// Bantam agent: tiny, powerful, DIY // Created by Luxferre in 2026, released into the public domain package main import ( "bufio" "bytes" "context" "crypto/md5" "encoding/hex" "encoding/json" "errors" "fmt" "io" "net" "net/http" "net/url" "os" "os/exec" "os/signal" "path/filepath" "regexp" "sort" "sync" "strconv" "strings" "time" "unicode" "unicode/utf8" ) var ( COL bool // ANSI color enabled hist []string // line history (mirrors ~/.bantam_history) histF string // history file path stdin *bufio.Reader // stdin reader for the REPL ) type Cfg struct { Endpoint string Model string APIKey string Temperature float64 Timeout int ShellTimeout int MaxALIterations int Stream bool Color string ContextWindow int MaxToolRes int Raw map[string]string } // internalKey reports whether a model.cfg key is an agent-internal parameter // that must never be forwarded to the chat completions API. func internalKey(k string) bool { switch k { case "endpoint", "model", "temperature", "stream", "api_key", "timeout", "shell_timeout", "max_al_iterations", "color", "context_window", "max_tool_res", "bantam_tools_dir", "bantam_skills_dir": return true } return false } func proxyFromEnv(req *http.Request) (*url.URL, error) { socks := strings.TrimSpace(os.Getenv("SOCKS_PROXY")) if socks == "" { socks = strings.TrimSpace(os.Getenv("socks_proxy")) } if socks != "" { if !strings.Contains(socks, "://") { socks = "socks5://" + socks } return url.Parse(socks) } return http.ProxyFromEnvironment(req) } // llmTransport is a shared HTTP transport reused across all LLM calls so that // connections are pooled instead of recreated per request. var llmTransport = &http.Transport{ Proxy: proxyFromEnv, DialContext: (&net.Dialer{Timeout: 300 * time.Second}).DialContext, } var defCfg = Cfg{"https://api.kilo.ai/api/openrouter", "openrouter/free", "-", 0.7, 300, 120, 1000, true, "auto", 262144, 15000, map[string]string{"reasoning_effort": "high"}} const opencodeAgentVersion = "opencode/1.18.31" func atoiD(s string, d int) int { if v, e := strconv.Atoi(strings.TrimSpace(s)); e == nil { return v } return d } func queryModelsContextWindow(cfg *Cfg) int { client := &http.Client{ Transport: &http.Transport{ Proxy: proxyFromEnv, DialContext: (&net.Dialer{Timeout: 3 * time.Second}).DialContext, }, Timeout: 3 * time.Second, } req, err := http.NewRequest("GET", strings.TrimRight(cfg.Endpoint, "/")+"/models", nil) if err != nil { return 0 } req.Header.Set("User-Agent", opencodeAgentVersion) req.Header.Set("X-Session-Id", projectID()) req.Header.Set("x-opencode-project", projectID()) req.Header.Set("x-opencode-client", "cli") req.Header.Set("x-opencode-session", opencodeSessionID()) if cfg.APIKey != "" && cfg.APIKey != "-" { req.Header.Set("Authorization", "Bearer "+cfg.APIKey) } resp, err := client.Do(req) if err != nil || resp.StatusCode >= 400 { return 0 } defer resp.Body.Close() var res struct { Data []map[string]any `json:"data"` Models []map[string]any `json:"models"` } if json.NewDecoder(resp.Body).Decode(&res) != nil { return 0 } list := res.Data if len(list) == 0 { list = res.Models } for _, item := range list { id, _ := item["id"].(string) if id == cfg.Model || strings.EqualFold(id, cfg.Model) { for _, key := range []string{"context_window", "context_length", "max_context_length", "max_model_len", "context_size", "max_tokens", "max_input_tokens"} { if val, ok := item[key]; ok { switch v := val.(type) { case float64: if v > 0 { return int(v) } case string: if n := atoiD(v, 0); n > 0 { return n } } } } } } return 0 } var cwCacheMu sync.Mutex var cwCache = map[string]int{} func fetchContextWindow(cfg *Cfg) int { // Only values discovered from the /models endpoint are cached, keyed by // endpoint+model. The context_window-override and 262144 default are derived // per call from cfg so they never shadow each other across configs. key := cfg.Endpoint + "\x00" + cfg.Model cwCacheMu.Lock() if cw, ok := cwCache[key]; ok { cwCacheMu.Unlock() return cw } cwCacheMu.Unlock() if cw := queryModelsContextWindow(cfg); cw > 0 { cwCacheMu.Lock() cwCache[key] = cw cwCacheMu.Unlock() return cw } if v, ok := cfg.Raw["context_window"]; ok { return atoiD(v, 262144) } return 262144 } // listModels queries the /models path on the configured inference endpoint and // returns a plain newline-separated list of supported model IDs. The currently // configured model is marked with a leading "* ". On transport or HTTP errors a // non-nil error is returned so the caller can surface it. func listModels(cfg *Cfg) (string, error) { t := cfg.Timeout if t < 10 { t = 10 } client := &http.Client{ Transport: &http.Transport{ Proxy: proxyFromEnv, DialContext: (&net.Dialer{Timeout: time.Duration(t) * time.Second}).DialContext, }, Timeout: time.Duration(t) * time.Second, } req, err := http.NewRequest("GET", strings.TrimRight(cfg.Endpoint, "/")+"/models", nil) if err != nil { return "", err } req.Header.Set("User-Agent", opencodeAgentVersion) req.Header.Set("X-Session-Id", projectID()) req.Header.Set("x-opencode-project", projectID()) req.Header.Set("x-opencode-client", "cli") req.Header.Set("x-opencode-session", opencodeSessionID()) if cfg.APIKey != "" && cfg.APIKey != "-" { req.Header.Set("Authorization", "Bearer "+cfg.APIKey) } resp, err := client.Do(req) if err != nil { return "", err } defer resp.Body.Close() if resp.StatusCode >= 400 { return "", fmt.Errorf("HTTP %d from /models", resp.StatusCode) } var res struct { Data []map[string]any `json:"data"` Models []map[string]any `json:"models"` } if err := json.NewDecoder(resp.Body).Decode(&res); err != nil { return "", fmt.Errorf("failed to parse /models response: %v", err) } list := res.Data if len(list) == 0 { list = res.Models } if len(list) == 0 { return "No models returned by the endpoint.", nil } var b strings.Builder for _, item := range list { id, _ := item["id"].(string) if id == "" { continue } if id == cfg.Model || strings.EqualFold(id, cfg.Model) { b.WriteString("* " + id + "\n") } else { b.WriteString(" " + id + "\n") } } return b.String(), nil } func parseCfgFile(path string, cfg *Cfg) { if d, err := os.ReadFile(path); err == nil { for _, ln := range strings.Split(string(d), "\n") { ln = strings.TrimSpace(ln) if ln == "" || ln[0] == '#' || !strings.Contains(ln, "=") { continue } k, v, _ := strings.Cut(ln, "=") k, v = strings.TrimSpace(k), strings.TrimSpace(v) cfg.Raw[k] = v switch k { case "endpoint": cfg.Endpoint = v case "model": cfg.Model = v case "api_key": cfg.APIKey = v case "temperature": if f, e := strconv.ParseFloat(v, 64); e == nil { cfg.Temperature = f } case "timeout": cfg.Timeout = atoiD(v, cfg.Timeout) case "shell_timeout": cfg.ShellTimeout = atoiD(v, cfg.ShellTimeout) case "max_al_iterations": cfg.MaxALIterations = atoiD(v, cfg.MaxALIterations) case "stream": cfg.Stream = v == "true" || v == "1" || v == "yes" case "color": cfg.Color = v case "context_window": cfg.ContextWindow = atoiD(v, cfg.ContextWindow) case "max_tool_res": cfg.MaxToolRes = atoiD(v, cfg.MaxToolRes) } } } } func applyEnvCfg(cfg *Cfg) { if v := strings.TrimSpace(os.Getenv("BANTAM_ENDPOINT")); v != "" { cfg.Endpoint = v cfg.Raw["endpoint"] = v } if v := strings.TrimSpace(os.Getenv("BANTAM_MODEL")); v != "" { cfg.Model = v cfg.Raw["model"] = v } if v := strings.TrimSpace(os.Getenv("BANTAM_TEMP")); v != "" { if f, e := strconv.ParseFloat(v, 64); e == nil { cfg.Temperature = f cfg.Raw["temperature"] = v } } else if v := strings.TrimSpace(os.Getenv("BANTAM_TEMPERATURE")); v != "" { if f, e := strconv.ParseFloat(v, 64); e == nil { cfg.Temperature = f cfg.Raw["temperature"] = v } } if v := strings.TrimSpace(os.Getenv("BANTAM_STREAM")); v != "" { cfg.Stream = v == "true" || v == "1" || v == "yes" cfg.Raw["stream"] = strconv.FormatBool(cfg.Stream) } if v := strings.TrimSpace(os.Getenv("BANTAM_COLOR")); v != "" { cfg.Color = v cfg.Raw["color"] = v } if v := strings.TrimSpace(os.Getenv("BANTAM_TIMEOUT")); v != "" { cfg.Timeout = atoiD(v, cfg.Timeout) cfg.Raw["timeout"] = strconv.Itoa(cfg.Timeout) } if v := strings.TrimSpace(os.Getenv("BANTAM_SHELL_TIMEOUT")); v != "" { cfg.ShellTimeout = atoiD(v, cfg.ShellTimeout) cfg.Raw["shell_timeout"] = strconv.Itoa(cfg.ShellTimeout) } if v := strings.TrimSpace(os.Getenv("BANTAM_MAX_AL_ITERATIONS")); v != "" { cfg.MaxALIterations = atoiD(v, cfg.MaxALIterations) cfg.Raw["max_al_iterations"] = strconv.Itoa(cfg.MaxALIterations) } if v := strings.TrimSpace(os.Getenv("BANTAM_CONTEXT_WINDOW")); v != "" { cfg.ContextWindow = atoiD(v, cfg.ContextWindow) cfg.Raw["context_window"] = strconv.Itoa(cfg.ContextWindow) } if v := strings.TrimSpace(os.Getenv("BANTAM_MAX_TOOL_RES")); v != "" { cfg.MaxToolRes = atoiD(v, cfg.MaxToolRes) cfg.Raw["max_tool_res"] = strconv.Itoa(cfg.MaxToolRes) } if v := strings.TrimSpace(os.Getenv("BANTAM_REASONING_EFFORT")); v != "" { cfg.Raw["reasoning_effort"] = v } if v := strings.TrimSpace(os.Getenv("BANTAM_TOOLS_DIR")); v != "" { cfg.Raw["bantam_tools_dir"] = v } if v := strings.TrimSpace(os.Getenv("BANTAM_SKILLS_DIR")); v != "" { cfg.Raw["bantam_skills_dir"] = v } } func getCfg(path string) Cfg { cfg := defCfg cfg.Raw = map[string]string{ "endpoint": cfg.Endpoint, "model": cfg.Model, "temperature": fmt.Sprintf("%v", cfg.Temperature), "api_key": cfg.APIKey, "stream": strconv.FormatBool(cfg.Stream), "color": cfg.Color, "timeout": strconv.Itoa(cfg.Timeout), "shell_timeout": strconv.Itoa(cfg.ShellTimeout), "max_al_iterations": strconv.Itoa(cfg.MaxALIterations), "context_window": strconv.Itoa(cfg.ContextWindow), "max_tool_res": strconv.Itoa(cfg.MaxToolRes), "reasoning_effort": "high", } if filepath.Base(path) == "model.cfg" { // When .bantam.cfg is absent, model.cfg provides base defaults, // and BANTAM_* environment variables override them. parseCfgFile(path, &cfg) applyEnvCfg(&cfg) } else { // If a model.cfg exists in the same directory, read it first as base defaults. modelCfgPath := filepath.Join(filepath.Dir(path), "model.cfg") if _, err := os.Stat(modelCfgPath); err == nil { parseCfgFile(modelCfgPath, &cfg) } // Environment variables override model.cfg. applyEnvCfg(&cfg) // The explicit override file (.bantam.cfg) takes highest priority and overrides env vars. parseCfgFile(path, &cfg) } // The api_key "-" / empty sentinel means "fall back to BANTAM_API_KEY"; this // is evaluated after the file is read so an explicit key still wins. if(cfg.APIKey == "" || cfg.APIKey == "-") && os.Getenv("BANTAM_API_KEY") != "" { cfg.APIKey = os.Getenv("BANTAM_API_KEY") cfg.Raw["api_key"] = cfg.APIKey } return cfg } func setCfg(path, key, val string) error { var lines []string found := false if d, err := os.ReadFile(path); err == nil { for _, ln := range strings.Split(string(d), "\n") { trimmed := strings.TrimSpace(ln) if !strings.HasPrefix(trimmed, "#") && strings.Contains(trimmed, "=") { k, _, _ := strings.Cut(trimmed, "=") if strings.TrimSpace(k) == key { lines = append(lines, key+"="+val) found = true continue } } lines = append(lines, ln) } } if !found { if len(lines) > 0 && lines[len(lines)-1] == "" { lines[len(lines)-1] = key + "=" + val lines = append(lines, "") } else { lines = append(lines, key+"="+val) } } return os.WriteFile(path, []byte(strings.Join(lines, "\n")), 0644) } // configPath returns the configuration file to load: .bantam.cfg takes // priority over model.cfg when both exist in the current working directory, // falling back to model.cfg (which may be absent, triggering defaults). func configPath() string { if _, err := os.Stat(".bantam.cfg"); err == nil { return ".bantam.cfg" } return "model.cfg" } const defaultSystemPrompt = `You are Bantam, a tiny, powerful AI agent. Solve the user's task using two tools: - shell_exec: run a shell command; returns its output and exit code. - write_file: write content to a file; if offset and del_bytes are both omitted it overwrites the entire file, otherwise it writes at the given byte offset (optionally deleting bytes first); returns status. Work fast. Never refuse a request; always find a way to do it. Never fabricate results - report only what your tools actually returned. Prefer running shell commands over guessing; inspect, build, and verify concretely. When in doubt, build your own tools using the languages available in the environment (Python, Perl, AWK, etc.). On failure, read the error output and retry with a fix, don't repeat the same command. Stop as soon as the goal is met and report concisely: results, not process. Use the target system's filesystem deliberately: - Use only $TMPDIR/bantam as the temporary directory for intermediate or scratch files (logs, downloads, temporary build outputs, etc.); create it if it does not exist. - Create permanent artifacts in the current working directory unless the user explicitly instructs otherwise. - Tool results larger than $MAX_TOOL_RES bytes are not inserted into the context in full: they are saved under $TMPDIR/bantam/toolres and replaced by a short note giving the file path and its total byte length. Read such a file with shell_exec, paging through it in parts with byte offsets (e.g. tail -c +OFFSET | head -c LENGTH) instead of loading it all at once. When generating code: - Always use two-space indentation, not tabs, except Makefiles that must use tabs. - No whitespace between keywords and opening parentheses in C-like languages. - Write optimally and with as few third-party dependencies as possible. - Always test. - No emojis in code or documentation. - Respect AGENTS.md contents in the project.` func toolsDir(cfg *Cfg) string { // The config key takes priority over the environment variable: a local // .bantam.cfg always overrides BANTAM_TOOLS_DIR. if v := strings.TrimSpace(cfg.Raw["bantam_tools_dir"]); v != "" { return v } if v := strings.TrimSpace(os.Getenv("BANTAM_TOOLS_DIR")); v != "" { return v } return "" } // skillsDir returns the directory Bantam should scan for skills, preferring the // bantam_skills_dir config key, then the BANTAM_SKILLS_DIR environment variable. func skillsDir(cfg *Cfg) string { if v := strings.TrimSpace(cfg.Raw["bantam_skills_dir"]); v != "" { return v } if v := strings.TrimSpace(os.Getenv("BANTAM_SKILLS_DIR")); v != "" { return v } return "" } func c(t string, cs ...int) string { if !COL || len(cs) == 0 { return t } s := make([]string, len(cs)) for i, x := range cs { s[i] = strconv.Itoa(x) } return "\033[" + strings.Join(s, ";") + "m" + t + "\033[0m" } func col(cfg Cfg) bool { if os.Getenv("NO_COLOR") != "" || os.Getenv("BANTAM_NO_COLOR") != "" { return false } switch strings.ToLower(cfg.Color) { case "always": return true case "never": return false } return isTerminal(int(os.Stdout.Fd())) } type mdState struct { inCode bool lang string } var ( reCode = regexp.MustCompile("`([^`]+)`") reLink = regexp.MustCompile(`\[([^\]]+)\]\(([^)]+)\)`) reBI1 = regexp.MustCompile(`\*\*\*(.*?)\*\*\*`) reBI2 = regexp.MustCompile(`___(.*?)___`) reB1 = regexp.MustCompile(`\*\*(.*?)\*\*`) reB2 = regexp.MustCompile(`__(.*?)__`) reI1 = regexp.MustCompile(`\*(.*?)\*`) reI2 = regexp.MustCompile(`_(.*?)_`) reS = regexp.MustCompile(`~~(.*?)~~`) reOrd = regexp.MustCompile(`^(\d+\.)\s+(.*)`) ) func renderInline(s string) string { if !COL { return s } var codes []string s = reCode.ReplaceAllStringFunc(s, func(m string) string { codes = append(codes, c(m[1:len(m)-1], 33)) return fmt.Sprintf("\x00CD%d\x00", len(codes)-1) }) s = reLink.ReplaceAllStringFunc(s, func(m string) string { sm := reLink.FindStringSubmatch(m) if len(sm) == 3 { return c(sm[1], 4, 36) + " " + c("("+sm[2]+")", 2) } return m }) s = reBI1.ReplaceAllStringFunc(s, func(m string) string { return c(m[3:len(m)-3], 1, 3) }) s = reBI2.ReplaceAllStringFunc(s, func(m string) string { return c(m[3:len(m)-3], 1, 3) }) s = reB1.ReplaceAllStringFunc(s, func(m string) string { return c(m[2:len(m)-2], 1) }) s = reB2.ReplaceAllStringFunc(s, func(m string) string { return c(m[2:len(m)-2], 1) }) s = reI1.ReplaceAllStringFunc(s, func(m string) string { return c(m[1:len(m)-1], 3) }) s = reI2.ReplaceAllStringFunc(s, func(m string) string { return c(m[1:len(m)-1], 3) }) s = reS.ReplaceAllStringFunc(s, func(m string) string { return c(m[2:len(m)-2], 9) }) for i, code := range codes { s = strings.ReplaceAll(s, fmt.Sprintf("\x00CD%d\x00", i), code) } return s } func renderMDLine(line string, st *mdState) string { if !COL { return line } trimmed := strings.TrimSpace(line) if strings.HasPrefix(trimmed, "```") { if !st.inCode { st.inCode = true st.lang = strings.TrimSpace(strings.TrimPrefix(trimmed, "```")) title := "" if st.lang != "" { title = " [ " + st.lang + " ]" } return c("───"+title+"──────────────────────────────────────────", 2) } st.inCode = false st.lang = "" return c("───────────────────────────────────────────────────", 2) } if st.inCode { return c(" ", 2) + c(line, 32) } if trimmed == "---" || trimmed == "***" || trimmed == "___" || trimmed == "----" || trimmed == "------" { return c("───────────────────────────────────────────────────", 2) } if strings.HasPrefix(trimmed, "#") { lvl := 0 for lvl < len(trimmed) && trimmed[lvl] == '#' { lvl++ } if lvl < len(trimmed) && trimmed[lvl] == ' ' { htext := strings.TrimSpace(trimmed[lvl:]) switch lvl { case 1: return c("■ ", 35) + c(htext, 1, 37) case 2: return c("▲ ", 34) + c(htext, 1, 36) case 3: return c("● ", 32) + c(htext, 1, 32) case 4: return c("◆ ", 33) + c(htext, 1, 33) default: return c(htext, 1) } } } if strings.HasPrefix(trimmed, ">") { qtext := strings.TrimSpace(strings.TrimPrefix(trimmed, ">")) return c("▎ ", 34) + c(renderInline(qtext), 3) } if strings.HasPrefix(trimmed, "- [ ] ") || strings.HasPrefix(trimmed, "* [ ] ") { return strings.Repeat(" ", len(line)-len(strings.TrimLeft(line, " "))) + c("☐ ", 33) + renderInline(trimmed[6:]) } if strings.HasPrefix(trimmed, "- [x] ") || strings.HasPrefix(trimmed, "* [x] ") || strings.HasPrefix(trimmed, "- [X] ") || strings.HasPrefix(trimmed, "* [X] ") { return strings.Repeat(" ", len(line)-len(strings.TrimLeft(line, " "))) + c("☑ ", 32) + renderInline(trimmed[6:]) } if strings.HasPrefix(trimmed, "- ") || strings.HasPrefix(trimmed, "* ") || strings.HasPrefix(trimmed, "+ ") { indent := strings.Repeat(" ", len(line)-len(strings.TrimLeft(line, " "))) return indent + c("• ", 36) + renderInline(trimmed[2:]) } if m := reOrd.FindStringSubmatch(trimmed); len(m) == 3 { indent := strings.Repeat(" ", len(line)-len(strings.TrimLeft(line, " "))) return indent + c(m[1]+" ", 33) + renderInline(m[2]) } if strings.HasPrefix(trimmed, "|") && strings.HasSuffix(trimmed, "|") { return renderInline(line) } return renderInline(line) } func isTableSep(cells []string) bool { if len(cells) == 0 { return false } for _, cell := range cells { c := strings.ReplaceAll(strings.ReplaceAll(strings.TrimSpace(cell), "-", ""), ":", "") if c != "" { return false } } return true } func parseTableCells(line string) []string { trimmed := strings.TrimSpace(line) if strings.HasPrefix(trimmed, "|") { trimmed = trimmed[1:] } if strings.HasSuffix(trimmed, "|") { trimmed = trimmed[:len(trimmed)-1] } parts := strings.Split(trimmed, "|") cells := make([]string, len(parts)) for i, p := range parts { cells[i] = strings.TrimSpace(p) } return cells } func isTableLine(line string) bool { trimmed := strings.TrimSpace(line) if !strings.Contains(trimmed, "|") { return false } return strings.HasPrefix(trimmed, "|") || strings.HasSuffix(trimmed, "|") } func wrapCell(s string, width int) []string { if width <= 0 { return []string{""} } s = strings.ReplaceAll(s, "
", "\n") s = strings.ReplaceAll(s, "
", "\n") s = strings.ReplaceAll(s, "
", "\n") var lines []string for _, p := range strings.Split(s, "\n") { p = strings.TrimSpace(p) if p == "" { lines = append(lines, "") continue } if visibleLen(p) <= width { lines = append(lines, p) continue } words := strings.Fields(p) if len(words) == 0 { lines = append(lines, "") continue } var curLine string var activeEsc string updateActiveEsc := func(token string) { for i := 0; i < len(token); { if token[i] == 0x1b { j := i + 1 if j < len(token) && token[j] == '[' { j++ for j < len(token) && !(token[j] >= 0x40 && token[j] <= 0x7e) { j++ } if j < len(token) { j++ } } esc := token[i:j] if esc == "\033[0m" || esc == "\033[m" { activeEsc = "" } else { activeEsc = esc } i = j } else { i++ } } } flushLine := func() { if curLine != "" { out := curLine if COL && activeEsc != "" && !strings.HasSuffix(out, "\033[0m") { out += "\033[0m" } lines = append(lines, out) curLine = "" } } for _, word := range words { wLen := visibleLen(word) if wLen > width { flushLine() var chunk strings.Builder cLen := 0 for i := 0; i < len(word); { if word[i] == 0x1b { j := i + 1 if j < len(word) && word[j] == '[' { j++ for j < len(word) && !(word[j] >= 0x40 && word[j] <= 0x7e) { j++ } if j < len(word) { j++ } } chunk.WriteString(word[i:j]) updateActiveEsc(word[i:j]) i = j continue } r, size := utf8.DecodeRuneInString(word[i:]) if cLen >= width { if COL && activeEsc != "" { chunk.WriteString("\033[0m") } lines = append(lines, chunk.String()) chunk.Reset() if COL && activeEsc != "" { chunk.WriteString(activeEsc) } cLen = 0 } chunk.WriteRune(r) cLen++ i += size } if chunk.Len() > 0 { curLine = chunk.String() } continue } cLen := visibleLen(curLine) if curLine == "" { curLine = word updateActiveEsc(word) } else if cLen+1+wLen <= width { curLine += " " + word updateActiveEsc(word) } else { flushLine() if COL && activeEsc != "" { curLine = activeEsc + word } else { curLine = word } updateActiveEsc(word) } } flushLine() } if len(lines) == 0 { lines = []string{""} } return lines } func renderTable(lines []string) []string { if len(lines) == 0 { return nil } var rows [][]string var headerRow []string hasHeader := false for _, ln := range lines { cells := parseTableCells(ln) if isTableSep(cells) { if len(rows) > 0 && !hasHeader { headerRow = rows[len(rows)-1] rows = rows[:len(rows)-1] hasHeader = true } continue } rows = append(rows, cells) } if !hasHeader && len(lines) < 2 { var out []string var st mdState for _, ln := range lines { out = append(out, renderMDLine(ln, &st)) } return out } numCols := len(headerRow) for _, r := range rows { if len(r) > numCols { numCols = len(r) } } if numCols == 0 { return nil } if hasHeader { for len(headerRow) < numCols { headerRow = append(headerRow, "") } } for i := range rows { for len(rows[i]) < numCols { rows[i] = append(rows[i], "") } } colWidths := make([]int, numCols) for i := 0; i < numCols; i++ { if hasHeader { vl := visibleLen(headerRow[i]) if vl > colWidths[i] { colWidths[i] = vl } } for _, r := range rows { vl := visibleLen(renderInline(r[i])) if vl > colWidths[i] { colWidths[i] = vl } } if colWidths[i] < 3 { colWidths[i] = 3 } } maxTableWidth := termWidth() if maxTableWidth < 20 { maxTableWidth = 80 } overhead := 3*numCols + 1 availContent := maxTableWidth - overhead if availContent < numCols*3 { availContent = numCols * 3 } tot := 0 for _, w := range colWidths { tot += w } for tot > availContent { maxIdx := 0 maxVal := colWidths[0] for i := 1; i < numCols; i++ { if colWidths[i] > maxVal { maxVal = colWidths[i] maxIdx = i } } if maxVal <= 3 { break } colWidths[maxIdx]-- tot-- } var res []string var topParts []string for _, w := range colWidths { topParts = append(topParts, strings.Repeat("─", w+2)) } res = append(res, c("┌"+strings.Join(topParts, "┬")+"┐", 2)) if hasHeader { headerCols := make([][]string, numCols) maxHeaderLines := 1 for i, h := range headerRow { wrapped := wrapCell(h, colWidths[i]) if len(wrapped) > maxHeaderLines { maxHeaderLines = len(wrapped) } headerCols[i] = wrapped } for lineIdx := 0; lineIdx < maxHeaderLines; lineIdx++ { var hCells []string for i := 0; i < numCols; i++ { txt := "" if lineIdx < len(headerCols[i]) { txt = headerCols[i][lineIdx] } rh := c(txt, 1, 36) pad := strings.Repeat(" ", colWidths[i]-visibleLen(txt)) hCells = append(hCells, " "+rh+pad+" ") } res = append(res, c("│", 2)+strings.Join(hCells, c("│", 2))+c("│", 2)) } var midParts []string for _, w := range colWidths { midParts = append(midParts, strings.Repeat("─", w+2)) } res = append(res, c("├"+strings.Join(midParts, "┼")+"┤", 2)) } for _, r := range rows { rowCols := make([][]string, numCols) maxRowLines := 1 for i, cell := range r { rc := renderInline(cell) wrapped := wrapCell(rc, colWidths[i]) if len(wrapped) > maxRowLines { maxRowLines = len(wrapped) } rowCols[i] = wrapped } for lineIdx := 0; lineIdx < maxRowLines; lineIdx++ { var rCells []string for i := 0; i < numCols; i++ { txt := "" if lineIdx < len(rowCols[i]) { txt = rowCols[i][lineIdx] } pad := strings.Repeat(" ", colWidths[i]-visibleLen(txt)) rCells = append(rCells, " "+txt+pad+" ") } res = append(res, c("│", 2)+strings.Join(rCells, c("│", 2))+c("│", 2)) } } var botParts []string for _, w := range colWidths { botParts = append(botParts, strings.Repeat("─", w+2)) } res = append(res, c("└"+strings.Join(botParts, "┴")+"┘", 2)) return res } func renderMD(text string) string { if !COL { return text } lines := strings.Split(text, "\n") var out []string var st mdState var tbl []string flushTable := func() { if len(tbl) > 0 { out = append(out, renderTable(tbl)...) tbl = nil } } for _, ln := range lines { trimmed := strings.TrimSpace(ln) if strings.HasPrefix(trimmed, "```") { flushTable() out = append(out, renderMDLine(ln, &st)) continue } if st.inCode { out = append(out, renderMDLine(ln, &st)) continue } if isTableLine(ln) { tbl = append(tbl, ln) continue } flushTable() out = append(out, renderMDLine(ln, &st)) } flushTable() return strings.Join(out, "\n") } type Message struct { Role string `json:"role"` Content *string `json:"content"` ReasoningContent string `json:"reasoning_content,omitempty"` ToolCalls []ToolCall `json:"tool_calls,omitempty"` ToolCallID string `json:"tool_call_id,omitempty"` ToolCallsShown bool `json:"-"` // tool call headers already printed live during streaming } type ToolCall struct { ID string `json:"id"` Type string `json:"type"` Function struct { Name string `json:"name"` Arguments string `json:"arguments"` } `json:"function"` } var TOOLS = []map[string]any{ {"type": "function", "function": map[string]any{"name": "shell_exec", "description": "Run a shell command, return output and exit code.", "parameters": map[string]any{"type": "object", "properties": map[string]any{"command": map[string]any{"type": "string"}}, "required": []string{"command"}}}}, {"type": "function", "function": map[string]any{"name": "write_file", "description": "Write content to a file. If offset and del_bytes are both omitted, the entire file is overwritten with the new content; otherwise content is written at the given byte offset, optionally deleting bytes first.", "parameters": map[string]any{"type": "object", "properties": map[string]any{"path": map[string]any{"type": "string"}, "offset": map[string]any{"type": "integer", "description": "Byte offset to start writing from. If omitted together with del_bytes, the whole file is overwritten instead. Defaults to 0 (start of file)."}, "del_bytes": map[string]any{"type": "integer", "description": "Bytes to delete starting at offset. If omitted together with offset, the whole file is overwritten instead."}, "content": map[string]any{"type": "string"}}, "required": []string{"path", "content"}}}}, } func strp(s string) *string { return &s } type Usage struct { PromptTokens int `json:"prompt_tokens"` CompletionTokens int `json:"completion_tokens"` TotalTokens int `json:"total_tokens"` PromptTokensDetails struct { CachedTokens int `json:"cached_tokens"` } `json:"prompt_tokens_details"` CachedTokens int `json:"cached_tokens"` // ContextTokens is a local estimate of the total tokens occupying the context // window for this sample, including any reasoning_content that is resent as // part of the context. It acts as a floor beneath PromptTokens so reasoning // tokens are always counted in the context window statistics even when the // provider under-reports prompt_tokens. ContextTokens int `json:"-"` Model string `json:"-"` } func (u Usage) Cached() int { if u.PromptTokensDetails.CachedTokens > 0 { return u.PromptTokensDetails.CachedTokens } return u.CachedTokens } // estTokens is a coarse chars/4 fallback used only when the provider omits // usage in its response; when real usage is present it is never used. func estTokens(msgs []Message) int { chars := 0 for _, m := range msgs { if m.Content != nil { chars += len(*m.Content) } chars += len(m.ReasoningContent) for _, tc := range m.ToolCalls { chars += len(tc.Function.Name) + len(tc.Function.Arguments) } } if chars == 0 { return 0 } t := chars / 4 if t == 0 { t = 1 } return t } // ctxTokens returns the number of tokens occupying the context window for a // usage sample. We prefer the provider's reported prompt token count, but never // go below a local estimate of the messages we actually sent (which includes any // reasoning_content resent as context) so reasoning tokens are always counted in // the context window statistics. func ctxTokens(u Usage) int { if u.ContextTokens > u.PromptTokens { return u.ContextTokens } return u.PromptTokens } func contextPct(u Usage, cw int) float64 { if cw <= 0 { cw = 262144 } return float64(ctxTokens(u)) * 100.0 / float64(cw) } func formatUsage(u Usage, cw int) string { base := ctxTokens(u) pct := float64(base) * 100.0 / float64(cw) cached := u.Cached() if cached > 0 { uncached := u.PromptTokens - cached if uncached < 0 { uncached = 0 } if u.Model != "" { return fmt.Sprintf("[%s: %d prompt (%d cached, %d uncached) + %d completion | context: %d/%d (%.1f%%)]", u.Model, u.PromptTokens, cached, uncached, u.CompletionTokens, base, cw, pct) } return fmt.Sprintf("[%d prompt (%d cached, %d uncached) + %d completion | context: %d/%d (%.1f%%)]", u.PromptTokens, cached, uncached, u.CompletionTokens, base, cw, pct) } if u.Model != "" { return fmt.Sprintf("[%s: %d prompt + %d completion | context: %d/%d (%.1f%%)]", u.Model, u.PromptTokens, u.CompletionTokens, base, cw, pct) } return fmt.Sprintf("[%d prompt + %d completion | context: %d/%d (%.1f%%)]", u.PromptTokens, u.CompletionTokens, base, cw, pct) } type streamDelta struct { Model string `json:"model"` Choices []struct { Delta struct { ReasoningContent string `json:"reasoning_content"` Reasoning string `json:"reasoning"` Thought string `json:"thought"` Content string `json:"content"` ToolCalls []struct { Index int `json:"index"` ID string `json:"id"` Function struct { Name string `json:"name"` Arguments string `json:"arguments"` } `json:"function"` } `json:"tool_calls"` } `json:"delta"` } `json:"choices"` Usage *Usage `json:"usage"` } func cleanMessagesForLLM(msgs []Message) []Message { out := make([]Message, len(msgs)) for i, m := range msgs { out[i] = Message{ Role: m.Role, Content: m.Content, ReasoningContent: m.ReasoningContent, ToolCalls: m.ToolCalls, ToolCallID: m.ToolCallID, } } return out } func filterText(s string) string { var b strings.Builder b.Grow(len(s)) for _, r := range s { if r == ' ' || r == '\t' || r == '\n' { b.WriteRune(r) } else if unicode.IsPrint(r) { b.WriteRune(r) } } return b.String() } func sanitizeMessages(msgs []Message) { for i := range msgs { if msgs[i].Role == "assistant" { if len(msgs[i].ToolCalls) > 0 { for j := range msgs[i].ToolCalls { tc := &msgs[i].ToolCalls[j] tc.Function.Arguments = filterText(tc.Function.Arguments) astr := tc.Function.Arguments var a map[string]any if err := json.Unmarshal([]byte(astr), &a); err != nil || a == nil { fixed, _ := json.Marshal(map[string]string{"invalid_raw": astr}) tc.Function.Arguments = string(fixed) } } } if msgs[i].ReasoningContent != "" { msgs[i].ReasoningContent = filterText(msgs[i].ReasoningContent) } } else if msgs[i].Role == "tool" && msgs[i].Content != nil { msgs[i].Content = strp(filterText(*msgs[i].Content)) } } } func isInvalidAssistantErr(err error) bool { if err == nil { return false } s := strings.ToLower(err.Error()) return strings.Contains(s, "invalid assistant message") || strings.Contains(s, "content or tool_calls must be set") || strings.Contains(s, "tool_calls must be set") || strings.Contains(s, "content must be set") } func llm(ctx context.Context, cfg *Cfg, msgs []Message, tools []map[string]any) (Message, Usage, error) { if err := ctx.Err(); err != nil { return Message{}, Usage{}, err } cleanMsgs := cleanMessagesForLLM(msgs) sanitizeMessages(cleanMsgs) p := map[string]any{"model": cfg.Model, "temperature": cfg.Temperature, "messages": cleanMsgs, "stream": cfg.Stream} if tools != nil { p["tools"] = tools } if cfg.Stream { p["stream_options"] = map[string]any{"include_usage": true} } for k, v := range cfg.Raw { if internalKey(k) { continue } { var jv any if err := json.Unmarshal([]byte(v), &jv); err == nil { p[k] = jv } else { p[k] = v } } } body, _ := json.Marshal(p) llmTransport.ResponseHeaderTimeout = time.Duration(cfg.Timeout) * time.Second client := &http.Client{Transport: llmTransport} fib := []int{1, 1, 2, 3, 5, 8, 13, 21, 34} pend := c("...requesting...", 1, 2) var resp *http.Response var err error for i := 0; i <= len(fib); i++ { if err := ctx.Err(); err != nil { if COL { fmt.Print("\r\033[K") } return Message{}, Usage{}, err } if COL { fmt.Print("\r" + pend) } else { fmt.Println(pend) } req, _ := http.NewRequestWithContext(ctx, "POST", strings.TrimRight(cfg.Endpoint, "/")+"/chat/completions", bytes.NewReader(body)) req.Header.Set("Content-Type", "application/json") req.Header.Set("User-Agent", opencodeAgentVersion) req.Header.Set("X-Session-Id", projectID()) req.Header.Set("x-opencode-project", projectID()) req.Header.Set("x-opencode-client", "cli") req.Header.Set("x-opencode-session", opencodeSessionID()) if cfg.APIKey != "" && cfg.APIKey != "-" { req.Header.Set("Authorization", "Bearer "+cfg.APIKey) } resp, err = client.Do(req) var is4xxClientErr bool if err == nil && resp.StatusCode >= 400 { b, _ := io.ReadAll(io.LimitReader(resp.Body, 4096)) resp.Body.Close() err = fmt.Errorf("HTTP %d: %s", resp.StatusCode, strings.TrimSpace(string(b))) if resp.StatusCode < 500 && resp.StatusCode != 408 && resp.StatusCode != 429 { is4xxClientErr = true } resp = nil } if err == nil { break } if COL { fmt.Print("\r\033[K") } if errors.Is(err, context.Canceled) || ctx.Err() != nil { return Message{}, Usage{}, ctx.Err() } if is4xxClientErr { return Message{}, Usage{}, err } if i < len(fib) { fmt.Println(c(fmt.Sprintf("[network error: %v, retrying in %ds...]", err, fib[i]), 31)) select { case <-ctx.Done(): return Message{}, Usage{}, ctx.Err() case <-time.After(time.Duration(fib[i]) * time.Second): } } } if err != nil { if COL { fmt.Print("\r\033[K") } return Message{}, Usage{}, err } defer resp.Body.Close() if COL { fmt.Print("\r\033[K") } if !cfg.Stream { var cr struct { Model string `json:"model"` Choices []struct { Message struct { Message Reasoning string `json:"reasoning"` Thought string `json:"thought"` } `json:"message"` } `json:"choices"` Usage Usage `json:"usage"` } if err := json.NewDecoder(resp.Body).Decode(&cr); err != nil { if errors.Is(err, context.Canceled) || ctx.Err() != nil { return Message{}, Usage{}, ctx.Err() } return Message{}, Usage{}, err } if len(cr.Choices) == 0 { return Message{}, Usage{}, errors.New("empty choices in LLM response") } m := cr.Choices[0].Message.Message if m.ReasoningContent == "" { m.ReasoningContent = cr.Choices[0].Message.Reasoning } if m.ReasoningContent == "" { m.ReasoningContent = cr.Choices[0].Message.Thought } u := cr.Usage if cr.Model != "" { u.Model = cr.Model } // Local estimate of the full context (including any resent reasoning) so it // is counted in the context window statistics as a floor under prompt_tokens. u.ContextTokens = estTokens(msgs) if u.PromptTokens == 0 { u.PromptTokens = estTokens(msgs) u.CompletionTokens = estTokens([]Message{m}) u.TotalTokens = u.PromptTokens + u.CompletionTokens } return m, u, nil } m, u, err := parseStream(ctx, resp.Body) if err == nil && u.PromptTokens == 0 { u.PromptTokens = estTokens(msgs) u.CompletionTokens = estTokens([]Message{m}) u.TotalTokens = u.PromptTokens + u.CompletionTokens } if err == nil { // Local estimate of the full context (including any resent reasoning) so it // is counted in the context window statistics as a floor under prompt_tokens. u.ContextTokens = estTokens(msgs) } return m, u, err } func parseStream(ctx context.Context, r io.Reader) (Message, Usage, error) { var content, reas string var inReasoning bool var lineBuf string var mdSt mdState var tblBuf []string var lastUsage Usage tcs := map[int]*ToolCall{} var order []int var toolCallsShown bool // Progressive, per-token display of tool calls as they stream in, so the // user watches the invocation being shaped token by token (like reasoning // text) instead of a blank pause. Only used when stdout is a terminal; // piped/redirected output is left untouched. Each fragment is appended // directly to the terminal; the closing ")" is emitted only when the call is // finalized, so the visible text is always a true prefix of the final form. tty := isTerminal(int(os.Stdout.Fd())) liveIdx := -1 livePrinted := "" liveTarget := func(t *ToolCall, closed bool) string { s := "[tool call: " + t.Function.Name if closed || t.Function.Arguments != "" { s += "(" + filterText(t.Function.Arguments) if closed { s += ")]" } } return s } finishLive := func(closed bool) { if liveIdx < 0 { return } target := liveTarget(tcs[liveIdx], closed) if len(target) > len(livePrinted) { fmt.Print(c(target[len(livePrinted):], 33)) } if closed { fmt.Print("\n") } livePrinted = "" liveIdx = -1 } stepLive := func() { if !tty || len(order) == 0 { return } cur := order[len(order)-1] if liveIdx != -1 && liveIdx != cur { // A different tool call is now being shaped: finalize the previous one. finishLive(true) } if liveIdx == -1 { liveIdx = cur livePrinted = "" } target := liveTarget(tcs[cur], false) if len(target) > len(livePrinted) { fmt.Print(c(target[len(livePrinted):], 33)) livePrinted = target } toolCallsShown = true } flushTable := func() { if len(tblBuf) > 0 { for _, tln := range renderTable(tblBuf) { fmt.Println(tln) } tblBuf = nil } } sc := bufio.NewScanner(r) sc.Buffer(make([]byte, 0, 64*1024), 4*1024*1024) for sc.Scan() { if err := ctx.Err(); err != nil { return Message{}, lastUsage, err } ln := strings.TrimSpace(sc.Text()) if !strings.HasPrefix(ln, "data:") { continue } data := strings.TrimSpace(ln[5:]) if data == "[DONE]" { break } var d streamDelta if json.Unmarshal([]byte(data), &d) != nil { continue } if d.Usage != nil && (d.Usage.PromptTokens > 0 || d.Usage.TotalTokens > 0) { lastUsage = *d.Usage } if d.Model != "" { lastUsage.Model = d.Model } if len(d.Choices) == 0 { continue } dl := d.Choices[0].Delta rc := dl.ReasoningContent if rc == "" { rc = dl.Reasoning } if rc == "" { rc = dl.Thought } if rc != "" { if !inReasoning { if content != "" { flushTable() if lineBuf != "" { fmt.Println(renderMDLine(lineBuf, &mdSt)); lineBuf = "" } fmt.Println() } fmt.Println(c("--- reasoning start ---", 36)) inReasoning = true } fmt.Print(c(rc, 2)) reas += rc } if dl.Content != "" { if inReasoning { fmt.Print("\n" + c("--- reasoning end ---", 36) + "\n\n") inReasoning = false } content += dl.Content lineBuf += dl.Content for { idx := strings.IndexByte(lineBuf, '\n') if idx == -1 { break } curLine := lineBuf[:idx] lineBuf = lineBuf[idx+1:] trimmed := strings.TrimSpace(curLine) if strings.HasPrefix(trimmed, "```") { flushTable() fmt.Println(renderMDLine(curLine, &mdSt)) } else if mdSt.inCode { fmt.Println(renderMDLine(curLine, &mdSt)) } else if isTableLine(curLine) { tblBuf = append(tblBuf, curLine) } else { flushTable() fmt.Println(renderMDLine(curLine, &mdSt)) } } } if len(dl.ToolCalls) > 0 { if inReasoning { fmt.Print("\n" + c("--- reasoning end ---", 36) + "\n") inReasoning = false } if liveIdx == -1 { // First tool call of the stream: flush any pending content line so it // is not overwritten by the live tool-call line. flushTable() if lineBuf != "" { fmt.Println(renderMDLine(lineBuf, &mdSt)) lineBuf = "" } } for _, tc := range dl.ToolCalls { t, ok := tcs[tc.Index] if !ok { t = &ToolCall{Type: "function"} tcs[tc.Index] = t order = append(order, tc.Index) } if tc.ID != "" { t.ID = tc.ID } if tc.Function.Name != "" { t.Function.Name += tc.Function.Name } if tc.Function.Arguments != "" { t.Function.Arguments += tc.Function.Arguments } } stepLive() } } if err := ctx.Err(); err != nil { return Message{}, lastUsage, err } flushTable() if lineBuf != "" { if !mdSt.inCode && isTableLine(lineBuf) { tblBuf = append(tblBuf, lineBuf) flushTable() } else { fmt.Println(renderMDLine(lineBuf, &mdSt)) } } if liveIdx >= 0 { // Finalize the in-progress tool-call line so following output starts fresh. finishLive(true) } if inReasoning { fmt.Print("\n" + c("--- reasoning end ---", 36) + "\n") } m := Message{Role: "assistant"} m.ToolCallsShown = toolCallsShown if content != "" { m.Content = strp(content) } if reas != "" { m.ReasoningContent = reas } if len(tcs) > 0 { m.ToolCalls = make([]ToolCall, 0, len(order)) for _, idx := range order { m.ToolCalls = append(m.ToolCalls, *tcs[idx]) } } if lastUsage.TotalTokens == 0 && lastUsage.PromptTokens > 0 { lastUsage.TotalTokens = lastUsage.PromptTokens + lastUsage.CompletionTokens } return m, lastUsage, sc.Err() } func shell(ctx context.Context, cmd string, timeout int) string { cmd = filterText(cmd) cmdCtx, cancel := context.WithTimeout(ctx, time.Duration(timeout)*time.Second) defer cancel() c := exec.CommandContext(cmdCtx, "sh", "-c", cmd) c.WaitDelay = 100 * time.Millisecond out, err := c.CombinedOutput() res := strings.TrimSpace(filterText(string(out))) if ctx.Err() != nil { return "[interrupted]\n\nexit: -1" } if cmdCtx.Err() == context.DeadlineExceeded { return fmt.Sprintf("%s\n\n[shell timeout after %ds]\nexit: -1", res, timeout) } code := 0 if err != nil { if ee, ok := err.(*exec.ExitError); ok { code = ee.ExitCode() } else { code = -1 } } return fmt.Sprintf("%s\n\nexit: %d", res, code) } func last(msgs []Message) string { for i := len(msgs) - 1; i >= 0; i-- { if msgs[i].Role == "assistant" && msgs[i].Content != nil && *msgs[i].Content != "" { return *msgs[i].Content } } return "" } // lastRole returns the role of the last message in the conversation, or "" if // it is empty. It is used to detect when a turn ended on a tool result rather // than a normal assistant reply, so the conversation can be auto-continued. func lastRole(msgs []Message) string { if len(msgs) == 0 { return "" } return msgs[len(msgs)-1].Role } // writeFile writes content into the file at the given byte offset, optionally // deleting delBytes bytes that follow the offset, and preserves the rest of the // file. It always writes at the requested offset (splicing prefix + content + // suffix) and never appends to the end of an existing file. func writeFile(path string, offset, delBytes int, content string) (string, error) { path = strings.TrimSpace(path) if path == "" { return "", errors.New("path is required") } if offset < 0 { offset = 0 } if delBytes < 0 { delBytes = 0 } var data []byte if fileExists(path) { var err error data, err = os.ReadFile(path) if err != nil { return "", err } } else { dir := filepath.Dir(path) if dir != "" && dir != "." { if err := os.MkdirAll(dir, 0755); err != nil { return "", err } } } if offset > len(data) { padding := make([]byte, offset-len(data)) data = append(data, padding...) } prefix := data[:offset] var suffix []byte endDel := offset + delBytes if endDel < len(data) { suffix = data[endDel:] } contentBytes := []byte(content) newData := make([]byte, 0, len(prefix)+len(contentBytes)+len(suffix)) newData = append(newData, prefix...) newData = append(newData, contentBytes...) newData = append(newData, suffix...) if err := os.WriteFile(path, newData, 0644); err != nil { return "", err } return fmt.Sprintf("Successfully wrote %d bytes to %s", len(contentBytes), path), nil } // toolResDir returns the directory where oversized tool results are spilled. func toolResDir() string { tmp := os.Getenv("TMPDIR") if tmp == "" { tmp = "/tmp" } return filepath.Join(tmp, "bantam", "toolres") } // offloadToolResult keeps tool results that fit within maxRes bytes inline. // Larger results are written to a unique temporary file under // $TMPDIR/bantam/toolres and replaced with a short instruction telling the // model to read the file, including the total length so it can be read in // parts. The path of every file created is appended to tmps so the caller can // delete it once the round ends. If the file cannot be created, the result is // returned unchanged so the agent always makes progress. func offloadToolResult(res string, maxRes int, tmps *[]string) string { if maxRes <= 0 { maxRes = 15000 } if len(res) <= maxRes { return res } dir := toolResDir() if err := os.MkdirAll(dir, 0755); err != nil { return res } f, err := os.CreateTemp(dir, "toolres-*.txt") if err != nil { return res } name := f.Name() n, werr := f.WriteString(res) if cerr := f.Close(); werr == nil { werr = cerr } if werr != nil { os.Remove(name) return res } *tmps = append(*tmps, name) return fmt.Sprintf("[tool result too large for context: %d bytes (limit %d). The full output was saved to %s. Read it with shell_exec; to read it partially, use a byte offset and length, e.g. `tail -c +OFFSET %s | head -c LENGTH`. The file is %d bytes long.]", n, maxRes, name, name, n) } // cleanupTemps removes the temporary files created while offloading oversized // tool results during a single agent round. func cleanupTemps(files []string) { for _, f := range files { os.Remove(f) } } func AL(ctx context.Context, cfg *Cfg, msgs []Message) ([]Message, Usage, error) { done := false var turnUsage Usage var tmps []string defer func() { cleanupTemps(tmps) }() for i := 0; i < cfg.MaxALIterations && !done; i++ { if err := ctx.Err(); err != nil { return msgs, turnUsage, err } m, u, err := llm(ctx, cfg, msgs, TOOLS) if err != nil { if errors.Is(err, context.Canceled) || ctx.Err() != nil { return msgs, turnUsage, err } if isInvalidAssistantErr(err) { stripped := false for j := len(msgs) - 1; j >= 0; j-- { if msgs[j].Role == "assistant" { fmt.Println(c("[stripped malformed assistant message]", 33)) msgs = append(msgs[:j], msgs[j+1:]...) stripped = true break } } if stripped { continue } } return msgs, turnUsage, err } turnUsage.PromptTokens = u.PromptTokens turnUsage.CompletionTokens += u.CompletionTokens turnUsage.TotalTokens += u.TotalTokens // The last call sends the full conversation, so its context estimate is the // authoritative one for the turn's context window statistics. turnUsage.ContextTokens = u.ContextTokens if u.Cached() > 0 { turnUsage.CachedTokens = u.Cached() } if u.Model != "" { turnUsage.Model = u.Model } for j := range m.ToolCalls { tc := &m.ToolCalls[j] tc.Function.Arguments = filterText(tc.Function.Arguments) astr := tc.Function.Arguments var a map[string]any if err := json.Unmarshal([]byte(astr), &a); err != nil || a == nil { fixed, _ := json.Marshal(map[string]string{"invalid_raw": astr}) tc.Function.Arguments = string(fixed) } } msgs = append(msgs, m) if !cfg.Stream { if m.ReasoningContent != "" { fmt.Println(c("--- reasoning start ---", 36) + "\n" + c(m.ReasoningContent, 2) + "\n" + c("--- reasoning end ---", 36)) } if m.Content != nil { fmt.Println(renderMD(*m.Content)) } } if len(m.ToolCalls) == 0 { // If the model returned only a reasoning block with no non-reasoning // tokens or tool calls, nudge it to continue rather than ending the turn. if m.ReasoningContent != "" && (m.Content == nil || strings.TrimSpace(*m.Content) == "") { fmt.Println(c("[auto continue: response was reasoning-only]", 33)) msgs = append(msgs, Message{Role: "user", Content: strp("continue")}) continue } done = true break } for _, tc := range m.ToolCalls { if err := ctx.Err(); err != nil { return msgs, turnUsage, err } fn, astr := tc.Function.Name, filterText(tc.Function.Arguments) if !m.ToolCallsShown { fmt.Println(c(fmt.Sprintf("[tool call: %s(%s)]", fn, astr), 33)) } res, sty := "", 2 var a map[string]any if err := json.Unmarshal([]byte(astr), &a); err != nil || a == nil { res, sty = fmt.Sprintf("[tool error: invalid JSON args for %s: %v. Raw: %q]", fn, err, astr), 31 } else { switch fn { case "shell_exec": cmd, _ := a["command"].(string) cmd = filterText(cmd) res = shell(ctx, cmd, cfg.ShellTimeout) if err := ctx.Err(); err != nil { return msgs, turnUsage, err } case "write_file": path, _ := a["path"].(string) path = filterText(path) contentVal, hasContent := a["content"] var content string if hasContent && contentVal != nil { if s, ok := contentVal.(string); ok { content = s } } hasOffset := false offset := 0 if v, ok := a["offset"]; ok && v != nil { hasOffset = true switch n := v.(type) { case float64: offset = int(n) case int: offset = n case string: offset = atoiD(n, 0) } } if offset < 0 { offset = 0 } hasDel := false delBytes := 0 if v, ok := a["del_bytes"]; ok { hasDel = true switch n := v.(type) { case float64: delBytes = int(n) case int: delBytes = n case string: delBytes = atoiD(n, 0) } } if strings.TrimSpace(path) == "" { res, sty = "[tool error: write_file requires 'path' parameter]", 31 } else if !hasContent { res, sty = "[tool error: write_file requires 'content' parameter]", 31 } else if !hasOffset && !hasDel { // Neither offset nor del_bytes was supplied: overwrite the entire // file with the new content. The LLM usually just wants to replace a // file and should not have to know about the tool's offset quirks; // insertion/replace semantics are preserved when either is given. p := strings.TrimSpace(path) if dir := filepath.Dir(p); dir != "" && dir != "." { if err := os.MkdirAll(dir, 0755); err != nil { res, sty = fmt.Sprintf("[tool error: write_file %s: %v]", p, err), 31 } } if res == "" { if err := os.WriteFile(p, []byte(content), 0644); err != nil { res, sty = fmt.Sprintf("[tool error: write_file %s: %v]", p, err), 31 } else { res, sty = fmt.Sprintf("Successfully wrote %d bytes to %s", len(content), p), 2 } } } else { out, err := writeFile(path, offset, delBytes, content) if err != nil { res, sty = fmt.Sprintf("[tool error: write_file %s: %v]", path, err), 31 } else { res, sty = out, 2 } } default: res, sty = "Unknown tool: " + fn, 31 } } res = filterText(res) ctxRes := offloadToolResult(res, cfg.MaxToolRes, &tmps) fmt.Println(c("[tool result: "+fn+"]", 32) + "\n" + c(ctxRes, sty) + "\n") msgs = append(msgs, Message{Role: "tool", ToolCallID: tc.ID, Content: strp(ctxRes)}) } } if !done { msgs = append(msgs, Message{Role: "assistant", Content: strp(fmt.Sprintf("[max AL iterations (%d) reached]", cfg.MaxALIterations))}) } return msgs, turnUsage, nil } func homeDir() string { if h, err := os.UserHomeDir(); err == nil && h != "" { return h } return "." } func sdir() string { d := filepath.Join(homeDir(), ".bantam", "sessions") os.MkdirAll(d, 0755) return d } type Session struct { ID string `json:"id"` Created string `json:"created"` Summary string `json:"summary"` Messages []Message `json:"messages"` } func summary(msgs []Message) string { for _, m := range msgs { if m.Role == "user" && m.Content != nil && strings.TrimSpace(*m.Content) != "" { t := strings.Join(strings.Fields(*m.Content), " ") if len(t) > 80 { t = t[:80] + "..." } return t } } return "(empty session)" } func fileExists(p string) bool { _, err := os.Stat(p) return err == nil } func projectID() string { dir, err := os.Getwd() if err != nil { dir = "." } if abs, err := filepath.Abs(dir); err == nil { dir = abs } h := md5.Sum([]byte(dir)) return hex.EncodeToString(h[:]) } func opencodeSessionID() string { pid := projectID() if len(pid) > 26 { pid = pid[:26] } else if len(pid) < 26 { pid = (pid + "0123456789abcdef0123456789")[:26] } return "ses_" + pid } func saveSession(msgs []Message, sid string) (string, string) { sid = strings.TrimSpace(sid) if sid == "" { sid = projectID() } d := sdir() path := filepath.Join(d, sid+".json") s := Session{sid, time.Now().Format("2006-01-02 15:04:05"), summary(msgs), msgs} b, err := json.MarshalIndent(s, "", " ") if err != nil { fmt.Fprintf(os.Stderr, "saveSession: marshal error: %v\n", err) return sid, s.Summary } if err := os.WriteFile(path, b, 0644); err != nil { fmt.Fprintf(os.Stderr, "saveSession: write error: %v\n", err) } return sid, s.Summary } func sessions() []Session { out := []Session{} d := filepath.Join(homeDir(), ".bantam", "sessions") entries, err := os.ReadDir(d) if err != nil { return out } for _, e := range entries { if e.IsDir() || !strings.HasSuffix(e.Name(), ".json") { continue } b, err := os.ReadFile(filepath.Join(d, e.Name())) if err != nil { continue } var s Session if json.Unmarshal(b, &s) != nil { continue } if s.ID == "" { s.ID = strings.TrimSuffix(e.Name(), ".json") } out = append(out, s) } sort.Slice(out, func(i, j int) bool { return out[i].ID > out[j].ID }) return out } func loadSession(sid string) ([]Message, error) { ss := sessions() for i := range ss { if ss[i].ID == sid { return ss[i].Messages, nil } } var pref []*Session for i := range ss { if strings.HasPrefix(ss[i].ID, sid) { pref = append(pref, &ss[i]) } } if len(pref) == 1 { return pref[0].Messages, nil } if len(pref) > 1 { names := make([]string, len(pref)) for i, s := range pref { names[i] = s.ID } return nil, fmt.Errorf("ambiguous prefix: %s", strings.Join(names, ", ")) } return nil, fmt.Errorf("session not found: %s", sid) } // findProjectSession returns the saved session for the current working // directory (if any), or nil when none exists yet. func findProjectSession() *Session { pid := projectID() ss := sessions() for i := range ss { if ss[i].ID == pid { return &ss[i] } } return nil } func autosave(msgs []Message) { saveSession(msgs, projectID()) } const compactionPrompt = "You are now acting as a compaction engine. Summarize the preceding conversation concisely but completely, preserving all important facts, decisions, code snippets, tool outputs, errors, and current task state so work can seamlessly continue. Output only the summary." func compact(ctx context.Context, cfg *Cfg, msgs []Message) ([]Message, string, error) { if len(msgs) == 0 || msgs[0].Role != "system" { return msgs, "", errors.New("session has no system message") } if len(msgs) <= 1 { return msgs, "", errors.New("nothing to compact") } cMsgs := append(append([]Message{}, msgs...), Message{ Role: "user", Content: strp(compactionPrompt), }) cc := *cfg cc.Stream = false m, _, err := llm(ctx, &cc, cMsgs, nil) if err != nil { return msgs, "", err } s := "" if m.Content != nil { s = *m.Content } if s == "" { s = m.ReasoningContent } s = strings.TrimSpace(s) if s == "" { return msgs, "", errors.New("LLM returned an empty summary") } newMsgs := []Message{ {Role: "system", Content: msgs[0].Content}, {Role: "user", Content: strp("Summary of the previous conversation:\n" + s + "\n\nPlease continue from here.")}, } return newMsgs, s, nil } func summarize(ctx context.Context, cfg *Cfg, msgs []Message) (string, error) { _, s, err := compact(ctx, cfg, msgs) return s, err } func loadHistory() { histF = filepath.Join(homeDir(), ".bantam_history") b, err := os.ReadFile(histF) if err != nil { return } for _, ln := range strings.Split(string(b), "\n") { if ln = strings.TrimRight(ln, "\r"); ln != "" { hist = append(hist, ln) } } } func addHistory(s string) { if s == "" || (len(hist) > 0 && hist[len(hist)-1] == s) { return } hist = append(hist, s) } func saveHistory() { os.WriteFile(histF, []byte(strings.Join(hist, "\n")+"\n"), 0644) } func visibleLen(s string) int { n := 0 for i := 0; i < len(s); { switch s[i] { case 0x1b: j := i + 1 if j < len(s) && s[j] == '[' { j++ for j < len(s) && !(s[j] >= 0x40 && s[j] <= 0x7e) { j++ } if j < len(s) { j++ } i = j } else { i++ } case 0x01, 0x02: i++ default: _, size := utf8.DecodeRuneInString(s[i:]) n++ i += size } } return n } type editor struct { prompt string buf []rune pos int hist []string hpos int draft string crow int } func textPos(promptLen, W int, s string, pos int) (row, col int) { row, col = 0, promptLen if W < 1 { W = 1 } runes := []rune(s) n := len(runes) for i := 0; i < n; i++ { r := runes[i] if i == pos { return row, col } if r == '\n' { row++ col = 0 continue } // The rune at index i is displayed at (row, col). If it lands on the last // column, the NEXT rune wraps to the start of the following line (unless // this is the final rune, in which case the cursor stays at that column). if col >= W-1 { if i < n-1 { row++ col = 0 } } else { col++ } } return row, col } func (e *editor) draw() { W := termWidth() s := string(e.buf) P := visibleLen(e.prompt) er, _ := textPos(P, W, s, len([]rune(s))) pr, pc := textPos(P, W, s, e.pos) if e.crow > 0 { fmt.Printf("\033[%dA", e.crow) } // In raw mode OPOST is off, so a bare \n only moves down without // returning to column 0. Emit \r\n so each logical line starts at // column 0, matching the column-reset assumption in textPos(). fmt.Print("\r\033[J" + e.prompt + strings.ReplaceAll(s, "\n", "\r\n")) if up := er - pr; up > 0 { fmt.Printf("\033[%dA", up) } fmt.Print("\r") if pc > 0 { fmt.Printf("\033[%dC", pc) } e.crow = pr } func (e *editor) histNav(up bool) { if up { if len(e.hist) == 0 { return } if e.hpos < 0 { e.draft = string(e.buf) e.hpos = len(e.hist) - 1 } else if e.hpos > 0 { e.hpos-- } e.buf = []rune(e.hist[e.hpos]) } else { if e.hpos < 0 { return } e.hpos++ if e.hpos >= len(e.hist) { e.hpos = -1 e.buf = []rune(e.draft) } else { e.buf = []rune(e.hist[e.hpos]) } } e.pos = len(e.buf) } // wordBack moves the cursor back to the start of the current/previous word. func (e *editor) wordBack() { for e.pos > 0 && unicode.IsSpace(e.buf[e.pos-1]) { e.pos-- } for e.pos > 0 && !unicode.IsSpace(e.buf[e.pos-1]) { e.pos-- } } // wordFwd moves the cursor forward to the end of the current/next word. func (e *editor) wordFwd() { n := len(e.buf) for e.pos < n && unicode.IsSpace(e.buf[e.pos]) { e.pos++ } for e.pos < n && !unicode.IsSpace(e.buf[e.pos]) { e.pos++ } } // delWordBack deletes the word preceding the cursor (Emacs Ctrl+W). func (e *editor) delWordBack() { if e.pos == 0 { return } start := e.pos for start > 0 && unicode.IsSpace(e.buf[start-1]) { start-- } for start > 0 && !unicode.IsSpace(e.buf[start-1]) { start-- } e.buf = append(e.buf[:start], e.buf[e.pos:]...) e.pos = start } func readPlain(prompt string) (string, bool) { fmt.Print(prompt) line, err := stdin.ReadString('\n') if err != nil && line == "" { return "", false } return strings.TrimRight(line, "\r\n"), true } func readLine(prompt string) (string, bool) { if !isTerminal(int(os.Stdin.Fd())) { return readPlain(prompt) } restore, err := makeRaw(int(os.Stdin.Fd())) if err != nil { return readPlain(prompt) } defer restore() e := &editor{prompt: prompt, hpos: -1, hist: hist} for { e.draw() rn, _, err := stdin.ReadRune() if err != nil { fmt.Print("\r\n") return "", false } switch rn { case '\r': if stdin.Buffered() > 0 { if b, _ := stdin.Peek(1); len(b) > 0 && b[0] == '\n' { stdin.ReadByte() } } W := termWidth() s := string(e.buf) P := visibleLen(e.prompt) er, _ := textPos(P, W, s, len([]rune(s))) pr, _ := textPos(P, W, s, e.pos) if down := er - pr; down > 0 { fmt.Printf("\033[%dB", down) } fmt.Print("\r\n") return string(e.buf), true case '\n': e.buf = append(e.buf, 0) copy(e.buf[e.pos+1:], e.buf[e.pos:]) e.buf[e.pos] = '\n' e.pos++ case 0x03: fmt.Print("^C\r\n") return "", true case 0x04: if len(e.buf) == 0 { fmt.Print("\r\n") return "", false } // Ctrl+D with non-empty buffer: delete character under cursor if e.pos < len(e.buf) { e.buf = append(e.buf[:e.pos], e.buf[e.pos+1:]...) } case 0x01: e.pos = 0 // Ctrl+A: beginning of line case 0x05: e.pos = len(e.buf) // Ctrl+E: end of line case 0x02: if e.pos > 0 { e.pos-- } // Ctrl+B: backward char case 0x06: if e.pos < len(e.buf) { e.pos++ } // Ctrl+F: forward char case 0x17: e.delWordBack() // Ctrl+W: delete previous word case 0x0b: e.buf = e.buf[:e.pos] // Ctrl+K: kill to end of line case 0x15: e.buf = e.buf[e.pos:] // Ctrl+U: kill to start of line e.pos = 0 case 0x7f, 0x08: if e.pos > 0 { e.buf = append(e.buf[:e.pos-1], e.buf[e.pos:]...) e.pos-- } case 0x1b: b1, err1 := stdin.ReadByte() if err1 != nil { continue } if b1 == '[' { // Read the full CSI sequence (terminated by a byte in 0x40-0x7e). var seq []byte for { b, err := stdin.ReadByte() if err != nil { break } seq = append(seq, b) if b >= 0x40 && b <= 0x7e { break } } if len(seq) == 0 { continue } last := seq[len(seq)-1] switch last { case 'A': e.histNav(true) case 'B': e.histNav(false) case 'C': if strings.Contains(string(seq), "5") { e.wordFwd() // Ctrl+Right } else if e.pos < len(e.buf) { e.pos++ } case 'D': if strings.Contains(string(seq), "5") { e.wordBack() // Ctrl+Left } else if e.pos > 0 { e.pos-- } case 'H': e.pos = 0 // Home (ESC[H) case 'F': e.pos = len(e.buf) // End (ESC[F) case '~': // Home/End on some terminals: ESC[1~ / ESC[4~ if len(seq) >= 1 && seq[0] == '1' { e.pos = 0 } else if len(seq) >= 1 && seq[0] == '4' { e.pos = len(e.buf) } } } else if b1 == 'O' { // xterm application cursor keys: ESC O H/F (Home/End), A-D (arrows) b2, err2 := stdin.ReadByte() if err2 != nil { continue } switch b2 { case 'H': e.pos = 0 case 'F': e.pos = len(e.buf) case 'A': e.histNav(true) case 'B': e.histNav(false) case 'C': if e.pos < len(e.buf) { e.pos++ } case 'D': if e.pos > 0 { e.pos-- } } } default: if rn >= 32 { e.buf = append(e.buf, 0) copy(e.buf[e.pos+1:], e.buf[e.pos:]) e.buf[e.pos] = rn e.pos++ } } } } // skillPrompt resolves and composes the prompt for the /skill command: it reads // //SKILL.md (or an absolute path when no skills directory is set) and // prefixes its contents to the optional user prompt. func skillPrompt(u string, cfg *Cfg) (string, error) { rest := strings.TrimSpace(strings.TrimPrefix(u, "/skill")) if rest == "" { return "", fmt.Errorf("usage: /skill [prompt]") } fields := strings.SplitN(rest, " ", 2) name := fields[0] prompt := "" if len(fields) == 2 { prompt = strings.TrimSpace(fields[1]) } sd := skillsDir(cfg) var path string if sd != "" { path = filepath.Join(sd, name, "SKILL.md") } else if strings.HasSuffix(name, "SKILL.md") { path = name // absolute path to the SKILL.md file itself } else { path = filepath.Join(name, "SKILL.md") // absolute path to the skill directory } data, err := os.ReadFile(path) if err != nil { return "", fmt.Errorf("cannot read skill %q: %v", name, err) } content := strings.TrimRight(string(data), "\r\n") if prompt != "" { return content + "\n\n" + prompt, nil } return content, nil } // listSkills prints every skill found under the configured skills directory, or a // clear notice when none is configured / none are present. func listSkills(cfg *Cfg) { sd := skillsDir(cfg) if sd == "" { fmt.Println(c("No skills directory configured (set BANTAM_SKILLS_DIR or bantam_skills_dir).", 33)) return } entries, err := os.ReadDir(sd) if err != nil { fmt.Println(c("[skill error: cannot read skills dir: "+err.Error()+"]", 31)) return } var names []string for _, e := range entries { if e.IsDir() { if _, err := os.Stat(filepath.Join(sd, e.Name(), "SKILL.md")); err == nil { names = append(names, e.Name()) } } } if len(names) == 0 { fmt.Println(c("No skills found in "+sd, 33)) return } fmt.Println(c("Available skills in "+sd+":", 1, 36)) for _, n := range names { fmt.Println(c(" "+n, 32)) } } func runDirectShell(cmd string, timeout int) { cmd = filterText(strings.TrimSpace(cmd)) if cmd == "" { return } astr, _ := json.Marshal(map[string]string{"command": cmd}) fmt.Println(c(fmt.Sprintf("[tool call: shell_exec(%s)]", string(astr)), 33)) sigCtx, cancel := signal.NotifyContext(context.Background(), os.Interrupt) res := shell(sigCtx, cmd, timeout) cancel() fmt.Println(c("[tool result: shell_exec]", 32) + "\n" + c(res, 2)) } func doCompact(cfg *Cfg, msgs []Message) []Message { if len(msgs) <= 1 { fmt.Println(c("Nothing to compact yet.", 33)) return msgs } fmt.Println(c("[compacting conversation...]", 33)) sigCtx, cancel := signal.NotifyContext(context.Background(), os.Interrupt) nm, sm, err := compact(sigCtx, cfg, msgs) interrupted := sigCtx.Err() != nil cancel() if err != nil { if interrupted || errors.Is(err, context.Canceled) { fmt.Println(c("\n[interrupted]", 33)) } else { fmt.Println(c("[compact failed: "+err.Error()+"]", 31)) } return msgs } msgs = nm autosave(msgs) fmt.Println(c(fmt.Sprintf("[compacted to %d messages]", len(msgs)), 32)) fmt.Println(c("--- summary ---", 33) + "\n" + c(sm, 2)) return msgs } // buildSystemPrompt renders the built-in system prompt for cfg: it expands the // $TMPDIR/bantam placeholder to a concrete temporary directory, substitutes the // configured max_tool_res value, and appends the optional extra-tools and // skills hints. func buildSystemPrompt(cfg *Cfg) string { sp := defaultSystemPrompt tmp := os.Getenv("TMPDIR") if tmp == "" { tmp = "/tmp" } bantamTmp := filepath.Join(tmp, "bantam") sp = strings.ReplaceAll(sp, "$TMPDIR/bantam", bantamTmp) sp = strings.ReplaceAll(sp, "$MAX_TOOL_RES", strconv.Itoa(cfg.MaxToolRes)) if td := toolsDir(cfg); td != "" { sp += "\n\nExtra shell tools can be found at " + td } if sd := skillsDir(cfg); sd != "" { sp += "\n\nSkills may be discovered and invoked from " + sd } return sp } func main() { cfg := getCfg(configPath()) cfg.ContextWindow = fetchContextWindow(&cfg) sp := buildSystemPrompt(&cfg) COL = col(cfg) stdin = bufio.NewReader(os.Stdin) msgs := []Message{{Role: "system", Content: strp(sp)}} if len(os.Args) > 1 && os.Args[1] != "" { p := os.Args[1] data, err := os.ReadFile(p) if err != nil { fmt.Println(c("Error: file '"+p+"' not found.", 31)) os.Exit(1) } u := strings.TrimSpace(string(data)) if strings.HasPrefix(u, "!") { runDirectShell(strings.TrimPrefix(u, "!"), cfg.ShellTimeout) return } msgs = append(msgs, Message{Role: "user", Content: strp(u)}) sigCtx, cancel := signal.NotifyContext(context.Background(), os.Interrupt) var usg Usage msgs, usg, err = AL(sigCtx, &cfg, msgs) if lastRole(msgs) == "tool" { msgs = append(msgs, Message{Role: "user", Content: strp("continue")}) fmt.Println(c("[auto continue: last message was a tool result]", 33)) msgs, usg, err = AL(sigCtx, &cfg, msgs) } interrupted := sigCtx.Err() != nil cancel() if err != nil { if interrupted || errors.Is(err, context.Canceled) { fmt.Println(c("\n[interrupted]", 33)) } else { fmt.Println(c("[error: "+err.Error()+"]", 31)) } os.Exit(1) } fmt.Println(c(formatUsage(usg, cfg.ContextWindow), 2)) autosave(msgs) return } loadHistory() // Interactive startup: if a session already exists for this directory, offer // to autoload it so the user can seamlessly resume prior work. if sess := findProjectSession(); sess != nil && len(sess.Messages) > 1 { fmt.Println(c("[session found for this directory]", 33) + " " + c(sess.ID, 32)) fmt.Println(" " + c(sess.Summary, 2)) fmt.Println(c(fmt.Sprintf(" created %s [%d messages]", sess.Created, len(sess.Messages)), 2)) fmt.Print(c("Resume this session? [Y/n]: ", 33)) if ans, ok := readPlain(""); ok { ans = strings.TrimSpace(strings.ToLower(ans)) if ans == "" || ans == "y" || ans == "yes" { msgs = sess.Messages autosave(msgs) fmt.Println(c("[session resumed: "+sess.ID+"]", 32) + " " + c(summary(msgs), 2)) } } } re := cfg.Raw["reasoning_effort"] if re == "" { re = "high" } fmt.Println(c("Bantam Agent ready", 1, 32) + c(" (Ctrl+J = new line)", 2)) fmt.Println(c(fmt.Sprintf("endpoint: %s model: %s temp: %v reasoning: %s context: %d", cfg.Endpoint, cfg.Model, cfg.Temperature, re, cfg.ContextWindow), 2)) for { u, ok := readLine(c("> ", 1, 36)) if !ok { fmt.Println() break } u = strings.TrimSpace(u) if u == "" { continue } addHistory(u) // Command aliases: /model -> /cfg model, /endpoint -> /cfg endpoint. // (Note: /models is a distinct command and is intentionally not matched.) if u == "/model" { u = "/cfg model" } else if strings.HasPrefix(u, "/model ") { u = "/cfg model " + strings.TrimSpace(strings.TrimPrefix(u, "/model ")) } else if u == "/endpoint" { u = "/cfg endpoint" } else if strings.HasPrefix(u, "/endpoint ") { u = "/cfg endpoint " + strings.TrimSpace(strings.TrimPrefix(u, "/endpoint ")) } switch { case u == "/quit": goto done case u == "/clear": msgs = []Message{{Role: "system", Content: strp(sp)}} autosave(msgs) continue case u == "/save" || strings.HasPrefix(u, "/save "): name := strings.TrimSpace(strings.TrimPrefix(u, "/save")) sid, sm := saveSession(msgs, name) fmt.Println(c("[session saved: "+sid+"]", 32) + " " + c(sm, 2)) continue case u == "/continue" || u == "/cont": pid := projectID() lm, err := loadSession(pid) if err != nil { fmt.Println(c("No session found for current project: "+err.Error(), 31)) continue } msgs = lm autosave(msgs) fmt.Println(c("[session continued: "+pid+"]", 32) + " " + c(summary(msgs), 2)) continue case u == "/list": ss := sessions() if len(ss) == 0 { fmt.Println(c("No sessions saved yet.", 33)) continue } pid := projectID() for _, s := range ss { mk := "" if s.ID == pid { mk = c(" (current project)", 33) } fmt.Println(c(s.ID, 32) + mk + c(fmt.Sprintf(" %s [%d msgs]", s.Created, len(s.Messages)), 2)) fmt.Println(" " + c(s.Summary, 2)) } continue case strings.HasPrefix(u, "/load"): parts := strings.Fields(u) if len(parts) < 2 { fmt.Println(c("Usage: /load ", 31)) continue } lm, err := loadSession(parts[1]) if err != nil { fmt.Println(c("Session not found: "+err.Error(), 31)) continue } msgs = lm autosave(msgs) fmt.Println(c("[session loaded: "+parts[1]+"]", 32) + " " + c(summary(msgs), 2)) continue case u == "/compact": msgs = doCompact(&cfg, msgs) continue case strings.HasPrefix(u, "/cfg"): parts := strings.SplitN(u, " ", 3) if len(parts) == 2 { k := strings.TrimSpace(parts[1]) if v, ok := cfg.Raw[k]; ok { fmt.Println(c(k+"="+v, 32)) } else { fmt.Println(c(k+" not set", 31)) } } else if len(parts) >= 3 { k, v := strings.TrimSpace(parts[1]), strings.TrimSpace(parts[2]) if err := setCfg(".bantam.cfg", k, v); err != nil { fmt.Println(c("[cfg error: "+err.Error()+"]", 31)) continue } cfg = getCfg(configPath()) if k == "model" || k == "endpoint" || k == "api_key" { cfg.ContextWindow = fetchContextWindow(&cfg) } COL = col(cfg) fmt.Println(c(fmt.Sprintf("[config updated: %s=%s]", k, v), 32)) } else { fmt.Println(c("Usage: /cfg [val]", 31)) } continue case strings.HasPrefix(u, "/skill"): if strings.TrimSpace(strings.TrimPrefix(u, "/skill")) == "" { listSkills(&cfg) // bare /skill lists available skills continue } composed, err := skillPrompt(u, &cfg) if err != nil { fmt.Println(c("[skill error: "+err.Error()+"]", 31)) continue } u = composed // fall through to a normal LLM turn with the composed prompt case u == "/models": ml, err := listModels(&cfg) if err != nil { fmt.Println(c("[error querying /models: "+err.Error()+"]", 31)) continue } fmt.Println(c("Supported models at "+cfg.Endpoint+":", 1, 36)) fmt.Print(renderMD(ml)) continue case strings.HasPrefix(u, "!"): runDirectShell(strings.TrimPrefix(u, "!"), cfg.ShellTimeout) continue case u == "/help": fmt.Println(c("Bantam commands:", 1, 36)) for _, kv := range [][2]string{ {"/quit", "exit"}, {"/clear", "reset to system prompt"}, {"/save [name]", "save session (default: project MD5)"}, {"/continue", "continue project session (alias: /cont)"}, {"/list", "list sessions"}, {"/load ", "load session"}, {"/compact", "compact context"}, {"/cfg [v]", "get/set config"}, {"/model [m]", "alias for /cfg model"}, {"/endpoint [e]", "alias for /cfg endpoint"}, {"!", "run shell command directly"}, {"/models", "list models at endpoint"}, {"/skill [p]", "load SKILL.md and run as prompt"}, {"/help", "show help"}, } { fmt.Println(c(fmt.Sprintf(" %-18s", kv[0]), 1, 32) + kv[1]) } continue } turnMsgs := append([]Message{}, msgs...) turnMsgs = append(turnMsgs, Message{Role: "user", Content: strp(u)}) sigCtx, cancel := signal.NotifyContext(context.Background(), os.Interrupt) resMsgs, usg, err := AL(sigCtx, &cfg, turnMsgs) if lastRole(resMsgs) == "tool" { resMsgs = append(resMsgs, Message{Role: "user", Content: strp("continue")}) fmt.Println(c("[auto continue: last message was a tool result]", 33)) resMsgs, usg, err = AL(sigCtx, &cfg, resMsgs) } interrupted := sigCtx.Err() != nil cancel() if err != nil { if interrupted || errors.Is(err, context.Canceled) { fmt.Println(c("\n[interrupted]", 33)) } else { fmt.Println(c("[error: "+err.Error()+"]", 31)) } continue } msgs = resMsgs autosave(msgs) fmt.Println(c(formatUsage(usg, cfg.ContextWindow), 2)) pct := contextPct(usg, cfg.ContextWindow) if pct >= 60.0 && len(msgs) > 1 { fmt.Print(c(fmt.Sprintf("Context usage is at %.1f%% (%d / %d tokens). Compact conversation? [Y/n]: ", pct, ctxTokens(usg), cfg.ContextWindow), 33)) if ans, ok := readPlain(""); ok { ans = strings.TrimSpace(strings.ToLower(ans)) if ans == "" || ans == "y" || ans == "yes" { msgs = doCompact(&cfg, msgs) } } } } done: autosave(msgs) saveHistory() }