v0.8.2 — context diet

Cuts tokens wasted on every full-agent turn.

Root CLAUDE.md → DEVELOPER.md
- The Agent SDK auto-loads every CLAUDE.md from the agent's CWD up to
  the git root. Runtime agents run from groups/main/, so they were
  inheriting ~1,330 tokens of developer documentation (key-file table,
  architecture notes, NanoClaw heritage) into every system prompt. That
  context belongs in DEVELOPER.md for Claude Code reading the repo, not
  in a Telegram chat turn.
- File unchanged, only name moved. Per-group CLAUDE.md files
  (groups/{name}/CLAUDE.md) keep their name — those are the runtime
  instructions and SHOULD load.

appendMemoryLog deleted
- The orchestrator was writing "- [auto] Handled: <raw Telegram XML
  preview>" to memory/log.md after every full-agent turn. The agent
  then re-read that log on every session start, paying tokens to
  re-ingest message previews from days past.
- Logging is now the agent's responsibility. groups/{name}/CLAUDE.md
  already instructs it to prepend meaningful events. Andy is in the
  loop for chat turns — trust him + the agent. Removes ~30 lines of
  index.ts and an unused GROUPS_DIR import.

Existing [auto] noise pruned from groups/main/memory/log.md
- 74 [auto] entries dropped. File 204 → 86 lines.
- Real entries (versions, decisions, debugging notes) preserved.
- Gitignored, so this prune is local-only.

Verified
- 425 tests pass.
- Typecheck and build clean.
- v0.8.1 was 33 minutes ago — no architectural change here, just
  context diet.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: ziggy

CHANGELOG.md

@@ -1,5 +1,36 @@
 # Changelog
 
+## v0.8.2 (2026-05-03) — Context diet
+
+Cuts tokens wasted on every full-agent turn.
+
+### Changed
+- **Root `CLAUDE.md` renamed to `DEVELOPER.md`.** The Anthropic SDK auto-loads
+  every `CLAUDE.md` from the agent's CWD up to the git root. The runtime
+  agent runs from `groups/main/`, so it was inheriting ~1,330 tokens of
+  developer documentation (architecture notes, key-file table, NanoClaw
+  heritage) into every system prompt — context that has no business in a
+  Telegram chat turn. The file is unchanged; only its name moved. Per-group
+  `CLAUDE.md` files (`groups/{name}/CLAUDE.md`) keep their name and continue
+  to load as runtime instructions.
+- **Orchestrator-side `appendMemoryLog` deleted.** After every Telegram
+  turn the orchestrator was writing `- [auto] Handled: <raw Telegram XML
+  preview>` to `memory/log.md`. The agent re-read the log on every session
+  start, so it was paying tokens to re-ingest message previews from days
+  past. Logging is now the agent's responsibility — `groups/{name}/CLAUDE.md`
+  already instructs it to prepend meaningful events. Andy is in the loop
+  for chat turns; trust him + the agent.
+- **Existing `[auto]` noise pruned** from `groups/main/memory/log.md` —
+  74 entries dropped, file went from 204 → 86 lines. Real entries
+  (shipped versions, decisions, debugging notes) preserved.
+
+### Notes
+- Gmail MCP tool definitions (~1,500–3,000 tokens per full-agent turn) is
+  the next-biggest waste, but no quick fix without an architectural change
+  (move Gmail to a dedicated email group). SDK doesn't expose
+  `cache_control` on the system prompt option, so we can't cache MCP tool
+  defs the way we cached the fast-path system prompt in v0.8.1.
+
 ## v0.8.1 (2026-05-03) — Fast-path cost levers + reliability nits
 
 ### New

CLAUDE.md DEVELOPER.mdCLAUDE.md renamed to DEVELOPER.md

@@ -1,6 +1,6 @@
 {
   "name": "ghostclaw",
-  "version": "0.8.1",
+  "version": "0.8.2",
   "description": "Personal AI assistant. Bare metal, Telegram-only, no containers.",
   "type": "module",
   "main": "dist/index.js",

package.json

@@ -9,7 +9,6 @@ import {
   ASSISTANT_NAME,
   getDailyBudgetUsd,
   DATA_DIR,
-  GROUPS_DIR,
   IDLE_TIMEOUT,
   MAIN_GROUP_FOLDER,
   MAX_MESSAGES_PER_PROMPT,
@@ -449,38 +448,6 @@ async function handleAgentError(
   return false;
 }
 
-/**
- * Append `- [auto] Handled: ...` under today's heading in the group's
- * memory/log.md after a successful turn that actually produced output.
- * Best-effort — memory logging must never bring down a conversation.
- */
-function appendMemoryLog(ctx: BatchContext, outcome: AgentTurnOutcome): void {
-  if (!outcome.outputSentToUser) return;
-  try {
-    const logFile = path.join(GROUPS_DIR, ctx.group.folder, 'memory', 'log.md');
-    if (!fs.existsSync(logFile)) return;
-    const today = new Date().toISOString().slice(0, 10);
-    const logContent = fs.readFileSync(logFile, 'utf-8');
-    const taskSummary = currentTasks[ctx.chatJid] || ctx.prompt.slice(0, 120);
-    const entry = `- [auto] Handled: ${taskSummary.replace(/\n/g, ' ')}`;
-    if (logContent.includes(`## ${today}`)) {
-      const updated = logContent.replace(
-        `## ${today}`,
-        `## ${today}\n${entry}`,
-      );
-      fs.writeFileSync(logFile, updated);
-    } else {
-      const updated = logContent.replace(
-        '---\n',
-        `---\n\n## ${today}\n${entry}\n`,
-      );
-      fs.writeFileSync(logFile, updated);
-    }
-  } catch {
-    // Non-critical — don't fail the session over a log write.
-  }
-}
-
 async function processGroupMessages(chatJid: string): Promise<boolean> {
   const ctx = prepareBatch(chatJid);
   if (!ctx) return true;
@@ -501,7 +468,11 @@ async function processGroupMessages(chatJid: string): Promise<boolean> {
   if (outcome.hadError) return handleAgentError(ctx, outcome);
 
   consecutiveFailures[ctx.chatJid] = 0;
-  appendMemoryLog(ctx, outcome);
+  // Memory logging is the agent's responsibility — see groups/{group}/CLAUDE.md
+  // for the instruction to prepend meaningful events to memory/log.md. The
+  // previous orchestrator-side auto-write produced "- [auto] Handled: <raw
+  // Telegram XML preview>" entries that were noise the agent re-read on
+  // every session start. Removed in v0.8.2.
   return true;
 }