Lab #2498: workflow resume of a suspended run crashes mid-run when the workflow was edited during the approval window

Problem

swamp workflow resume <wf> --run <id> on a suspended run walks the current workflow YAML against the stored run. If the workflow was edited during the approval window, the resume does not detect it. The mismatch is only found after work has started. The run is then left with steps and jobs stuck in running, and nothing clears them. In the moved shape, a real side effect has already happened when the resume crashes.

This was split out of swamp-club#2433, which fixes the same gap for resuming a failed run (--from and the #2409 retry). The approach for failed runs (a structure check before any change) does not carry over directly: suspended runs hold unfinished records that are legitimate, and some edits that work today have to keep working.

Reproduction

Installed swamp 20260924.141756.0-sha.5b8a2673. Job main runs prep → gate (manual_approval) → deploy → notify, all in one job. Run it; it suspends at gate with deploy and notify pending and main running. Edit the YAML, then run swamp workflow approve <wf> gate --run <id> followed by swamp workflow resume <wf> --run <id>.

Added step. Add lint to main, depending on gate. The resume fails with Error: Workflow resume failed: Step run not found: lint and exits 1. History: the run is failed and main is running. deploy is left running with no completedAt, although it never ran. notify is pending.
Removed step. Remove notify. The resume exits 0 and the run succeeded, with notify left pending. This works today and must keep working; only the leftover pending record is cosmetic.
Moved step. The original YAML also has job post (dependsOn main) with step announce. Move notify from main into post. deploy runs (its data artifacts are written), then the resume fails with Step run not found: notify and exits 1. History: main succeeded with notify pending. post and announce are left running.

Scripts: ~/swamp/repro/issue-2433/rerun.sh suspended-added|suspended-removed|suspended-moved on the triager's machine. They build a fresh repo, run, swap the YAML, approve and resume.

Cause

resume() (src/domain/workflows/execution_service.ts) walks every job of the current workflow whose stored record is not terminal. runJob then walks every step of the current job and looks up that step's record in the same stored job. A step with no record there throws Step run not found mid-run. A record the workflow no longer puts in that job stays pending, and the job and the run still complete as succeeded.

Design constraints, carried over from the #2433 adversarial review

Tell removed apart from moved. A pending record whose step exists in no current job (removed) must not be refused, because the removed shape succeeds today. A record whose step the current workflow places in another job (moved) must be refused. (#2433 ADV-1)
Narrowing a forEach at resume must keep working. resume --input can shrink a forEach collection. Iterations that were already expanded stay pending, and today the job and the run succeed. A blanket "no success while a step is pending" rule would break this. (#2433 ADV-2)
Refuse in the resolver. Serve registers the run and charges the principal's cap before resume() runs (src/serve/resume_launcher.ts), so the check has to run in resolveResumableRun and the suspended resolver. A refusal also leaves the run suspended and awaiting resume, so the message has to name the way out: swamp workflow cancel --run <id>, or a new run that supersedes it. (#2433 ADV-4)
recover() re-enters through this path. recover() calls resetUnknownStepsForRecovery(), which sets the run to suspended, and saves the run before it calls resume(). A check on the suspended path has to run before that mutation, or accept that recovery is refused afterwards. The fingerprint check usually prevents this, but runs started with --last-evaluated record no run plan. (#2433 ADV-11)
Allow evaluated names. Runs started with --last-evaluated, and legacy forEach records, carry no forEachTemplate and have expression-named templates. Their unfinished records must not be refused. (#2433 ADV-3)

A nested structural swamp still waits on its run's lock when the lock holder is not a same-host ancestor (cross-host worker, --server into a non-ancestor serve)

A nested structural swamp that outlives the run that started it keeps skipping that run's lock while the holder writes

Show signal waits to remote clients and the dashboard

Settle expired signal waits from swamp serve

Resume signalled workflow runs automatically under swamp serve

Deliver workflow signals through a write-once outcome record instead of writing the run record

Forward held-lock identity over --server so a loopback nested swamp does not wait on its caller's lock

A nested structural swamp on a same-host remote worker waits on its own step's lock held by swamp serve

Fail fast when nested structural commands under parallel runs wait on each other's locks

Docs: remove followUpActions from the MethodResult reference in the extension model manual

signal_test: already-settled test fails by chance when a random UUID spells the sender name

Add a wait_for_signal workflow step that pauses a run for a JSON message

Deliver workflow signals through swamp serve

run doctor through serve interrupts a live run of another serve instance when no heartbeats are recorded

Flaky test: WalSink replayed segments are delivered before events written after replay leaves one WAL segment

Remove followUpActions: no model or extension can produce them

Workflow schedule watcher and workflow edit symlink lookup ignore the managed config workflows dir

main is red: repository_dirty_coverage_test fails for UnifiedData.collectGarbage(orphaned deferred write)

isProcessGone always reports alive on Linux under the test task permissions, failing 9 tests on main

Flaky test: WalSink replayed segments are delivered before events written after replay

serve: relative tls cert-file/key-file and --config resolve against the working directory, not the repository

findBySpec/findByTag lose a workflow step's output after a version is deleted, pruned or rolled back

serve HA: worker enrollment on a replica that lacks the enrollment token's definition may create a second definition (unverified)

serve HA: pull a token's definition on an auth miss so a peer accepts a new token at once

Run tracker keeps some interrupted workflow rows forever: retention waits for markSettled, which several paths never call

run doctor reads every run record on each run to rebuild the run indexes

Serve audit: two sinks with the same name share one emitter cursor, so the second never receives events

Serve audit: a durable sink write that times out is retried while the first call is still pending, so the store can hold duplicate sequences

Flaky test: ModelResolver data accessors for a renamed model return the earlier id's record from findBySpec

data query --limit returns fewer live records than exist and reports limited: false when stale catalog rows fall inside the limit

data get: help example passes --run latest, which is not a recognised run id

workflow reject leaves sibling gates waiting_approval and dependents pending in a failed, retryable run

Show expired approval gates as expired, with a Cancel action in the dashboard

digitalocean codegen leaves an orphaned security_secret.ts model file after its endpoint left the spec

workflow cancel cannot cancel a suspended run whose workflow file was deleted, and cancel --all silently skips it

A job alone in its level starts after the abort, and its unstarted step is recorded as a real failure instead of settledByAbort

workflow approve racing workflow cancel loses the cancel: the run returns to suspended after cancel reported cancelled

data query: does not follow data rename forwarding that data get follows

Unit tests share fixed /tmp catalog paths, so concurrent runs corrupt them and fail every later run

data query: renamed data items are not found by their old name

data query: model lookup by id and orphan recovery parity with data get

workflow approve on a run that is not suspended prints a fatal error with a stack trace

data query: no way to target a workflow's latest run

data query: binary content lacks contentEncoding/base64 parity with data get

data query: no fallback from spec name to data instance name

data query: no definitionHash or garbageCollection fields in query results

A nested structural swamp spawned by swamp serve skips every lock serve holds, including other runs' locks

data query: confirm catalog parity with data get for custom datastores

data query: single-result --json mode for scripts migrating from data get

Docs: document data query --single and fix the stale --limit default in reference/data.md

data get --workflow silently returns an arbitrary step's data when several steps share a data name

swamp doctor workflows misses nested and .yml extension workflows that the loader reads

auth whoami says the stored API key is no longer valid when auth.json only holds an env-key identity cache

Docs: swamp skill says cancelling a parent run cancels its nested child runs; the child is left suspended

Namespace advice in the slow-lock warning is wrong for a single repo on a local datastore

Method-run records left running by a dead owner are never reaped

Flaky: usecase_sync_characterization_data_test 'data gc (serve)' asserts markDirty order that varies under parallel load

Worker dispatch rebuilds the definition with Definition.create, so a legacy-named model cannot run on a worker

A cancelled workflow exits before its steps stop, leaving their method-run records at running

Nested swamp structural command in a shell step times out on its parent's per-model lock (SWAMP_LOCK_HOLDER_PID is stripped)

Auth gate blocks nested swamp in workflow shell steps when the credential is in SWAMP_API_KEY

swamp workflow create crashes with an [FTL] ZodError stack trace on an invalid (uppercase) workflow name

swamp model create crashes with an [FTL] ZodError stack trace on an invalid (uppercase) model name

model cancel SIGTERMs a running swamp serve process when the method run belongs to a serve-run workflow

Follow-up action retries ignore the run's abort signal and can outlast a cancel

model cancel --all on a running workflow leaves its method-run records at running, or records them failed instead of cancelled

model cancel SIGKILLs the owner 2 s after SIGTERM, before the step executor's own 3 s kill grace, so the run never records its own cancellation

workflow cancel and supersede of a suspended run leave its jobs running in the cancelled record

Tell timeouts apart from cancels in method-run records, and review the hidden 30 s fallback timer for step-called model methods

A force-exited workflow run stays running with a dead pid: resume, recover and run doctor --fix cannot clear it

A nested workflow that suspends on manual approval forwards its suspended event into the parent run's stream

workflow cancel SIGKILLs the run 2 s after SIGTERM, cutting off always/completed cleanup jobs mid-step

Warn when datastore setup keeps managedConfig but the new datastore has no config tier

Docs: run doctor --fix and workflow recover settle runs left running by a force-exited process

serve: a relative grants-dir resolves against the working directory, not the repository

workflow resume cleanup mode can run an always-gated teardown before a suspended job's approved work

Docs: serve-flags.md should state that relative grants-dir and grants-file resolve against the repository

managedConfig: extension writes update the shared lockfile from a stale cache and overwrite peers' entries

managedConfig: model/vault create and other bulk-push commands re-upload a stale extension lockfile and erase peers' entries

Cancel, reject and supersede of a parent run should settle its suspended nested child runs