path: root/web/transcript.go

// transcript.go — the conversation a session's reading happened inside.
//
// The notevi log says a session read a file at 10:04; the harness that ran that
// session kept the whole conversation on disk, in its own format. Here we find
// that file by session id and flatten both known formats into Turns, so a note
// can be opened at the moment it was written.
//
// Two rules keep this cheap and safe. Tool calls and their output are reduced
// to a line each — transcripts reach megabytes and none of that belongs in a
// page — and nothing outside a configured root is ever opened, with session
// ids checked before they touch the filesystem.
package web

import (
	"encoding/json"
	"fmt"
	"os"
	"path/filepath"
	"sort"
	"strings"
	"sync"
	"time"
)

const (
	maxTranscript      = 32 << 20 // refuse to parse more than this in one file
	transcriptCacheMax = 8        // parsed transcripts are large; keep few
	briefLine          = 160      // one line of a tool call or its output
	maxText            = 6000     // one message
	convWindow         = 80       // turns rendered around the anchored moment
	convMax            = 1500     // turns rendered by "show all"
)

// A Turn is one moment in a session: something said, a tool called or answered,
// or — woven in from the notevi log — a read, a grep, a note.
type Turn struct {
	Role   string // user | assistant | thinking | tool | result | notevi
	Kind   string // qualifies Role: the notevi op, or "error" on a failed tool
	When   time.Time
	Text   string // what was said
	Tool   string // tool name, on tool and result turns
	Detail string // one compact line: the call's arguments, or its output
	File   string // notevi turns: the file the op touched
	Note   *Note  // notevi note turns: the note itself, rendered as a card
	Here   bool   // the anchored moment
}

func (t Turn) Clock() string {
	if t.When.IsZero() {
		return ""
	}
	return t.When.Local().Format("15:04:05")
}

// A Transcript is one harness conversation, already normalized.
type Transcript struct {
	Path    string
	Harness string // claude-code | codex
	Turns   []Turn
	Err     string
}

// transcripts finds and parses harness transcripts, remembering what it has
// already parsed. Roots are the only places it will look.
type transcripts struct {
	roots []string

	mu    sync.Mutex
	cache map[string]*Transcript // path|mtime|size -> parsed
}

func newTranscripts(roots []string) *transcripts {
	tx := &transcripts{cache: map[string]*Transcript{}}
	for _, r := range roots {
		r = strings.TrimSpace(r)
		if r == "" {
			continue
		}
		if abs, err := filepath.Abs(r); err == nil {
			r = abs
		}
		tx.roots = append(tx.roots, filepath.Clean(r))
	}
	return tx
}

// defaultTranscriptRoots are where the two harnesses we know about keep their
// conversations.
func defaultTranscriptRoots() []string {
	home, err := os.UserHomeDir()
	if err != nil {
		return nil
	}
	return []string{filepath.Join(home, ".claude", "projects"), filepath.Join(home, ".codex", "sessions")}
}

// safeSession refuses ids that could name something we did not mean to open:
// the id goes into a filesystem path, so it must be one plain component.
func safeSession(id string) error {
	if id == "" {
		return fmt.Errorf("no session id")
	}
	if len(id) > 128 {
		return fmt.Errorf("session id is too long")
	}
	for _, r := range id {
		ok := r == '-' || r == '_' || r == ':' || r == '.' ||
			'0' <= r && r <= '9' || 'a' <= r && r <= 'z' || 'A' <= r && r <= 'Z'
		if !ok {
			return fmt.Errorf("session id contains %q", r)
		}
	}
	if strings.Contains(id, "..") {
		return fmt.Errorf("session id contains %q", "..")
	}
	return nil
}

// mangleDir is how Claude Code names a project directory after its cwd:
// /home/u/0x4200.cafe -> -home-u-0x4200-cafe.
func mangleDir(dir string) string {
	return strings.NewReplacer("/", "-", ".", "-").Replace(dir)
}

// under states the rule the globs below already obey: a transcript we open
// lies inside a configured root. It is checked anyway, so the rule has one
// place to fail rather than living in the shape of two patterns.
func under(root, path string) bool {
	rel, err := filepath.Rel(root, path)
	if err != nil {
		return false
	}
	return rel != ".." && !strings.HasPrefix(rel, ".."+string(filepath.Separator))
}

// find locates a session's transcript. Claude Code writes
// <root>/<mangled cwd>/<session>.jsonl; codex writes
// <root>/YYYY/MM/DD/rollout-<time>-<thread>.jsonl and notevi's session id is that
// thread id, so it is matched on the suffix. Both patterns are tried under
// every root — a root knows its own layout, we do not have to.
func (tx *transcripts) find(session, dir string) string {
	if safeSession(session) != nil {
		return ""
	}
	for _, root := range tx.roots {
		pats := []string{
			filepath.Join(root, "*", session+".jsonl"),
			filepath.Join(root, "*", "*", "*", "rollout-*-"+session+".jsonl"),
		}
		// the cwd we logged names the directory directly, when it still exists
		if dir != "" {
			pats = append([]string{filepath.Join(root, mangleDir(dir), session+".jsonl")}, pats...)
		}
		for _, pat := range pats {
			hits, _ := filepath.Glob(pat)
			for _, hit := range hits {
				if under(root, hit) {
					return hit
				}
			}
		}
	}
	return ""
}

// load returns the parsed transcript for a session, or nil when there is none.
// Parses are cached by path, mtime and size, so a 5 MB file is read once.
func (tx *transcripts) load(session, dir string) *Transcript {
	path := tx.find(session, dir)
	if path == "" {
		return nil
	}
	st, err := os.Stat(path)
	if err != nil {
		return nil
	}
	key := fmt.Sprintf("%s|%d|%d", path, st.ModTime().UnixNano(), st.Size())

	tx.mu.Lock()
	t, ok := tx.cache[key]
	tx.mu.Unlock()
	if ok {
		return t
	}

	t = readTranscript(path, st.Size())
	tx.mu.Lock()
	if len(tx.cache) >= transcriptCacheMax {
		tx.cache = map[string]*Transcript{}
	}
	tx.cache[key] = t
	tx.mu.Unlock()
	return t
}

// ── parsing ───────────────────────────────────────────────────────────────

// One line of either format. Codex wraps every record in "payload"; Claude
// Code puts the model turn in "message". Sniffing per line keeps the two
// parsers independent of any file header.
type rawRecord struct {
	Type      string          `json:"type"`
	Timestamp string          `json:"timestamp"`
	Payload   json.RawMessage `json:"payload"`
	Message   json.RawMessage `json:"message"`
}

type parser struct {
	t     *Transcript
	tools map[string]string // tool call id -> name, so results keep their name
	last  time.Time         // records without a stamp inherit the previous one
}

func readTranscript(path string, size int64) *Transcript {
	t := &Transcript{Path: path}
	if size > maxTranscript {
		t.Err = fmt.Sprintf("transcript is %d MB — too large to parse", size>>20)
		return t
	}
	b, err := os.ReadFile(path)
	if err != nil {
		t.Err = err.Error()
		return t
	}
	p := &parser{t: t, tools: map[string]string{}}
	for _, line := range strings.Split(string(b), "\n") {
		if line == "" {
			continue
		}
		var rec rawRecord
		if json.Unmarshal([]byte(line), &rec) != nil {
			continue
		}
		at := parseTime(rec.Timestamp)
		if at.IsZero() {
			at = p.last
		} else {
			p.last = at
		}
		switch {
		case len(rec.Payload) > 0:
			t.Harness = "codex"
			p.codex(rec, at)
		case len(rec.Message) > 0:
			t.Harness = "claude-code"
			p.claude(rec, at)
		}
	}
	return t
}

// add drops turns that would render as nothing — empty thinking blocks,
// tool results whose output was pure whitespace.
func (p *parser) add(t Turn) {
	if t.Text == "" && t.Detail == "" && t.Tool == "" {
		return
	}
	t.Text = clip(t.Text, maxText)
	p.t.Turns = append(p.t.Turns, t)
}

// ── claude code ───────────────────────────────────────────────────────────

type claudeMsg struct {
	Role    string          `json:"role"`
	Content json.RawMessage `json:"content"` // a string, or blocks
}

type claudeBlock struct {
	Type      string          `json:"type"`
	Text      string          `json:"text"`
	Thinking  string          `json:"thinking"`
	ID        string          `json:"id"`
	Name      string          `json:"name"`
	Input     json.RawMessage `json:"input"`
	ToolUseID string          `json:"tool_use_id"`
	Content   json.RawMessage `json:"content"` // tool_result: a string, or blocks
	IsError   bool            `json:"is_error"`
}

// claude turns records of type user and assistant into Turns; the rest of the
// file (modes, titles, file history, attachments) is harness bookkeeping.
func (p *parser) claude(rec rawRecord, at time.Time) {
	if rec.Type != "user" && rec.Type != "assistant" {
		return
	}
	var m claudeMsg
	if json.Unmarshal(rec.Message, &m) != nil {
		return
	}
	var text string
	if json.Unmarshal(m.Content, &text) == nil { // plain prompts arrive as a string
		p.add(Turn{Role: rec.Type, When: at, Text: text})
		return
	}
	var blocks []claudeBlock
	if json.Unmarshal(m.Content, &blocks) != nil {
		return
	}
	for _, b := range blocks {
		switch b.Type {
		case "text":
			p.add(Turn{Role: rec.Type, When: at, Text: b.Text})
		case "thinking":
			p.add(Turn{Role: "thinking", When: at, Text: b.Thinking})
		case "tool_use":
			p.tools[b.ID] = b.Name
			p.add(Turn{Role: "tool", When: at, Tool: b.Name, Detail: callDetail(b.Input)})
		case "tool_result":
			t := Turn{Role: "result", When: at, Tool: p.tools[b.ToolUseID], Detail: resultDetail(b.Content)}
			if b.IsError {
				t.Kind = "error"
			}
			p.add(t)
		}
	}
}

// ── codex ─────────────────────────────────────────────────────────────────

type codexPayload struct {
	Type      string          `json:"type"`
	Role      string          `json:"role"`
	Name      string          `json:"name"`
	CallID    string          `json:"call_id"`
	Arguments json.RawMessage `json:"arguments"` // a JSON string holding JSON
	Input     json.RawMessage `json:"input"`
	Output    json.RawMessage `json:"output"`
	Content   []codexText     `json:"content"`
	Summary   []codexText     `json:"summary"`
}

type codexText struct {
	Type string `json:"type"`
	Text string `json:"text"`
}

// codex reads response_items — the model's own history. The event_msg stream
// beside it is the TUI's view of the same turns, so taking both would double
// everything.
func (p *parser) codex(rec rawRecord, at time.Time) {
	if rec.Type != "response_item" {
		return
	}
	var pl codexPayload
	if json.Unmarshal(rec.Payload, &pl) != nil {
		return
	}
	switch pl.Type {
	case "message":
		// developer messages are the harness's own instructions, and the first
		// user messages are context it injects; neither was said by anyone
		if pl.Role == "developer" {
			return
		}
		text := joinText(pl.Content)
		if injected(text) {
			return
		}
		role := "assistant"
		if pl.Role == "user" {
			role = "user"
		}
		p.add(Turn{Role: role, When: at, Text: text})
	case "reasoning":
		// summaries are usually encrypted and come back empty; add drops those
		p.add(Turn{Role: "thinking", When: at, Text: joinText(pl.Summary)})
	case "function_call", "custom_tool_call", "web_search_call":
		name := pl.Name
		if name == "" {
			name = strings.TrimSuffix(pl.Type, "_call")
		}
		p.tools[pl.CallID] = name
		detail := callDetail(pl.Arguments)
		if detail == "" {
			detail = callDetail(pl.Input)
		}
		p.add(Turn{Role: "tool", When: at, Tool: name, Detail: detail})
	case "function_call_output", "custom_tool_call_output", "web_search_output":
		p.add(Turn{Role: "result", When: at, Tool: p.tools[pl.CallID], Detail: resultDetail(pl.Output)})
	}
}

func joinText(items []codexText) string {
	var parts []string
	for _, it := range items {
		if it.Text != "" {
			parts = append(parts, it.Text)
		}
	}
	return strings.Join(parts, "\n\n")
}

// injected recognizes the wrappers a harness puts around a user turn to carry
// its own state.
func injected(text string) bool {
	for _, tag := range []string{"<environment_context>", "<user_instructions>", "<permissions instructions>"} {
		if strings.HasPrefix(text, tag) {
			return true
		}
	}
	return false
}

// ── summarizing ───────────────────────────────────────────────────────────

// callFields are the arguments worth a line, most telling first. Anything that
// carries file content (new_string, content) is deliberately not here.
var callFields = []string{"command", "cmd", "file_path", "path", "pattern", "query", "url", "prompt", "description"}

// callDetail reduces a tool call's arguments to their one interesting field.
func callDetail(raw json.RawMessage) string {
	if len(raw) == 0 {
		return ""
	}
	var s string
	if json.Unmarshal(raw, &s) == nil { // codex passes arguments as a JSON string
		raw = json.RawMessage(s)
	}
	var m map[string]any
	if json.Unmarshal(raw, &m) != nil {
		return brief(string(raw))
	}
	for _, k := range callFields {
		if v, ok := m[k].(string); ok && strings.TrimSpace(v) != "" {
			return brief(v)
		}
	}
	b, err := json.Marshal(m)
	if err != nil {
		return ""
	}
	return brief(string(b))
}

// resultDetail summarizes what a tool answered, whether that came back as a
// string, as content blocks, or as an image nobody wants inlined.
func resultDetail(raw json.RawMessage) string {
	if len(raw) == 0 {
		return ""
	}
	var s string
	if json.Unmarshal(raw, &s) == nil {
		return brief(s)
	}
	var items []codexText
	if json.Unmarshal(raw, &items) == nil {
		var parts []string
		for _, it := range items {
			switch {
			case it.Text != "":
				parts = append(parts, it.Text)
			case it.Type != "":
				parts = append(parts, "["+it.Type+"]")
			}
		}
		return brief(strings.Join(parts, "\n"))
	}
	return brief(string(raw))
}

// brief keeps a blob's first line and says how much was left behind. Whole
// files and command output pass through here; a page gets their shape, never
// their bytes.
func brief(s string) string {
	s = strings.TrimRight(s, "\n")
	if strings.TrimSpace(s) == "" {
		return ""
	}
	first, rest, more := strings.Cut(s, "\n")
	first = clip(strings.TrimSpace(first), briefLine)
	if !more || strings.TrimSpace(rest) == "" {
		return first
	}
	return fmt.Sprintf("%s … +%d lines", first, strings.Count(rest, "\n")+1)
}

func clip(s string, n int) string {
	if len(s) <= n { // bytes first: the common case never allocates
		return s
	}
	r := []rune(s)
	if len(r) <= n {
		return s
	}
	return strings.TrimRight(string(r[:n]), " \t") + "…"
}

// ── one session: the transcript with the notevi log woven in ──────────────────

// A Conv is what the session page shows: the conversation, the log entries
// that happened during it, and where in a long one we are looking.
type Conv struct {
	Session, Agent, Model, Dir string
	Bit                        uint8
	Harness, Path              string
	Roots                      []string
	Err                        string
	Reads, Greps, Notes        int

	Turns      []Turn
	Total      int    // moments in the whole conversation
	From, To   int    // 1-based position of Turns within it
	Anchor     string // element id to scroll to, "" when nothing is anchored
	AllHref    string // same page, unwindowed
	Windowed   bool
	Incomplete bool // even "show all" had to stop
}

// sessionTurns are the notevi log's own moments for a session: what it read,
// grepped and noted.
func (s *Site) sessionTurns(id string) []Turn {
	var out []Turn
	for _, e := range s.Entries {
		if e.Session != id || e.Op == "note" {
			continue
		}
		out = append(out, Turn{Role: "notevi", Kind: e.Op, When: parseTime(e.Time),
			Text: describe(e), File: e.File})
	}
	for _, n := range s.Notes {
		if n.Session == id {
			out = append(out, Turn{Role: "notevi", Kind: "note", When: n.At, File: n.File, Note: n})
		}
	}
	return out
}

// conversation merges a session's transcript with its log entries. Both sides
// are compared as instants — the log stamps carry an offset, transcripts are
// UTC — so ordering is the real one, and only the rendering is local.
func (sv *server) conversation(site *Site, id string, note int, all bool) *Conv {
	c := &Conv{Session: id, Roots: sv.tx.roots}
	if err := safeSession(id); err != nil {
		c.Err = err.Error()
		return c
	}
	for _, s := range site.Sessions(time.Now(), liveWindow, 0) {
		if s.ID == id {
			c.Agent, c.Model, c.Dir, c.Bit = s.Agent, s.Model, s.Dir, s.Bit
			c.Reads, c.Greps, c.Notes = s.Reads, s.Greps, s.Notes
			break
		}
	}

	logged := site.sessionTurns(id)
	// the parsed transcript is shared with other requests, so the merge copies
	// it rather than appending onto its slice
	var turns []Turn
	if t := sv.tx.load(id, c.Dir); t != nil {
		c.Path, c.Harness, c.Err = t.Path, t.Harness, t.Err
		turns = make([]Turn, 0, len(t.Turns)+len(logged))
		turns = append(turns, t.Turns...)
	}
	turns = append(turns, logged...)
	sort.SliceStable(turns, func(i, j int) bool { return turns[i].When.Before(turns[j].When) })

	at := -1
	if note > 0 {
		for i := range turns {
			if turns[i].Note != nil && turns[i].Note.ID == note {
				at, turns[i].Here = i, true
				c.Anchor = fmt.Sprintf("note-%d", note)
				break
			}
		}
	}
	c.Total = len(turns)
	from, window := convWindowAt(turns, at, all)
	c.Turns, c.From, c.To = window, from+1, from+len(window)
	c.Windowed = len(window) < len(turns)
	c.Incomplete = c.Windowed && all
	c.AllHref = "/session/" + escPath(id) + "?all=1"
	if note > 0 {
		c.AllHref += fmt.Sprintf("&note=%d#note-%d", note, note)
	}
	return c
}

// convWindowAt cuts a conversation down to what a page may render: a stretch
// around the anchored moment, or the tail of it when nothing is anchored.
// "show all" raises the bound but does not remove it.
func convWindowAt(turns []Turn, at int, all bool) (int, []Turn) {
	n := convWindow
	if all {
		n = convMax
	}
	if len(turns) <= n {
		return 0, turns
	}
	if at < 0 {
		return len(turns) - n, turns[len(turns)-n:]
	}
	lo := at - n/2
	if lo < 0 {
		lo = 0
	}
	if lo+n > len(turns) {
		lo = len(turns) - n
	}
	return lo, turns[lo : lo+n]
}