Zip a run

One command produces the bundle; the level decides how deep it reaches.

bun .claude/skills/zip-run/scripts/zip-run.ts <run-dir | campaign-dir> --light|--medium|--verbose \
  [--run <runId>] [--out <zip>] [--max-mb N] [--json]

<run-dir> is the launched worktree (holds .scratch/quick-run/launch.json, for example ~/Developer/hb4-run-<runId>). Its campaigns/ tree also carries campaigns copied from other runs, so the script keys on launch.json's runId and takes the one campaign whose controller/<runId>/opening.json exists. A campaign directory is accepted directly, and --run names the run when it holds more than one. The batteries are the ones batteryRunDirs matches by this run's canonical iteration ids, so a sibling run whose id merely starts with this one never enters, and MANIFEST.json records the controller's strict reading of the run (controller, controllerError) without letting a refusal stop the handover. The zip lands at --out, default ./<runId>-<level>.zip; put it in the session scratchpad and send it with SendUserFile.

Pick the level from the request

operator says	level	measured on the Sol truss run of 2026-09-12 (11 epochs, 450 cases)
"traces", "reasoning", "logs", "under 2 MB", handover to a reviewing model	--light	1.34 MB, 544 files
"what each case did", "tool calls", "verifier and Judge output", "the whole evaluator"	--medium	6.9 MB, 3,020 files
"everything", "everything the model saw", "full tool outputs"	--verbose	44 MB, 45,149 files

A run bundle of 20 MB or more is verbose by definition. If a lighter request produces that, the level is wrong, not the cap. --light fails when the zip exceeds 2 MB; --max-mb sets any other cap. When a cap fails, drop the largest group the script prints or choose a lighter level; do not strip files by hand inside the zip.

What each level holds

Levels are cumulative. Every level includes README.md (levels, groups, what was left out, then the level's reading guidelines) and MANIFEST.json (every file with byte size and group, run id, campaign, source commit). The guidelines are presets in readme/<level>.md: when to choose that level, the reading order, what the bundle cannot answer, and what not to conclude from it. Edit the preset, not the script, to change what a recipient is told.

light: how the run started and ended, what each round decided, and what the Builder was thinking.

• launch/: launch.json, stop.json, probe.json, fullrun.log.
• controller/: opening.json, terminal.json.
• campaign ledgers: epochs.json, budget.json, case-record.jsonl, isolation/*.json.
• claims/, difficulty-decisions/, analysis/ (rebuild advice packets, epoch reviews, admissions), promotions/, safeguards/.
• epochs/<epoch>/: builder-prose.jsonl (reasoning per turn), builder-execution.json (tool tallies, failed calls, submits), builder-session.json, backends.json, campaign.json.
• versions/<version>/: version.json, conformance.json, claim-stages.json, agent/ (tools.ts, tools-spec.json, BUILT_AGENTS.md), correctness-model/brief.json and evaluator.ts.
• versions/<version>/runs/<runId>/: battery.json, run-manifest.json, backends.json, and cases/<case>/case-result.json only.

medium adds the per-case and per-check material.

• cases/<case>/: trace.json (Built Harness turns and tool calls with argument digests and result previews), verifier.json, judge.json, public-task.json, final-submission.json, artifact.json.
• runs/<runId>/judge/: census sample, bait corpus, controls, standing.
• whole correctness-model/: tasks.json with hidden expectations, controls.json, reference/, evaluator tests.
• epochs/<epoch>/: builder-path-record.jsonl (every guarded file access), trials/ and the NN-<slug>/ preview receipts (census.json, conformance.json, solvability.json, iteration.json), verifier-workshop.jsonl.

verbose adds everything the models saw and everything the host recorded, minus binaries.

• observability/<runId>.jsonl: every start prompt the Builder, Built Harness and Judge received, in full (44 MB uncompressed on the Sol run: 256 solver prompts, 745 Judge prompts, 11 Builder prompts).
• controller/verifier-lifetime/: intent.json, spawned.json, settlement.json per verifier process (13,702 processes on the Sol run).
• every remaining case file: built-runtime.json, built-registration.json, draft-checkpoints.json.
• epochs/<epoch>/workspace/ and .oss/ metadata: the Builder's authoring tree, STARTER.md, generated scripts. .toolchain/, .oss/downloads/, .oss/build/, node_modules/, .bundle-snapshots/ (preview candidates; the accepted ones are under versions/) and any file over 8 MB are excluded and listed under excluded in the manifest.
• controller.sqlite.
• transcripts/: full session transcripts with complete tool outputs, when the backend left them on disk. Claude-backed runs keep them under ~/.claude/projects/<epoch workspace path with / replaced by ->/*.jsonl; the script finds those per epoch. Codex-backed runs keep rollouts only if the frozen CODEX_HOME (.scratch/quick-run/codex/sessions/) retained them; the Sol run of 2026-09-12 did not, so its verbose bundle carries no Codex rollouts and the README says so. Do not go looking in ~/.codex/sessions for them: rollouts there with the run workspace as cwd are later review sessions, not the Builder.

Never included at any level: installed toolchain binaries, dependency trees, .git, provider auth.json, .DS_Store.

After zipping

Report the zip path, the compressed size, the file count and the group table the script prints. Name what was left out in one line; the README inside repeats it. When the operator asked for a size, say whether the cap held. If a verbose request finds no transcripts, say that the run recorded prompts, reasoning and tool-call previews but not full tool outputs, and where that limit comes from (Codex rollouts not retained under the frozen home).

Focused check

test/zip-run-script.test.ts builds a synthetic run tree and proves the three levels select and exclude the named groups, that --light reports its cap, and that launch.json's runId selects the campaign among copied ones.

scripts/worktree.sh run <dir> bun run test -- test/zip-run-script.test.ts