Track whether tool execution occurred during each turn via approval
request notifications. If a turn completes with only text (no commands
or file changes), automatically start a follow-up turn nudging Codex
to execute rather than just describe plans. Max 5 auto-continues per
user message to prevent infinite loops.
Also strengthen instructions.md with execution-first policy.
- Remove commands/ sync from agent-runner (skills/ is SSOT for both agents)
- Remove ~/.codex/skills/ as source (use ~/.claude/skills/ only)
- Codex uses $skill prefix, Claude Code uses /skill — both read from skills/
- Minor formatting fixes in config.ts, group-queue.test.ts, index.ts
Codex sessions now also receive skills and commands from ~/.claude/
in addition to ~/.codex/, so both agents share the same tool set
(patch, linear, sentry, coolify, exa, playwright, sprite, etc.)
- Rewrite codex-runner to use `codex app-server` JSON-RPC instead of
`codex exec`. Supports streaming (item/agentMessage/delta), session
persistence (thread/start, thread/resume), and mid-turn message
injection (turn/steer with IPC polling during execution).
- Fix token refresh not syncing to per-group session directories,
causing 401 errors on running agents.
- Fix typing indicator staying on when codex returns null result
by always clearing typing on any streaming callback.
- Update README and CLAUDE.md to reflect dual-service architecture,
host process mode, and codex app-server integration.
- Add token-refresh module that auto-renews Claude OAuth tokens 30min before expiry
- Fix Codex session loss on service restart by resuming previous sessions
- Inject nanoclaw MCP server into Codex config.toml for send_message/schedule_task
- Sanitize ANTHROPIC_API_KEY and CLAUDE_CODE_OAUTH_TOKEN from Codex subprocess env
CODEX_CONFIG_DIR is not recognized by Codex CLI. CODEX_HOME is the correct
env var that controls the root directory for sessions, state, and config.
Without this, all groups sharing the same workDir would have their sessions
collide, causing resume --last to pick up wrong group's session.
- Make STORE_DIR, DATA_DIR, GROUPS_DIR configurable via env vars
(NANOCLAW_STORE_DIR, NANOCLAW_DATA_DIR, NANOCLAW_GROUPS_DIR)
for running two instances from one codebase
- Remove broken 'both' agent type sequential loop from processGroupMessages
- Delete dead code: container-runtime, mount-security, credential-proxy,
Dockerfiles, and related config constants (~1,085 lines removed)
- Add codex instance launchd plist template
- AgentType now supports 'both' — runs both agents sequentially
- Each agent gets its own folder suffix (_cc / _codex) for isolated sessions
- Claude Code bot handles message storage for 'both' channels
- Responses routed through the matching bot for each agent type
- Enrich spawned process PATH with ~/.npm-global/bin so codex CLI is
findable when running under launchd
- Clear typing indicator in streaming callback after sending response,
instead of waiting for runner process to fully exit
- Read/write agent_type column in getAllRegisteredGroups/setRegisteredGroup
(was missing, causing all groups to fall back to claude-code)
- Strip CLAUDECODE env var to prevent nested session detection
- Add node binary path to spawned process PATH for launchd compatibility
- Clear stale container-era sessions that reference /workspace/group
- Rewrite container-runner.ts to spawn node processes directly instead
of Docker/Apple Container
- Runner paths configurable via env vars (NANOCLAW_GROUP_DIR, etc.)
with container-path defaults for backwards compat
- Pass real credentials directly (no credential proxy needed)
- Remove ensureContainerSystemRunning() and startCredentialProxy()
from startup
- Add build:runners script for building agent-runner and codex-runner
- Fix TS strict mode errors in agent-runner ipc-mcp-stdio.ts
- Add Codex as alternative agent type (agentType: 'codex' per group)
- Codex container runner with session resume and 100-turn limit
- Dual Discord bot support (DISCORD_BOT_TOKEN + DISCORD_CODEX_BOT_TOKEN)
- Each bot handles only its agentType's registered groups
- Discord typing indicator refreshes every 8s (was single-shot, expired after ~10s)
- Agent runner 100-turn limit to prevent infinite bot-to-bot loops
- Apple Container runtime support (convert-to-apple-container skill)
- Session commands (/compact) support
- DB migration for agent_type column
Move skill definitions from the nanoclaw-skills marketplace plugin
into .claude/skills/ so they're available as unprefixed slash commands
(e.g. /add-whatsapp instead of /nanoclaw-skills:add-whatsapp).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace the custom skills engine with standard git operations.
Feature skills are now git branches (on upstream or channel forks)
applied via `git merge`. Channels are separate fork repos.
- Remove skills-engine/ (6,300+ lines), apply/uninstall/rebase scripts
- Remove old skill format (add/, modify/, manifest.yaml) from all skills
- Remove old CI (skill-drift.yml, skill-pr.yml)
- Add merge-forward CI for upstream skill branches
- Add fork notification (repository_dispatch to channel forks)
- Add marketplace config (.claude/settings.json)
- Add /update-skills operational skill
- Update /setup and /customize for marketplace plugin install
- Add docs/skills-as-branches.md architecture doc
Channel forks created: nanoclaw-whatsapp (with 5 skill branches),
nanoclaw-telegram, nanoclaw-discord, nanoclaw-slack, nanoclaw-gmail.
Upstream retains: skill/ollama-tool, skill/apple-container, skill/compact.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat: implement credential proxy for enhanced container environment isolation
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: address PR review — bind proxy to loopback, scope OAuth injection, add tests
- Bind credential proxy to 127.0.0.1 instead of 0.0.0.0 (security)
- OAuth mode: only inject Authorization on token exchange endpoint
- Add 5 integration tests for credential-proxy.ts
- Remove dangling comment
- Extract host gateway into container-runtime.ts abstraction
- Update Apple Container skill for credential proxy compatibility
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: scope OAuth token injection by header presence instead of path
Path-based matching missed auth probe requests the CLI sends before
the token exchange. Now the proxy replaces Authorization only when
the container actually sends one, leaving x-api-key-only requests
(post-exchange) untouched.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: bind credential proxy to docker0 bridge IP on Linux
On bare-metal Linux Docker, containers reach the host via the bridge IP
(e.g. 172.17.0.1), not loopback. Detect the docker0 interface address
via os.networkInterfaces() and bind there instead of 0.0.0.0, so the
proxy is reachable by containers but not exposed to the LAN.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: bind credential proxy to loopback on WSL
WSL uses Docker Desktop with the same VM routing as macOS, so
127.0.0.1 is correct and secure. Without this, the fallback to
0.0.0.0 was triggered because WSL has no docker0 interface.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: detect WSL via /proc instead of env var
WSL_DISTRO_NAME isn't set under systemd. Use
/proc/sys/fs/binfmt_misc/WSLInterop which is always present on WSL.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Scheduled tasks that send messages via send_message (IPC) instead of
returning text as result left the container idle for ~30 minutes until
the hard timeout killed it (exit 137). This blocked new messages for
the group during that window.
Root cause: scheduleClose() was only called inside the
`if (streamedOutput.result)` branch. Tasks that communicate solely
through IPC (e.g. heartbeat check-ins) complete with result=null,
so the 10s close timer was never set.
Fix: also call scheduleClose() on status==='success', covering both
result-based and IPC-only task completions.
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* feat(skill): add reactions skill (emoji reactions + status tracker)
* refactor(reactions): minimize overlays per upstream review
Address gavrielc's review on qwibitai/nanoclaw#509:
- SKILL.md: remove all inline code, follow add-telegram/add-whatsapp pattern (465→79 lines)
- Rebuild overlays as minimal deltas against upstream/main base
- ipc-mcp-stdio.ts: upstream base + only react_to_message tool (8% delta)
- ipc.ts: upstream base + only reactions delta (14% delta)
- group-queue.test.ts: upstream base + isActive tests only (5% delta)
- Remove group-queue.ts overlay (isActive provided by container-hardening)
- Remove group-queue.ts from manifest modifies list
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Two bugs in the DM with dedicated bot number setup:
1. The skill asked for the bot's own phone number to use as the JID.
But from the bot's perspective, incoming DMs appear with the SENDER's
JID (the user's personal number), not the bot's own number. The
registration must use the user's personal number as the JID.
2. DM with bot (1:1 conversation) should use --no-trigger-required, same
as self-chat. A trigger prefix is unnecessary in a private DM.
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
The pairing code was only emitted to stdout, which is buffered by the
calling process and not visible until the auth command exits (~120s).
By also writing to store/pairing-code.txt the moment the code is ready,
callers can poll that file and display the code to the user within seconds
instead of after the 60s expiry window.
Update the add-whatsapp skill instructions to use the background +
file-poll pattern instead of waiting on buffered stdout.
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Thanks @glifocat! Clean skill package — good docs, solid tests, nice intent files. Pushed a small fix for path traversal on the PDF filename before merging.
getNewMessages() and getMessagesSince() loaded all rows after a
checkpoint with no cap, causing growing memory and token costs.
Both queries now use a DESC LIMIT subquery to return only the
most recent N messages, re-sorted chronologically.
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
