Shell jobs · 32 KiB output

Raw snapshots are private and are not served by this website. Only filtered structural findings are published.

Raw idle observations. These observations are separate from the slope fit; a lower heap after idle does not by itself establish that completed outputs were released.

Records are matched by the Core Job.Active reference fields and fixture session, not by class name alone. Fixture marker and child-PID prefixes identify their output nodes. The two control jobs are the real preflight commands.

Snapshot limitation. Bun exports these job objects without incoming edges, so no complete GC-root retaining path is available. Missing paths do not establish collectibility. A two-process calibration using the same embedded Bun reproduces this for a live Map-only value, while a direct object property has an incoming edge.

The actual record → info → output reference edges are preserved in the filtered data. Bun truncates string labels to 1,024 characters, so the snapshot identifies output prefixes rather than verifying all payload bytes. Exact output size was checked during execution. Output-node shallow bytes are not total retained size.

Filtered object counts and reference edges

The process-global Job service stores completed jobs, their deferred results, metadata, and output in its Map. The shipped shell/subagent executors create jobs, but Job has no removal or bounded-exit cache, unlike Shell's 25-entry cache.

• Invoke the installed built-in shell executor, not a replacement.
• Each real command returns exit 0 and exactly 32768 bytes with the fixture prefix and a child PID. The command text stays constant across cycles.
• The fixed session's message history stays empty; the fixture retains only first/last shell IDs, never output history.

control-1

Functional preflight passed (2 case + 2 control cycles).
{"completed":2,"controlCompleted":132,"progressCalls":2,"outputBytes":65536,"bytesPerCommand":32768,"sessionMessages":0,"sessionID":"ses_f9bdd9fb7ffeLoIky5kCuYjEER","firstShellID":"sh_064226076001gopBdIh0I0Vnd2","lastShellID":"sh_0642260bc003ei8m2kfelSXjRi","outputMarker":"JOBMEM_LARGE:"}

case-1

Functional preflight passed (2 case + 2 control cycles).
{"completed":132,"controlCompleted":2,"progressCalls":132,"outputBytes":4325376,"bytesPerCommand":32768,"sessionMessages":0,"sessionID":"ses_f9bdcd07affeQYoLv2UZAJ0Nr6","firstShellID":"sh_064232fc5001KpAAbslOM7SXvF","lastShellID":"sh_064239bb50015tSUYQSv54kWKN","outputMarker":"JOBMEM_LARGE:"}

Control cadence is calibrated from two real preflight cycles in each process and recorded as controlCadenceMs. Both arms get at least 20 seconds of warm-up/settling before the baseline. This approximately matches elapsed time, not every internal allocation or polling read.

Source inspected at 4772b6a3e8; this is the investigation reference, not a claim about the binary’s build commit.

• packages/core/src/job.ts:187 — Completion replaces the Map entry with a completed record and output; it does not remove it.
• packages/core/src/tool/plugin/shell.ts:242 — Every shipped shell-tool execution starts a Job and waits for it.
• packages/core/src/shell.ts:29 — The separate shell-process cache is capped at 25 exits.

One fixed session and one inspector transform remain. Commands finish before measurement; the fixture retains neither results nor a growing ID list. Session context remains empty.

Control. Run the same executable and output through a plain subprocess, without OpenCode's shell/job bookkeeping. Both arms are cadence-matched and warmed beyond the shell service's 25-exit cache.

Metric: target process.memoryUsage().heapUsed after Bun.gc(true). Warm-up excluded. Theil–Sen slope; later-half slope must remain positive. Subtract positive control drift only. A growing control, large baseline drift, or a noise allowance above max(1 MiB, 5% of baseline) that prevents detecting a trend makes attribution inconclusive. Net growth and slope × cycles must both exceed max(128 KiB, 6 × combined robust residual noise). Require agreement across at least two fresh-process repetitions. Heuristic, not a significance test.

OpenCode

0.0.0-dev-18911 · dev channel · installed binary

Drive

1.4.4 · installed package

Runtime

Bun 1.4.0 · linux / x64

Isolation

Private server + file-backed database. TUI closed before warm-up. Measurement plugin remains loaded; no object-level instrumentation.

Checkpoints

Case-specific cleanup checks pass before GC; 300 ms settling allowance. This is not a global background-work barrier.

Snapshots

Baseline and final, plus post-idle snapshots enabled. Raw heaps are private under the run's .cache directory, not served by this website.

Run

2026-09-02T21-59-26-177Z-95f1bd · 2026-09-02T21:59:26.177Z

cd /root/projects/opencode-memory-lab && bun run run --case shell-jobs-large --warmup 30 --batches 5 --batch-size 20 --repetitions 1 --settle-ms 300 --warmup-ms 20000 --idle-ms 15000 --snapshots

• A single repetition is a smoke test, not a repeatability check

Includes native allocations and allocator behavior. Not used to classify JS heap growth.