feat(training): JSONL exporter for ATIF trajectories

[anderskev]

Summary

• Adds daydream export-jsonl — a new CLI subcommand that turns archived ATIF trajectories into a versioned, schema-validated JSONL corpus suitable for training/eval pipelines.
• Introduces the daydream/training/ package: schema v1 definition, exclusion + copyleft filtering, record/span builders, deterministic stack-stratification, and an end-to-end export orchestrator.
• Extends the archive manifest with a code_context block (base_sha, changed_files) so newly archived runs carry the git context the exporter needs end-to-end.

Motivation

The trajectory archive captures per-run ATIF data, but there was no supported way to turn it into a training-ready dataset. Downstream consumers need:

• A stable, versioned JSONL schema (v1) they can validate against.
• Deterministic, byte-identical output across reruns so corpora can be diffed and cached.
• Built-in filtering for license-incompatible (copyleft) and explicitly excluded sources.
• Stack-aware stratification so a single dominant stack does not swamp the corpus.

This PR delivers that pipeline in one focused milestone so the training side can consume archived runs without bespoke glue.

Changes

Added

• daydream/training/ package: schema.py, exclusion.py, export.py, schema artifacts (schema/v1.json, schema/copyleft.txt, schema/exclusion.txt).
• daydream export-jsonl --out <path> CLI subcommand with the full filter, stratification, opt-in, and diagnostic flag surface from the plan (validates --max-stack-share in (0, 1] and --min-grounding in [0, 1]).
• ExportConfig + run_export orchestrator: schema-only short-circuit, filter+stratify pipeline, dry-run summary, atomic JSONL write via tempfile + Path.replace, and schema.json side-car emission.
• GitContext + Manifest carry base_sha and changed_files; git_ops.diff_name_only populates them at archive time; manifest.to_dict() emits a code_context block.
• archive.index.count_runs() for cheap unfiltered totals (no row materialization).
• Test fixtures under tests/fixtures/training/ plus suites covering exclusion, query/filter, record/span builders, stratification, and the end-to-end export.

Changed

• Exporter sources head_sha / branch / base_branch from the manifest git block (was inadvertently reading from code_context); falls back to manifest_row when older archives have no code_context block.
• is_copyleft() accepts a pre-loaded list so the per-row hot loop in _query_index doesn't reopen the file N times.
• _build_query() accepts an exclusion kwarg (typed frozenset | set | None) so callers can inject the set without re-reading from disk.
• Step IDs are cast to int to match the v1 schema integer constraint.
• JSONL output uses sort_keys=True and compact JSON separators for deterministic, byte-identical reruns.
• code_context field order in the manifest now matches the schema definition (head_sha then base_sha).
• Skip-record warning copy clarified from "missing" to "corrupt or unreadable"; stack=None warning is now gated behind stratify_by == \"stack\".

Fixed

• Atomic tempfile writes are wrapped in try/except so a failed write doesn't leave an orphan .tmp file beside the output.
• Multi-outcome sessions go through _single_outcome_label(), which warns instead of silently dropping labels.
• Schema v1 drops the unused test_outcome field.

Test Plan

• uv run pytest — full suite passes locally (343 existing + new training/archive/git_ops coverage).
• uv run ruff check and uv run mypy daydream pass; the _build_query exclusion-param typing fix was added specifically to unblock mypy.
• Exporter determinism verified via a fixture-driven byte-identical rerun test.
• Schema-only short-circuit, dry-run no-op, and missing-trajectory skip paths each covered by a dedicated test.
• git_ops.diff_name_only covered for happy path, multi-file output, empty-line filtering, and bad-ref soft-failure.

Checklist

• Tests pass locally (uv run pytest)
• Linting passes (uv run ruff check)
• Type checking passes (uv run mypy daydream)
• Documentation updated (if applicable) — CLI surface documented inline; no user-facing docs site to update for this milestone.

Additional Context

Diff size: ~2,100 LOC added across 22 files; the bulk lives in daydream/training/export.py and its test suite. The code_context rollout is forward-compatible: older archives that predate the block surface base_sha=None / changed_files=[], which the v1 schema explicitly allows.

---

Generated with Claude Code

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: f8c5f62f-886e-4f86-a5d1-ff678ba276cb

📥 Commits

Reviewing files that changed from the base of the PR and between 2d1bc51 and e12fcce.

⛔ Files ignored due to path filters (1)

• uv.lock is excluded by !**/*.lock

📒 Files selected for processing (5)

• daydream/cli.py
• daydream/git_ops.py
• daydream/training/export.py
• pyproject.toml
• tests/fixtures/training/build_archive.py

🚧 Files skipped from review as they are similar to previous changes (4)

• daydream/git_ops.py
• daydream/cli.py
• tests/fixtures/training/build_archive.py
• daydream/training/export.py

---

Walkthrough

This PR adds a training-record JSONL exporter: archive metadata now includes merge-base SHA and changed files; training package, v1 JSON Schema, and file-backed exclusion/copyleft lists are added; trajectories are converted to schema v1 records (spans, labels, code_context); a parameterized SQL query/filter pipeline with copyleft handling and skill->stack mapping is implemented; stack-based stratification is supported; run_export orchestrates counting, querying, stratifying, and atomic JSONL+schema emission; a synchronous CLI subcommand exposes the flow; fixture-backed tests cover the end-to-end behavior.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 61.54% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title 'feat(training): JSONL exporter for ATIF trajectories' accurately summarizes the main change: introducing a JSONL export pipeline for archived ATIF trajectories.
Description check	✅ Passed	The description comprehensively details the PR's purpose, changes, and context related to the JSONL exporter, archive manifest extensions, filtering, stratification, and test coverage.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Actionable comments posted: 7

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@daydream/cli.py`:
- Around line 379-383: Detect when both args.include_all_labels and args.label
are provided and fail fast instead of silently overriding: in the CLI handling
code where labels is computed (the block referencing args.include_all_labels,
args.label, and labels), add a check that if args.include_all_labels is true and
args.label is non-empty, call the argument parser's error/exit (e.g.,
parser.error or raise SystemExit with a clear message) or refactor the flags
into an argparse mutually exclusive group so the parser prevents both from being
set; ensure the error message clearly states that --include-all-labels cannot be
used with --label.

In `@daydream/git_ops.py`:
- Around line 502-505: The call to _run_git inside diff_name_only can raise
GitError and must be caught to preserve the function's contract of returning []
on subprocess failure; wrap the _run_git call in a try/except that catches
GitError (the exception type raised by _run_git) and return [] from the except
block, keeping the existing behavior that also returns [] when proc.returncode
!= 0 and leaving the timeout and arguments to _run_git unchanged.

In `@daydream/training/export.py`:
- Around line 63-69: The step_id parsing in the loop over
trajectory.get("steps", []) (inside daydream/training/export.py) can raise
ValueError/TypeError when calling int(step.get("step_id", i + 1)); guard this by
wrapping the int(...) conversion in a try/except (catching ValueError and
TypeError) and on failure fall back to a safe default (e.g., use i + 1 or None)
and optionally log/debug the malformed step; update the code around the step_id
assignment so malformed step_id values do not abort export.
- Around line 164-169: The parsing of raw_labels (outcome_labels) silently
swallows JSON errors and drops data; modify the try/except around
json.loads(raw_labels) so that on JSONDecodeError/TypeError you log the error
and the offending raw_labels (and any record identifier available) before
falling back to an empty list, or re-raise if that better fits upstream
handling; update the except block to call the module logger (e.g.,
logging.exception or logger.warning with exc_info=True) referencing raw_labels
and keep the subsequent isinstance(labels, list) check to enforce type safety.

In `@tests/fixtures/training/build_archive.py`:
- Around line 27-37: The FixtureSession dataclass docstring is missing a
Google-style "Attributes:" section and the public function build_fixture_archive
lacks a "Returns:" section; update the FixtureSession docstring to include an
Attributes: block that lists session_id (str), repo_slug (str), skill (str),
grounding_rate (float), outcome_labels (tuple[str, ...]), status (str, default
"complete"), and notes (str, default ""), and update the build_fixture_archive
docstring to include a Google-style "Returns:" section describing the return
type and what the returned value represents (and add or complete Args: and
Raises: sections if applicable), ensuring wording matches existing docstring
style and uses the exact symbol names FixtureSession and build_fixture_archive
to locate the spots to edit.

In `@tests/test_archive.py`:
- Around line 85-93: Update the inline `# noqa` comments on the subprocess.run
calls in tests/test_archive.py (the git init/config/commit invocations) to
include an explicit rationale for S607 in addition to S603; e.g., augment each
`# noqa: S603, S607` comment to state that arguments are not user-controlled and
that the `git` command is a hardcoded, trusted command (same change also for the
later subprocess.run calls around lines 113-137).

In `@tests/test_training_export.py`:
- Line 16: The test suite imports the jsonschema package (seen in
test_training_export.py and test_training_record.py) but jsonschema is not
declared as a dependency; add jsonschema to the test/dev dependencies in
pyproject.toml (e.g., under [tool.poetry.dev-dependencies] or
[project.optional-dependencies."test"]) so CI installs it before running tests
and import errors are avoided.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: 5e0e1b36-f6a8-44c7-bd46-f8ad8718001b

📥 Commits

Reviewing files that changed from the base of the PR and between f25b1ba and 2d1bc51.

⛔ Files ignored due to path filters (1)

• uv.lock is excluded by !**/*.lock

📒 Files selected for processing (21)

• daydream/archive/git_context.py
• daydream/archive/index.py
• daydream/archive/manifest.py
• daydream/cli.py
• daydream/git_ops.py
• daydream/training/__init__.py
• daydream/training/exclusion.py
• daydream/training/export.py
• daydream/training/schema.py
• daydream/training/schema/copyleft.txt
• daydream/training/schema/exclusion.txt
• daydream/training/schema/v1.json
• tests/fixtures/training/__init__.py
• tests/fixtures/training/build_archive.py
• tests/test_archive.py
• tests/test_git_ops.py
• tests/test_training_exclusion.py
• tests/test_training_export.py
• tests/test_training_query.py
• tests/test_training_record.py
• tests/test_training_stratify.py