Skip to content

Codex Decision 0.11.1

Choose a tag to compare

@PhilippElhaus PhilippElhaus released this 07 Oct 15:49

Codex Decision 0.11.1 reduces CPU and memory costs from intercepted hook input through classification, relevance assessment, exact output rendering, and private evidence publication. The control package is 0.10.6.

  • Fit relevance batches with exact prefix costs and retain the independent final request-size guard.
  • Count tool-input and provider budgets without temporary encoded buffers. Borrow command metadata and source rows.
  • Serialize borrowed provider requests, receipts, and panel rows. Prepare bulk bytes before the session lock and stamp publication time under the lock.
  • Normalize provider replies without cloning their full JSON trees. Preserve exact types, question order, usage, and error checks.
  • Create private directories with mode 700 at creation. Reject nonregular event and lock files promptly. Bound log-tail reads and prune only recognized private records.
  • Stage private batch bytes before publication, clean recognized stale pending files, and preserve rollback on caught publication errors.
  • Render omission markers directly and combine duplicate representatives in one map. Preserve exact source spans and complete readable and typed originals.
  • Reuse panel row and header nodes across results. Clear cached rows when the panel clears, preserve protected labels and animations, and avoid redundant snapshot serialization.
  • Add complete wire and rendering parity checks, interruption recovery, and concurrent protocol-fault checks to CI.
  • Add panel shrink/growth, score changes, empty restoration, and resident-message reload checks.
  • Update architecture, installation, test, and measured-performance documentation. Keep new Rust modules concern-specific and within 200 lines.

The 5% relevance cutoff, eight classification classes, 95% branch gate, provider selection, request questions, timeouts, minimum savings rules, and protected evidence remain compatible. Live trials of shorter or alternative questions did not establish a consistent accuracy improvement, so the established questions remain in production.

Local benchmarks reduced OpenAI packing CPU by 2.4–10 times for ordinary fixtures. A 10,000-line receipt went from 80,043 allocations to two and approximately 16.16 ms to 4.13 ms. The panel serializer went from 290,049 allocations to two and approximately 40.40 ms to 7.09 ms. These are local measurements; they do not measure remote API latency or imply reduced API tokens.

Validation before delivery:

  • 143 Rust tests, format checks, and Clippy passed. Linux Node checks passed 68 tests with one optional platform check skipped; four Python checks passed.
  • The frozen corpus passed 312 cases across both providers, covering 5,124 required labels with zero actual or proposed evidence loss and unchanged saved-byte totals.
  • Optimized pressure checks passed 10,128 OpenAI invocations in one hour and 5,368 TypeSafe invocations in 30 minutes, including deliberate protocol faults and eight concurrent hooks on one CPU. Both recorded zero deadline fallbacks.
  • The revised panel passed Chromium and native Edge checks. A one-hour Chromium run passed 1,638 updates with 242- and 10,000-row snapshots, 81 completed-animation checks, 163 toolbar-fallback checks, and zero page errors, without reloads or forced garbage collection.
  • The short resident-panel CI command passed historical and current protected snapshots, reloads, animation checks, and fallback checks. Exact release archives contain nine plugin files and nineteen control files.

Limits: VS Code's running Windows updater blocked the revised renderer's native VS Code rerun. Earlier native checks cover the preceding renderer. Full snapshots sent repeatedly through the debugger exceeded a 3 GiB container with both renderers; resident-message endurance passed. A renderer-free diagnostic confirmed substantial debugger-driver heap growth. These failures remain documented rather than counted as passes. The hook declares glibc versions through 2.34 and the controller through 2.39, matching the preceding release; older controller runtime compatibility was not established.

Released from commit faa35d9c21ca0c24b5c1ed91aeb706ed15e6dca5. CI passed. Hook 0.11.1, control 0.10.6, and the pinned Codex 26.1002.51308 composer patch are installed. The nine plugin files, seventeen installed control payload files, and seven composer files match their verified package or patch bytes; VS Code adds installation metadata to its control manifest. Hook trust passed. An eligible installed hook call used the real OpenAI provider, processed all 83 synthetic rows in one API request, shortened the result, and verified exact readable and typed originals, independent sessions, and unchanged shared settings. The preceding plugin cache and composer originals remain available for rollback. Reload VS Code and start a new Codex thread to activate the updated packages in existing windows.

Asset SHA-256 checksums:

80b6c8b6a664b6a6d9c26947ed70426f30d82129ac1958d0b0a2ba3baf8c1460  codex-decision-0.11.1.zip
e9ea5785322edd1003c160bace3d8592c40ae7e219fdd1ea70d3c8d1fba9bdc8  codex-decision-control-0.10.6.vsix