Skip to content

Add CI, lint and pre-commit, and repair the full test suite - #166

Merged
ProfSynapse merged 23 commits into
feat/submodule-cloud-api-v1from
claude/ci-and-test-repair
Oct 2, 2026
Merged

Add CI, lint and pre-commit, and repair the full test suite#166
ProfSynapse merged 23 commits into
feat/submodule-cloud-api-v1from
claude/ci-and-test-repair

Conversation

@ProfSynapse

@ProfSynapse ProfSynapse commented Oct 2, 2026

Copy link
Copy Markdown
Owner

Summary

PR 5 of 5. Stacked on #165.

Adds a CI workflow that runs the full test suite, and fixes the 84 failures it exposed. Before this, only a path-filtered conformance subset ran in CI.

CI, lint and pre-commit

  • .github/workflows/ci.yml:
    • A ruff lint job, plus three test shards on ubuntu-24.04 with Python 3.12 and CPU torch: core, modal-inference-a-l and modal-inference-m-z.
    • A gate job that requires all three shards.
    • Superseded runs are cancelled, every job has a timeout, and pip is cached.
    • The existing conformance workflow is untouched.
  • scripts/ci_pytest_shard.sh: runs a shard locally exactly the way CI does.
  • Ruff: pyflakes rules only (F). All 1,462 existing findings are fixed. 25 hash-locked files keep per-file-ignores, so the inference and packaged-worker locks didn't need re-qualifying. Those entries are to be removed at the next deliberate lock refresh.
  • .pre-commit-config.yaml: ruff, check-yaml, check-toml, merge-conflict, large-files and actionlint. Nothing rewrites files.

Real bugs fixed along the way

  • Trainers/sft/train_sft.py: an undefined _REPO_ROOT silently disabled per-example loss computation for relative dataset paths. Fixed, with a regression test.
  • shared/cloud_artifacts.py::sync_file_to_hf_bucket: used repo_type="bucket", which huggingface_hub rejects, so every single-file bucket push failed. It now uses batch_bucket_files, like the provider adapter does.
  • Evaluator/config/scenarios/vault_gym.yaml: the allowlist named CLI command strings, but the executor checks the concrete tool names. Every vault gym tool call was being blocked.
  • SynthChat/workspace/sections.py: wrapper-less tool formats rendered "Use the `` wrapper for tool calls."
  • SynthChat/scenarios/destructive.yaml: a tool rename had left each of four tools: maps with a duplicated storageManager_archive key. Each pair is collapsed to one entry.
  • shared/flywheel/experiment_loop.py: the surrogate's feature_importance returned integer split counts instead of floats.
  • Evaluator/ui.py: the "... and N more failures" footer could never print.

Test repairs, by cause

  • Order-dependent:
    • data_loader module-name collisions across trainers: new tests/trainers/_trainer_import.py loads each trainer's modules under unique names.
    • A leaked torch stub.
    • A hub-import boundary check that now runs in a fresh interpreter.
    • A flaky DARE statistical bound: six standard errors over 1,000 trials, which is stricter than before.
    • A fixed-timeout worker test that now waits for the actual condition.
  • Out of date: tests are updated to match intentional code changes, with each change traced to its commit:
    • the useTools default removal;
    • PromptCase.expected_tools removal;
    • durable HF provisioning;
    • the call-site census;
    • a stale token-profile lock pin;
    • a CRLF fixture, now stored as -text.
  • Platform:
    • Tests that only used Windows paths or process groups now run on every OS.
    • Two of them previously sent real SIGKILLs on Linux.
    • Genuinely Windows-only behaviour is marked skipif(os.name != "nt"), the repo's existing convention.
  • No test was skipped, xfailed, quarantined or deleted to get green.

Locks

train_sft.py, runtime_v1.py and shared/cloud_artifacts.py changed. Only the offline SFT worker closure and the Modal runtime lock were refreshed. The inference lock and the packaged-worker closure are untouched. All four checks report CURRENT.

Test plan

Run with the CI shard script, on Python 3.12 with CPU torch:

  • core: 11,481 passed, 6 failed, 48 skipped.
    • 4 failures are test_capture_hf_training_smoke_launcher_lock, which correctly refuses this sandbox's proxy and credential variables. All 7 tests in that file pass with env -i.
    • 2 failures depend on the interpreter layout (console_entrypoint, sft runtime_v1). They were shown to pass with a runner-style interpreter.
  • modal-inference-a-l: 350 passed.
  • modal-inference-m-z: 390 passed.
  • ruff check . and pre-commit run --all-files are clean.
  • The first GitHub Actions run confirms the 6 environment-only failures pass on a runner.

The workflow doesn't mark itself required. Make it a required check in branch protection once the first run is green.

🤖 Generated with Claude Code

https://claude.ai/code/session_013etFNjLuo22CCoQmUGYqCT

Generated by Claude Code

claude added 23 commits October 2, 2026 12:42
@claude
lint: add ruff pyflakes (F) config and fix existing findings
8591a02 
Add [tool.ruff] to pyproject.toml selecting only the F family, excluding
Datasets/, scratch/, notebooks and the generated .agents/.claude skill
mirrors. No formatter or other rule family is enabled.

Fixes that change behaviour (found via F821/F811/F402):
- Trainers/sft/train_sft.py: the post-training loss hook referenced an
  undefined _REPO_ROOT, so a relative train_dataset raised NameError that
  the surrounding except swallowed and per-example losses were silently
  skipped. Resolve against Path(__file__).parent.parent.parent, the
  repo-root expression the module already uses for its sys.path bootstrap,
  with a source-level regression test.
- Evaluator/ui.py: the "... and N more failures" footer sat unreachable
  after a return in _expected_for_display; move it back to the end of
  rich_failure_details.
- tests/cloud/test_cloud_eval_handler.py: _canonical_repo_source() fell off
  the end returning None; its return statement had been displaced after
  _source_preparation()'s return.
- tests/execution/providers/test_modal_packaged_dispatch.py: drop a
  byte-identical duplicate of three tests/helpers and duplicate imports.
- Redundant local re-imports (SynthChat engine/renderer), loop variables
  shadowing dataclasses.field, an unused field import.

Mechanical: unused imports/variables (F401/F841, reviewed; calls with side
effects kept as bare calls), f-strings without placeholders (F541), and
star imports in tests replaced by explicit imports. Availability probes,
registration imports, pytest fixture imports, re-exported script helpers
and a monkeypatch seam keep their imports with a targeted noqa and reason.

Hashed source inventories: Trainers/sft/train_sft.py and
Trainers/sft/runtime_v1.py are fixed and the offline SFT worker closure and
Modal runtime lock were refreshed with their checked-in scripts (hash-only
changes; inference lock and packaged-worker closure untouched, all four
checks CURRENT). The remaining locked members are not edited; their
findings stay in per-file-ignores until the next deliberate refresh that
covers them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013etFNjLuo22CCoQmUGYqCT
@claude
ci: add lint + full CPU test workflow and check-only pre-commit hooks
2432cf7 
.github/workflows/ci.yml (the path-filtered conformance workflow is
unchanged):
- lint: ruff check with the pyproject.toml configuration (F rules only).
- tests: the whole pytest suite on ubuntu-24.04 / CPython 3.12, CPU-only
  torch, in three shards whose union is every test file. The Modal
  inference provider tests take most of the runtime and are split
  alphabetically into two shards; "core" runs everything else. A gate job
  requires all shards. Shard selection lives in scripts/ci_pytest_shard.sh
  so it runs identically locally.
- Triggers on pull requests and pushes to main and
  feat/submodule-cloud-api-v1; superseded runs are cancelled, every job has
  a timeout, pip downloads are cached and JUnit XML is uploaded.

requirements-ci-tests.txt holds the CPU test environment. GPU stacks
(unsloth, vllm, bitsandbytes, mlx) and the modal SDK are left out; tests
needing them skip themselves.

tests/trainers/sft/__init__.py: without it the full suite stops at
collection because tests/trainers/sft/test_runtime_v1.py and
tests/ingestion/test_runtime_v1.py share a module basename. Sibling
tests/trainers/* packages already have __init__.py.

.pre-commit-config.yaml: ruff-check, check-yaml (multi-document allowed),
check-toml, check-merge-conflict, check-added-large-files (2048 KB, tracked
Datasets/ artifacts excluded) and actionlint. No hook rewrites files.
check-yaml skips Datasets/ (three invalid legacy rubric drafts) and
SynthChat/scenarios/destructive.yaml, which repeats storageManager_archive
with different weights inside single tools maps and needs an owner fix.

The suite is not green yet. Measured locally with these shard commands:
core shard on this base: 11366 passed, 84 failed, 43 skipped, 2 collection
errors; both Modal inference shards (740 tests) passed on the previous
base. Failures are stale tests after config-first refactors and stricter
config validation, order-dependent tests (Trainers/*/src/data_loader
module-name collisions), Windows-only simulations, a huggingface_hub>=1
requirement that conflicts with transformers<5, and a few
local-environment artifacts.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013etFNjLuo22CCoQmUGYqCT
@claude
test(cloud): give the stage-runner mock config CloudTrainingConfig's …
30b2dbe 
…split defaults

TestTrainingRunnerHappyPath (and the other HFTrainingStageRunner.run tests
that use _mock_backend) failed with "dataset.validation_group_key requires
dataset.split_dataset: true" after the grouped-split work started validating
split settings in the HF training stage.

The validation is correct: real configs come from HFJobsBackend.load_config,
where split_dataset / test_size / validation_group_key default to None, and
validation_split_flags returns [] for that. The mock config was a bare
MagicMock, so those attributes read back as truthy, non-None MagicMock
children: a "set" group key without split_dataset: true. The test double now
models the documented defaults instead of inventing an invalid combination.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013etFNjLuo22CCoQmUGYqCT
@claude
test(trainers): load trainer src modules under unique module names
f133e3d 
Trainers/{sft,dpo,grpo,kto,embedding,...}/src all ship a data_loader.py, and
the tests imported them the way the trainer entrypoints do: put src on
sys.path, then a bare "import data_loader". In one pytest process the first
trainer collected wins sys.modules["data_loader"], so the SFT data-loader,
label-fix and mask-doctor tests (and the GRPO chat-template tests) ran against
Trainers/dpo/src/data_loader.py and failed or errored, while passing alone.

tests/trainers/_trainer_import.py loads Trainers/<method>/src/<name>.py as
trainer_src_<method>_<name>. While the file executes, its bare sibling
imports (from data_loader import ..., from preprocessing import ...) resolve
only to the same trainer's src files, each also under its unique name; the
bare sys.modules entries are restored afterwards and sys.path is untouched.
Trainer runtime imports are unchanged.

The GRPO pivot-profiler test also installed a MagicMock as sys.modules["torch"]
whenever torch had not been imported yet (even when installed) and never
removed it, which broke every later datasets-based test. The helper now
stubs a module only when it is not installed, only while the file loads.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013etFNjLuo22CCoQmUGYqCT
@claude
test(coordinator): check the SDK-free import in a fresh interpreter
d0d5154 
test_coordinator_import_is_provider_sdk_and_storage_neutral asserted that no
huggingface_hub/modal/runpod module was in this pytest process's sys.modules,
which fails whenever any earlier test (e.g. anything importing datasets) has
loaded huggingface_hub. Import the coordinator model, ports and state machine
in a subprocess and assert on that interpreter's modules, which is what the
test means.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013etFNjLuo22CCoQmUGYqCT
@claude
test(lora_surgery): bound the DARE expected-value check by its standa…
19f730a 
…rd error

The check averaged 100 drop-and-rescale trials and required every element to
be within an absolute 0.5 of the original. The per-element error scales with
|x| * sqrt(p / (1 - p) / N), so large randn weights exceeded 0.5 by chance:
about 6% of RNG states failed (measured over 3000 replicates), which made the
outcome depend on how much global RNG earlier tests consumed.

Use 1000 trials and a six-sigma bound on that standard error per element.
False failures drop to ~1e-6 for the tensor, and the bound is below the old
0.5 for every |x| < 4, so it is stricter in practice (it also catches a 5%
rescale bias the old check never caught).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013etFNjLuo22CCoQmUGYqCT
@claude
test(cloud): wait for the stage-event workers instead of a fixed timeout
660077b 
test_cloud_stage_event_appends_are_serialized_across_processes raised
queue.Empty in the full run: each of 16 spawned children imported the whole
experiment-handler test module (torch and the tracking stack) to find its
target, and the parent gave each result a fixed 40 s.

The target now lives in tests/cloud/_stage_event_worker.py, so a child imports
only that and shared.cloud_stage_logging (the test drops from ~16 s to ~2 s
here). The parent waits for the condition itself, every child reporting or
every child gone with nothing left to read, under a 300 s bound, and the
assertions report which children never answered.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013etFNjLuo22CCoQmUGYqCT
@claude
Import the Buckets client only where the Space sync uses it
c541dfb 
sync_bucket_prefix.py runs inside the warm vLLM Space with its own
huggingface_hub>=1.0 venv (/opt/bucket-sync-venv, see Dockerfile.tmpl),
because Buckets need a newer Hub client than the transformers 4.x stack
allows. Its module-level `from huggingface_hub import sync_bucket` made the
pure parse_bucket_uri helper unimportable next to the engine's pinned
huggingface-hub<1.0, which broke collection of tests/cloud/test_manage_space.py.
manage_space.py itself only uses APIs present in hub 0.x.

Defer the import into sync_bucket_prefix(), the only code that needs it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013etFNjLuo22CCoQmUGYqCT
@claude
Skip the precloud-gate audit test when the Modal SDK is absent
6614885 
Past the capability gate, launch.execute() imports modal before it reaches
local storage. Use pytest.importorskip("modal") like the other Modal chat
example tests, since CI deliberately installs no provider SDKs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013etFNjLuo22CCoQmUGYqCT
@claude
Run platform-specific process and path tests on every host
54ea349 
- docker_archive: the exclusive-output, byte-bound and stuck-reader tests
  are cross-platform; only the process-group mechanism differs. Assert the
  platform's own group kwargs (CREATE_NEW_PROCESS_GROUP / start_new_session)
  and stub the platform's tree stop. On POSIX the fake pid previously
  reached a real os.killpg(os.getpgid(4321), SIGKILL).
- image_lock: assert platform group kwargs in the subprocess runner test.
  The two taskkill.exe tests are Windows-only behaviour and are marked
  os.name != "nt" like the repo's other Windows-only tests; the fallback one
  had been passing on Linux only by sending SIGKILL to pid 1234's real
  process group. Add POSIX counterparts that stub getpgid/killpg.
- registry: the replace retry is deliberately gated on os.name == "nt"
  (POSIX os.replace has no sharing violations) and register_run locks with
  msvcrt vs fcntl by os.name, so the two retry tests are Windows-only.
- snapshot_nexus: fake Windows for the module under test only. Patching the
  global os.name made pathlib build WindowsPath objects on POSIX. The
  envelope and batch-failure tests also stub executable resolution, as the
  argv test already does; they stubbed subprocess.run but still needed a
  real `nexus` CLI on PATH, so they failed with NEXUS_RUNTIME_UNAVAILABLE.
- derived_training_image: set the module's _WINDOWS flag, as sibling tests
  do, for the two tests that exercise the Docker Desktop Buildx authority.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013etFNjLuo22CCoQmUGYqCT
@claude
Return float feature importances from the experiment surrogate
63a8994 
SurrogateModel.feature_importance() is declared Dict[str, float], but
LightGBM's default "split" importance type yields integer split counts, so
the method returned ints and TestSurrogateModel::test_feature_importance
failed wherever lightgbm is installed. Convert each importance to float.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013etFNjLuo22CCoQmUGYqCT
@claude
Bring stale tests in line with the code they cover
395b107 
- token_capture: 0cbc018 deliberately replaced the {conversations, reward,
  token ids} GRPO row with the static-GRPO trainer schema and made
  test_stager assert it exactly, but left these tests on the old shape.
  Assert the current contract: captured token fields stay addressable via
  the filter view's content.* namespace and are not part of the exact row.
- handlers versions: 8e58851 moved `flywheel versions` from repo_root (now
  a compatibility shim) to engine_root; patch engine_root so the fixture
  datasets directory is actually consulted.
- hf_smoke parent-chain swap: c10120f replaced the re-run link check with
  held, handle-relative O_NOFOLLOW opens. A swapped parent still fails
  closed, now as "could not be read safely" (POSIX ELOOP) or "reparse
  points" (Windows); assert those instead of the retired message.
- tracking provenance: restore the contained URI before the symlink case;
  the test still pointed at tracking://../outside.json, so the escape check
  fired first and the symlink refusal was never exercised.
- cloud_artifacts non-git: bound git discovery with GIT_CEILING_DIRECTORIES
  so a basetemp under the repository's scratch/ cannot supply a repo.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013etFNjLuo22CCoQmUGYqCT
@claude
Fix single-file HF bucket upload
ae8bfd9 
sync_file_to_hf_bucket called HfApi.upload_file with repo_type="bucket",
which huggingface_hub rejects ("Invalid repo type"), so every single-file
bucket push failed. Use batch_bucket_files, the bucket API the provider
adapter already uses, and name the qualified hub version when the installed
hub lacks it. Refresh the offline SFT worker closure and Modal runtime lock
for the one changed member.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013etFNjLuo22CCoQmUGYqCT
@claude
Collapse duplicate storageManager_archive keys in destructive scenarios
a7517e8 
A tool rename mapped vaultManager_deleteFolder and vaultManager_deleteNote
onto the single storageManager_archive tool, leaving each tools map with a
repeated key (YAML kept the last weight). Collapse each pair to one entry,
fix the matching prompt text, and stop excluding the file from check-yaml.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013etFNjLuo22CCoQmUGYqCT
@claude
Render no wrapper instruction for wrapper-less tool formats
ab8013c 
Since 6dcb543 the built-in tool-call format has no wrapper; useTools is
only the configured default in SynthChat/config/tool_call_formats.yaml.
The available_tools section still always rendered "Use the `<wrapper>`
wrapper for tool calls.", producing "Use the `` wrapper" for native
formats such as openai_native. Emit that line only when a wrapper is
configured.

The workspace tests still expected useTools as the code default. They now
assert that the built-in default has no wrapper, that the configured
default format supplies its wrapper, and that the rendered section
carries the configured wrapper line only when one exists.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013etFNjLuo22CCoQmUGYqCT
@claude
Allow the concrete tools vault gym CLI commands expand to
2f29bb0 
f5e9126 rewrote the vault gym environment allowlist as CLI command
strings ("content read", "search content", ...). The environment
executor expands each wrapped CLI command into its concrete tool
(contentManager_read, searchManager_content, ...) before checking the
allowlist, so every vault gym tool call was blocked and no
environment-backed vault gym case could pass. List the concrete tool
names from cli-first-tool-schemas.json instead; scoring paths keep the
emitted command strings, which is what they match against.

The vault gym tests still built the pre-CLI {"calls": [...]} payload and
passed PromptCase.expected_tools, which no longer exists. They now emit
one call in the configured default tool-call format (wrapper and
required fields from tool_call_formats.yaml, IDs from the case's
expected context) with the scenario's preferred CLI command paths, and
assert the executed concrete tools.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013etFNjLuo22CCoQmUGYqCT
@claude
Move SynthChat and Evaluator tests onto the configured tool-call format
dee9fb3 
These tests predate be1220a/6dcb543f (CLI wrapper payloads, wrapper-less
built-in default) and aabe1f4 (assertion-driven evaluation, no
PromptCase.expected_tools); all of them have failed since those commits.

- Assistant payloads use the configured default format from
  SynthChat/config/tool_call_formats.yaml: its wrapper name, its
  required argument fields and CLI command strings, instead of the
  removed context/calls shape and a hard-coded useTools.
- Tool scenarios generate assistant turns through structured output, so
  the fake clients serve them as structured responses; the empty-reply
  retry is exercised on the structured path.
- Multistep evaluator cases declare `correct` assertions on the final
  response, since a case without them can no longer pass, and assert the
  executed concrete tools.
- Schema tests assert the current contract: the built-in default is
  native, the configured wrapper takes a JSON-string arguments payload
  that names the configured fields, and allowed tools are named in the
  generation prompt.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013etFNjLuo22CCoQmUGYqCT
@claude
Drive the HF stage-runner fixture through durable SUCCEEDED provisioning
be804e0 
c10120f made every protected HF stage require durable SUCCEEDED
provisioning (claim event plus terminal event) in addition to a
CONSUMABLE source transport. The shared _install_hf_source_transport
fixture still attached evidence through record_provisioning_acknowledged,
which never reaches SUCCEEDED, so the eval and loss stage runner tests
stopped at "HF protected operation requires durable SUCCEEDED
provisioning" before reaching the behaviour they check.

The fixture now records the claim and its SUCCEEDED terminal event under
the provisioning execution lock, the same chain the tracking-service
tests use, binding the same evidence. No provider call is made.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013etFNjLuo22CCoQmUGYqCT
@claude
Test Qwen3.5 SFT support as this lineage implements it
5c401e1 
test_model_loader_source.py was added in 97e529e as a cherry-pick of the
test half of 7a4e17c ("Add Qwen3.5 4B SFT launch path",
origin/codex/qwen35-4b-launch). The implementation half (a FastVisionModel
branch in Trainers/sft/src/model_loader.py and a Qwen3.5 4-bit warning in
Trainers/sft/train_sft.py) never landed here, so both source-grep tests
have failed ever since. Both files are now hash-locked members of the
offline SFT worker closure, and this lineage supports Qwen3.5 another
way.

Replace the source greps with checks of that behaviour:
- the locked loader loads Qwen3.5 through FastLanguageModel, keeps the
  processor a vision-language checkpoint returns, and the
  memory-efficient loss guard rejects a stock ForConditionalGeneration
  loss fallback;
- every checked-in Qwen3.5 recipe and cloud experiment disables 4-bit
  loading, the configuration-level form of "no QLoRA for Qwen3.5".

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013etFNjLuo22CCoQmUGYqCT
@claude
Classify the structured-document bakeoff in the completion census
98043d6 
SynthChat/scripts/structured_document_bakeoff.py (7648833, after the
census in d0c2b1a) calls structured_output on BaseLLMClient instances
and was never classified, so the exact census failed.

It follows the rules of the other classified consumers: clients come
only from the shared shared.llm.create_client factory, which resolves
the provider key by environment-variable name; the script reads no
credential itself and makes no raw HTTP calls; and every completion is
read only through .value and .usage. List it under
BASE_CLIENT_CONSUMERS so the field-only consumption rule now covers it
(it passes).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013etFNjLuo22CCoQmUGYqCT
@claude
Keep the legacy job-path scan from matching its own module
2146e70 
TestReferenceCompleteness greps tests/ (among other roots) for
Trainers/local/jobs and Trainers/cloud/jobs. The only hits are in
tests/discovery/test_recipes.py itself, which has to spell the legacy
paths to search for them, so both checks have failed since e671cdf.

Drop hits from this module, both the repo-relative git-grep form and the
absolute fallback form, and nothing else. A new check pins that the filter
keeps a real hit elsewhere while dropping this module and docs/plans.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013etFNjLuo22CCoQmUGYqCT
@claude
Store the CRLF ingestion fixture with its CRLF bytes
b12d6f9 
The markdown ingestion acceptance test needs nested/beta.md to be a
BOM + CRLF file, so it can check that ingestion normalises line endings.
Under "* text=auto" the file was committed with LF line endings
(64da4cc). It only had CRLF in a checkout that converts to CRLF, so the
test failed on Linux and in CI.

Mark that one path -text so Git stores and checks out its exact bytes on
every platform, and renormalise it to BOM + CRLF.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013etFNjLuo22CCoQmUGYqCT
@claude
Pin the token-profile example to the current Modal runtime lock
52998ab 
The qwen35_4b_token_profile.yaml example commits to the SHA-256 of
modal-runtime-v1.lock.json, and the profiler refuses to run when the
two differ. The example still carried the digest from fd28ab2; every
reviewed lock refresh since 33dc682 has changed the lock, so the
checked-in example could not run and
test_checked_in_qwen_example_matches_current_profiler_contract failed.

The lock is the authority: scripts/regenerate_modal_runtime_lock.py
reports it CURRENT for the locked sources. No checked-in tool
regenerates the example, so the new digest is computed from the lock
bytes, the same digest the profiler checks, rather than typed in. The
skill mirrors are re-synced with .skills/scripts/sync_skill_trees.py.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013etFNjLuo22CCoQmUGYqCT
This was referenced Oct 2, 2026
Merged
Merged
@ProfSynapse
ProfSynapse changed the base branch from claude/sft-mask-doctor-port to feat/submodule-cloud-api-v1 October 2, 2026 23:28
@ProfSynapse
ProfSynapse merged commit 1d6a337 into feat/submodule-cloud-api-v1 Oct 2, 2026
9 checks passed
@ProfSynapse
ProfSynapse deleted the claude/ci-and-test-repair branch October 6, 2026 19:17
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@ProfSynapse @claude