LegoFlow

BlocksTerminal Curator

Output Format

Published task pool

runtime_info.output.harbor_tasks_dir.path defaults to artifacts/tasks/, a symlink to an immutable artifacts/tasks-versions/<batch-id>/ publication:

artifacts/tasks/
├── verifiable_tasks.txt
├── registry.json
├── validation.json
└── <task-id>/
    ├── instruction.md
    ├── task.toml
    ├── provenance.json
    ├── environment/Dockerfile
    ├── solution/solve.sh
    ├── tests/test.sh
    └── tests/test_outputs.py

verifiable_tasks.txt lists accepted task IDs, one per line. Tracer uses it as its allowlist. registry.json describes the same IDs as [email protected], using absolute paths into that immutable batch for Evaluator. Both switch together; keep versions referenced by existing jobs. See Downstream Handoff for consumer configuration.

Verified, rejected and unfinished tasks

Both Curator and Terminal Curator use verifiable_tasks.txt as the authoritative list of runtime-verified tasks. Directory existence alone never proves acceptance. Their raw layouts differ: Curator's per-language pool can also contain candidates; Terminal Curator's artifacts/tasks/ contains only published tasks, like Curator's verified artifacts/merged_swe_tasks/ export.

LocationMeaning
artifacts/tasks/<task-id>/ plus its manifest entryPublished after fresh Oracle=1 and NOP=0; validation.json binds its identity.
artifacts/runs/<run-id>/candidates/Generated candidates, including partial or rejected attempts when files exist.
artifacts/runs/<run-id>/normalized/Candidates that passed normalization; Oracle/NOP may still fail or be unfinished.
artifacts/runs/<run-id>/report.jsonverified_tasks, rejected reasons, task_progress and per-role validation results.

There is no separate failed/ pool. Imported originals stay at source.path; failed generation may leave only a report entry. If no task passes, the previous publication stays available. Match run IDs before comparing its manifest with the latest report; Dashboard's retained batches are historical publications.

Newly generated candidates also retain generation-test-review.json: the structured approval and hashes of the instruction, reference and tests before normalization. Preserve it as generation history when auditing a corrected copy.

Reports and checksums

PathContents
artifacts/state.jsonLatest phase, counters and per-task progress.
artifacts/reports/latest.jsonLatest run report, including failures and validation outcomes.
artifacts/runs/<run-id>/report.jsonExact run report and links to its local evidence.
artifacts/runs/<run-id>/validation/Raw Oracle/NOP trial results and JUnit reports.

generation_completed includes failed attempts; generation_failed is that subset. Imports use collected_questions: null. Candidate counts are not verified-task counts.

Report fieldIdentity
harbor_task_checksumsNative Harbor checksum; must match each trial's task_checksum.
tree_fingerprintsFile contents, types, permissions and empty directories.
task_checksumsSeparate block file digest; not interchangeable with Harbor's checksum.

These identities are checked around validation and on the copied publication. validation.json binds the accepted IDs, Oracle/NOP gate outcomes and task identities to that batch. Tracer and Evaluator check it before consumption; changed task contents, modes, directory trees or manifests are rejected. Use trial results and the semantic review record together before downstream use.

Run archives

start.sh archives on success, failure or interruption and appends artifacts/index.yaml:

artifacts/archives/run_NNN/
├── metadata.yaml
├── config.yaml
├── report.json
└── scripts/

The archive stores redacted configuration and code frozen at run start, exit status and repository SHAs. Its sequential id: run_NNN differs from the runtime run_id; report.json is included only for that exact run. Failed preflight may have no config/report snapshot. No vendor tree or credentials are copied.

For manual archival, use --run-id <id> to link a specific report. A fresh checkout has no index. Raw candidates, traces and archives are local ignored artifacts; review them before sharing and retain diagnostic evidence before cleanup.

On this page