool calls into plain text so the compaction request
|
||||
// doesn't need tool schemas.
|
||||
flattened := flattenToolMessages(s.messages)
|
||||
flattened = append(flattened, backend.Message{
|
||||
Role: "user",
|
||||
Content: compactionPrompt,
|
||||
})
|
||||
|
||||
ch, err := b.ChatStream(ctx, flattened, nil, backend.GenerationParams{})
|
||||
if err != nil {
|
||||
return 0, "", fmt.Errorf("compact: %w", err)
|
||||
}
|
||||
|
||||
var summary strings.Builder
|
||||
for ev := range ch {
|
||||
if ev.Content != "" {
|
||||
summary.WriteString(ev.Content)
|
||||
}
|
||||
if ev.Done {
|
||||
break
|
||||
}
|
||||
}
|
||||
|
||||
summaryText := strings.TrimSpace(summary.String())
|
||||
if summaryText == "" {
|
||||
return 0, "", fmt.Errorf("compact: empty summary")
|
||||
}
|
||||
|
||||
// Parse structured task state from the model's JSON response.
|
||||
var ts TaskState
|
||||
raw := extractJSON(summaryText)
|
||||
if err := json.Unmarshal([]byte(raw), &ts); err != nil {
|
||||
// Fallback: treat as unstructured summary if JSON parse fails.
|
||||
ts = TaskState{Objective: summaryText}
|
||||
}
|
||||
s.TaskState = &ts
|
||||
|
||||
beforeCount := len(s.messages)
|
||||
s.messages = buildCompactedHistory(ts, s.messages)
|
||||
return beforeCount - len(s.messages), summaryText, nil
|
||||
}
|
||||
|
||||
// buildCompactedHistory constructs a three-zone post-compaction message list:
|
||||
// - Cold: structured task state summary (everything older than warm+hot)
|
||||
// - Warm: brief index of recent decisions/actions (warmIndexSize messages)
|
||||
// - Hot: last hotTailSize messages verbatim
|
||||
func buildCompactedHistory(ts TaskState, allMessages []backend.Message) []backend.Message {
|
||||
total := len(allMessages)
|
||||
|
||||
// Determine zone boundaries.
|
||||
hotStart := total - hotTailSize
|
||||
if hotStart < 0 {
|
||||
hotStart = 0
|
||||
}
|
||||
|
||||
// If hotStart lands on a tool message, walk backward to include the
|
||||
// preceding assistant message with tool_calls. Without this, the hot
|
||||
// zone starts with an orphaned tool result which violates the OpenAI
|
||||
// API constraint that tool messages must follow an assistant message
|
||||
// containing the corresponding tool_calls.
|
||||
for hotStart > 0 && allMessages[hotStart].Role == "tool" {
|
||||
hotStart--
|
||||
}
|
||||
|
||||
warmStart := hotStart - warmIndexSize
|
||||
if warmStart < 0 {
|
||||
warmStart = 0
|
||||
}
|
||||
|
||||
// Cold zone: structured task state.
|
||||
stateJSON, _ := json.Marshal(ts)
|
||||
cold := backend.Message{
|
||||
Role: "user",
|
||||
Content: "[compacted context — cold zone]\n```json\n" + string(stateJSON) + "\n```",
|
||||
}
|
||||
|
||||
var out []backend.Message
|
||||
out = append(out, cold)
|
||||
|
||||
// Warm zone: one-line summaries of decisions in the warm window.
|
||||
warmSlice := allMessages[warmStart:hotStart]
|
||||
if len(warmSlice) > 0 {
|
||||
var sb strings.Builder
|
||||
sb.WriteString("[warm zone — recent decisions]\n")
|
||||
for _, m := range warmSlice {
|
||||
line := summarizeMessage(m)
|
||||
if line != "" {
|
||||
sb.WriteString("- ")
|
||||
sb.WriteString(line)
|
||||
sb.WriteByte('\n')
|
||||
}
|
||||
}
|
||||
out = append(out, backend.Message{Role: "user", Content: sb.String()})
|
||||
}
|
||||
|
||||
// Hot zone: verbatim recent messages.
|
||||
out = append(out, allMessages[hotStart:]...)
|
||||
return out
|
||||
}
|
||||
|
||||
// summarizeMessage produces a one-line summary of a message for the warm index.
|
||||
func summarizeMessage(m backend.Message) string {
|
||||
switch m.Role {
|
||||
case "assistant":
|
||||
if len(m.ToolCalls) > 0 {
|
||||
names := make([]string, len(m.ToolCalls))
|
||||
for i, tc := range m.ToolCalls {
|
||||
names[i] = tc.Name
|
||||
}
|
||||
return "called: " + strings.Join(names, ", ")
|
||||
}
|
||||
text := m.Content
|
||||
if len(text) > 120 {
|
||||
text = text[:120] + "…"
|
||||
}
|
||||
return "said: " + text
|
||||
case "user":
|
||||
text := m.Content
|
||||
if len(text) > 120 {
|
||||
text = text[:120] + "…"
|
||||
}
|
||||
return "user: " + text
|
||||
case "tool":
|
||||
text := m.Content
|
||||
if len(text) > 80 {
|
||||
text = text[:80] + "…"
|
||||
}
|
||||
return "result(" + m.ToolCallID + "): " + text
|
||||
}
|
||||
return ""
|
||||
}
|
||||
|
||||
// extractJSON finds the first JSON object in s (handling markdown fences).
|
||||
func extractJSON(s string) string {
|
||||
// Strip markdown code fence if present.
|
||||
if i := strings.Index(s, "```json"); i >= 0 {
|
||||
s = s[i+7:]
|
||||
if j := strings.Index(s, "```"); j >= 0 {
|
||||
s = s[:j]
|
||||
}
|
||||
} else if i := strings.Index(s, "```"); i >= 0 {
|
||||
s = s[i+3:]
|
||||
if j := strings.Index(s, "```"); j >= 0 {
|
||||
s = s[:j]
|
||||
}
|
||||
}
|
||||
// Find first { ... }
|
||||
start := strings.Index(s, "{")
|
||||
if start < 0 {
|
||||
return s
|
||||
}
|
||||
end := strings.LastIndex(s, "}")
|
||||
if end < start {
|
||||
return s
|
||||
}
|
||||
return s[start : end+1]
|
||||
}
|
||||
|
||||
// flattenToolMessages converts tool call/result sequences into plain text
|
||||
// so the compaction request doesn't include tool-specific structures that
|
||||
// the API may reject when no tools are defined.
|
||||
func flattenToolMessages(messages []backend.Message) []backend.Message {
|
||||
out := make([]backend.Message, 0, len(messages))
|
||||
for _, m := range messages {
|
||||
switch {
|
||||
case m.Role == "assistant" && len(m.ToolCalls) > 0:
|
||||
var sb strings.Builder
|
||||
if m.Content != "" {
|
||||
sb.WriteString(m.Content)
|
||||
sb.WriteString("\n\n")
|
||||
}
|
||||
for _, tc := range m.ToolCalls {
|
||||
fmt.Fprintf(&sb, "[Tool call: %s(%s)]\n", tc.Name, string(tc.Arguments))
|
||||
}
|
||||
out = append(out, backend.Message{Role: "assistant", Content: sb.String()})
|
||||
case m.Role == "tool":
|
||||
text := m.Content
|
||||
if len(text) > 4000 {
|
||||
text = text[:4000] + "..."
|
||||
}
|
||||
out = append(out, backend.Message{
|
||||
Role: "user",
|
||||
Content: fmt.Sprintf("[Tool result for %s]:\n%s", m.ToolCallID, text),
|
||||
})
|
||||
default:
|
||||
out = append(out, m)
|
||||
}
|
||||
}
|
||||
return out
|
||||
}
|
||||
|
||||
func (s *History) appendUserMessage(content string) {
|
||||
s.messages = append(s.messages, backend.Message{Role: "user", Content: content})
|
||||
}
|
||||
|
||||
// cloneMessages returns a deep copy of the message slice.
|
||||
func cloneMessages(msgs []backend.Message) []backend.Message {
|
||||
out := make([]backend.Message, len(msgs))
|
||||
for i, m := range msgs {
|
||||
out[i] = backend.Message{
|
||||
Role: m.Role,
|
||||
Content: m.Content,
|
||||
ToolCallID: m.ToolCallID,
|
||||
}
|
||||
if len(m.ToolCalls) > 0 {
|
||||
out[i].ToolCalls = make([]backend.ToolCall, len(m.ToolCalls))
|
||||
for j, tc := range m.ToolCalls {
|
||||
out[i].ToolCalls[j] = backend.ToolCall{
|
||||
ID: tc.ID,
|
||||
Name: tc.Name,
|
||||
Arguments: append(json.RawMessage(nil), tc.Arguments...),
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
return out
|
||||
}
|
||||
|
||||
// estimateTokens returns a rough token count (~4 chars per token).
|
||||
func (s *History) estimateTokens() int {
|
||||
chars := 0
|
||||
for _, m := range s.messages {
|
||||
chars += len(m.Content)
|
||||
for _, tc := range m.ToolCalls {
|
||||
chars += len(tc.Name) + len(tc.Arguments)
|
||||
}
|
||||
}
|
||||
return chars / 4
|
||||
}
|
||||
|
||||
// stripColdResults summarizes large tool-result messages outside the hot tail
|
||||
// using an LLM call. Messages in the last hotTailSize slots are left verbatim.
|
||||
func (s *History) stripColdResults(ctx context.Context, b backend.Backend) {
|
||||
hot := len(s.messages) - hotTailSize
|
||||
for i := range s.messages {
|
||||
if i >= hot {
|
||||
break
|
||||
}
|
||||
m := &s.messages[i]
|
||||
if m.Role == "tool" && len(m.Content) > 200 {
|
||||
m.Content = llmSummarizeToolResult(ctx, b, m.ToolCallID, m.Content)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// stripCold implements the state interface.
|
||||
func (s *History) stripCold(ctx context.Context, b backend.Backend) {
|
||||
s.stripColdResults(ctx, b)
|
||||
}
|
||||
|
||||
// contextDebug returns a multi-line breakdown of the history.
|
||||
func (s *History) contextDebug() string {
|
||||
var sb strings.Builder
|
||||
sb.WriteString(fmt.Sprintf("=== %d messages ===\n", len(s.messages)))
|
||||
for i, m := range s.messages {
|
||||
preview := m.Content
|
||||
if len(preview) > 80 {
|
||||
preview = preview[:80] + "..."
|
||||
}
|
||||
chars := len(m.Content)
|
||||
for _, tc := range m.ToolCalls {
|
||||
chars += len(tc.Name) + len(tc.Arguments)
|
||||
}
|
||||
sb.WriteString(fmt.Sprintf(" [%d] role=%-10s chars=%-6d %q\n", i, m.Role, chars, preview))
|
||||
}
|
||||
return sb.String()
|
||||
// DetachedInfo describes a detached process for external consumers.
|
||||
type DetachedInfo struct {
|
||||
PID int
|
||||
Command string
|
||||
Started int64 // unix timestamp
|
||||
Exited bool
|
||||
ExitCode int
|
||||
}
|
||||
|
|
|
|||
187
session/types.go
187
session/types.go
|
|
@ -1,187 +0,0 @@
|
|||
package session
|
||||
|
||||
import (
|
||||
"context"
|
||||
"errors"
|
||||
|
||||
"github.com/simonfxr/pubsub"
|
||||
"ollie/backend"
|
||||
)
|
||||
|
||||
// WatchField names supported by Core.WaitChange.
|
||||
const (
|
||||
WatchState = "state"
|
||||
WatchUsage = "usage"
|
||||
WatchCtxSz = "ctxsz"
|
||||
WatchCWD = "cwd"
|
||||
WatchAgent = "agent"
|
||||
)
|
||||
|
||||
// ErrInterrupted is returned when the user cancels an agent turn (Ctrl-C).
|
||||
var ErrInterrupted = errors.New("interrupted")
|
||||
|
||||
// Event is a typed output event emitted during an agent turn or in response
|
||||
// to a command.
|
||||
type Event struct {
|
||||
Role string
|
||||
Name string
|
||||
Content string
|
||||
ResponseID string
|
||||
}
|
||||
|
||||
// EventHandler receives events from the agent.
|
||||
type EventHandler func(Event)
|
||||
|
||||
// Core is the interface between a frontend (TUI, HTTP handler, etc.) and the
|
||||
// agent engine. All output from the agent is delivered via the event bus.
|
||||
type Core interface {
|
||||
// Submit processes one line of user input. Slash commands and shell
|
||||
// shortcuts are dispatched synchronously; any other input starts an agent
|
||||
// turn that publishes events to the bus until the turn is complete.
|
||||
// After the turn, any queued prompts are drained sequentially.
|
||||
Submit(ctx context.Context, input string)
|
||||
|
||||
// Interrupt cancels the current in-progress agent turn.
|
||||
// Returns true if an action was running and was cancelled.
|
||||
Interrupt(cause error) bool
|
||||
|
||||
// Inject sends a message that will be appended to the next tool result
|
||||
// as a user interruption. If no turn is running, it is silently dropped.
|
||||
Inject(prompt string)
|
||||
|
||||
// Queue pushes a prompt onto the FIFO for execution after the current
|
||||
// turn completes.
|
||||
Queue(prompt string)
|
||||
|
||||
// Bus returns the session event bus.
|
||||
Bus() *pubsub.Bus
|
||||
|
||||
// PopQueue removes and returns the next queued prompt.
|
||||
// Returns ("", false) if the queue is empty.
|
||||
PopQueue() (string, bool)
|
||||
|
||||
// IsRunning returns true if an agent turn is currently in progress.
|
||||
IsRunning() bool
|
||||
|
||||
// State returns the current agent state: "idle", "thinking", or "calling: <tool>".
|
||||
State() string
|
||||
|
||||
// Reply returns the assistant text from the most recently completed turn.
|
||||
// Cleared when a new prompt is submitted.
|
||||
Reply() string
|
||||
|
||||
// AgentName returns the name of the active agent.
|
||||
AgentName() string
|
||||
|
||||
// BackendName returns the name of the active backend (e.g. "anthropic", "ollama").
|
||||
BackendName() string
|
||||
|
||||
// ModelName returns the name of the active model.
|
||||
ModelName() string
|
||||
|
||||
// CtxSz returns the estimated context size as a one-line summary.
|
||||
CtxSz() string
|
||||
|
||||
// Usage returns billed token counts as a one-line summary.
|
||||
Usage() string
|
||||
|
||||
// Cost returns per-turn and session cost as a two-line key=value summary.
|
||||
Cost() string
|
||||
|
||||
// ListModels returns available model names, one per line.
|
||||
ListModels() string
|
||||
|
||||
// CWD returns the current working directory used for tool execution.
|
||||
CWD() string
|
||||
|
||||
// SetCWD changes the working directory for tool execution and
|
||||
// updates the system prompt. Returns an error if the path does not exist.
|
||||
SetCWD(dir string) error
|
||||
|
||||
// SetSessionID renames the session: updates the in-memory ID, renames
|
||||
// persisted files on disk, and propagates to the execute server env.
|
||||
SetSessionID(newID string) error
|
||||
|
||||
// Context returns the current message history as it would be sent to the
|
||||
// backend: system prompt prepended, stale reads pruned. Does not include
|
||||
// tool definitions (see Tools) or generation params (see GenerationParams).
|
||||
Context() []backend.Message
|
||||
|
||||
// SystemPrompt returns the fully rendered system prompt for this session.
|
||||
SystemPrompt() string
|
||||
|
||||
// GenerationParams returns the current sampling parameters.
|
||||
GenerationParams() backend.GenerationParams
|
||||
|
||||
// SetGenerationParams replaces the current sampling parameters.
|
||||
// Returns an error if the agent is currently running.
|
||||
SetGenerationParams(params backend.GenerationParams) error
|
||||
|
||||
// CompactionModel returns the model override used for context compaction.
|
||||
CompactionModel() string
|
||||
|
||||
// SetCompactionModel changes the model used for context compaction.
|
||||
SetCompactionModel(model string)
|
||||
|
||||
// SetEnv injects a session-scoped environment variable into the shell
|
||||
// subprocesses. Does not affect the daemon process environment.
|
||||
SetEnv(key, value string)
|
||||
|
||||
// WaitChange blocks until the named field changes from current, then returns
|
||||
// the new value. Returns ("", false) if ctx is cancelled before a change.
|
||||
// Supported fields: WatchState, WatchUsage, WatchCtxSz, WatchCWD.
|
||||
WaitChange(ctx context.Context, field, current string) (string, bool)
|
||||
|
||||
// ToolCallCount returns the total number of tool calls executed in this
|
||||
// session since the agent was created. The counter is monotonically
|
||||
// increasing and never resets. Blocked calls (pre-tool hook exit 2) are
|
||||
// not counted. Timed-out or erroring calls are counted because execution
|
||||
// was attempted. Use modulo arithmetic in hooks to fire every N calls:
|
||||
// [ $(($(cat tcct) % 10)) -eq 0 ] && ...
|
||||
ToolCallCount() int64
|
||||
|
||||
// SaveSession writes the current session state to the given path.
|
||||
// The file includes all messages, agent/backend/model metadata, and task state.
|
||||
SaveSession(path string) error
|
||||
|
||||
// Close releases resources associated with the session, including its
|
||||
// temporary directory under /tmp/ollie/.
|
||||
Close()
|
||||
|
||||
// Detach detaches the currently running process from the agent.
|
||||
// The process continues running; returns false if nothing is executing.
|
||||
Detach() bool
|
||||
|
||||
// ListDetached returns info about all detached processes.
|
||||
ListDetached() []DetachedInfo
|
||||
|
||||
// SignalDetached sends a signal to a detached process by PID.
|
||||
SignalDetached(pid, signal int) error
|
||||
|
||||
// GetDetachedOutput returns the ring buffer output for a detached process.
|
||||
GetDetachedOutput(pid int) (string, error)
|
||||
|
||||
// DismissDetached removes an exited process from the list.
|
||||
DismissDetached(pid int) bool
|
||||
|
||||
// InjectSystemEvent appends a system-originated message to the session
|
||||
// context and emits it on the event bus. The agent sees it on its next
|
||||
// turn; it appears in the chat log immediately.
|
||||
InjectSystemEvent(content string)
|
||||
|
||||
// React records an emoji reaction to the most recent assistant response.
|
||||
React(emoji string)
|
||||
// ReactTo records an emoji reaction to a specific assistant response.
|
||||
ReactTo(responseID, emoji string) error
|
||||
// Reactions returns the current response ID to emoji mapping.
|
||||
Reactions() map[string]string
|
||||
}
|
||||
|
||||
// DetachedInfo describes a detached process for external consumers.
|
||||
type DetachedInfo struct {
|
||||
PID int
|
||||
Command string
|
||||
Started int64 // unix timestamp
|
||||
Exited bool
|
||||
ExitCode int
|
||||
}
|
||||
Reference in New Issue