Director: a coordination ledger for coding agents

$ claude
  … deep into the pagination rework …
▸ decision   recorded · cursor pagination, not offset; offsets break under deletes
  …
▸ open-item  recorded · timezone edge case before the backfill merges
  … later …
▸ handoff    parked · cursor rework done · next: the backfill script · watch the p99

──────────── session ends ────────────

$ codex
> where were we?
▸ Director: acme-api-main-7c21e9d4 · 1 open-item(s), 1 need-you

## open-items
01KWJ4X2…  open-item  timezone edge case before the backfill merges  [risk:escalate]

## handoffs
01KWJ9RM…  handoff    cursor rework done · next: the backfill script · watch the p99

## decisions
01KWJ4W8…  decision   cursor pagination, not offset; offsets break under deletes

recorded while it works · loaded when the next one starts

how it works

An append-only log per repo. Deterministic projections. A push at session start.

A Claude Code session records decisions, open items, and handoffs to an append-only log; a Codex session a day later starts with them injected and appends to the same log. — sessions end · the log doesn’t · one file, any harness

$ director emit

Sessions record as they work: decision · open-item · handoff · note. Four kinds, one append-only NDJSON log per repo, lossless under concurrent writers.

$ director resolve

Loops close consciously. An open item stays visible until someone resolves it; nothing scrolls away, nothing silently vanishes.

$ director brief

The log collapses deterministically into the machine digest, the human brief, and the fleet cockpit. Every session and the human read the same picture, byte-identical, verifiable.

SessionStart hook

Every new session starts from the injected CHARTER + digest: ground truth pushed at session start, not recalled by similarity. A cold re-entry starts from your parked handoff.

across harnesses

One ledger, four harnesses. Zero copy-paste.

the standing review panel

Build in Claude Code with Codex, OpenCode, and Copilot CLI reviewing the same repo. Every harness opens on the same injected ground truth, and a review verdict lands back in the log for the next session to pick up.

the limit wall

Hit a usage limit mid-task? Open another harness on the same repo: it starts from the same digest. The wall costs you the tool, not the thread.

neutral ground

The state of the work lives in a plain NDJSON log on your machine, not in any vendor’s session format. Different vendors, different models, one state of the work.

$ director install --all

Same boundary commands, same ground-truth injection at the top of every session, on all four. Wire one agent or all of them; they read and write the same log, live.

where it fits

Keep your tools. This is a different layer.

vs orchestrators: Nothing here spawns, routes, or commands agents; Director refuses that job on principle. The name names your seat, not the tool’s: your agents keep the ledger, and you direct the work.
vs compaction: Compaction compresses the transcript at an arbitrary moment and hopes the right things survive: a rot vector, not a cure. Director records the work itself: typed facts, written when they happen, that outlive the session.
vs memory tools: They answer what does the agent know; Director answers what is the state of the work. Facts accumulate; loops open and close. Run both: they don’t overlap.
vs issue trackers: The tracker holds the work items. Director holds the narrative between tasks: the decision that reframed the task, the loop deferred while doing it, the handoff parked when the block ended.
vs a markdown file: Append-only integrity under concurrent writers, byte-identical verifiable views, lifecycle semantics, push-injection. Willpower is not a file format.