Skip to main content
← Back to list
01Issue
BugOpenSwamp CLIPublic
AssigneesNone

Relationships

#4998 v3: a swamp child process started from inside a method can't join its parent's writer lease, so it waits 60 s and fails

Opened by skunk-ape · 11/22/2026

What happens

On a local v3 datastore with no serve, a method that runs swamp as a child process (for example to run another behavior or workflow) deadlocks against its own parent. The parent CLI command holds the datastore's writer lease for its whole life (#4993). The child is a new process, asks for the lease, waits the full lock timeout (60 s by default) and exits 75 with datastore_v3_writer_busy.

Per-behavior locks already handle this case. A parent passes the locks it holds to its children through SWAMP_LOCK_HOLDER_TOKENS, so a child joins them and doesn't wait. The v3 writer lease isn't part of that hand-off.

On v2, and on v3 through swamp serve, the same child process works.

We found this while making stagecraft publish to a tracker from inside a write: its child-process publish route can't work on local v3 without serve. Stagecraft now uses an in-process route where context.runModel() allows it.

Steps to reproduce

  1. swamp repo init and swamp datastore setup filesystem --path ~/ds3 --format 3. Don't start serve.
  2. Write an extension method that runs swamp behavior run <other> <method> as a child process and waits for it.
  3. Run that method with swamp behavior run.

Actual: the child waits about 60 s, then exits 75, datastore_v3_writer_busy, naming the parent's pid. Expected: the child joins the parent's lease and runs, as it does for per-behavior locks.

Measured on swamp 20261120.161258 (sha 7b758e05).

Cause

  • The token hand-off lives in src/domain/datastore/lock_holder_marker.ts (SWAMP_LOCK_HOLDER_TOKENS, line 41; childLockEnv). It carries per-behavior lock tokens only. Nothing there covers the writer lease.
  • The writer lease is reference-counted per process (src/infrastructure/datastore_v3/v3_writer_lock.ts:138, :153-178), so a child process can't share it today.

Suggested fix

Extend the hand-off to the writer lease: the parent passes a lease token in the child's environment, and the child's holdV3Writer joins the parent's hold rather than taking a new one, while the parent is alive and still holds it. Check against the checkout protection, the writer marker and interrupted-writer recovery, which assume one writing process per checkout.

This is part of the v3 write-model decision in #4993 (the whole-command lease hold). If #4993 moves to a per-commit lease, this changes shape, so decide them together.

Version

swamp 20261120.161258 (sha 7b758e05). Code references to main at e52c839ba.

02Bog Flow
◉OPEN○TRIAGED○IN PROGRESS○SHIPPED

Open

11/22/2026, 4:49:52 PM

No activity in this phase yet.

03Sludge Pulse

Sign in to post a ripple.