This repository has been archived on 2026-08-16. You can view files and clone it, but cannot push or open issues or pull requests.
ollie-core/agent/core.go

1590 lines
45 KiB
Go

package agent
import (
"context"
"crypto/rand"
"encoding/json"
"errors"
"fmt"
"io"
"os"
"path/filepath"
"runtime/debug"
"slices"
"strconv"
"strings"
"sync"
"sync/atomic"
"syscall"
"time"
"github.com/simonfxr/pubsub"
"ollie/backend"
olog "ollie/log"
"ollie/paths"
"ollie/tools"
)
// toolClassifier reports whether a named tool is safe to run concurrently
// with other read-class tools. nil means treat all tools as serial.
type toolClassifier func(name string) bool
// BuildRuntime constructs a Runtime from a pre-configured Dispatcher and
// optional agent config. cwd sets the working directory reported in the
// system prompt; if empty, the process working directory is used.
// 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 {
var messages []string
var allToolInfos []tools.ToolInfo
var allTools []backend.Tool
serverOf := make(map[string]string)
if cfg == nil || cfg.ToolsEnabled() {
var listErr error
allToolInfos, listErr = d.ListTools()
if listErr != nil {
messages = append(messages, fmt.Sprintf("list tools: %v", listErr))
}
for _, t := range allToolInfos {
serverOf[t.Name] = t.Server
}
// Only built-in executors (with InputSchema) become backend tools.
// Named tool scripts are promoted via the tool registry and appear in the preamble.
allTools = toolInfosToBackend(allToolInfos)
// Append named tool scripts for preamble listing only.
allToolInfos = append(allToolInfos, tools.DiscoverTools()...)
}
hooks := Hooks{}
var preamble string
var genParams backend.GenerationParams
var maxSteps int
if cfg != nil {
for k, v := range cfg.Hooks {
hooks[k] = []string(v)
}
if resolved, err := resolvePrompt(cfg.Prompt, cwd, env); err != nil {
fmt.Fprintf(os.Stderr, "resolve prompt: %v\n", err)
} else {
preamble = resolved
}
genParams = backend.GenerationParams{
MaxTokens: cfg.MaxTokens,
MaxCompletionTokens: cfg.MaxCompletionTokens,
Temperature: cfg.Temperature,
TopP: cfg.TopP,
TopK: cfg.TopK,
MinP: cfg.MinP,
TopA: cfg.TopA,
FrequencyPenalty: cfg.FrequencyPenalty,
PresencePenalty: cfg.PresencePenalty,
RepetitionPenalty: cfg.RepetitionPenalty,
ThinkingBudget: cfg.Reasoning,
ReasoningEffort: cfg.ReasoningEffort,
IncludeReasoning: cfg.IncludeReasoning,
ResponseFormat: cfg.ResponseFormat,
Stop: cfg.Stop,
Verbosity: cfg.Verbosity,
}
maxSteps = cfg.MaxSteps
if len(cfg.AllowTools) > 0 {
if srv, ok := d.GetServer("execute"); ok {
if rs, ok := srv.(tools.ToolRestrictionSetter); ok {
rs.SetAllowTools(cfg.AllowTools)
}
}
}
}
exec := func(ctx context.Context, name string, args json.RawMessage) (string, []backend.ContentBlock, error) {
// Dynamically resolve the server for this tool — handles
// lazy-promoted tools that appeared after BuildRuntime.
infos, listErr := d.ListTools()
if listErr != nil {
return "", nil, listErr
}
server := ""
for _, t := range infos {
if t.Name == name {
server = t.Server
break
}
}
if server == "" {
return "", nil, fmt.Errorf("unknown tool: %s", name)
}
raw, err := d.Dispatch(ctx, server, name, args)
if err != nil {
return "", nil, err
}
text, blocks, isErr := extractToolResult(raw)
if isErr {
return "", nil, fmt.Errorf("%s", text)
}
return text, blocks, nil
}
var classify toolClassifier
if srv, ok := d.GetServer("execute"); ok {
if pc, ok := srv.(tools.ParallelClassifier); ok {
classify = pc.IsParallelRead
}
}
var tierFn func(string, json.RawMessage) ResultTier
if srv, ok := d.GetServer("execute"); ok {
if tc, ok := srv.(tools.TierClassifier); ok {
tierFn = func(name string, args json.RawMessage) ResultTier {
switch tc.ResultTierArgs(name, args) {
case "cold":
return TierCold
case "warm":
return TierWarm
default:
return TierHot
}
}
}
}
var backendName, modelName, compactionModel string
if cfg != nil {
backendName = cfg.Backend
modelName = cfg.Model
compactionModel = cfg.CompactionModel
}
// Prepend base layers (system prompt, operational model, environment)
// before the agent-specific preamble.
if len(baseLayers) > 0 {
var prefix strings.Builder
for _, layer := range baseLayers {
if layer != "" {
prefix.WriteString(layer)
prefix.WriteByte('\n')
}
}
if prefix.Len() > 0 {
preamble = prefix.String() + preamble
}
}
// Append compact tool surface listing (name + description).
// Full prompts are available on-demand via /tools write (net/dns pattern).
var toolListing strings.Builder
for _, ti := range allToolInfos {
if ti.Description != "" && ti.Server == "" {
// Only named tool scripts (Server==""), not built-in executors
fmt.Fprintf(&toolListing, "- **%s** — %s\n", ti.Name, ti.Description)
}
}
if toolListing.Len() > 0 {
preamble += "\n# Available Tools\n\n" + toolListing.String()
}
return &Runtime{
Dispatcher: d,
Tools: allTools,
Exec: exec,
ClassifyTool: classify,
ClassifyTier: tierFn,
Hooks: hooks,
Preamble: preamble,
GenParams: genParams,
MaxSteps: maxSteps,
CfgBackend: backendName,
CfgModel: modelName,
CompactionModel: compactionModel,
Messages: messages,
}
}
// DefaultPromptsDir returns the default directory for prompt templates.
func DefaultPromptsDir() string {
return paths.CfgDir() + "/prompts"
}
// PromptsDirs returns all prompt directories from OLLIE_PROMPTS_PATH (colon-separated).
func PromptsDirs() []string {
if p := os.Getenv("OLLIE_PROMPTS_PATH"); p != "" {
return strings.Split(p, ":")
}
return []string{DefaultPromptsDir()}
}
// AgentsDirs returns all agent directories from OLLIE_AGENTS_PATH (colon-separated).
func AgentsDirs() []string {
if p := os.Getenv("OLLIE_AGENTS_PATH"); p != "" {
return strings.Split(p, ":")
}
return []string{paths.CfgDir() + "/agents"}
}
// AgentConfigPath resolves the config file path for a named agent.
// It searches all directories in OLLIE_AGENTS_PATH, falling back to agentsDir.
func AgentConfigPath(agentsDir, name string) string {
for _, dir := range AgentsDirs() {
p := dir + "/" + name + ".json"
if _, err := os.Stat(p); err == nil {
return p
}
}
return agentsDir + "/" + name + ".json"
}
// NewSessionID generates a unique, lexicographically sortable session identifier.
// Format: <unix-nanoseconds>-<random-hex> — sortable by creation time, unique
// even if two sessions are created within the same nanosecond.
func NewSessionID() string {
b := make([]byte, 3)
rand.Read(b) //nolint:errcheck
return strconv.FormatInt(time.Now().UnixNano(), 10) + "-" + fmt.Sprintf("%06x", b)
}
// NewResponseID generates a unique identifier for a single assistant response.
func NewResponseID() string {
b := make([]byte, 3)
rand.Read(b) //nolint:errcheck
return "resp_" + strconv.FormatInt(time.Now().UnixNano(), 10) + "_" + fmt.Sprintf("%06x", b)
}
// NewReactionID generates a unique identifier for a user reaction.
func NewReactionID() string {
b := make([]byte, 3)
rand.Read(b) //nolint:errcheck
return "react_" + strconv.FormatInt(time.Now().UnixNano(), 10) + "_" + fmt.Sprintf("%06x", b)
}
// infoEvent wraps a plain-text message as an info Event.
func infoEvent(text string) Event {
return Event{Role: "info", Content: text + "\n"}
}
// actionHandle holds the cancel function for the current agent turn.
type actionHandle struct {
cancel context.CancelCauseFunc
}
// AgentCoreConfig is the configuration for creating an agent.
type AgentCoreConfig struct {
Backend backend.Backend
ModelName string // if non-empty, overrides backend's default model
AgentName string
AgentsDir string
SessionsDir string
SessionID string
Uname string // immutable user principal for 9P identity
CWD string // working directory for tool execution and system prompt
Session *Session
Runtime *Runtime
NewDispatcher func() tools.Dispatcher
NewBackend func(string) (backend.Backend, error) // if nil, defaults to backend.NewWithName
Log *olog.Logger // if nil, logging is disabled
// MaxSteps overrides the agent JSON maxSteps when non-zero.
// 0 means use the value from the agent config (or unlimited if absent).
MaxSteps int
// ReadPlanStep, if non-nil, is called to get the next unchecked plan step.
// When provided, it is used directly instead of reading from the filesystem.
ReadPlanStep func() string
// ListHandlers provides harness-specific list implementations for slash
// commands such as /skills, /tools, /agents. Keyed by command name without
// the leading slash. Commands with no entry emit nothing.
ListHandlers map[string]func() []string
// PromptEnvExtra holds PRIME_* env vars for prompt resolution.
// For remote sessions these come from HostInfo; for local, from PromptEnv().
// Stored so /agent reloads use the correct values.
PromptEnvExtra []string
// Remote is the SSH target for remote execution (empty = local).
Remote string
// BaseLayers are the system prompt, operational model, and environment
// block prepended to the agent preamble. Stored so /agent reloads
// preserve the server-injected context.
BaseLayers []string
}
// agent is the Core implementation. It owns all agent and session state
// but has no knowledge of how output is rendered.
type agent struct {
session *Session
runtime *Runtime
cfg agentConfig // per-turn config built from runtime; set in executeTurn
log *olog.Logger
agentName string
agentsDir string
sessionsDir string
sessionID string
uname string
newDispatcher func() tools.Dispatcher
newBackend func(string) (backend.Backend, error)
readPlanStep func() string
listHandlers map[string]func() []string
turnError func(ctx context.Context, errType, errMsg string) HookResult // overridable for tests
cwd string
remote string // SSH target for remote execution
promptEnvExtra []string // PRIME_* vars for prompt resolution
baseLayers []string // system prompt, operational model, env block — for /agent reloads
startupMessages []string
currentAction atomic.Pointer[actionHandle]
toolCallCount atomic.Int64
fifo PromptFIFO
bus *pubsub.Bus
pendingInject atomic.Pointer[string]
mu sync.RWMutex
state string // "idle", "thinking", "calling: <tool>"
reply string // assistant text from the most recently completed turn
envMu sync.RWMutex
env map[string]string // session-scoped env vars
changeMu sync.Mutex
changeCond *sync.Cond
submitMu sync.Mutex // serializes Submit calls (commands + turns)
warnedContext bool // true after a context-usage warning; cleared on compaction
auditLog *olog.Logger
// Debounced session persistence
saveMu sync.Mutex
saveDirty bool
saveTimer *time.Timer
// Cross-turn tool result cache (read-safe tools only)
resultCache sync.Map
}
// ToolCallCount returns the total number of tool calls executed in this
// session. The counter is monotonically increasing and never resets.
// Blocked calls (pre-tool hook exit 2) are not counted.
func (s *agent) ToolCallCount() int64 {
return s.toolCallCount.Load()
}
// SetEnv stores a session-scoped variable and propagates it to the execute server.
func (s *agent) SetEnv(key, value string) {
s.envMu.Lock()
if s.env == nil {
s.env = make(map[string]string)
}
s.env[key] = value
s.envMu.Unlock()
if s.runtime == nil || s.runtime.Dispatcher == nil {
return
}
if srv, ok := s.runtime.Dispatcher.GetServer("execute"); ok {
if es, ok := srv.(tools.EnvSetter); ok {
es.SetEnv(key, value)
}
}
}
// pushSessionEnv injects OLLIE_SESSION_ID into the execute server subprocess env.
func (s *agent) pushSessionEnv() {
if s.runtime == nil || s.runtime.Dispatcher == nil || s.sessionID == "" {
return
}
if srv, ok := s.runtime.Dispatcher.GetServer("execute"); ok {
if es, ok := srv.(tools.EnvSetter); ok {
es.SetEnv("OLLIE_SESSION_ID", s.sessionID)
if s.uname != "" {
es.SetEnv("OLLIE_UNAME", s.uname)
}
}
}
}
// pushLockDir sets the flock directory on the execute server to the session tmpdir.
func (s *agent) pushLockDir() {
if s.runtime == nil || s.runtime.Dispatcher == nil || s.sessionID == "" {
return
}
}
var _ Core = (*agent)(nil) // compile-time interface check
var sweepTmpOnce sync.Once
// ollieTmpDir returns the base temp directory for session tmpdirs.
func ollieTmpDir() string {
if p := os.Getenv("OLLIE_TMP_PATH"); p != "" {
return p
}
return filepath.Join(os.TempDir(), "ollie")
}
// readNextUnchecked reads a plan file and returns the first unchecked step.
func readNextUnchecked(path string) string {
data, err := os.ReadFile(path)
if err != nil || len(data) == 0 {
return ""
}
return NextUncheckedStep(data)
}
// NextUncheckedStep returns the first unchecked step from plan bytes.
func NextUncheckedStep(data []byte) string {
for _, line := range strings.Split(string(data), "\n") {
trimmed := strings.TrimSpace(line)
if strings.HasPrefix(trimmed, "- [ ]") {
return strings.TrimSpace(trimmed[5:])
}
}
return ""
}
// sweepStaleTmpDirs removes session tmpdirs left by a previous crash.
// All session tmpdirs are owned by a single process, so anything
// present at startup is stale.
func sweepStaleTmpDirs() {
sweepTmpOnce.Do(func() {
base := ollieTmpDir()
os.RemoveAll(base) //nolint:errcheck
os.MkdirAll(base, 0700) //nolint:errcheck
})
}
// NewAgentCore creates an agent from the given configuration.
func NewAgentCore(cfg AgentCoreConfig) Core {
sweepStaleTmpDirs()
if cfg.ModelName != "" {
cfg.Backend.SetModel(cfg.ModelName)
}
if cfg.NewBackend == nil {
cfg.NewBackend = backend.NewWithName
}
rt := cfg.Runtime
if rt == nil {
rt = &Runtime{}
}
// 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
// value loaded from the agent JSON.
if cfg.MaxSteps > 0 {
rt.MaxSteps = cfg.MaxSteps
}
var readPlanStep func() string
if cfg.SessionID != "" {
os.MkdirAll(filepath.Join(ollieTmpDir(), cfg.SessionID), 0700) //nolint:errcheck
}
if cfg.ReadPlanStep != nil {
readPlanStep = cfg.ReadPlanStep
}
log := cfg.Log
if log == nil {
log = olog.NewWriter("core", olog.LevelError+1, io.Discard, io.Discard)
}
a := &agent{
session: cfg.Session,
runtime: rt,
log: log,
auditLog: log.Sub("audit"),
agentName: cfg.AgentName,
agentsDir: cfg.AgentsDir,
sessionsDir: cfg.SessionsDir,
sessionID: cfg.SessionID,
uname: cfg.Uname,
cwd: paths.ExpandHome(cfg.CWD),
remote: cfg.Remote,
promptEnvExtra: cfg.PromptEnvExtra,
baseLayers: cfg.BaseLayers,
startupMessages: rt.Messages,
newDispatcher: cfg.NewDispatcher,
newBackend: cfg.NewBackend,
readPlanStep: readPlanStep,
listHandlers: cfg.ListHandlers,
state: "idle",
bus: pubsub.NewBus(),
}
a.changeCond = sync.NewCond(&a.changeMu)
a.turnError = func(_ context.Context, errType, errMsg string) HookResult {
hookCtx, cancel := context.WithTimeout(context.Background(), time.Duration(hookTimeout)*time.Second)
defer cancel()
return a.runtime.Hooks.Run(hookCtx, HookTurnError, map[string]string{
"session_id": a.sessionID,
"cwd": a.CWD(),
"model": a.runtime.Backend.Model(),
"error_type": errType,
"error": errMsg,
}, a.log)
}
a.pushSessionEnv()
a.pushLockDir()
return a
}
// Close releases resources for this session, including its tmpdir.
func (s *agent) Close() {
s.log.Debug("Close() session=%q", s.sessionID)
s.flushSave()
if s.runtime != nil && s.runtime.Dispatcher != nil {
if srv, ok := s.runtime.Dispatcher.GetServer("execute"); ok {
if c, ok := srv.(interface{ Close() }); ok {
s.log.Debug("Close() calling execute.Close()")
c.Close()
}
}
}
if s.sessionID != "" {
os.RemoveAll(filepath.Join(ollieTmpDir(), s.sessionID)) //nolint:errcheck
}
}
// execServer returns the execute server if available, or nil.
func (s *agent) execServer() interface{} {
if s.runtime == nil || s.runtime.Dispatcher == nil {
return nil
}
srv, _ := s.runtime.Dispatcher.GetServer("execute")
return srv
}
func (s *agent) Detach() bool {
if srv := s.execServer(); srv != nil {
if d, ok := srv.(interface{ Detach() bool }); ok {
return d.Detach()
}
}
return false
}
func (s *agent) ListDetached() []DetachedInfo {
if srv := s.execServer(); srv != nil {
type listDetacher interface {
ListDetachedRaw() []any
}
if ld, ok := srv.(listDetacher); ok {
raw := ld.ListDetachedRaw()
out := make([]DetachedInfo, 0, len(raw))
for _, r := range raw {
if m, ok := r.(map[string]any); ok {
di := DetachedInfo{}
if v, ok := m["pid"].(int); ok {
di.PID = v
}
if v, ok := m["command"].(string); ok {
di.Command = v
}
if v, ok := m["started"].(int64); ok {
di.Started = v
}
if v, ok := m["exited"].(bool); ok {
di.Exited = v
}
if v, ok := m["exit_code"].(int); ok {
di.ExitCode = v
}
out = append(out, di)
}
}
return out
}
}
return nil
}
func (s *agent) SignalDetached(pid, signal int) error {
if srv := s.execServer(); srv != nil {
type signaler interface {
SignalDetached(int, syscall.Signal) error
}
if sg, ok := srv.(signaler); ok {
return sg.SignalDetached(pid, syscall.Signal(signal))
}
}
return fmt.Errorf("no execute server available")
}
func (s *agent) GetDetachedOutput(pid int) (string, error) {
if srv := s.execServer(); srv != nil {
type outputGetter interface {
GetDetachedOutput(int) (string, error)
}
if og, ok := srv.(outputGetter); ok {
return og.GetDetachedOutput(pid)
}
}
return "", fmt.Errorf("no execute server available")
}
func (s *agent) DismissDetached(pid int) bool {
if srv := s.execServer(); srv != nil {
type dismisser interface {
DismissDetached(int) bool
}
if d, ok := srv.(dismisser); ok {
return d.DismissDetached(pid)
}
}
return false
}
func (s *agent) InjectSystemEvent(content string) {
s.Queue("<detached-process-result>\n" + content + "\n</detached-process-result>")
}
// classifyReaction returns a category and description for a reaction emoji.
func classifyReaction(emoji string) (category, description string, positive bool) {
switch emoji {
case "👍", "✅":
return "positive", "The response was good. Keep doing what you're doing.", true
case "🚀", "🎉":
return "excellent", "The response was exactly what was wanted.", true
case "👎", "❌":
return "negative", "The response was wrong or unhelpful.", false
case "💩", "🤬":
return "terrible", "The response was fundamentally wrong. Stop this approach entirely and reassess from scratch.", false
case "🤔":
return "confused", "The response was unclear or confusing.", false
default:
return "unknown", "", false
}
}
func (s *agent) Reactions() map[string]string {
result := make(map[string]string)
if s.session == nil {
return result
}
for _, reaction := range s.session.Reactions {
result[reaction.ResponseID] = reaction.Emoji
}
return result
}
func (s *agent) React(emoji string) {
_ = s.ReactTo("", emoji)
}
func (s *agent) ReactTo(responseID, emoji string) error {
if s.session == nil {
return fmt.Errorf("no active session")
}
category, _, _ := classifyReaction(emoji)
if category == "unknown" {
return fmt.Errorf("unsupported reaction: %s", emoji)
}
if responseID == "" {
for i := len(s.session.messages) - 1; i >= 0; i-- {
if s.session.messages[i].Role == "assistant" {
responseID = s.session.messages[i].ID
break
}
}
}
if responseID == "" {
return fmt.Errorf("no assistant response to react to")
}
found := false
for i := range s.session.messages {
if s.session.messages[i].Role == "assistant" && s.session.messages[i].ID == responseID {
found = true
break
}
}
if !found {
return fmt.Errorf("assistant response not found: %s", responseID)
}
reaction := Reaction{ID: NewReactionID(), ResponseID: responseID, Emoji: emoji, Category: category, CreatedAt: time.Now()}
replaced := false
for i := range s.session.Reactions {
if s.session.Reactions[i].ResponseID == responseID {
if s.session.Reactions[i].Emoji == emoji {
return nil
}
s.session.Reactions[i] = reaction
replaced = true
break
}
}
if !replaced {
s.session.Reactions = append(s.session.Reactions, reaction)
}
s.session.recomputeReactionCounts()
s.saveSession()
return nil
}
func (s *agent) AgentName() string {
v := s.agentName
s.log.Debug("AgentName() = %q", v)
return v
}
func (s *agent) BackendName() string {
v := s.runtime.Backend.Name()
s.log.Debug("BackendName() = %q", v)
return v
}
func (s *agent) ModelName() string {
v := s.runtime.Backend.Model()
s.log.Debug("ModelName() = %q", v)
return v
}
func (s *agent) State() string {
s.mu.RLock()
defer s.mu.RUnlock()
return s.state
}
func (s *agent) notifyChange() {
s.changeMu.Lock()
s.changeCond.Broadcast()
s.changeMu.Unlock()
}
// WaitChange blocks until the named field changes from current, then returns
// the new value. Returns ("", false) if ctx is cancelled.
func (s *agent) WaitChange(ctx context.Context, field, current string) (string, bool) {
read := func() string {
switch field {
case WatchState:
return s.State()
case WatchUsage:
return s.Usage()
case WatchCtxSz:
return s.CtxSz()
case WatchCWD:
return s.CWD()
case WatchAgent:
return s.AgentName()
}
return ""
}
// context.AfterFunc fires in a separate goroutine when ctx is done,
// broadcasting to unblock any waiters.
stop := context.AfterFunc(ctx, func() {
s.changeMu.Lock()
s.changeCond.Broadcast()
s.changeMu.Unlock()
})
defer stop()
s.changeMu.Lock()
defer s.changeMu.Unlock()
for ctx.Err() == nil {
if v := read(); v != current {
return v, true
}
s.changeCond.Wait()
}
return "", false
}
func (s *agent) setState(state string) {
s.mu.Lock()
s.state = state
s.mu.Unlock()
s.log.Debug("state -> %q", state)
s.notifyChange()
}
func (s *agent) Reply() string {
s.mu.RLock()
defer s.mu.RUnlock()
s.log.Debug("Reply() len=%d", len(s.reply))
return s.reply
}
// CWD returns the current working directory for tool execution.
func (s *agent) CWD() string {
if s.cwd != "" {
s.log.Debug("CWD() = %q", s.cwd)
return s.cwd
}
wd, _ := os.Getwd()
s.log.Debug("CWD() = %q (from getwd)", wd)
return wd
}
// SetCWD changes the working directory for tool execution and updates the
// system prompt. Returns an error if the path does not exist.
func (s *agent) SetCWD(dir string) error {
s.log.Debug("SetCWD(%q)", dir)
dir = paths.ExpandHome(dir)
if dir != "" {
if _, err := os.Stat(dir); err != nil {
return fmt.Errorf("cwd: %w", err)
}
}
oldCwd := s.cwd
s.cwd = dir
// Update cwd references in the system prompt.
if oldCwd != "" && dir != "" && oldCwd != dir {
s.runtime.Preamble = strings.ReplaceAll(s.runtime.Preamble, oldCwd, dir)
}
// Propagate to any tool server that knows how to handle it (e.g. execute).
if s.runtime != nil && s.runtime.Dispatcher != nil {
if srv, ok := s.runtime.Dispatcher.GetServer("execute"); ok {
if ws, ok := srv.(tools.CWDSetter); ok {
ws.SetCWD(dir)
}
}
}
s.notifyChange()
return nil
}
// SetSessionID renames the session. It updates the in-memory ID, renames
// persisted files on disk, and propagates to the execute server env.
func (s *agent) SetSessionID(newID string) error {
s.log.Debug("SetSessionID(%q) old=%q", newID, s.sessionID)
oldID := s.sessionID
if oldID == newID {
return nil
}
// Rename active persisted files on disk.
if s.sessionsDir != "" && oldID != "" {
for _, suffix := range []string{".json", ".compaction.jsonl"} {
oldPath := s.activeSessionPath(oldID, suffix)
if _, err := os.Stat(oldPath); err == nil {
if err := os.Rename(oldPath, s.activeSessionPath(newID, suffix)); err != nil {
return fmt.Errorf("rename %s: %w", suffix, err)
}
}
}
}
s.sessionID = newID
// Update session ID references in the system prompt.
s.runtime.Preamble = strings.ReplaceAll(s.runtime.Preamble, oldID, newID)
// Rename tmpdir so isread markers remain valid after rename.
oldTemp := filepath.Join(ollieTmpDir(), oldID)
newTemp := filepath.Join(ollieTmpDir(), newID)
if _, err := os.Stat(oldTemp); err == nil {
os.Rename(oldTemp, newTemp) //nolint:errcheck
}
s.pushSessionEnv()
return nil
}
// defaultContextLength is used when the backend cannot report the model's
// actual context window (e.g. CodeWhisperer). 128k tokens is a safe default
// for modern models.
const defaultContextLength = 128000
const defaultToolResultMaxBytes = 131072
// autoCompactLimit returns the token threshold for auto-compaction (75%).
func (s *agent) autoCompactLimit(ctx context.Context) int {
ctxLen := s.runtime.Backend.ContextLength(ctx)
if ctxLen <= 0 {
ctxLen = defaultContextLength
}
return ctxLen * 3 / 4
}
// autoWarnLimit returns the token threshold for a context-usage warning (60%).
func (s *agent) autoWarnLimit(ctx context.Context) int {
ctxLen := s.runtime.Backend.ContextLength(ctx)
if ctxLen <= 0 {
ctxLen = defaultContextLength
}
return ctxLen * 3 / 5
}
// spawnContext assembles the agent context injected at each session refresh
// point (session start, post-clear, post-compaction). It combines the
// agent-specific prompt with any agentSpawn hook output.
func (s *agent) spawnContext(ctx context.Context) string {
result := s.runtime.Hooks.Run(ctx, HookAgentSpawn, map[string]string{
"session_id": s.sessionID,
"agent": s.agentName,
"cwd": s.CWD(),
"model": s.runtime.Backend.Model(),
}, s.log)
if result.Warning != "" {
s.emit(infoEvent(result.Warning))
}
if sum := result.Summary(); sum != "" {
s.emit(infoEvent("agentSpawn: " + sum))
}
var parts []string
if result.Context != "" {
parts = append(parts, result.Context)
}
return strings.Join(parts, "\n\n---\n\n")
}
// runCompact executes a full compaction cycle: pre-hook, compact, spawn-context
// re-injection, post-hook. Returns (n compacted, error). Returns (0, nil) if
// the pre-hook blocked or there was nothing to compact. Caller manages setState.
func (s *agent) runCompact(ctx context.Context, trigger string) (int, error) {
payload := map[string]string{"session_id": s.sessionID, "trigger": trigger, "cwd": s.CWD()}
pre := s.runtime.Hooks.Run(ctx, HookPreCompact, payload, s.log)
if pre.Warning != "" {
s.emit(infoEvent(pre.Warning))
}
if sum := pre.Summary(); sum != "" {
s.emit(infoEvent("preCompact: " + sum))
}
if pre.Blocked {
s.emit(infoEvent("compact cancelled by hook"))
return 0, nil
}
if pre.Context != "" {
s.session.appendUserMessage(pre.Context)
}
// Use a cheaper model for compaction if configured.
compactModel := resolveCompactionModel(s.runtime.CompactionModel, s.runtime.Backend)
origModel := s.runtime.Backend.Model()
if compactModel != "" && compactModel != origModel {
s.runtime.Backend.SetModel(compactModel)
defer s.runtime.Backend.SetModel(origModel)
}
n, _, err := s.session.compact(ctx, s.runtime.Backend)
if err != nil {
return 0, err
}
if n > 0 {
s.auditLog.Debug("compact: removed %d messages trigger=%s session=%s", n, trigger, s.sessionID)
s.warnedContext = false
if sc := s.spawnContext(ctx); sc != "" {
s.session.appendUserMessage(sc)
}
}
post := s.runtime.Hooks.Run(ctx, HookPostCompact, payload, s.log)
if post.Warning != "" {
s.emit(infoEvent(post.Warning))
}
if sum := post.Summary(); sum != "" {
s.emit(infoEvent("postCompact: " + sum))
}
if post.Context != "" {
s.session.appendUserMessage(post.Context)
}
return n, nil
}
func (s *agent) activeSessionPath(id, suffix string) string {
return filepath.Join(s.sessionsDir, "active", id+suffix)
}
func (s *agent) saveSession() {
s.saveMu.Lock()
s.saveDirty = true
if s.saveTimer == nil {
s.saveTimer = time.AfterFunc(2*time.Second, s.flushSave)
}
s.saveMu.Unlock()
}
// flushSave immediately persists the session if dirty.
func (s *agent) flushSave() {
s.saveMu.Lock()
dirty := s.saveDirty
s.saveDirty = false
if s.saveTimer != nil {
s.saveTimer.Stop()
s.saveTimer = nil
}
s.saveMu.Unlock()
if !dirty {
return
}
if s.session == nil || s.sessionID == "" || s.sessionsDir == "" {
return
}
path := s.activeSessionPath(s.sessionID, ".json")
if err := os.MkdirAll(filepath.Dir(path), 0700); err != nil {
s.log.Error("session save: %v", err)
return
}
if err := s.session.saveToFull(path, s.sessionID, s.agentName,
s.runtime.Backend.Name(), s.runtime.Backend.Model(), s.CWD(), s.remote); err != nil {
s.log.Error("session save: %v", err)
}
}
// SaveSession writes the current session state to the given path, including
// backend and model metadata for external restore.
func (s *agent) SaveSession(path string) error {
s.mu.RLock()
defer s.mu.RUnlock()
if s.session == nil {
return fmt.Errorf("no active session")
}
return s.session.saveToFull(path, s.sessionID, s.agentName,
s.runtime.Backend.Name(), s.runtime.Backend.Model(), s.CWD(), s.remote)
}
func (s *agent) getActionCancel() context.CancelCauseFunc {
if a := s.currentAction.Load(); a != nil {
return a.cancel
}
return nil
}
// Interrupt cancels the current in-progress agent turn.
// Returns true if an action was running and was cancelled.
func (s *agent) Interrupt(cause error) bool {
s.log.Debug("Interrupt() cause=%v", cause)
if cancel := s.getActionCancel(); cancel != nil {
cancel(cause)
return true
}
return false
}
func (s *agent) Inject(prompt string) {
// If an inject is already pending, fall back to the normal FIFO so nothing
// is lost. Use CompareAndSwap to avoid a race between the nil check and store.
if !s.pendingInject.CompareAndSwap(nil, &prompt) {
s.fifo.Push(prompt)
return
}
s.emit(Event{Role: "info", Content: "\n"})
s.emit(Event{Role: "user", Content: prompt})
}
func (s *agent) injectRewrite(prompt string) {
s.pendingInject.Store(&prompt)
s.emit(Event{Role: "info", Content: "\n"})
s.emit(Event{Role: "user", Content: prompt})
}
func (s *agent) Queue(prompt string) {
s.log.Debug("Queue() len=%d", len(prompt))
s.fifo.Push(prompt)
s.bus.Publish("queued", prompt)
if !s.IsRunning() {
go s.drainQueue()
}
}
func (s *agent) drainQueue() {
if prompt, ok := s.fifo.Pop(); ok {
s.Submit(context.Background(), prompt)
}
}
func (s *agent) Bus() *pubsub.Bus {
return s.bus
}
func (s *agent) emit(ev Event) {
s.bus.Publish("event", ev)
}
func (s *agent) PopQueue() (string, bool) {
v, ok := s.fifo.Pop()
s.log.Debug("PopQueue() ok=%v len=%d", ok, len(v))
return v, ok
}
func (s *agent) IsRunning() bool {
v := s.currentAction.Load() != nil
s.log.Debug("IsRunning() = %v", v)
return v
}
func (s *agent) CtxSz() string {
if s.session == nil {
s.log.Debug("CtxSz() no session")
return "no active session"
}
ctxLen := s.runtime.Backend.ContextLength(context.Background())
if ctxLen <= 0 {
ctxLen = defaultContextLength
}
estimated := s.session.estimateTokens()
pct := estimated * 100 / ctxLen
v := fmt.Sprintf("%d / %d (%d%%)", estimated, ctxLen, pct)
s.log.Debug("CtxSz() = %q", v)
return v
}
func (s *agent) Cost() string {
if s.session == nil {
return "no active session"
}
return fmt.Sprintf("costLast=$%.4f\ncostSession=$%.4f\n",
s.session.LastTurnCostUSD, s.session.SessionCostUSD)
}
func (s *agent) Usage() string {
if s.session == nil {
s.log.Debug("Usage() no session")
return "no active session"
}
str := fmt.Sprintf("%d in, %d out, %d requests",
s.session.TotalInputTokens, s.session.TotalOutputTokens,
s.session.TotalRequests)
if s.session.TotalCachedInputTokens > 0 {
str += fmt.Sprintf(", %d cached", s.session.TotalCachedInputTokens)
}
if s.session.Estimated {
str += " [estimated]"
}
s.log.Debug("Usage() = %q", str)
return str
}
func (s *agent) Context() []backend.Message {
s.mu.RLock()
var msgs []backend.Message
if s.session != nil {
msgs = slices.Clone(s.session.history())
}
s.mu.RUnlock()
if s.runtime.Preamble != "" {
msgs = append([]backend.Message{{Role: "system", Content: s.runtime.Preamble}}, msgs...)
}
return msgs
}
func (s *agent) SystemPrompt() string {
s.log.Debug("SystemPrompt() len=%d", len(s.runtime.Preamble))
return s.runtime.Preamble
}
func (s *agent) GenerationParams() backend.GenerationParams {
s.mu.RLock()
defer s.mu.RUnlock()
return s.runtime.GenParams
}
func (s *agent) CompactionModel() string {
s.mu.RLock()
defer s.mu.RUnlock()
return s.runtime.CompactionModel
}
func (s *agent) SetCompactionModel(model string) {
s.mu.Lock()
defer s.mu.Unlock()
s.runtime.CompactionModel = model
}
func (s *agent) SetGenerationParams(params backend.GenerationParams) error {
if s.IsRunning() {
return fmt.Errorf("cannot change params while agent is running")
}
s.mu.Lock()
s.runtime.GenParams = params
s.mu.Unlock()
return nil
}
func (s *agent) ListModels() string {
s.log.Debug("ListModels()")
models := s.runtime.Backend.Models(context.Background())
slices.Sort(models)
return strings.Join(models, "\n")
}
// firstSentence returns the first sentence of s (up to the first period or
// newline), trimmed. Falls back to s truncated at 80 chars if no sentence end
// is found.
func firstSentence(s string) string {
for i, r := range s {
if r == '.' || r == '\n' {
return strings.TrimSpace(s[:i+1])
}
}
if len(s) > 80 {
return s[:77] + "..."
}
return s
}
// Submit implements Core. It processes one line of user input: slash commands
// and shell shortcuts are dispatched immediately; any other input
// starts an agent turn that streams events to the bus. If a turn is already
// in progress the prompt is queued as an in-stream interruption instead.
//
// Continuations (post-turn hook context, unconsumed inject, FIFO drain) are
// handled via an explicit loop rather than recursion to avoid stack growth.
func (s *agent) Submit(ctx context.Context, input string) {
defer func() {
if r := recover(); r != nil {
s.log.Error("panic: %v\n%s", r, debug.Stack())
if a := s.currentAction.Swap(nil); a != nil {
a.cancel(fmt.Errorf("%v", r))
}
s.setState("idle")
s.emit(Event{Role: "error", Content: fmt.Sprintf("%v", r)})
}
}()
s.log.Debug("Submit() input_len=%d running=%v", len(input), s.IsRunning())
if input == "" {
return
}
// Fast path: inject and FIFO push use atomics and are safe without
// the submit lock. Handle them before acquiring submitMu so they
// don't block behind a long-running turn or command.
if s.IsRunning() {
if s.handleCommand(ctx, input) {
return
}
s.fifo.Push(input)
return
}
// Serialize commands and turns so that e.g. a /compact arriving via
// ctl cannot race with an executeTurn arriving via prompt.
s.submitMu.Lock()
defer s.submitMu.Unlock()
if s.handleCommand(ctx, input) {
return
}
if s.IsRunning() {
s.fifo.Push(input)
return
}
for input != "" && ctx.Err() == nil {
input = s.executeTurn(ctx, input)
}
}
// executeTurn runs a single agent turn and returns the next prompt to execute,
// or "" if there is nothing more to do.
func (s *agent) executeTurn(ctx context.Context, input string) string {
s.emit(Event{Role: "user", Content: input})
hookResult := s.runtime.Hooks.Run(ctx, HookPreTurn, map[string]string{
"session_id": s.sessionID,
"cwd": s.CWD(),
"prompt": input,
}, s.log)
if hookResult.Blocked {
s.emit(infoEvent("hook blocked prompt"))
return ""
}
if hookResult.Warning != "" {
s.emit(infoEvent(hookResult.Warning))
}
if sum := hookResult.Summary(); sum != "" {
s.emit(infoEvent("preTurn: " + sum))
}
if hookResult.Context != "" {
input += "\n" + hookResult.Context
}
// Snapshot session state before this turn modifies it. Restored on failure
// so the session is clean for the next attempt.
snapSession := s.session
var snapMessages []backend.Message
if s.session != nil {
snapMessages = cloneMessages(s.session.messages)
}
if s.session == nil {
for _, msg := range s.startupMessages {
s.log.Debug("startup: %s", msg)
s.emit(infoEvent(msg))
}
s.startupMessages = nil
s.session = newSession(input)
if sc := s.spawnContext(ctx); sc != "" {
s.session.appendUserMessage(sc)
}
s.session.appendUserMessage(input)
} else {
s.session.appendUserMessage(input)
}
actCtx, actCancel := context.WithCancelCause(ctx)
handle := &actionHandle{cancel: actCancel}
s.currentAction.Store(handle)
s.setState("thinking")
s.auditLog.Debug("turn: start input=%s session=%s", auditTruncate(input), s.sessionID)
// Build per-turn agentConfig from the current runtime.
s.cfg = agentConfig{
Backend: s.runtime.Backend,
preamble: s.runtime.Preamble,
Tools: s.runtime.Tools,
Exec: s.runtime.Exec,
ClassifyTool: s.runtime.ClassifyTool,
ClassifyTier: s.runtime.ClassifyTier,
GenerationParams: s.runtime.GenParams,
MaxSteps: s.runtime.MaxSteps,
ReadPlanStep: s.readPlanStep,
TurnError: s.turnError,
}
var replyBuf strings.Builder
s.cfg.Output = func(ev Event) {
switch ev.Role {
case "assistant":
replyBuf.WriteString(ev.Content)
case "call":
s.setState("calling: " + ev.Name)
s.auditLog.Debug("call: %s %s", ev.Name, auditTruncate(string(ev.Content)))
case "tool":
s.auditLog.Debug("result: %s %s", ev.Name, auditTruncate(ev.Content))
case "state":
s.setState(ev.Content)
case "limitretry":
s.setState("limitretry")
case "error":
s.auditLog.Debug("error: %s", ev.Content)
}
if ev.Role == "usage" && s.session != nil {
var in, out, est, cached, creation int
var costUSD float64
fmt.Sscanf(ev.Content, "%d %d %d %g %d %d", &in, &out, &est, &costUSD, &cached, &creation)
s.session.addUsage(backend.Usage{
InputTokens: in,
CachedInputTokens: cached,
CacheCreationTokens: creation,
OutputTokens: out,
CostUSD: costUSD,
}, est != 0)
s.notifyChange()
if maxCostStr := os.Getenv("OLLIE_MAX_SESSION_COST"); maxCostStr != "" {
if limit, ferr := strconv.ParseFloat(maxCostStr, 64); ferr == nil && limit > 0 && s.session.SessionCostUSD >= limit {
s.emit(infoEvent(fmt.Sprintf("spending cap $%.2f reached — stopping", limit)))
s.Interrupt(ErrInterrupted)
}
}
}
s.emit(ev)
}
s.cfg.PopInject = func() string {
if p := s.pendingInject.Swap(nil); p != nil {
return *p
}
return ""
}
s.cfg.PreTool = func(ctx context.Context, name string, args json.RawMessage) HookResult {
return s.runtime.Hooks.Run(ctx, HookPreTool, map[string]string{
"session_id": s.sessionID,
"cwd": s.CWD(),
"tool": name,
"args": string(args),
}, s.log)
}
s.cfg.PostTool = func(ctx context.Context, name string, args json.RawMessage, result string) HookResult {
return s.runtime.Hooks.Run(ctx, HookPostTool, map[string]string{
"session_id": s.sessionID,
"cwd": s.CWD(),
"tool": name,
"args": string(args),
"result": result,
}, s.log)
}
s.cfg.IncrToolCallCount = func() int64 {
return s.toolCallCount.Add(1)
}
s.cfg.SaveSession = func() { s.saveSession() }
s.cfg.ResultCache = &s.resultCache
s.cfg.AutoCompact = func(ctx context.Context) {
if ctx.Err() != nil || s.session == nil {
return
}
limit := s.autoCompactLimit(ctx)
if limit <= 0 || s.session.estimateTokens() < limit {
return
}
s.emit(Event{Role: "info", Content: "auto-compacting context...\n"})
s.setState("compacting")
if _, err := s.runCompact(ctx, "auto"); err != nil {
panic(fmt.Sprintf("mid-turn auto-compact: %v", err))
}
s.setState("thinking")
}
// Warn once when context usage crosses 60%; compact at 75%.
if s.session != nil {
tokens := s.session.estimateTokens()
if compactLimit := s.autoCompactLimit(ctx); compactLimit > 0 && tokens >= compactLimit {
s.emit(Event{Role: "info", Content: "auto-compacting context...\n"})
s.setState("compacting")
if _, err := s.runCompact(ctx, "auto"); err != nil {
panic(fmt.Sprintf("auto-compact: %v", err))
}
s.setState("thinking")
} else if warnLimit := s.autoWarnLimit(ctx); warnLimit > 0 && tokens >= warnLimit && !s.warnedContext {
ctxLen := s.cfg.Backend.ContextLength(ctx)
if ctxLen <= 0 {
ctxLen = defaultContextLength
}
pct := tokens * 100 / ctxLen
s.emit(Event{Role: "info", Content: fmt.Sprintf("context at %d%% — will auto-compact at 75%%\n", pct)})
s.warnedContext = true
}
}
// Spending cap: reject before spending more tokens.
if maxCostStr := os.Getenv("OLLIE_MAX_SESSION_COST"); maxCostStr != "" && s.session != nil {
if limit, err := strconv.ParseFloat(maxCostStr, 64); err == nil && limit > 0 {
if s.session.SessionCostUSD >= limit {
s.emit(Event{Role: "error", Content: fmt.Sprintf("spending cap $%.2f reached (session total $%.4f)", limit, s.session.SessionCostUSD)})
s.setState("idle")
actCancel(nil)
s.currentAction.CompareAndSwap(handle, nil)
if snapSession == nil {
s.session = nil
} else {
s.session.messages = snapMessages
}
return ""
}
}
}
if s.session != nil {
s.session.resetTurnAccumulators()
}
// Run the turn, retrying once after compaction on context overflow.
var (
overflowRetried bool
err error
)
for {
err = run(actCtx, s.cfg, s.session)
actCancel(nil)
s.currentAction.CompareAndSwap(handle, nil)
if err == nil {
break
}
if errors.Is(err, context.Canceled) || errors.Is(err, ErrInterrupted) {
break
}
var ctxErr *backend.ContextOverflowError
if !overflowRetried && errors.As(err, &ctxErr) && s.session != nil {
overflowRetried = true
s.session.messages = snapMessages
s.emit(Event{Role: "info", Content: "context overflow — compacting and retrying...\n"})
s.setState("compacting")
if _, cerr := s.runCompact(ctx, "overflow"); cerr != nil {
break
}
s.session.appendUserMessage(input)
s.setState("thinking")
s.session.resetTurnAccumulators()
replyBuf.Reset()
actCtx, actCancel = context.WithCancelCause(ctx)
handle = &actionHandle{cancel: actCancel}
s.currentAction.Store(handle)
continue
}
break
}
s.mu.Lock()
s.reply = replyBuf.String()
s.mu.Unlock()
replyBuf.Reset()
s.setState("idle")
s.flushSave()
if err != nil {
// Keep completed work — only remove cancelled tool results.
// Error results are valuable feedback for the agent.
if s.session != nil {
s.session.removeCancelledToolResults()
}
if errors.Is(err, context.Canceled) || errors.Is(err, ErrInterrupted) {
s.auditLog.Debug("turn: interrupted session=%s", s.sessionID)
s.saveSession()
return ""
}
s.emit(Event{Role: "error", Content: err.Error()})
// Drain one FIFO item — the turnError hook may have queued a recovery prompt.
if next, ok := s.fifo.Pop(); ok {
return next
}
return ""
}
stopResult := s.runtime.Hooks.Run(ctx, HookPostTurn, map[string]string{
"session_id": s.sessionID,
"cwd": s.CWD(),
}, s.log)
if stopResult.Warning != "" {
s.emit(infoEvent(stopResult.Warning))
}
if sum := stopResult.Summary(); sum != "" {
s.emit(infoEvent("postTurn: " + sum))
}
if !stopResult.Blocked && stopResult.Context != "" && s.session != nil {
s.session.appendUserMessage(stopResult.Context)
}
if s.session != nil {
s.session.recordTurnCost(s.cfg.Backend.Model())
appendUsageLog(s.sessionID, s.cfg.Backend.Name(), s.cfg.Backend.Model(), s.session)
if s.session.LastTurnCostUSD > 0 {
s.emit(Event{Role: "info", Content: fmt.Sprintf("costLast=$%.4f\n", s.session.LastTurnCostUSD)})
}
s.auditLog.Debug("turn: end reply=%s cost=$%.4f session_total=$%.4f session=%s",
auditTruncate(s.reply), s.session.LastTurnCostUSD, s.session.SessionCostUSD, s.sessionID)
s.notifyChange()
}
s.saveSession()
// Post-turn hook said "continue" — its context becomes the next prompt.
if stopResult.Blocked && stopResult.Context != "" {
return stopResult.Context
}
// Inject that was pending but never consumed (text-only response with no
// tool calls) — treat it as the next user message.
if p := s.pendingInject.Swap(nil); p != nil {
return *p
}
// Drain one item from the FIFO; the outer loop handles the rest.
if next, ok := s.fifo.Pop(); ok {
return next
}
return ""
}
func toolInfosToBackend(infos []tools.ToolInfo) []backend.Tool {
out := make([]backend.Tool, len(infos))
for i, t := range infos {
out[i] = backend.Tool{
Name: t.Name,
Description: t.Description,
Parameters: t.InputSchema,
}
}
return out
}
func extractToolResult(raw json.RawMessage) (text string, contentBlocks []backend.ContentBlock, isError bool) {
var result struct {
IsError bool `json:"isError"`
Content []struct {
Type string `json:"type"`
Text string `json:"text"`
MediaType string `json:"media_type"`
Data string `json:"data"`
} `json:"content"`
}
if err := json.Unmarshal(raw, &result); err != nil {
return string(raw), nil, false
}
var parts []string
for _, c := range result.Content {
switch c.Type {
case "text":
parts = append(parts, c.Text)
case "image":
contentBlocks = append(contentBlocks, backend.ContentBlock{
Type: "image",
ImageSource: &backend.ImageSource{
Type: "base64",
MediaType: c.MediaType,
Data: c.Data,
},
})
}
}
return strings.Join(parts, "\n"), contentBlocks, result.IsError
}