Repository navigation
Reduce durable-state duplication and heartbeat volume #1147
Description
Suggested labels: enhancement, priority:P2, storage, performance
Problem
Long campaigns retain repeated workspaces, trusted CLI closures, continuation snapshots, adapter data, and sub-second heartbeat events. One observed run consumed roughly 13 GB, and more than 90% of its large event stream consisted of heartbeats.
Generic reproduction
- Run a detached campaign for several hours.
- Continue it multiple times.
- Measure the run directory and count event types.
Expected behavior
Durable state grows mainly with unique artifacts and meaningful transitions.
Proposed fix
- Content-address immutable closures and workspace snapshots.
- Reference shared objects instead of copying them per continuation.
- Coalesce unchanged heartbeats into intervals.
- Rotate or compact event logs while preserving terminal events.
- Apply retention policies to superseded attempts.
- Expose a storage breakdown in the CLI.
Acceptance criteria
- Continuation does not duplicate unchanged closures.
- Heartbeat storage scales with reporting intervals rather than polling frequency.
- Compaction preserves terminal and verification records.
- The CLI identifies the largest retained categories.
This issue intentionally omits the audited repository and all audit results.
Activity
fix(runtime): stop Smithers persisting a TaskHeartbeat event per hear…
fix(runtime): stop Smithers persisting per-pulse heartbeat events and…
feat(runtime)!: resume runs the pnpm-patched runner instead of instal…
aviggiano commentedon Sep 29, 2026
Status update. Heartbeat volume is fixed, and the duplicated per-run trees are the remaining cost. Those are tracked in #921's plan, so I'm keeping this open.
Landed
TaskHeartbeatevent. It carried no data, andNodeStarted/NodeFinishedalready bound the interval. It was 83% of event rows (3,925 of 4,740) in a tiny-vault campaign database.Draft
Measured on current
main(tiny-vault smoke campaign)These two dominate durable state. #921's plan replaces them with a slim run build (~20 MB / ~1.8k files of first-party code, with third-party packages linked into the install) and a
PATHshim. It is waiting on the owner questions there.Not built
_smithers_eventswould break Smithers' seq pagination and Ultrafuzz's full-history event reads.feat(runtime)!: resume runs the pnpm-patched runner instead of instal…
feat(runtime)!: resume runs the pnpm-patched runner instead of instal…
feat(runtime)!: resume runs the pnpm-patched runner instead of instal…
feat(runtime)!: resume runs the pnpm-patched runner instead of instal…
feat(runtime)!: resume runs the pnpm-patched runner instead of instal…
feat(runtime)!: run lifecycle commands from the pnpm-patched install …
aviggiano commentedon Oct 2, 2026
Closing as a duplicate of #921. The heartbeat half is fixed: #1166 (aa58130) stopped Smithers persisting the per-pulse
TaskHeartbeatevent, #1201 (f1789d7) removed the ~595 MB TMPDIR controller root left by each resume, and #1245 stopped copying Frog into every run. The dominant remaining cost, the per-generation execution snapshot and the per-run trusted CLI closure, is #921's slim-run-build andPATH-shim steps; unreaped task worktrees are tracked in #1227. Event-log compaction (it would break Smithers' seq pagination and full-history event reads), pruning of superseded generations/attempts, and a storage-breakdown command are not planned; refile any of them with a measurement if they still matter after #921 lands.