gitdone-agent 0.6.0 → 0.6.3

package/index.js

@@ -12,10 +12,10 @@
 import { execSync, spawn } from 'node:child_process'
 import {
   existsSync, writeFileSync, readFileSync, unlinkSync,
-  mkdirSync, copyFileSync, appendFileSync, readdirSync,
+  mkdirSync, copyFileSync, appendFileSync, readdirSync, rmSync,
 } from 'node:fs'
 import { resolve, join } from 'node:path'
-import { homedir, hostname } from 'node:os'
+import { homedir, hostname, tmpdir } from 'node:os'
 import { randomUUID } from 'node:crypto'
 
 // ─── Stable agent dir, config + logging ────────────────────────────────────────
@@ -27,7 +27,7 @@ import { randomUUID } from 'node:crypto'
 // Reported to the server on every sync so the web UI can flag outdated agents.
 // Keep in lockstep with packages/agent/package.json "version" AND
 // src/lib/agentVersion.ts LATEST_AGENT_VERSION.
-const AGENT_VERSION = '0.6.0'
+const AGENT_VERSION = '0.6.3'
 
 const AGENT_DIR = join(homedir(), '.gitdone-agent')
 const CONFIG_PATH = join(AGENT_DIR, 'config.json')
@@ -239,6 +239,18 @@ function git(cmd, cwd) {
   }
 }
 
+// Like git(), but returns the RAW bytes instead of a UTF-8 string. Needed for
+// `git diff`, whose payload can be Windows-1251 (CP1251) Cyrillic — decoding
+// those bytes as UTF-8 up front would corrupt them to „�"/„?" irreversibly
+// (gd-276). Callers decode per-file via decodeDiffText().
+function gitRaw(cmd, cwd) {
+  try {
+    return execSync(cmd, { cwd, encoding: 'buffer', stdio: ['pipe', 'pipe', 'pipe'] })
+  } catch {
+    return Buffer.alloc(0)
+  }
+}
+
 // Like git(), but surfaces failures (with stderr) instead of swallowing them —
 // used for push/pull where we need to detect auth rejection.
 function gitTry(cmd, cwd) {
@@ -354,14 +366,57 @@ function discardFile(repoPath, file) {
   return 'discarded (new file removed)'
 }
 
-function parseDiffByFile(diffOutput) {
+// ─── Windows-1251 (CP1251) aware diff decoding (gd-276) ───────────────────────
+// Legacy Cyrillic sources are often saved in CP1251. Their high bytes (0x80–0xFF,
+// single-byte) are invalid UTF-8, so reading the diff as UTF-8 shows „?"/„�"
+// instead of кирилица. We keep the raw bytes and decode each file's diff section
+// on its own: valid UTF-8 stays UTF-8; the rest goes through the CP1251 table
+// below — so a UTF-8 repo is untouched while CP1251 files finally read correctly.
+
+// CP1251 high range 0x80–0xFF → Unicode code points (0x00–0x7F is plain ASCII).
+// 0x98 is unassigned in CP1251 → U+FFFD.
+const CP1251_HIGH = [
+  0x0402, 0x0403, 0x201A, 0x0453, 0x201E, 0x2026, 0x2020, 0x2021, 0x20AC, 0x2030, 0x0409, 0x2039, 0x040A, 0x040C, 0x040B, 0x040F,
+  0x0452, 0x2018, 0x2019, 0x201C, 0x201D, 0x2022, 0x2013, 0x2014, 0xFFFD, 0x2122, 0x0459, 0x203A, 0x045A, 0x045C, 0x045B, 0x045F,
+  0x00A0, 0x040E, 0x045E, 0x0408, 0x00A4, 0x0490, 0x00A6, 0x00A7, 0x0401, 0x00A9, 0x0404, 0x00AB, 0x00AC, 0x00AD, 0x00AE, 0x0407,
+  0x00B0, 0x00B1, 0x0406, 0x0456, 0x0491, 0x00B5, 0x00B6, 0x00B7, 0x0451, 0x2116, 0x0454, 0x00BB, 0x0458, 0x0405, 0x0455, 0x0457,
+  0x0410, 0x0411, 0x0412, 0x0413, 0x0414, 0x0415, 0x0416, 0x0417, 0x0418, 0x0419, 0x041A, 0x041B, 0x041C, 0x041D, 0x041E, 0x041F,
+  0x0420, 0x0421, 0x0422, 0x0423, 0x0424, 0x0425, 0x0426, 0x0427, 0x0428, 0x0429, 0x042A, 0x042B, 0x042C, 0x042D, 0x042E, 0x042F,
+  0x0430, 0x0431, 0x0432, 0x0433, 0x0434, 0x0435, 0x0436, 0x0437, 0x0438, 0x0439, 0x043A, 0x043B, 0x043C, 0x043D, 0x043E, 0x043F,
+  0x0440, 0x0441, 0x0442, 0x0443, 0x0444, 0x0445, 0x0446, 0x0447, 0x0448, 0x0449, 0x044A, 0x044B, 0x044C, 0x044D, 0x044E, 0x044F,
+]
+
+function decodeCp1251(buf) {
+  let out = ''
+  for (const b of buf) out += String.fromCharCode(b < 0x80 ? b : CP1251_HIGH[b - 0x80])
+  return out
+}
+
+// A strict UTF-8 decoder (always available even on small-ICU builds) — throws on
+// any invalid sequence, which is exactly how we tell UTF-8 apart from CP1251.
+const UTF8_STRICT = new TextDecoder('utf-8', { fatal: true })
+function decodeDiffText(buf) {
+  try {
+    return UTF8_STRICT.decode(buf)
+  } catch {
+    return decodeCp1251(buf)
+  }
+}
+
+function parseDiffByFile(diffBuf) {
   const files = {}
-  if (!diffOutput) return files
-  const sections = diffOutput.split(/(?=^diff --git )/m)
+  if (!diffBuf || diffBuf.length === 0) return files
+  // Latin1 is a lossless 1-byte⇢1-char mapping, so the ASCII "diff --git" split
+  // markers match while every original byte survives for the per-section decode.
+  const raw = Buffer.isBuffer(diffBuf) ? diffBuf.toString('latin1') : String(diffBuf)
+  const sections = raw.split(/(?=^diff --git )/m)
   for (const section of sections) {
     if (!section.trim()) continue
-    const match = section.match(/^diff --git a\/.+ b\/(.+)$/m)
-    if (match) files[match[1]] = section
+    // Decode THIS file's bytes on their own (UTF-8 or CP1251), then read the
+    // filename from the decoded header.
+    const text = decodeDiffText(Buffer.from(section, 'latin1'))
+    const match = text.match(/^diff --git a\/.+ b\/(.+)$/m)
+    if (match) files[match[1]] = text
   }
   return files
 }
@@ -391,7 +446,7 @@ function getSnapshot(repoPath) {
     lastCommit = { sha, message, author, date }
   }
 
-  const diffs = parseDiffByFile(git('git diff HEAD', repoPath))
+  const diffs = parseDiffByFile(gitRaw('git diff HEAD', repoPath))
 
   // Origin remote → lets the server auto-detect the GitHub repo for the History tab.
   const remoteUrl = git('git config --get remote.origin.url', repoPath) || null
@@ -466,19 +521,24 @@ async function postRunEvents(cfg, runId, events, status, result) {
 
 // Post transcript lines (and optional turn status / Claude session id) for an
 // interactive AiSession chat turn.
-async function postSessionEvents(cfg, sessionId, events, status, claudeSessionId) {
+async function postSessionEvents(cfg, sessionId, events, status, claudeSessionId, streamingText) {
   await api(cfg, '/api/v1/agent/ai-session/events', {
     sessionId,
     events,
     ...(status ? { status } : {}),
     ...(claudeSessionId ? { claudeSessionId } : {}),
+    // Explicit '' clears the live preview; `undefined` leaves it untouched.
+    ...(streamingText !== undefined ? { streamingText } : {}),
   }).catch((e) => log(`✗ ai-session events post failed: ${e.message}`))
 }
 
 // Turn one stream-json line from `claude -p` into console events. Phase 1 shows
 // what the model writes (TEXT) and which tools it calls (TOOL_CALL); raw tool
 // results (Bash stdout etc.) come in Phase 2 via a PostToolUse hook.
-function parseStreamLine(line, push, onInit) {
+// With `--include-partial-messages`, Claude also emits `stream_event` lines
+// carrying incremental text deltas — onDelta gets those so chat sessions can
+// show the reply being written live (gd-302).
+function parseStreamLine(line, push, onInit, onDelta) {
   let ev
   try { ev = JSON.parse(line) } catch { return }
   if (ev.type === 'system') {
@@ -489,6 +549,16 @@ function parseStreamLine(line, push, onInit) {
     }
     return
   }
+  // Partial-message stream: only the model's visible text deltas feed the live
+  // preview. Tool-input JSON deltas and thinking deltas are ignored here — the
+  // full block still arrives as a normal `assistant` event below.
+  if (ev.type === 'stream_event' && typeof onDelta === 'function') {
+    const e = ev.event
+    if (e?.type === 'content_block_delta' && e.delta?.type === 'text_delta' && e.delta.text) {
+      onDelta(e.delta.text)
+    }
+    return
+  }
   if (ev.type === 'assistant' && ev.message?.content) {
     for (const block of ev.message.content) {
       if (block.type === 'text' && block.text?.trim()) {
@@ -711,15 +781,42 @@ function runAiCommand(cfg, cmd, repoPath) {
   })
 }
 
+// Download the images attached to a chat turn to a per-session temp dir and
+// return their local paths. `claude -p` can't take image URLs on stdin, so we
+// fetch them locally and point Claude at the files (its Read tool renders
+// images). Best-effort: a failed download is skipped, not fatal.
+async function downloadSessionImages(sessionId, urls) {
+  const dir = join(tmpdir(), `gitdone-ai-${sessionId}`)
+  mkdirSync(dir, { recursive: true })
+  const paths = []
+  for (let i = 0; i < urls.length; i++) {
+    try {
+      const res = await fetch(urls[i])
+      if (!res.ok) { log(`✗ image download HTTP ${res.status}: ${urls[i]}`); continue }
+      const buf = Buffer.from(await res.arrayBuffer())
+      const m = /\.(png|jpe?g|gif|webp|bmp)(?:\?|$)/i.exec(urls[i])
+      const ext = m ? m[1].toLowerCase() : 'png'
+      const file = join(dir, `image-${i + 1}.${ext}`)
+      writeFileSync(file, buf)
+      paths.push(file)
+    } catch (e) {
+      log(`✗ image download failed (${urls[i]}): ${e.message}`)
+    }
+  }
+  return { dir, paths }
+}
+
 // Run one turn of an interactive chat session: `claude -p` (resuming the prior
 // Claude session when known) with the user's message on stdin, streaming the
 // reply back to the session console. Long-running + fire-and-forget like ai_run.
-function runAiChat(cfg, cmd, repoPath) {
+async function runAiChat(cfg, cmd, repoPath) {
   const sessionId = cmd.payload?.sessionId
   const prompt = cmd.payload?.prompt ?? ''
+  const images = Array.isArray(cmd.payload?.images) ? cmd.payload.images.filter((u) => typeof u === 'string' && u) : []
   const claudeSessionId = cmd.payload?.claudeSessionId || null
   const allowCommit = cmd.payload?.allowCommit === true
-  if (!sessionId || !prompt) {
+  // A turn may be text-only, image-only, or both — but needs at least one.
+  if (!sessionId || (!prompt && images.length === 0)) {
     reportCommandResult(cfg, cmd.id, 'error', 'invalid ai_chat payload')
     return
   }
@@ -740,6 +837,9 @@ function runAiChat(cfg, cmd, repoPath) {
     '-p',
     '--output-format', 'stream-json',
     '--verbose',
+    // Stream the reply token-by-token so the console can show it being written
+    // live, not only once each block is complete (gd-302).
+    '--include-partial-messages',
     '--permission-mode', 'acceptEdits',
     '--allowedTools', 'Read,Edit,Write,Bash,mcp__gitdone__*',
     // Block git commit/push unless this repo opted in (per-repo aiAutoCommit).
@@ -755,24 +855,58 @@ function runAiChat(cfg, cmd, repoPath) {
   let pending = []
   let capturedSession = null // Claude's session id, sent (idempotently) for --resume
   let flushing = false
+  let liveText = '' // in-progress text of the current assistant block (live preview)
+  let sentLive = '' // last streamingText we posted — so we only push on change
+  let lastPostAt = 0 // Date.now() of the last post, for the keep-alive heartbeat
   const flush = async () => {
-    if (flushing || pending.length === 0) return
+    if (flushing) return
+    const hasEvents = pending.length > 0
+    const liveChanged = liveText !== sentLive
+    // Heartbeat: with nothing new for a while, still post (bumps lastActivityAt)
+    // so the console shows honest "still working" and can spot a real stall.
+    const heartbeat = !hasEvents && !liveChanged && Date.now() - lastPostAt > 2500
+    if (!hasEvents && !liveChanged && !heartbeat) return
     flushing = true
     const batch = pending.map((e) => ({ role: roleFor(e.kind), text: e.text })); pending = []
-    await postSessionEvents(cfg, sessionId, batch, 'running', capturedSession || undefined)
+    const streamingText = liveChanged ? liveText : undefined
+    sentLive = liveText
+    await postSessionEvents(cfg, sessionId, batch, 'running', capturedSession || undefined, streamingText)
+    lastPostAt = Date.now()
     flushing = false
   }
-  const push = (kind, text) => { if (text != null && String(text) !== '') pending.push({ kind, text: String(text) }) }
+  const push = (kind, text) => {
+    if (text == null || String(text) === '') return
+    // A completed text block becomes a real event — drop its live preview so the
+    // finalised bubble and the cleared preview swap in on the same flush.
+    if (kind === 'TEXT') liveText = ''
+    pending.push({ kind, text: String(text) })
+  }
   const onInit = (sid) => { if (sid && !claudeSessionId) capturedSession = sid }
-  const timer = setInterval(flush, 800)
+  const onDelta = (chunk) => { liveText += chunk }
+  const timer = setInterval(flush, 500)
+
+  // Fetch any attached images locally and fold their paths into the prompt so
+  // Claude reads them with its Read tool (URLs on stdin don't render).
+  let imgDir = null
+  let fullPrompt = prompt
+  if (images.length > 0) {
+    const dl = await downloadSessionImages(sessionId, images)
+    imgDir = dl.dir
+    if (dl.paths.length > 0) {
+      const list = dl.paths.map((p) => `- ${p}`).join('\n')
+      const note = `Потребителят прикачи ${dl.paths.length} изображени${dl.paths.length === 1 ? 'е' : 'я'}. Прегледай ги с Read tool от следните локални пътища:\n${list}`
+      fullPrompt = prompt ? `${prompt}\n\n[${note}]` : `[${note}]`
+    }
+  }
 
-  log(`▶ ai_chat ${sessionId} @ ${repoPath} (resume=${claudeSessionId ? 'yes' : 'no'}) via ${claudePath}`)
+  log(`▶ ai_chat ${sessionId} @ ${repoPath} (resume=${claudeSessionId ? 'yes' : 'no'}, images=${images.length}) via ${claudePath}`)
 
   let child
   try {
     child = spawn(claudePath, args, { cwd: repoPath, shell, windowsHide: true, env: childEnv })
   } catch (err) {
     clearInterval(timer)
+    if (imgDir) { try { rmSync(imgDir, { recursive: true, force: true }) } catch { /* best-effort */ } }
     postSessionEvents(cfg, sessionId, [{ role: 'SYSTEM', text: `Грешка при стартиране: ${err.message}` }], 'error')
     reportCommandResult(cfg, cmd.id, 'error', err.message)
     return
@@ -780,7 +914,7 @@ function runAiChat(cfg, cmd, repoPath) {
 
   if (child.stdin) {
     child.stdin.on('error', () => {})
-    try { child.stdin.write(prompt); child.stdin.end() }
+    try { child.stdin.write(fullPrompt); child.stdin.end() }
     catch (e) { push('SYSTEM', `stdin write failed: ${e.message}`) }
   }
 
@@ -791,14 +925,16 @@ function runAiChat(cfg, cmd, repoPath) {
     while ((nl = buf.indexOf('\n')) >= 0) {
       const line = buf.slice(0, nl).trim()
       buf = buf.slice(nl + 1)
-      if (line) parseStreamLine(line, push, onInit)
+      if (line) parseStreamLine(line, push, onInit, onDelta)
     }
   })
   child.stderr.on('data', (d) => { const s = d.toString().trim(); if (s) push('SYSTEM', s) })
   child.on('error', (err) => push('SYSTEM', `Процесна грешка: ${err.message}`))
   child.on('close', async (code) => {
     clearInterval(timer)
-    if (buf.trim()) parseStreamLine(buf.trim(), push, onInit)
+    if (imgDir) { try { rmSync(imgDir, { recursive: true, force: true }) } catch { /* best-effort */ } }
+    if (buf.trim()) parseStreamLine(buf.trim(), push, onInit, onDelta)
+    liveText = '' // turn is over — drop any lingering live preview
     await flush()
     const ok = code === 0
     await postSessionEvents(
@@ -806,6 +942,7 @@ function runAiChat(cfg, cmd, repoPath) {
       ok ? [] : [{ role: 'SYSTEM', text: `✗ Ходът приключи с код ${code}.` }],
       ok ? 'idle' : 'error',
       capturedSession || undefined,
+      '', // clear the live preview on the terminal post
     )
     reportCommandResult(cfg, cmd.id, ok ? 'done' : 'error', `exit ${code}`)
     log(`■ ai_chat ${sessionId} приключи (code ${code})`)

package/package.json

@@ -1,6 +1,6 @@
 {
     "name":  "gitdone-agent",
-    "version":  "0.6.0",
+    "version":  "0.6.3",
     "description":  "Local git agent for gitdone — watches a local repo and sends snapshots to gitdone.eu",
     "type":  "module",
     "bin":  {