Swarm Roadmap¶

Known gaps and planned improvements for the attoswarm hybrid swarm orchestrator, organized by priority.

Tier 1: Interpretability (highest ROI)¶

Per-agent execution trace streaming — stream tool calls, cost deltas, and reasoning from each agent into TUI; currently only terminal events (heartbeat/done/failed) are visible
Merge conflict visualization — show which files conflicted, merge strategy attempted, merger's resolution in TUI events table
Task output preview — when a task completes, show a summary of what files were created/modified and key content (first N lines of new files)
Budget projection — extrapolate remaining budget vs estimated tasks; warn at 80%, graceful shutdown at 95%
Failure attribution chain — when a task fails, trace the root cause: was it timeout, cost, agent crash, dependency failure, or coordination error?

Context propagation for dependent tasks — when task t0 finishes in worktree-1, sync its file changes into worktree-2 before starting t1 (which depends on t0)
Shared read-only context directory — broadcast completed task outputs to a shared .agent/hybrid-swarm/<run>/shared/ directory that all agents can read
Inter-agent message channel — allow agents to post structured messages (e.g., "I created api.py with these exports") that dependent agents receive as context
Incremental index updates — re-run CodeIndex.build() after each task completion and make the updated snapshot available to subsequent agents

Merge conflict detection — before assigning merge task, run git merge --no-commit to check for conflicts and report them
Auto-rebase fallback — if merge fails, try git rebase before marking as failed
Quality gate unit tests — test judge/critic quorum logic, threshold computation, rejection/retry loops (currently only roundtrip serialization tested)
Budget enforcement tests — test hard-limit shutdown, reserve ratio, cost-based termination
Cascade failure detection — if t0 fails, immediately mark t1 (depends on t0) as blocked instead of letting it wait indefinitely
Agent crash recovery tests — test watchdog restart + task reassignment end-to-end

Multi-file conflict smoke test — two workers edit the same file in parallel, verify merge handles it
Budget exhaustion test — set very low budget, verify swarm terminates gracefully with proper state
Timeout cascade test — one agent hangs past task_silence_timeout, verify cleanup + retry
Mixed backend test — Claude worker + Codex judge, verify cross-backend protocol works
Large DAG test — 10+ tasks with complex dependency graph, verify ordering and parallel dispatch
Resume fidelity test — run halfway, kill, resume, verify no duplicate work or lost state
Worktree isolation test — verify agents can't see each other's uncommitted changes

LLM-based task decomposition — use the orchestrator model to decompose goals into concrete tasks (currently falls back to parallel)
Dynamic task splitting — if a task is too large mid-execution, split it into subtasks
Cost-aware scheduling — assign cheaper models to simple tasks, expensive models to complex ones
Agent capability matching — match tasks to agents by capability, not just role