Skip to content

feat: pulsing-forge orchestration tier — MCP-native mailboxes, journal, orchestrators - #82

Open
Damon6666 wants to merge 1 commit into
DeepLink-org:mainfrom
Damon6666:feat/forge-orchestration
Open

Damon6666 wants to merge 1 commit into
DeepLink-org:mainfrom
Damon6666:feat/forge-orchestration

Conversation

@Damon6666

@Damon6666 Damon6666 commented Nov 9, 2026 •

Copy link
Copy Markdown

What this adds

The existing pulsing-forge crate is the per-agent tool runtime (codex-compatible sandbox/handlers, Python bindings). This PR adds the missing team coordination tier on top of it — three new workspace crates:

Crate Role
pulsing-forge-core gates, closure semantics, Orchestrator trait + six built-ins (relay / pipeline / debate / saboteur / broadcast / review). 29 tests drawn from a 16-entry agent-team failure taxonomy (docs/forge/failure-taxonomy.md)
pulsing-forge-hub TideHub — per-agent mailboxes, exactly-once journaled ledger (replayable flight recorder), MCP control port exposing pulsing_send / pulsing_broadcast / pulsing_drain_mail / pulsing_health / pulsing_recall, fan-in-debounced scheduler, forced closure. 10 tests
pulsing-forge-runner forge run <scenario.toml> one-command scenarios, forge attach MCP stdio bridge, swimlane dashboard; deterministic / sleep-benchmark / codex (live-LLM) adapters

Any MCP-capable agent runtime joins a team with zero modifications — the whole agent-side contract is those four tools. Agents hosted on the existing pulsing-forge runtime call the same tools.

Related: #79 prototypes actor-mailbox collaboration via MCP; this PR is the productionized mailbox/orchestrator layer in that direction, hardened by 100-scenario testing and fleet benchmarks (10 sandboxes × 50 agents, 38/40 closed loops, real-LLM decision-making).

Verified on this branch

  • cargo build / cargo test -p pulsing-forge-core -p pulsing-forge-hub: 39/39 green
  • End-to-end: forge run examples/forge/leak-review.toml completes a full closed loop (12 turns, 21 letters, dual-channel closure)

Contents

  • crates/pulsing-forge-{core,hub,runner} — Rust sources, unit tests, scenario schema
  • examples/forge/ — six runnable scenarios + a sanitized Responses-to-Chat proxy for live-LLM runs (no endpoints or credentials embedded)
  • docs/forge/ — failure taxonomy, scenario schema, ledger format, adapters, design notes (orchestration.md positions the tier; mkdocs nav wiring left to maintainers)

Notes for reviewers

  • Naming: -hub (mailbox hub) and -runner (scenario CLI) deliberately avoid runtime, which the existing pulsing-forge crate owns conceptually.
  • The hub keeps mailboxes in-process with a file-backed journal by design (single-binary deploy); rebasing it onto pulsing-actor transports for multi-node mailboxes is a natural follow-up — trait boundaries were drawn to allow it.
  • 不再单独维护分仓:上游合并后,独立的 pulsing-forge 仓库将归档,后续演进全部在本仓库进行。

Adds the agent-team coordination layer on top of the existing
pulsing-forge tool runtime (which serves a single agent):

- pulsing-forge-core: gates, closure semantics, Orchestrator trait with
  six built-ins (relay/pipeline/debate/saboteur/broadcast/review);
  29 tests drawn from a 16-entry agent-team failure taxonomy.
- pulsing-forge-hub: TideHub — per-agent mailboxes, exactly-once
  journaled ledger, MCP control port (pulsing_send / pulsing_broadcast /
  pulsing_drain_mail / pulsing_health / pulsing_recall), fan-in debounced
  shop scheduler, forced closure; 10 tests.
- pulsing-forge-runner: forge run <scenario.toml>, forge attach MCP
  bridge, swimlane dashboard; deterministic + sleep-benchmark + codex
  (live LLM) adapters.

examples/forge/ ships six scenarios plus a sanitized Responses-to-Chat
proxy for live-LLM runs; docs/forge/orchestration.md positions the tier
(agents hosted on the pulsing-forge runtime participate via the four
MCP tools — zero runtime modifications).

Verified on this branch: cargo build/test green (39 tests) and the
deterministic leak-review scenario completes a full closed loop.

This branch has not been deployed

No deployments
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