[emanuilo]

Summary

Each chat now keeps one Claude Code session, reused across messages and resumed after a restart, instead of a fresh CLI per message with a rolling history window.

 handle(message, chat)
-  system prompt = AGENTS.md + MEMORY.md + time + memories + last N messages   # changes every turn, never cached
-  spawn Claude CLI (loads ~/.claude settings, plugins, MCP servers)
-  query → close CLI
-  consolidate before replying
+  context = <context> time, relevant memories, delivered reminders,
+            AGENTS.md / MEMORY.md if changed since the session saw them </context>
+  client = live client for chat
+        ?? connect(resume = sessions.json[chat].session_id)   # fingerprint mismatch / failed resume → fresh session
+  client.query(context + message)                            # system prompt static per session → prompt-cached
+  sessions.json[chat] = session_id + hashes of AGENTS.md / MEMORY.md seen
+  consolidate in the background, after the reply

Locked-down CLI: setting_sources=[], strict_mcp_config, tools=[] (Memclaw tools only), permission_mode="dontAsk" (was bypassPermissions). claude-agent-sdk → 0.2.164.

Telegram:

+ /new                    fresh conversation (also a "/new" message in Slack, WhatsApp, terminal)
+ /model, /effort         inline button menus; saved to ~/.memclaw/.env, conversation kept
+ "Working… step N: …"    silent status message following tool calls, deleted when the reply arrives

Code structure:

ClaudeAgentBackend
  _options(model, effort, …)   # the lockdown above, in one place, for sessions and one-shots
  _session_options(…)          # + cwd, MCP server, allowed tools, resume
  run_turn → _turn_cost(…)     # running total → this turn's cost, or a token estimate
MemclawAgent.handle            # read/hash files → context → run_turn → _log_turn,
                               #   _ack_notes, _save_session

Defaults: claude-sonnet-5-5 at medium. CONVERSATION_HISTORY_* config removed. Per-turn cost is the turn's, not the session's running total.

Evidence

• Before: pytest → 252 passed
  After: pytest → 317 passed, ruff clean (also clears 11 findings that were already on main). New coverage includes:

  first turn starts a session, saved to sessions.json
  next turn and a restart resume that session
  fingerprint change → old session not resumed
  AGENTS.md / MEMORY.md change → sent once in <context>; own writes not echoed
  /new drops the session; a turn running during /new doesn't re-save it,
    and a message sent while /new is still disconnecting starts fresh
  /model refuses to switch when the model list can't be loaded
  /model, /effort menus and button callbacks
  

• Latency (live, subscription): follow-up turns ~1.6–3 s, vs a cold CLI start of ~4–8 s per message before.
• Deployed to the Mac mini (Telegram). Restart resume and /new were verified end to end there.

Merge Danger

Door: two-way

Revert restores per-message CLIs. The new sessions.json is ignored by the old code, and memories are untouched.

Blast Radius: conversations

• On first run, every chat starts a fresh conversation, since there's no saved session yet.
• The CLI no longer loads the user's ~/.claude settings, plugins or MCP servers. That's intended, but anyone relying on them inside Memclaw loses them.
• Cursor keeps an interim in-memory transcript here; native Cursor conversations land in feat: native Cursor conversations, /model and /effort for Cursor #13.

🤖 Generated with Claude Code