Skip to content

Phase 22 — Shared containment follow-up

This is a bounded follow-up to the workspace increment. It does not complete shared discovery or the Phase 22 technical gate. Initial measured source is commit cee0283; constructor-boundary correction e67365a follows it; native reports remain 1.6.0 and the schema 1.7.0 draft stays separate.

Implemented behavior

  • Python containment identifies imported SDK Context annotations separately from caller parameters, including Annotated wrappers. Similarly named or rebound annotations remain caller-controlled. TypeScript registrations treat the first handler argument as caller input and the second as injected context.
  • Python containment follows a simple factory returning a statically constructed local instance into its method. Return annotations alone do not establish the implementation. Replaced instance methods remain unresolved. Custom constructors, decorated classes and unresolved bases do not establish instance identity in either registration discovery or containment flow (e67365a).
  • Discovery follows a single recoverable implementation in local base classes. Conflicting base implementations and replaced inherited members are unresolved.
  • pathlib division retains the path type through an anchored join. Canonical resolution and an enforced related boundary check remain required.
  • TypeScript helpers retain path identity through native normalization of an already canonical path. An enforced component-safe prefix or equality against a canonical operator root can establish containment. Bare prefixes, unrelated inputs and arbitrary suffixes cannot.

The regression tests first reproduced the misses and false alarms. They retain source/sink locations and paired safe/unsafe controls. The focused containment and discovery suite passes 75 tests after the constructor correction. The initial source full suite passes 776 tests, with 36 opt-in Docker skips and 86.86% branch coverage. Ruff, formatting, strict mypy, native schema checks and the strict documentation build pass. Exposed-input measurements are retained separately under artifacts/phase22/shared-containment/; only completed records establish their respective checks. The constructor correction passes Ruff, formatting and mypy and the full suite: 780 passed, 36 skipped, 86.86% branch coverage. All 30 applicable hosted checks pass on 6ec5201, including Windows, installed-wheel Docker replay and network isolation.

Both initial-source runs completed 45/45 inputs, with identical stable findings and outcomes and no model calls. No findings changed relative to the workspace measurement; these regressions do not demonstrate improved exposed-case recall. The 273 raw evidence files are retained in measurements.tar.gz with per-file SHA-256 hashes in measurement-files.json, verified against the archive. Extract with tar -xzf artifacts/phase22/shared-containment/measurements.tar.gz -C artifacts/phase22/shared-containment. The initial-source measurements are not measurements of the later constructor correction or the eventual full implementation.

Limits and remaining work

These changes do not establish the complete Atlassian service-factory/upload flow or Mastra's security-failure return/fallback behavior. Filesystem collection guards, non-existent-path/symlink branches and all approved independent vulnerable/fixed conditions still need condition-level adjudication. Arbitrary factory results, reflection, custom construction and dynamic instance state are not validated by the simple-factory regressions.

The new source support currently feeds SENT-012. Recognition does not prove that the other detectors follow the same flows. SENT-013 through SENT-016, Kubernetes SENT-002, complete compatibility coverage, candidate-bound review, runtime campaigns and coordinated schema migration remain pending. The technical contract remains authoritative. No paid model calls, holdout-driven tuning, merge, publication or pilot acceptance occurred.