- Add 'tool' param to read tool code from 9P namespace
- Add 'args' as array (passed as positional params via set --)
- Tools from 9P skip dangerous pattern validation (trusted)
- Add Go language support
- Update all skills to use tool/args syntax
- Flatten agent/tools/ to serve tools directly (no capability subdirs)
- Restore SKILLS_PROMPT.md with proactive skill discovery
- Remove discover_role.sh and discover_tool.sh
- Remove help endpoint from roles and tools
- Update all SKILL.md files to use flat tool paths
- Clear agent prompt in anvillm-agent.json
Reuse the same layered config approach as anvilsrv:
- global.yaml provides base permissions (including /usr, /lib, /bin)
- sandbox/<name>.yaml adds sandbox-specific permissions
- Merge uses most-permissive-wins strategy
This fixes the fallback to default sandbox, which now has proper
system paths from global.yaml.
When sandbox is not explicitly specified, retry with 'default' sandbox
if the initial execution fails with permission denied or missing file errors.
This allows third-party skills designed for execute_bash to work without
modification while preserving isolation when sandbox is explicitly set.
- Add sandbox parameter to execute_code (default: anvilmcp)
- Load sandbox config from ~/.config/anvillm/sandbox/<name>.yaml
- Fallback to ./cfg/sandbox/<name>.yaml for development
- Add cwd config option: 'temp' for isolated workspace, '{CWD}' for caller's dir
- Create cfg/sandbox/anvilmcp.yaml with restricted defaults
- Update cfg/sandbox/default.yaml with cwd: '{CWD}'
Dangerous pattern fixes (bash, not JS/Python):
- Remove curl/wget block (network controlled by landrun)
- Remove chmod 777, eval(), exec() (wrong syntax)
- Allow /etc/passwd (world-readable, needed for user lookups)
- Fix /dev/ write check to only block /dev/sd* (not /dev/null)
- Fix eval pattern for actual bash usage
Update skills with explicit sandbox specifications:
- anvilmcp sandbox for normal operations
- default sandbox for spawn_agent.sh (needs full permissions)
This allows me to assign a role via `context` to an agent, it discovers what that role means, then it can run the tools and load the skills required to do its work. In theory...
- Add ClearHandler, CompactHandler, ResumeHandler callbacks to tmux.Config
- kiro-cli: clear/resume need 250ms delay + confirmation key
- claude: clear/compact are simple /command[ENTER], resume not implemented
- Add Resume tag and handler to Assist prompt window
- Update server.go to use new Session methods
- Wire up roles filesystem in server.go
- Add /agent/roles directory with focus areas and help
- Support ANVILLM_TOOLS_DIR env var for multiple tool directories
- Update pcloudcc-testing skill documentation
- Add DisplayName field to RoleMeta for frontmatter name
- Name field now always contains physical filename (never overwritten)
- Use physical filename for 9P paths and deduplication
- Use DisplayName in help output for discovery
- Add unique filtering (sort -u) to discover_role/skill/tool scripts
- Add CommentCount field to BeadWithBlockers struct in readListFromStore and readBeadPropertyFromStore
- Call GetIssueComments to populate comment count when > 0
- Update beads skill documentation to guide checking comment_count and reading comments for additional context
Add w.Addr("0")/dot=addr/show after content writes in refreshList,
refreshMailboxWindow, refreshDaemonWindow, and refreshViewBeadWindow
so windows no longer scroll to the bottom on open or refresh.
Closes anv-hww
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Context was prepended to every Send() call. Now it is consumed
on the first Send() and cleared, acting as a one-shot spawn
initialization rather than a per-prompt overhead.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Changed NewServer to accept *BeadsFS instead of bd.Storage
- BeadsFS is always created but with no initial mounts
- Each mount opens its own store on-demand
- Added nil check in supervisor to handle missing beads gracefully
The beads store was being opened unconditionally at startup, causing
crashes when the database was corrupted or incompatible. Now it's only
opened on-demand when actually needed.
- Add mount param to github_to_beads.sh, jira_to_beads.sh, batch_create_beads.sh, sync_beads.sh
- Update all 9p paths to use mount-specific paths (agent/beads/<mount>/...)
- Fix SKILLS_PROMPT.md to correctly document mount requirements
- Only exceptions: mount_beads.sh, umount_beads.sh, list_mounts.sh
- Use proper shift-based parameter handling instead of positional
- Build command string that preserves newlines in description
- Fixes issue where multi-line descriptions were incorrectly parsed as parent-id
- All bead tools now require explicit mount parameter as first arg
- Multi-line values (descriptions, comments, reasons) properly quoted with printf
- Fixes issue where tools wrote to read-only agent/beads/list instead of mount-specific ctl
- Distinguish control files from bead IDs in mount walk logic
- Add bead directories to mount listings via Read(mountName/list)
- Verify mount exists before allowing walk into it
- Add encoding/json import to server.go
- Move SKILLS_PROMPT.md from kiro-cli/ to repo root (shared across backends)
- Replace verbose agent instructions with concise discovery protocol
- Enforce execute_code for all AnviLLM tool usage; forbid raw 9p commands
- Add agentspawn hook for claude backend
- Clean up install-mcp.sh tool list (now auto-discovered)
- Clarify SKILL.md wording for beads and anvillm-communication
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Add DISPLAY, WAYLAND_DISPLAY, USER to landrun env passthrough
- Mount /etc/passwd read-only for getuser() to work
- Update mcptools scripts to check /etc/shadow instead of /etc/passwd for sandbox detection
Levi Neely
Levi Neely