Companion candidate readiness#

This is the historical recovery report. The Windows qualification addendum subsequently passed the actual Python 3.11 gate; its remaining gates still apply. See development source/package provenance for the source publication. 0.5.0.dev0 remains not release-qualified.

Status: cross-database process-crash recovery blocker resolved; Windows release gates remain open. The focused repair persists authoritative packets/transition receipts and reconciles downstream results without replay. Unsaved output, incomplete stages and ambiguous legacy incidents still require explicit intervention; they are never reported as recovered success.

Final Windows suite: 411 passed, 1 skipped in 644.98 seconds. The skip needs Windows symlink privileges; the separate reparse-point guard regression passes. Final installed-wheel recovery suite: 50 passed in 490.61 seconds. The old wheel's 13 failure characterizations are retained. Actual v0.4.1/schema-1 wheel upgrades preserve historical rows; copies of the original checkpoint incidents repair automatically, while lost hash-only output remains unavailable.

One packaged repository review used exact GPT-OSS routing through the existing workflow engine (1,595 input / 72 output tokens), with independent Codex review, an explicit checkpoint and three separately recorded reviews. The incorrect local finding was rejected; Codex supplied corrected analysis. Local completion was observed before review completion. The final compatibility correction preserves the original non-companion ledger method contract; the local step's installed runtime was retained unchanged, and the final wheel handled the subsequent checkpoint after all earlier workers exited. See contributions.

Native dock/model-selector and missing-provider/model guidance checks pass. The GPT-OSS preference, exact Qwen tag/digest, stable checkout and frozen runtimes are preserved. No models were downloaded/deleted and no callback was sent. Supervised local use is supported within recovery's restrictions. Windows RC gates remain: authorized real authenticated callback testing, a fresh Windows machine/VM installation, Python 3.11, human visual/interaction acceptance, and the privilege-dependent native symlink check. New virtual environments on this host are isolated installs, not fresh-machine tests. The existing active-resource/perceptual measurement checklist is also retained; no performance certification or other-platform certification follows from this repair.

Historical feature-preservation follow-up#

The following records are the baseline for the focused repair above. Their byte-equality claims and counts describe that earlier candidate; necessary recovery changes to the workflow/ledger are documented separately, without repeating the broad preservation audit.

The matrix compares the actual pre-companion v0.5 implementation, not merely v0.4.1. Of 70 implementation/configuration/test files compared with the frozen manifest, only __main__.py, bridge.py, config.py and widget.py differ (additive CLI/transport options and the previously justified dead UI cleanup). Workflow, workflow ledger, catalog, delegation, adaptive routing, service, schemas, all 31 definitions, profiles and pre-companion v0.5 regressions remain byte-identical. All 20 compatible built-in role/skill pairs are now tested through executable custom project-catalog stages.

The missing companion integration is repaired with the existing Service and WorkflowEngine, full configuration, policy identity, one durable step/checkpoint per job, persisted bounded handoffs, exact model digest checks, local review tracking and dock budget/checkpoint details. See integration. No implemented feature was removed because it was unpublished/experimental.

Latest full Windows acceptance: 359 passed, 1 skipped, 148.24 seconds, including native Tk and real detached processes. Subsequently added exact-Qwen single-opinion and uncertain-local-review regressions pass in the 35-test workflow integration suite. A final retry-after-completion check exposed volatile planning estimates in the submission identity; comparison now excludes only those estimates while retaining task/configuration/policy/model identity. The subsequent companion and workflow suite passed 73 tests (one native test deselected, already covered by the full Windows run). Existing real MCP stdio tests retain 7-tool opt-out and 13-tool workflow-enabled behavior. Providers in automated tests are synthetic. The CLI integration test uses a disposable Git repository and exercises workflow submission/retry, worker restarts, status, next step, checkpoint, local review and successful finalization without initializing a Codex session.

The first broad run exposed a Git-owner-dependent CLI fixture and a transient Windows Access denied during research-worker atomic state replacement. The fixture now owns a disposable repository. State replacement now retries Windows sharing/access errors for at most 200 ms; persistent failure remains an error and retains old state plus temporary evidence. Both paths have regressions. The original coordinating implementation and frozen research evidence were not edited.

Live local verification used a frozen pre-companion multi-stage GPT-OSS workflow (1,861 input / 256 output), then a separately frozen candidate worker's durable workflow direct path (633 / 64) and fixed sequential consensus (GPT-OSS 508 / 60; exact Qwen3 14B 363 / 165). The launcher exited independently. The worker completed and exited; the next workflow step after restart stopped at the supervisor checkpoint. The two opinions disagreed, which is retained. No broad model-quality claim follows. Contribution accounting explicitly separates these paths from the initial direct Bridge calls.

The documentation checker validates 29 Markdown documents and generated CLI/configuration references. The public allowlist has 172 files, with zero pattern/ignore findings. Updated wheel/sdist and isolated-install verification cover 72 wheel entries, all 31 definitions and all 172 allowlisted source files; the extracted source's documentation and audit also pass. Isolated installation, upgrade sentinel preservation, uninstall data retention and CLI submission/retry/ cancellation are recorded separately from the older package evidence below. No authenticated callback, model download/deletion, stable checkout edit or publication occurred.

At that baseline, the cross-store crash window was explicit: an engine-ledger commit could outlive a crash before companion handoff persistence. Reservations remain spent; missing content requires supervisor recovery/new scope, never silent replay. Original v0.5 workflows are still continued through their original interface, not automatically imported into the companion. Fresh-machine, Python-minimum, other-platform and authenticated-callback gates remain open.

Initial candidate record (before the follow-up above)#

Repository: existing nishdel/LocalAgentBridge, clean stable checkout at f6df420181f21856cb421619e2c1310f6cf46685 (v0.4.1). Development branch: companion-release-candidate in a separate worktree. The original workspace root had no Git metadata and contained newer v0.5 work; commit cc88d4d preserves that source before candidate changes. The stable checkout and original runtime source remain unchanged. No models, historical evidence, user DBs or unrelated projects were deleted.

The previous Qwen terminal review was inspected before development: recorded started, reviewer and worker absent, final events were network reconnection failures, no response/report artifact. This is an unresolved interrupted review, not a successful review or confirmed graceful controller exit. No duplicate old review or Qwen benchmark was launched.

Staged work and validation#

1. Inspected state, preserved newer source, hashed frozen runtime/environment.
2. Added inventory, exact model selection, durable jobs/outbox and callbacks.
3. Tested lifecycle recovery and used genuine local code/test/doc contributions.
4. Added dock, contributor/user/security docs and evidence-backed cleanup.
5. Clean package/install/upgrade and end-to-end gates are recorded below as run.

Baseline: 276 passed, 1 skipped, 8 failed. Seven native tests failed because the restricted shell could not initialize Tcl; one lifecycle test asserted completion artifact existence before the worker finalized it. That race now waits for both terminal state and the artifact. Native tests run outside that restriction: 14 passed. New companion plus protected-worker targeted suite: 63 passed.

New automated tests cover duplicate submissions/completions, simultaneous review claims, busy session serialization, restart with uncertain delivery, seven-day quota retry metadata, authentication/network failure categories, cancellation, missing models, explicit final validation and deduplication. Clock advancement to six hours/seven days is simulated, not a multi-hour model run. Real detached Python processes with fake Ollama prove worker independence and result persistence.

Live contributions: GPT-OSS 20B code/test/doc proposals and one focused correction; see ledger. Their quality required corrections. No full Qwen evaluation was run. The genuine candidate worker completed independently (853 input / 129 output tokens, 92.35 seconds bridge execution); the result was rejected on supervising review. Its callback remains pending because automatic approval review rejected sending repository-derived content/result and local paths to authenticated Codex without specific permission. No completed live callback or successful run finalization is claimed.

Cleanup and compatibility#

• Removed unreachable registry controls, horizontal detail table, set_cap, toggle_details and refresh_details from LegacyDashboard. Factory dispatch sends all registry configs to RegistryDashboard; existing v0.1/v0.2 views remain supported and their native tests pass. No supported CLI/UI route was removed.
• Removed duplicate process/lock/Windows Job Object implementations from the research IO module, importing tested shared process_control primitives. Frozen historical jobs retain their old snapshots. New runner snapshots include the shared helper and fixture module.
• Removed the obsolete one-time freeze_v02.py from the development tree. It was already excluded from public packaging and had no current entry point, caller or test dependency; historical fixture/evidence files remain untouched.
• Removed the redundant benchmark dependency extra; psutil is core for durable process identity. Existing constraints already pin psutil/pywin32. No third-party runtime dependency was removed merely because a static scan found few imports.
• Replaced private-evidence-dependent lifecycle fixtures with explicit synthetic fixtures. Protected cancellation/ownership tests remain. The old assertion that current source must equal a historical snapshot was replaced by public profile compatibility plus per-experiment frozen-input verification.

Remaining research scripts have active tests/CLI/reference uses and are retained; this is not a claim that every remaining line is necessary. Research evidence is excluded from runtime packages, not deleted or relocated as a cleanup shortcut.

Known limits and release gates#

Callbacks are dedicated read-only reports. Automatic integration into the original desktop task is not implemented; the supervisor records decisions/validation. Managed sessions must be exclusive to this controller/store. Cross-client busy session coordination and a general app-server event transport are not implemented. Structured reset/error handling is tested with fixtures; CLI versions can emit unstructured reconnect errors, which require manual reconciliation.

Shared-server cancellation cancels the HTTP request, not guaranteed server-side generation. No exact token-position resume, inference during shutdown, model download manager, bundled executable installer, or non-Windows desktop certification is claimed. Models exposed through Ollama must truthfully report their remote metadata. Raw diagnostics require manual redaction before sharing.

Python/Tk is retained; measurements are recorded separately from Ollama inference in the local validation evidence. No language rewrite or inference speedup is claimed.

Final observed validation#

The package is a wheel/source distribution, not a bundled executable installer. Clean environment means a new venv on this Windows host, not a fresh Windows VM. Only Python 3.13 on Windows was exercised here; the declared Python 3.11 minimum and other operating systems still need a compatibility matrix run.

Tested dependencies include httpx 0.28.1, mcp 1.30.0, psutil 7.2.2, pywin32 312, Tk 8.6.15, Pillow 12.3.0, pytest 9.1.1 and Ruff 0.16.7. Build used setuptools 84.0.0. The frozen initial dogfood server was Ollama 0.34.3; shared-server correction and candidate calls used Ollama 0.32.5. Installed Codex CLI inspected was 0.155.0-alpha.16.4; live callback behavior on it remains unverified.

One warm controller sample (no inference) measured 1.51 seconds startup, 43.16 MiB resident memory, 0.78% of one CPU during a two-second idle sample and 9.76 ms mean SQLite setting transaction over 100 writes. The controller then exited with code 0. These are laptop samples, not performance guarantees. Active inference CPU/peak memory, true callback latency and perceptual UI responsiveness remain unmeasured; native UI functional checks pass. No C#/C++ rewrite was justified or attempted.

The development branch contains separate baseline, feature and cleanup commits. Final documentation and follow-up corrections remain staged while the pending live run retains base 8b02eba. The public stable checkout is still f6df420. Review the staged diff and source hashes alongside this report; do not confuse its Git HEAD with an assertion that the staged tree was already committed.