logicsrc/prd/0008-openfleet-reference-implementation.md
Anthony Ettinger ce8fc7c153 OpenFleet fix round: rebuild the ceiling from the ledger, once-markers, rule 6 in tree, lost only for what a roster can hold
The review of the reference implementation against moshcode found the two
readers disagreeing on the same files. This round applies the shared
rulings so both sides read a ledger the same way.

Ceiling (R-A, R-B, R-C, R1, R6, R10, R15, R17): memberCeiling rebuilds the
effective ceiling from the ledger on every read. The latest fleet-target
fleet.cap (else fleet.open, else the implicit fleet's) replaces the copy in
a record, so a sysop's widening cap reaches running members; then each
swarm.spawn narrowing down the path, then swarm caps last. In the implicit
fleet a parentless record's own approvals enters at the root; a ceiling a
writer left without the key is never read as native, and startMember fills
it with the engine's word while the record is unclaimed. A fleet.open or
cap with no hosts means the host it was written on (R23).

Once-markers (R-G, R28): member.start, member.end and swarm.end each take
an exclusive create under fleets/<fleet>/marks/<event>.<id> before the
append; a lost end takes <id>.lost so a real end can still supersede it.
The hooks let a real end follow a lost line (R9).

tree (R-F, R20): run by the sysop it enforces rule 6, stopping a member
past its effective until with state timeout and the members of a swarm or
fleet at its budget with state budget, then writes swarm.end for each swarm
touched once it is complete. An agent's tree stops nothing. lost is written
only for a member its engine's roster can hold: a claude-code background job
(8-hex member or session) or a moshcode pane, never an interactive session
claude agents does not list (R-E, R3, R14). A nested swarm is drawn under
the member that spawned it and its row shows the effective ceiling (R25).

stop and cap (R-D, R-H, R22, R27): swarm.end is written only once every
member and every nested swarm has an end line that counts; an engine that
will not end a member leaves it without an end line and the verb exits
non-zero. claude stop takes the job id: the member of a background job,
else the first eight characters of a session UUID; an interactive session
with no job id cannot be stopped and the tool says so. cap on a swarm
refuses a key that would widen. A derived claude-code job is named by its
job id and carries no pid.

Also: R-I (endMember ends only the engine-minted swarm of one), R35 (a
derived record's guessed approvals corrected at UserPromptSubmit), R32
(the UserPromptSubmit hook passes only exit 2 through), R31 (package
README), R36 (rule 13 says the launcher test is unimplemented in 0.1),
docs and PRD 0008 updated for lost, rule 6 and the markers. 113 openfleet
tests, 93 CLI tests, contract green.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RZV4zJ2pDZLNN3kE5jFCmV
2026-09-13 10:45:30 +00:00

154 lines
7.3 KiB
Markdown

---
openprd: "0.3"
id: "0008"
title: "Ship the OpenFleet reference implementation"
status: Draft
authors:
- anthony@profullstack.com
created: 2026-09-13
updated: 2026-09-13
repo: profullstack/logicsrc
discussion:
implementation: packages/openfleet
tags:
- openfleet
- agents
- fleet
- swarm
- claude-code
- moshcode
- cli
supersedes:
superseded-by:
---
## Problem
The OpenFleet specification (docs/openfleet.md) was published on 2026-09-13 from
one incident: a Claude Code background job asked moshcode to split a task across
two agents, and afterwards nobody, human or agent, could say who had started
either worker, why, under what ceiling, or with whose approval. The spec names
the record, the ledger and five sysop verbs, and named `logicsrc fleet` as the
reference sysop tool. It shipped with the line "None of the three ships yet".
A standard nothing implements is prose; the incident repeats every morning until
the files exist and something writes them.
## Goals
- A human can open a fleet, set its ceiling, see every agent session under them
as one tree, stop a swarm as one unit, and read afterwards what happened and
who did it, from one command: `logicsrc fleet`.
- Every Claude Code session on a box with the hooks installed becomes a
recorded member: it claims the record its starter wrote, or derives its own,
or is a root member of the implicit fleet, and it refuses to run above the
ceiling it was started under.
- moshcode and Claude Code write the same files the same way, so one tree
shows both engines' members without either reading the other's roster.
## Non-Goals
- No orchestration: splitting a task, choosing an engine, verifying, and
synthesising stay in `moshcode swarm` and `@logicsrc/agentswarm`.
- No change to Claude Code itself. The engine side ships as hooks over its own
settings file; FleetView grouping and `member.spend` at intervals wait on the
engine.
- No network surface, no signed ledger lines, no freeze, no `adopt`: the 0.2
questions stay open.
## Users
- The sysop: one developer answerable for every agent session on their box,
who wants to see the tree and stop the wrong part of it.
- An agent that spawns a swarm and wants its children to know who they are and
what they own.
- A later session inspecting a member's record to answer the five questions the
incident could not.
## Requirements
- R1 [P0] `@logicsrc/openfleet` 0.1.0: the record (write once, never
overwrite), the ledger (append-only, 0600, merged across `ledger*.jsonl` by
`at`), the ceiling rules (whole fleet ceiling, narrowed swarm keys, merge
that never widens, refusal by key), claim and derive exactly as the spec's
"Claiming and deriving", and `fold` into the tree the landing page shows.
- R2 [P0] `logicsrc fleet open|cap|tree|stop|log` with the spec's flags;
`open` and `cap` exit 4 when `OPENFLEET_MEMBER` is set; `stop` exits 4
outside the caller's subtree; `stop` ends nested swarms first and writes one
`swarm.end` per swarm, only once every member and every nested swarm has an
end line that counts, and exits non-zero when an engine would not end a
member; `cap` on a swarm refuses a key that would widen; every verb takes
`--json`.
- R2a [P0] Rule 6 lives in `tree`, run by the sysop: a working member past
its effective `until` is stopped through its engine and ends `timeout`; a
swarm or fleet whose summed `member.spend` in the budget's unit has reached
its budget has its members stopped, each ending `budget`; each swarm touched
gets its `swarm.end` when complete. An agent's `tree` stops nothing.
- R2b [P0] The effective ceiling is rebuilt from the ledger on every read:
the latest fleet-target `fleet.cap` (else `fleet.open`, else the implicit
fleet's) replaces the copy in a record, widening included; then each
`swarm.spawn` narrowing down the path, then swarm caps last. In the
implicit fleet a parentless record's own `approvals` enters at the root, and
the engine fills a ceiling a writer left without the key.
- R3 [P0] `stop` goes through the member's own engine: `claude stop` for
`claude-code`, `moshcode herd kill` for `moshcode/*`, `tmux kill-pane` for
`tmux`, a signal for `claude-p`. Never a shell string.
- R4 [P0] Claude Code hooks: `logicsrc fleet hook <Event>` for SessionStart,
UserPromptSubmit, PreToolUse, Stop and SessionEnd, and `logicsrc fleet hooks
install|remove|status` that merges into `~/.claude/settings.json` and never
clobbers it. A hook never fails the engine; a refused start exits 2 before
any `member.start`.
- R5 [P1] `tree` reads `claude agents --json --all` and
`~/.moshcode/herd/sessions.json` when it can, draws recordless sessions as
roots of the implicit fleet, and writes `member.end` state `lost` for a
recorded background job or pane its engine's roster can hold and no longer
lists. `claude agents` lists background jobs only, so an interactive or `-p`
session (a UUID member with no job id) is never marked lost by it.
- R6 [P1] The spec and the landing page say what ships, keep `Status: 0.1`,
and record the two verified Claude Code limits (no launcher environment
reaches a dispatched background job; exported variables reach tools but not
later hooks).
- R7 [P1] Tests cover record and ledger IO, every narrower case, the worked
example's claim and derive, the folded tree, hook install idempotence, and
each hook handler, including the exit-2 refusal and the PreToolUse deny.
## UX Notes
`logicsrc fleet tree` prints the tree the landing page shows: fleet header,
root members, swarms nested under their spawner, members with engine, state,
`[bypass]`, `owns`. `log` prints one line per event, oldest first, with who did
it. Refusals name the key, what was wanted and what was allowed.
## Tech Stack
TypeScript, NodeNext, commander 14, vitest 4. No workspace dependencies beyond
the CLI's `file:../openfleet` link. Node 18+ (the installer's floor), so the
tree is plain text rather than a TUI.
## Monetization
None. It is the reference implementation of an open standard.
## Success Metrics
- The worked example's morning can be replayed against a temp home and
`logicsrc fleet tree` prints the tree the spec shows.
- A Claude Code session started with the hooks installed appears in
`logicsrc fleet tree` with the right approvals mark without anyone editing a
file by hand.
## Risks & Open Questions
- A background job dispatched from `claude agents` gets no launcher
environment, so a launcher that wants it in a swarm must write its record and
pass the path another way (a `--settings` hook command, or a lookup by the
job's cwd and intent). Until then it is a root of the implicit fleet.
- User-level hooks fire for every `claude -p` a tool makes, so each becomes a
swarm of one and, at depth 1 in the implicit fleet, is refused on depth. The
spec lists this as an open question; the hooks enforce the letter of it.
- A ledger check before a write is not exclusion, so `member.start`,
`member.end` and `swarm.end` each take a once-marker first: an exclusive
create of `fleets/<fleet>/marks/<event>.<id>` (`.lost` suffixed for a lost
end, so a real end can still supersede it). moshcode uses the same paths.
A marker taken by a writer that then crashed before appending leaves the
line unwritten until someone clears the marker by hand; 0.1 accepts that
over a doubled audit line.