Skip to content

docs: improve agent commands with multi-word spaces, URL encoding, and tmux reference#7

Open
jsell-rh wants to merge 9 commits intomarkturansky:mainfrom
jsell-rh:docs/improve-agent-commands
Open

docs: improve agent commands with multi-word spaces, URL encoding, and tmux reference#7
jsell-rh wants to merge 9 commits intomarkturansky:mainfrom
jsell-rh:docs/improve-agent-commands

Conversation

@jsell-rh
Copy link

@jsell-rh jsell-rh commented Mar 6, 2026

Summary

  • Multi-word space name support: Both boss.check and boss.ignite now handle quoted arguments (e.g., "ProtocolDev" "Agent Boss Development") and URL-encode space names with %20 in all curl URLs.
  • Step 0 orientation (ignite only): New pre-flight step that grounds the agent — pwd, git branch, remote URL, coordinator health check, and reading CLAUDE.md.
  • Autonomous work loop (ignite only): After ignition, agents now have a documented read→check messages→work→POST→repeat loop instead of stopping after initial status.
  • Tmux quick reference (ignite only): Comprehensive section covering session spawning (interactive mode, not -p), observing, capturing pane output, tearing down, and sending commands.
  • Message awareness (check): Adds #### Messages scanning and a Step 3 "Act on messages" phase.
  • Sticky field documentation: Clarifies that repo_url and tmux_session are remembered server-side after first send.
  • Modernized examples: Updated from GitLab URLs to GitHub URLs.

Files changed

File Description
commands/boss.check.md Quoted arg parsing, URL encoding, message scanning, curl rule
commands/boss.ignite.md Step 0 orientation, work loop, tmux reference, sticky fields

jsell-rh added 2 commits March 6, 2026 14:07
…are empty

The callers of BroadcastCheckIn and SingleAgentCheckIn always pass empty
strings for checkModel and workModel, causing '/model ' (with no model
name) to be sent to agents. This resulted in broken model switch attempts
that would time out or fail.

Now when checkModel/workModel are empty, the model switch steps are
skipped entirely and the check-in runs on whatever model the agent is
already using. Model economy still works if actual model names are passed.
Rewrite tmuxIsIdle to use a comprehensive set of idle indicators instead
of only matching the Claude Code '>' prompt. New patterns recognized:

- Shell prompts: $, %, #, >, ❯, » (with false-positive guards for
  '=>', '->', '50%', 'line #3')
- Claude Code ❯ prompt with auto-suggested text after it
- Claude Code vim-mode status bar ('-- INSERT --', '-- NORMAL --')
- OpenCode status bar ('ctrl+p commands')
- Generic idle keywords ('waiting for input', 'ready', etc.)
- Heavy box-drawing character ┃ (U+2503) in addition to light │ (U+2502)

Also changes the default-on-error behavior from 'assume busy' to
'assume idle', preventing agents from being falsely skipped during
broadcasts when pane capture fails transiently.

Includes comprehensive table-driven tests for lineIsIdleIndicator (28
cases) and isShellPrompt (16 cases).
jsell-rh added 7 commits March 6, 2026 18:05
Fix broadcast model switching: skip /model when model names are empty
Improve tmux idle detection with robust prompt recognition
…protocol

Messages sent via the dashboard were stored in agent JSON but never
visible to agents because:
1. The /raw markdown endpoint did not render them
2. Agent status updates wiped the Messages field

This commit fixes both issues and makes messages a first-class feature:

Server changes:
- Render messages in /raw markdown under '#### Messages' with sender,
  timestamp, and content so agents see them during check-ins
- Preserve Messages and Documents fields when agents POST status updates
- Show pending messages in the ignition response with bold 'unread' header
- Add messaging protocol rule to ignition output

Protocol and command changes:
- Add 'Send message' endpoint to protocol.md endpoints table
- Add rule 9: check your messages on /raw, send via /message endpoint
- Update boss.check.md: scan for messages, acknowledge, act on instructions
- Update boss.ignite.md: document message checking and sending rules

Tests (9 new):
- TestMessagePostEndpoint: POST, verify delivery and retrieval
- TestMessagePreservedOnAgentUpdate: verify messages survive agent updates
- TestMessageRenderedInMarkdown: verify #### Messages in /raw output
- TestMessageValidation: missing sender, empty message, wrong method
- TestMessageLimit: verify 50-message cap
- TestMessageSSEBroadcast: verify agent_message SSE event
- TestMessageToNonexistentAgentCreatesAgent: auto-create on message
- TestIgnitionEndpoint: verify structure, tmux registration, /message docs
- TestIgnitionShowsPendingMessages: verify pending messages in ignition
Fix message system: render in /raw, preserve on updates, integrate into agent workflow
When a message is delivered to an agent via the /message endpoint, the
server now marks the agent for a nudge. The existing liveness loop
(which ticks every 1 second and already knows each agent's idle state)
picks up the pending nudge on the next tick where the agent is idle and
triggers a SingleAgentCheckIn in a goroutine.

This ensures agents read new messages promptly without:
- Polling (wasteful: ~95% of check-ins would be empty)
- Tmux keystroke injection (breaks agent's autonomous protocol discipline)
- Manual nudge button clicks from the dashboard

Debounce: multiple messages in quick succession coalesce into one nudge.
A nudgeInFlight flag prevents duplicate concurrent check-ins for the
same agent.
Auto-nudge agents when messages arrive via liveness loop
…d tmux reference

- Support quoted multi-word space names in argument parsing
- Add URL-encoding guidance for space names with spaces
- Add Step 0 orientation (pwd, git branch, health check, read CLAUDE.md)
- Add autonomous work loop after ignition
- Add comprehensive tmux quick reference (spawn, observe, teardown)
- Clarify sticky fields (repo_url, tmux_session)
- Add message awareness to check-in flow
- Modernize examples from GitLab to GitHub URLs
@jsell-rh jsell-rh force-pushed the docs/improve-agent-commands branch from 8f1e940 to 5030d9f Compare March 6, 2026 23:10
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant