Your coding agents.
One workbench.

The rollout ladder

1. Default first stage1 · shadowGates run, would-merges recorded, nothing merges
2. 2 · 2aFirst repos merge, low risk
3. 3 · 2bMore repos
4. 4 · 2cLeader class A
5. 5 · 3aMedium caps, Leader A and B
6. 6 · 3bMore repos
7. 7 · 3cMore repos
8. 8 · 3dLast signed stage

The default grant draft starts in shadow. The ladder advances when each stage's criteria are met in the ledger (shadow: 5 would-merges and 12 hours), drops back one on a breach, and cannot pass the last stage you signed. A separately signed elite-direct grant uses one rung instead: eligible elite models may land verified PRs immediately, and a breach restarts that rung's evidence window. Command shows the current stage, progress and the grant's countdown.

Every change, through the same gates

• G0Authority
• G1Protected paths
• G1bTamper
• G2Scope
• G3Verify
• G4Claims
• G5Blast radius
• G6Judge or signed elite-direct
• G7GitHub checks

Fleet lists every shadow decision with these chips and a one-sentence reason, so you can read what the fleet would have done before it does it.

Reversible merges

When a stage merges, the merge happens on GitHub, pinned to the head SHA, with grant, gate and ledger trailers. CI and a fresh re-run watch it for two hours; a red merge is reverted and the repository quarantined. Protected paths always go to you.

Host-verified

The fleet's GitHub App posts ashlr/verify only when verification passed on exactly the tree it proposes. Free-plan private repos, which cannot have rulesets, use local enforcement with the same check and low-risk scope.

Stop and revoke

Lowering never asks: the switch drops to Propose or Off at once, Stop halts running agents and armed merges, and Revoke needs a new grant to resume. A broken ledger chain halts everything.

Re-approval

A release that changes authority code pauses the grant until one more Touch ID, and the ladder continues from where it was. The resident daemon re-verifies grant, Stop and switch on every tick.

Your change-volume limits

Choose files and lines per change and merges per repo per day when reviewing the signed scope, or explicitly choose no volume cap. Existing grants keep their old effective limits until you edit and sign them. Risk, protected paths, spending and verification remain separate.

Codex is off for autonomy unless enabled. In an admitted run, a seat whose usage cannot be read is ineligible, because your reserve is not spent on a guess.

Six destinations, two work modes, one palette

Operator setup: Access, a private gateway config and Mac approval are required.Read the phone runbook

# the same wiki from the terminal
ashlr wiki build ashlr-hub
ashlr wiki ask "where are standing grants verified?"

Synthetic regression fixture: one mocked account takes 20 seconds; three others take 10 ms each. Both paths allow two clients. These are controlled test-clock observations, not provider latency or a live startup benchmark. Inspect the test.

Measure your own configuration with ashlr benchmark run; this explicitly starts local model work. ashlr benchmark --compare-reports BASE CANDIDATE is offline and refuses incompatible or uncontrolled receipts. It reports matched observations, not universal savings. Benchmark guide.

Missing deadlines and work history stay unknown. Jev can advise among eligible choices; normal routing checks current account capacity again. A percentage is never treated as a token allowance. Read the scheduling guide.

Recorded online qualification · October 1: one multi-file rename fixture passed its checker in 392.5 seconds, changing three files across 12 turns. The requested local model was Qwen3.8 27B Q8_0; the runtime reported its model name as unknown. Cache state was uncontrolled, so this is one observed run, not a baseline/head comparison or a general performance claim. Inspect the sanitized receipt.

Producer success, a recorded proposal and shipped work remain separate. Replayed records do not multiply lessons; incomplete data stays visible. Opening feedback uses a shared background read without calling a model. Inspect the feedback contract.

Use allowance before resets has global and account controls. Off disables reset priority and reserve shrinking. On still requires current task fit, signed floors and verified billing and execution-account boundaries; current native producer binding remains held. Purchased balances remain separate. Gray historical usage keeps its original timestamp while live readings load; captured dollar balances retain their capture time and expiration. Understand resource evidence.

Worth saying plainly, because most tools will not: on a single GPU, local parallelism buys throughput rather than speed — about 1.1× more work per minute than running the same agents one after another. That is a property of one piece of silicon, not a limit of the local console. Seats on different accounts can run in parallel when their providers and resources allow it. Separate machines can run independent Phantom workspaces; coordinated multi-machine control is not shipped.

The large win above was the cache fix, and it helps a single agent just as much.

The table shows the 3.10 result. In 3.11.2, first-chat-paint JavaScript fell to 349.7 KB, meeting the 350 KB target. The 3.10 runtime probe measured a 53–78 ms event-loop stall against a 20 ms budget; no newer result is reported here.

Chats and agent tools use the context you provide to your configured providers. Account readings keep their source and billing qualifiers; see resource evidence. This page adds no signup or tracking.

npm i -g @ashlr/phantom
phm verse

ashlr resources profile prepare --provider claude \
  --directory ~/.ashlr/native-profiles/claude-a \
  --executable /path/to/claude

ashlr leader say "focus: the flaky retry test"
ashlr comms setup-telegram

ashlr authority setup --dry-run
ashlr authority setup
ashlr authority resident start

ashlr devin connect
ashlr devin budget --acu 50