v0.4 verification evidence#

Reference checks ran on Windows with Python 3.13, Tcl/Tk, Ollama 0.32.5 and the dependencies constrained in requirements.lock.txt. A separate clean directory and virtual environment were created from the public source allowlist; no personal config or usage history was copied. This is a same-PC clean-install check, not a claim of testing a second physical Windows machine.

Installation and MCP#

• Installed the source with tray/dev/benchmark extras into the clean environment.
• Copied config.example.toml and generated a fresh SQLite database.
• doctor verified runtime dependencies, writable data, loopback Ollama and all three configured role defaults. Missing global registration is a diagnostic warning, not a failed Ollama connection.
• verify-models made three successful real calls: Qwen3 4B general, Coder7 coding, Phi4 Mini reasoning. No model substitution.
• smoke --all --consensus exercised all seven handlers and the additional two-participant consensus: nine successful local model calls, no handoffs/errors.
• Generated codex-config output was loaded by the installed Codex CLI with an isolated CODEX_HOME. codex mcp get localagentbridge --json confirmed the enabled stdio registration and exact clean-environment command. Global Codex settings were not changed.
• This proves configuration acceptance and real MCP-to-Ollama execution. It does not claim that a Codex reasoning turn chose to invoke a tool automatically. The README explains how to correlate an actual Codex invocation with local telemetry.

Final artifact verification is recorded in the public release checklist. The final clean source copy repeated the three-model verification and all-tool consensus smoke successfully: another 12 real local calls. Both clean runs retained provider counters and model names; they are separate from quality-scored benchmarks.

Regression and native UI#

Final workspace regression: 126 passed, 1 skipped, 29.04 seconds; lint passed. Final clean-copy regression: 126 passed, 1 skipped, 30.34 seconds; lint and pip check passed. The final package was also built as a non-editable wheel and inspected for packaged files and privacy-sensitive artifacts. The skip is the Windows Developer-Mode symlink case. One intermediate run reported an intermittent Tcl library lookup failure while heavy inference was running; the file was present, and the subsequent complete run passed. Its cause was not established.

Tests cover legacy single-model behavior, all seven MCP handlers, response budgets, repository boundaries/ignore rules, role/cap/resource/quality routing, bounded escalation, feedback, additive migrations, Codex counter reconciliation, setup diagnostics and release allowlist guards. GUI tests cover immediate tier updates, missing/disabled/unregistered models, a fresh database, routing facts, correct compression, local activity visibility and the first fixed-call failure popup. Native focus checks verify SW_SHOWNOACTIVATE; repeated polls do not repeat a failure popup. Existing v0.1–v0.3 tests remain in the suite.

Simulated Tk scaling of 100%, 125% and 150% passed. Synthetic screenshots at 100% and 150% were inspected for readable wrapping and no normal horizontal scrolling. This does not replace checks on every monitor, DPI configuration or Windows build.

Resource use and local boundaries#

The native stack uses Tk/pystray, without Electron or a web dashboard. A synthetic hidden widget measured about 0.47% of one CPU core and 47 MiB working set over 20 seconds. Actual tray-plus-collector samples after grouped-query optimization measured 2.60–4.10% of one core and 64–71 MiB over separate 45-second windows while Codex telemetry activity varied (roughly 0.22–0.34% of this 12-thread system). These short observations are not controlled proof of a fixed idle ceiling. The final refreshed tray/collector measured 2.53% of one core and 60.1 MiB over another 45-second window while Codex activity and separate smoke calls continued. Metadata inventory is infrequent; hidden aggregate rendering pauses.

The production telemetry listener was checked at 127.0.0.1:4318 only. MCP uses stdio. No cloud telemetry, model shell tool, vector store, autonomous loop or public server was introduced. Private benchmark/store/config material remains local and excluded.

Model evidence#

The v0.4 validation stores record 69 real local calls (historical baseline requests are excluded from this count):

There were 11 failures, all from the preserved GPT-OSS think=false protocol. The compatible low-thinking suite had no API/schema failures. These are local request counts, not cloud calls avoided. Measured cloud calls/token reduction remain unavailable. Benchmark payload compression is reported separately below.

See the complete paired comparison for exact tags, protocol compatibility, independently reviewed A/C/R grades, failures, token/time measurements, sampled memory, byte compression and routing recommendations. The five historical baseline results are preserved. New calls are separate records. No public release, GitHub push or unrelated project modification was performed.