6b5d3fd0b6
agents: align autoLoad with agent roles
...
Each agent now has tools matching its purpose:
- Read-only agents (explorer, navigator, observer, theo, copilot):
No shell, file_write, file_edit. Code analysis tools only.
- Workflow agents (author, reviewer, panelist, foreman):
9p client for peer messaging and goalstatus.
Analysis tools for verification.
- Orchestration agents (conductor):
subagent_spawn, 9p client, analysis tools.
No direct file editing (delegates to sub-agents).
- Full coding agents (default, driver):
Complete toolset including shell, file ops, LSP, code intel.
- Support agents (librarian, taskmanager):
Scoped to their domain (docs, task files).
Removed tools that violate agent constraints.
2026-08-21 10:14:42 +02:00
aad9c4f6e5
agents: move user-preferences to userPrompts
...
user-preferences.md is meant to be prepended to every user message,
not part of the system prompt. Moved from 'prompt' to 'userPrompts'
in: author, conductor, default, foreman, panelist, reviewer.
Other agents already had it in the correct location.
2026-08-20 11:48:10 +02:00
80a9f7ecc4
consensus workflow: foreman + panelist agents with peer messaging
...
New workflow that spawns N independent panelists (default 3) and a
foreman. Each panelist analyzes the goal from a different angle:
- correctness/completeness
- simplicity/maintainability
- edge cases/failure modes
Panelists report findings to foreman via peer/ links. The foreman
synthesizes agreement, resolves disagreements, and writes the verdict.
Files:
- data/workflows/consensus — orchestration script
- data/agents/foreman.json — low-temp synthesis profile
- data/agents/panelist.json — high-temp analysis profile
- data/prompts/agent-foreman.md — synthesis behavior
- data/prompts/agent-panelist.md — investigation behavior
Usage:
printf 'name=review\ncwd=%s\nworkflow=consensus\n' $PWD | ollie-9p write session/new
echo 'Review error handling in agent/loop.go' | ollie-9p write session/review/goal
2026-08-19 17:26:31 +02:00
Levi Neely