| name | ingest-l1 |
| description | L1 analysis loop for the abap_wiki knowledge base: for each batch it launches the abap-analyzer sub-agent in parallel, then the adversarial judge abap-deepcheck (separate session), applies only the analyses that pass the fail-closed gate, and commits. Resumes exactly after an interruption. Use this skill to document the code of a package, or invoke it via /loop for autonomous execution. |
Ingest L1 - code analysis with adversarial gate
Promotes objects from L0 stub to L1 page with verified code analysis.
The cycle is driven by you (main agent): you invoke sub-agents via the Task tool,
the scripts handle everything else (state, validation, gate, page writing).
Architecture (see core/docs/01-pipeline-l0-l1.md and 02-adversarial-gate.md):
- author = sub-agent
abap-analyzer (raw-only): reads the source and writes
output/runs/<run>/<task>/author.yaml (anchored claims + dependencies).
- deepcheck = sub-agent
abap-deepcheck (different model, separate session):
verifies that every claim is demonstrated by the cited lines and every dependency
is real. Writes deepcheck.json.
- gate fail-closed: no page is promoted without a valid, fresh verdict with full coverage.
Batch cycle
Generate run_id = run-<timestamp> and batch_id = b-<timestamp> at the start of each round.
- Recover (always, at the start - resumes interrupted tasks without repeating work):
python core/src/tools/pipeline.py recover
- Claim author (10-15 per batch):
python core/src/tools/pipeline.py claim --kind l1_author --limit 12 --worker <run_id>
Returns JSON with the tasks. If empty, skip to step 4; if empty there too, the loop is done.
- Fan-out author IN PARALLEL: for each task one
Task(subagent_type="abap-analyzer", ...)
passing in the prompt sap_name, sap_type, devclass, raw_source_path and
artifact_path = output/runs/<run_id>/<task_id>/author.yaml. On return from each:
python core/src/tools/pipeline.py submit-author --task <task_id> --run <run_id> --batch <batch_id>
- Claim deepcheck and fan-out IN PARALLEL of the judge (separate session):
python core/src/tools/pipeline.py claim --kind l1_deepcheck --limit 12 --worker <run_id>
For each task Task(subagent_type="abap-deepcheck", ...) with the prompt and verdict_path
as indicated. On return:
python core/src/tools/pipeline.py submit-verdict --task <task_id> --run <run_id> --batch <batch_id>
- Apply (ACCEPT only; idempotent):
python core/src/tools/pipeline.py apply --run <run_id> --batch <batch_id>
- Project + commit:
python core/src/tools/pipeline.py project
python core/src/tools/pipeline.py export-excel
python core/src/tools/pipeline.py git-commit --message "ingest L1 batch <batch_id>" --batch <batch_id>
- Progress and decide whether to continue:
python core/src/tools/pipeline.py progress --json
Autonomous execution with /loop
Prerequisite: Claude Code session open with working directory = repo root,
so the sub-agents abap-analyzer and abap-deepcheck (in .claude/agents/) are
registered and invocable via the Task tool. Verify with sync_agents.py --check.
Command (example for a single package; omit --package for all):
/loop Continue the L1 ingest of package ZPACKAGE until complete.
Each iteration, from the repo root with the venv (.venv\Scripts\python):
1) python core/src/tools/pipeline.py recover
2) Generate RUN=run-<timestamp> and BATCH=b-<timestamp>.
3) python core/src/tools/pipeline.py claim --kind l1_author --limit 10 --worker $RUN --run $RUN --batch $BATCH --package ZPACKAGE
If it returns [], skip to step 6.
4) For EACH task in the claim, IN PARALLEL: Task(subagent_type="abap-analyzer") passing in the prompt
sap_name, sap_type, devclass, raw_source_path and artifact_path=output/runs/$RUN/<task_id>/author.yaml.
On return from each: python core/src/tools/pipeline.py submit-author --task <task_id> --run $RUN --batch $BATCH
(if the agent fails: submit-author --task <id> --run $RUN --batch $BATCH --fail "msg").
5) python core/src/tools/pipeline.py claim --kind l1_deepcheck --limit 10 --worker $RUN --run $RUN --batch $BATCH --package ZPACKAGE
For EACH task IN PARALLEL: Task(subagent_type="abap-deepcheck") passing prompt_path=output/runs/$RUN/<author_task_id>/deepcheck-prompt.txt
and verdict_path=output/runs/$RUN/<deepcheck_task_id>/deepcheck.json (author_task_id is the latest l1_author task for the object).
On return: python core/src/tools/pipeline.py submit-verdict --task <deepcheck_task_id> --run $RUN --batch $BATCH
6) python core/src/tools/pipeline.py apply --run $RUN --batch $BATCH
7) python core/src/tools/pipeline.py project
8) python core/src/tools/pipeline.py git-commit --message "ingest L1 ZPACKAGE batch $BATCH" --batch $BATCH
9) python core/src/tools/pipeline.py progress --package ZPACKAGE --json
If "l1_remaining" is 0, STOP the loop. Otherwise continue with the next iteration.
Objects with REVERT status return to the queue automatically and are retried in subsequent rounds (max 3 attempts, then failed).
Note: if you invoke abap-deepcheck via the Task tool, use the registered subagent_type
(it already has model: sonnet in the frontmatter → automatic model independence).
To map deepcheck_task to the object's author_task, use the latest l1_author
for the same object_id (one per object).
Rules
- The author and the deepcheck MUST NEVER run in the same session/context:
judge independence is the anti-hallucination guarantee.
- Per-task errors do not block the batch: the failed object stays in the queue
(up to max-attempts) or goes to
failed; the loop continues.
- Gate outcomes: ACCEPT promotes; REVERT sends back to author with findings
(max 3 attempts, then
failed/escalation); BLOCKED re-runs only the deepcheck.
- Discovered SAP standards must be resolved with the separate MCP loop (
claim --kind mcp_lookup),
when the abap-fs server is running. See core/docs/05-runbook.md.
