Fulcrum Fleet Rewire

The board asks for four things: more agent types, a hookish and MCP-integrated backend, everything unified under tmux and A2A, and state that can leave GitHub Discussions. This is where each one lands in the code that already exists — which turns out to be most of it.

Written 2026-08-26 ·

Four forks, settled

Everything below follows from these. They are the load-bearing choices — change one and the plan changes shape.

Half of this board is already built

That is the shape of this whole design. It is not a greenfield build — it is mostly a wiring job plus one genuinely new subsystem. The spawn queue already enforces per-role concurrency caps and TTL reaping. The read side of the state migration already has a seam in discussion_cache.py. What is actually absent is the MCP layer and a cross-pane event path.

The invariant that has to change

The single-spawner invariant cannot survive this — persistent peers that message each other is the opposite of one top-level spawner. Replace it with a single-admitter invariant: any agent may request work over A2A, but only Fleet Lead may admit it into a pane, through af-fleet.spawn, which enforces the existing concurrency caps and budget. That keeps the property the old rule was protecting — bounded, accounted concurrency — while allowing the peer messaging the board asks for.

Sequence — and what it costs

Each phase leaves the system working. Nothing here requires a flag day, and the first two are small enough to do before committing to the rest.

1. 00

   Turn on what already exists

   $75

   backend/a2a_broker.py is a complete localhost A2A broker — per-agent bearer tokens, an authz matrix, task cards, SSE push, JSONL replay across restarts, and 204-on-empty so idle agents cost zero tokens. Grep every shell script, service file and module: nothing starts it. Start it from start-dashboard.sh with a health check and a watchdog. This is a day, and it de-risks everything downstream by proving the transport before anything depends on it.

   Sponsor this phase
2. 01

   The tmux substrate

   $150

   One session, fulcrum. Window 0 holds the persistent core panes; work panes open in window 1+ and close when their agent emits its envelope. fleet-up.sh creates each pane, registers it with the broker’s admin port, and injects the bearer token. Pane titles carry the agent id, so tmux list-panes becomes a free process registry and capture-pane a free screenshot path.

   Sponsor this phase
3. 02

   The MCP facade

   $350

   Six servers — af-threads, af-fleet, af-state, af-stats, af-ci, af-repo — each shelling to the script that already does the job. Nothing rewrites logic. What changes is who knows the shape of that script: afterwards, only the MCP server does. Freeze the contract and write the parity tests that guard phase 06.

   Sponsor this phase
4. 03

   The thread store

   $400

   Discussions are load-bearing: 207 non-test files mention them and 38 shell out to gh api graphql. So this is a seam, not a sweep. thread_store.py with github and sqlite drivers, five tables in state.db with an FTS5 index, backfill preserving original D#N numbers, a one-way mirror worker, and a reconciliation sweep — which is not optional, because GitHub stops being the backup the moment it stops being the truth.

   Sponsor this phase
5. 04

   The monitor agents

   $300

   Session Monitor, Display Agent and System Monitor come up as persistent panes on the event stream, and the cron sweeps they absorb get retired. The genuinely new part is error correction: watching a running agent and interrupting over A2A when a run goes off the rails, rather than filing a retro afterwards.

   Sponsor this phase
6. 05

   Review network and PR/CI manager

   $250

   A coordinator over the reviewers that already exist, deciding which of the chain a given PR needs instead of the phase state machine walking every PR through all of them. The label gate and merge path move under PR & CI Manager. The two-gate and CI-status blocks stay hard blocks — they move, they do not soften.

   Sponsor this phase
7. 06

   Migrate the shell control plane

   $475

   The knowledge graph puts 312 file-level nodes in the loop-orchestration layer, the largest non-test layer in the repo. Port it into Python behind the frozen MCP contract, one script at a time, each with a parity test. Open-ended on purpose: there is no deadline where the last shell script has to be gone.

   Sponsor this phase

Still open

• Does GitHub need write-back? One-way is simple and reliable. Two-way needs a poller, conflict rules and identity mapping — a materially larger build.
• What auth does the fleet run under? Org API key, or a Pro/Max subscription over OAuth? This decides whether the Usage and Cost API can replace the local token estimator or whether it stays and just gets labelled honestly.
• What is Reality Check? Either it re-verifies a reviewer’s verdict adversarially, or it verifies claims against the running system — executing the guarded path rather than trusting a report. The second is the more useful agent, but they are different builds.
• How do persistent agents handle context? One long-lived session per core pane grows context indefinitely. Restart-from-memory fits context_manager.py better than server-side compaction.
• Is Fleet Lead’s budget a hard cap or advisory? The circuit breaker already exists for the hard-stop case, so this is really about which one owns the decision.