feat(orchestration): DAG TaskGraph + pre-execution validation

[kuangmi-bit]

Summary

Multi-agent orchestration with dependency-graph execution and structured pre-flight validation, building on @georgex8001's work in #12436.

What's Included

TaskGraph — DAG Executor (src/orchestration/task_graph.py, 280 LOC)

• DFS cycle detection
• Parallel wave execution via asyncio
• Fail-fast propagation (dependent tasks → SKIPPED)

Pre-flight Validation (validate_dag())

Returns structured ValidationReport:

Check	Severity	Rule
Cycle detection	REJECT	Any cycle → reject
Orphan detection	WARNING	Zero in+out degree nodes
Depth limit	WARNING/REJECT	>4 warning, >8 reject (configurable)
Blast radius	INFO	BFS → risk level (routine/elevated/critical)

Supporting Modules

• AgentCommunicationBus — in-process async pub/sub for A2A messaging
• MultiAgentOrchestrator — coordinates via existing delegate_task
• Reflex hints — record/load failure patterns for self-improvement
• HonchoContext — long-term memory injection bridge

Design Decisions

• Stdlib only — asyncio, no new dependencies
• All validation is pre-flight — fails BEFORE any task runs
• Independent of transport — works with A2A, delegate_task, or direct calls

Tests

18 tests in tests/orchestration/test_task_graph.py:

• Node-level: cycle detection, fan-out, fan-in, deep chains
• Graph-level: multi-wave parallelism, fail-fast cascade, depth enforcement
• Boundary: empty DAG, single node, duplicate deps, unknown nodes

Related

• Closes feat(orchestration): DAG TaskGraph + delegate bridge + A2A bus + reflex hints #12436
• CI fix for upstream Discord test failures: fix(url_safety): add cdn.discordapp.com to trusted private-IP hosts #47008

Co-authored-by: @georgex8001

[teknium1]

Thanks for the DAG validation work. Current main needs substantial integration work before this can safely land.

Problems

• src/orchestration/delegate_runner.py:51 passes toolsets to delegate_task, but current main removed that argument in ba0bc01d1; this path will raise before a child starts.
• src/orchestration/multi_agent_orchestrator.py:29 builds TaskGraph without calling validate_dag, so the advertised depth/orphan/blast-radius pre-flight checks never gate execution.
• Parallel waves reach asyncio.to_thread (delegate_runner.py:81), while _temporary_environ mutates process-global os.environ (delegate_runner.py:47); concurrent nodes can cross-contaminate profile_env.
• The existing Kanban graph already rejects cycles and dependency-gates work (hermes_cli/kanban_db.py:2814-2841) and dispatches ready tasks (hermes_cli/kanban_db.py:7240-7309).

Suggested changes

• Rework against the current delegation contract, integrate validation into the actual execution path, and avoid global environment mutation.
• Re-scope around a concrete Kanban integration; remove unused bus/registry pieces, or wire and test them. broadcast_tool_hint also only constructs a message at agent_bus.py:53-65, as noted in the prior review.

Automated hermes-sweeper review.

[teknium1]

Blocking: current main removed the toolsets parameter from delegate_task in ba0bc01, so this call raises TypeError before any child starts. Child toolsets now inherit from the parent; rework this bridge around that contract.

[teknium1]

Blocking: this mutates process-global os.environ, but parallel graph roots reach this code concurrently through asyncio.to_thread. Two nodes with different profile_env values can observe or restore each other's environment; use explicit child configuration or process isolation instead.

[teknium1]

Blocking: this execution entry point never calls validate_dag, so depth, orphan, and blast-radius findings do not run before delegates are started. Validate here and define the reject behavior before constructing/running the graph.

[teknium1]

This method only builds an AgentMessage addressed to "*"; it never calls publish or enqueues messages for peers. Either implement real broadcast semantics with a delivery test or rename it to make the builder-only behavior explicit.

[kuangmi-bit]

@teknium1 thanks for the thorough analysis. The salvageability=low assessment is fair — this PR was an early prototype and main has moved significantly since.

Given the Kanban already handles cycle detection and dependency-gating, and the delegation contract has changed (toolsets removed), a full rewrite against current main is needed.

Question for you: is the DAG TaskGraph direction still interesting to the project, or should I close this and focus on more targeted contributions?

[kuangmi-bit]

Closing — no maintainer response on whether the DAG/AgentContext direction is still interesting to the project. The mainframe-to-AI-agent architecture work will continue in the standalone MACS project (github.com/deeparchi-ai/macs). Happy to reopen if a maintainer signals interest.

[kuangmi-bit]

Addressed knoal's blocking review:

1. broadcast_tool_hint now actually broadcasts to all known peers (iterates self._queues under lock, skips sender). No longer dead code — the function does what its name says.
2. TaskGraph.init documents why _validate_specs may run twice (constructor guard is intentional even if caller ran validate_dag).

Branch rebased onto current main (zero conflicts). Ready for re-review.

[kuangmi-bit]

Closing as announced on Jul 18 — this was an early prototype and main has moved significantly (overlapping direction also explored in #12436). If the DAG/AgentContext direction is still wanted, a fresh PR against current main is the cleaner path.