fix(engine): one version helper, locked tool pins and proven-dead members

[ecryptoguru]

Engine core fixes: version lookup, tool pins, proven-dead members

Summary

The engine looked up the distribution strix-agent for its own version. No
build of this project installs that name: pyproject.toml:2 declares
lyrashield-engine. The three wrong call sites therefore always raised
PackageNotFoundError, which silently omitted the SARIF tool.driver.version
field, showed dev in the interactive TUI header and reported unknown to
telemetry. One shared helper now resolves the real distribution name with an
explicit fallback, and the three call sites use it.

The pre-commit hook revisions are aligned to the versions uv.lock and CI
already resolve, so a local hook can no longer pass lint or formatting the
locked gate rejects. The bandit severity setting in pyproject.toml was inert,
and the fix states and enforces the floor the gate actually applies instead of
the medium floor the config only appeared to set.

Seven unreferenced members and files are deleted, each with whole-repo git grep proof. The image-pull deadline test gets a bounded daemon wait and CI
caches the Chromium download.

Per-change safety proof

Every deletion below was checked with git grep across the whole repo
(lyrashield, lyrashield_adapter, strix, tests, scripts, docs,
pyproject.toml, strix.spec, .github, Makefile, UPGRADES.md,
README.md, CONTRIBUTING.md, all *.mdx).

1. lyrashield/artifacts/usage.py — ancillary_cost_total property

$ git grep -n -w ancillary_cost_total
lyrashield/artifacts/usage.py:229:    def ancillary_cost_total(self) -> float:

One hit, the definition itself. total_cost at :234 computes the same sum
inline and is unaffected. _round_cost and _ancillary_costs remain used by
total_cost, to_record and reset, so no supporting symbol was orphaned.

2. lyrashield/artifacts/state.py — set_sandbox_cleanup_status

$ git grep -n -w set_sandbox_cleanup_status
lyrashield/artifacts/state.py:630:    def set_sandbox_cleanup_status(self, sandbox_removed: bool) -> None:

One hit, the definition itself. The method's own docstring calls it a
backward-compatible wrapper and no caller exists. The real method
set_cleanup_outcome keeps its callers (interface/cli.py:311,
lifecycle/finalize.py:126, plus tests), and CLEANUP_REMOVED and
CLEANUP_FAILED remain used inside set_cleanup_outcome itself.

3. lyrashield/lifecycle/agents.py — wait_kind_of

$ git grep -n -w wait_kind_of
lyrashield/lifecycle/agents.py:285:    async def wait_kind_of(self, agent_id: str) -> WaitKind | None:
strix/core/agents.py:214:    async def wait_kind_of(self, agent_id: str) -> WaitKind | None:

The only other hit is the upstream substrate twin, which is a controlled
derivative boundary and was not touched. No lyrashield/, test or script
caller exists. The wait_kinds dict and WaitKind type stay in use by
park_waiting, snapshot and restore.

4. lyrashield/interface/utils.py — process_pull_line shim

$ git grep -n 'process_pull_line' -- lyrashield tests scripts docs pyproject.toml strix.spec .github Makefile
lyrashield/interface/image_pull.py:169: ... process_pull_line(payload, layers_info, status, last_update)
lyrashield/interface/image_pull.py:320: ... process_pull_line(line, layers_info, status, last_update)
lyrashield/interface/image_pull.py:358: def process_pull_line(
lyrashield/interface/main.py:42:    process_pull_line,  # noqa: F401
lyrashield/interface/utils.py:586: def process_pull_line(
lyrashield/interface/utils.py:589:     from lyrashield.interface.image_pull import process_pull_line as _process_pull_line
lyrashield/interface/utils.py:591:     return _process_pull_line(line, layers_info, status, last_update)
tests/test_image_digest.py:13:    process_pull_line,
tests/test_image_digest.py:79:     update = process_pull_line(...)
tests/test_main_helper_exports.py:21:        ("process_pull_line", "image_pull", "process_pull_line"),

The only removed hits are the three utils.py lines. The test imports come
from lyrashield.interface.main, which re-exports the real
image_pull.process_pull_line at main.py:42. tests/test_main_helper_exports.py:21
asserts that main re-exports the image_pull implementation and still passes.
The other stub in the same file, update_layer_status, is a different function
and was left alone.

5. scripts/docker.sh

$ git grep -n -F 'docker.sh'
(no output)

Zero references anywhere in the repo. The script also builds strix-sandbox,
a different image name from the product one. The remaining 11 files in
scripts/ are referenced by CI, the Makefile, tests or docs.

6. .trivyignore.yaml — root Dockerfile path

$ ls Dockerfile
ls: cannot access 'Dockerfile': No such file or directory

Only containers/Dockerfile exists. The AVD-DS-0002 rule and its
containers/Dockerfile path are unchanged, so the reviewed trust-boundary
decision still applies to the real file.

7. strix.spec — tenacity hidden import

$ git grep -n -i 'tenacity' -- strix.spec pyproject.toml lyrashield strix
strix.spec:99:    # Tenacity retry
strix.spec:100:    'tenacity',
$ grep -n '^name = "tenacity"' uv.lock
(no hit)
$ ls .venv/lib/python3.14/site-packages/ | grep -i tenacity
(not installed)
$ git grep -n 'import tenacity\|from tenacity'
(no importer)

tenacity is absent from pyproject.toml, absent from uv.lock, not
installed, and imported nowhere. containers/python-requirements.txt pins it
for the sandbox image, which is a separate environment from the frozen engine
binary and is untouched. strix.spec is at the repo root, outside strix/**.

Tests added

All new tests were made to fail locally before their fix and pass after.

Test	Failed before with
tests/test_engine_version.py::test_sarif_write_path_emits_tool_driver_version	KeyError: 'version' — _write_report_projections resolved the wrong distribution, so sarif.py:262 omitted driver.version
tests/test_engine_version.py::test_sarif_write_path_still_omits_version_when_metadata_is_absent	passed only after the fallback existed; drives the real projection chain with metadata unavailable
tests/test_engine_version.py::test_engine_version_falls_back_when_distribution_is_missing	PackageNotFoundError escaped the old call sites
tests/test_engine_version.py::test_engine_version_survives_a_broken_metadata_backend	a frozen loader raises something other than PackageNotFoundError
tests/test_quality_environments.py::test_precommit_hooks_use_the_locked_tool_versions	AssertionError: 'v0.11.13' == 'v0.15.20'
tests/test_quality_environments.py::test_bandit_severity_floor_is_enforced_by_the_invocation	AssertionError: 'severity' in {... 'severity': 'medium'}
tests/test_release_build.py::test_binary_does_not_request_missing_hidden_imports	AssertionError — 'tenacity' present in the spec
tests/test_release_build.py::test_every_spec_hidden_import_is_a_declared_or_installed_distribution	AssertionError: ... ['tenacity']

The SARIF regression is the one that matters. tests/test_sarif.py:60 passes
tool_version explicitly, so it could never catch a broken lookup. The new test
drives ReportState._write_report_projections, the production call chain that
resolves the version itself, and reads the written findings.sarif.

Existing behaviour is unchanged for the four lookups that were already correct
(interface/arg_parser.py:40, lifecycle/runner.py:129,
tools/proxy/caido_api.py:678, lyrashield_adapter/cli.py:158). They were not
edited.

Bandit: how the intent was verified and why medium was rejected

The config key was inert. Verified with the installed bandit 1.9.4:

$ uv run python -c "from bandit.core.config import BanditConfig; c=BanditConfig('pyproject.toml'); print(repr(c.get_option('severity')))"
'medium'

$ uv run bandit -c pyproject.toml -f json /tmp/bandittest/mix.py | python -c "import json,sys; print(sorted({(r['test_id'], r['issue_severity']) for r in json.load(sys.stdin)['results']}))"
[('B311', 'LOW'), ('B602', 'LOW')]

$ uv run bandit -c pyproject.toml -ll -f json /tmp/bandittest/mix.py | python -c "import json,sys; print(sorted({(r['test_id'], r['issue_severity']) for r in json.load(sys.stdin)['results']}))"
[]

Bandit parses the key and never applies it as a filter. Only the -l/--level
CLI flag filters. The fixture is a temporary file outside the repo, and the
whole-repo run is unaffected:

$ uv run bandit -r strix lyrashield_adapter lyrashield -q -c pyproject.toml -f json   # default
0 findings
$ uv run bandit -r strix lyrashield_adapter lyrashield -q -c pyproject.toml -l -f json
0 findings

Enforcing medium would weaken the gate, which the brief forbids. The LOW rules
enabled by the current skip list are B311, B403, B405, B406, B407,
B408, B409. Their ruff equivalents S403 and S405-S409 are
preview-only in the locked ruff 0.15.20 and this repo does not enable
preview:

$ uv run ruff rule S403 | head -4
# suspicious-pickle-import (S403)
Derived from the **flake8-bandit** linter.
This rule is in preview and is not stable. The `--preview` flag is required for use.

$ grep -n 'preview' pyproject.toml
(no setting)

So a medium floor would drop seven rules and leave six of them with no
replacement check at all. The applied fix states the real floor and enforces it:

• pyproject.toml: the inert severity = "medium" key is removed, replaced by
  a comment recording the verified behaviour and why medium is rejected.
• Makefile security, .pre-commit-config.yaml and
  scripts/verify-controlled-derivative.sh: -l added.

No severity threshold was lowered, no rule was skipped and no suppression was
added. -l is behaviourally identical to bandit's default: no rule in bandit
1.9.4 carries UNDEFINED severity, so the default UNDEFINED floor and the
LOW floor select the same rules. -l makes the floor explicit rather than
changing it.

Gates run

Command	Result
uv sync --frozen	ok
uv run ruff check .	All checks passed
uv run ruff format --check .	448 files already formatted
uv run mypy strix lyrashield_adapter lyrashield	Success: no issues found in 251 source files
uv run bandit -r strix lyrashield_adapter lyrashield -q -c pyproject.toml -l	exit 0
uv run pytest tests/ (all 152 test modules plus tests/tui and tests/upstream, run in chunks)	3088 passed, 5 skipped, 2 failed — both environmental, see below
scripts/verify-controlled-derivative.sh footprint and digest steps	4 files changed, 22 insertions(+), 201 deletions(-); patch digest 3629e8f382fdd8eccf78553b102a25e99c73454c matches
python scripts/verify-customer-branding.py	Customer branding gate passed
uv run pre-commit validate-config .pre-commit-config.yaml	ok
python3 scripts/report_twin_drift.py	runs clean

Two failures, both pre-existing and environmental, neither touched by this
change:

1. tests/test_quality_environments.py::test_precommit_and_make_run_the_same_type_check
   — make is not installed in this sandbox.
2. tests/test_local_sources.py::test_clone_repository_checks_out_a_full_commit_sha_detached
   — this sandbox's git cannot resolve the detached test SHA.

The brief lists both as environmental. The three historical pytest failures
remain unreproduced and were not "fixed".

PROTECTED

• Pins: no pin bumped. ENGINE_REVISION, the engine
  .lyrashield-worker-pin and every dependency cap (openai, litellm,
  openai-agents, cryptography) are untouched. uv.lock is unchanged.
• strix footprint: git diff --shortstat "$(cat .lyrashield-upstream-base)" -- strix/
  returns 4 files changed, 22 insertions(+), 201 deletions(-), unchanged. The
  reviewed patch digest 3629e8f382fdd8eccf78553b102a25e99c73454c matches. No
  file under strix/** was edited. strix.spec is at the repo root, outside
  strix/**.
• Controlled-derivative boundary: AST-identical upstream twins were left
  alone. strix/core/agents.py:214 wait_kind_of is the upstream twin of the
  member removed on the product side; the upstream file was not touched. Nothing
  from the E2/E3 candidate lists beyond the seven proven-dead items was deleted.
• Ruff: 0.15.20 is what the lock already had. No upgrade, no new
  suppression, no per-file ignore added. ruff 0.16.6 (engine PR build(deps-dev): bump ruff from 0.15.20 to 0.16.6 #139) stays
  open and held.
• Customer branding gate: scripts/customer-branding-allowlist.json loses
  four entries that named the removed strix-agent lookups. Leaving them would
  have exempted the exact lines this change removes. The gate passes, and the
  remaining inherited-identifier exemptions are unchanged.
• Viewer, Windows, image and controlled-derivative coverage: the CI change
  adds a cache step only. Every step name, if condition and required check is
  unchanged, and the cache key falls back to a per-OS restore key so a miss
  still downloads.

Not done

• No broad flaky-test rewrite. The image-pull test keeps its intent; only
  the daemon-reach wait is bounded and documented.
• No enforcement of a medium bandit floor. Verified to weaken the gate, so
  it was rejected rather than applied. Details above.
• email_validator in strix.spec is also absent from uv.lock and the
  installed environment, like tenacity. It is a pre-existing pydantic optional
  extra, outside the reviewed deletion set, so it was recorded in the new test's
  known_absent set rather than removed. Flagging it for a separate decision.
• A real PyInstaller frozen build could not run in this sandbox: PyInstaller
  requires objdump from binutils, which is not installable here. The
  available substitute was run instead: parsing strix.spec and resolving all
  115 hiddenimports entries with importlib.util.find_spec under
  uv run --frozen --extra viewer leaves only email_validator unresolved and
  no tenacity. The frozen-build check therefore remains an operator step.
• The full verify-controlled-derivative.sh did not run to completion as one
  command: its pytest stage exceeds the 120-second sandbox command cap. Its
  footprint and digest invariants were run directly and pass, and its full pytest
  stage was run in chunks with the same suite and -W error::pydantic.PydanticDeprecatedSince211
  semantics; the bandit, ruff, mypy and format stages it wraps were each run
  individually and pass.
• Local TUI members (_encrypt, profile_for, update_run_status,
  iter_runs, KEYCHAIN_CHATGPT_TOKEN) were not deleted. They are in the E2/E3
  candidate list but not in this workstream's approved deletion set.

Rollback

Code only, forward-only for any schema. Revert this PR.

Generated with Claude Code

Summary by CodeRabbit

• Improvements
  • Version information is now reported consistently across the terminal interface, telemetry, and security reports.
• Maintenance
  • Updated quality and security checks, and improved the reliability of automated browser and Docker-related tests.

[coderabbitai]

Note

Currently processing new changes in this PR. This may take a few minutes, please wait...

⚙️ Run configuration

• Configuration used: defaults
• Review profile: CHILL
• Plan: Advanced
• Run ID: 4567cbb2-94f4-41d3-9e3c-b235474ef5ad

📥 Commits

Reviewing files that changed from the base of the PR and between d4f8a40 and 9b30f6d.

📒 Files selected for processing (20)

• .github/workflows/ci.yml
• .pre-commit-config.yaml
• .trivyignore.yaml
• Makefile
• lyrashield/artifacts/state.py
• lyrashield/artifacts/usage.py
• lyrashield/interface/tui/app.py
• lyrashield/interface/utils.py
• lyrashield/lifecycle/agents.py
• lyrashield/telemetry/_common.py
• lyrashield/version.py
• pyproject.toml
• scripts/customer-branding-allowlist.json
• scripts/docker.sh
• scripts/verify-controlled-derivative.sh
• strix.spec
• tests/test_engine_version.py
• tests/test_image_pull_deadline.py
• tests/test_quality_environments.py
• tests/test_release_build.py

 _______________________________________________
< Caches are bugs waiting to happen. - Rob Pike >
 -----------------------------------------------
  \
   \   (\__/)
       (•ㅅ•)
       / 　 づ

✨ Finishing Touches

📝 Generate docstrings

• Commit to this branch
• Create a new PR

🧪 Generate unit tests (beta)

• Commit to this branch
• Create a new PR

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

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