[okisdev]

closes #70
closes #72
closes #73

summary

• bench/lanes/ measures one routing lane at a time: run.mjs sends a task brief through one engine companion on an explicit model and effort, runs the task verifier outside the workspace (or saves a consult answer for scoring), and appends a jsonl record; summarize.mjs groups the records, and bench/lib.mjs holds the helpers moved unchanged out of bench/run.mjs. the first snapshot, bench/results/lanes-2026-10-13/, compares composer-2.5, composer-2.5-fast, and gpt-6-luna on T05, T06, and T09 and runs four blind consults through four cursor seats and deep-reasoner. the cursor rows are re-scored from it: kimi-k3 rises to intelligence 4, all five rows trade provisional notes for measurements, composer stays co-equal with grok burst, and muse-spark-1.3 stays the diversity default
• the codex companion's tested interval moves to [0.154.0, 0.163.0) with verified-versions.json at 0.162.1, and classifiedFailureKind maps the exact rust-v0.162.1 texts: Flex capacity unavailable. to rate_limited, and the content filter and stock cybersecurity and biological risk refusals to refused instead of error
• the grok child env pins GROK_TURN_SUMMARY=false (no extra grok-4.5 summary call per turn) and GROK_FILE_ACCELERATION=false, and drops inherited GROK_SANDBOX_MODE, GROK_ARTIFACTS_BIND_REMOUNT, GROK_FILE_ACCELERATION_ROUTES, GROK_CURSOR_WORKER_API_BASE, and GROK_CURSOR_WORKER_COMMAND. the grok tested interval stays [1.0.14, 1.0.42) until the live pass in live pass for grok 1.0.52 and the grok burst arm of the lane snapshot #75
• the docs stop claiming the bridge variables cover every compatibility import: grok enables the claude code plugins that ~/.claude/settings.json enables, and since 1.0.52 their hooks and mcp servers can start inside a managed run unless ~/.grok/config.toml sets [claude_compat] imported = true or lists them in [plugins] disabled. the same pass fixes the stale [1.0.14, 1.0.31) interval in three files, the "entire grok home" write root, and the modelUsage field names, and fusion:trivial-worker pins claude-haiku-5-5

test plan

already verified

• npm test -> 1438 tests, 1436 pass, 0 fail, 2 skipped; one earlier full run hit the timing flake timeout stops after INT when Codex gracefully reaps its group child, which passed 5 of 5 alone and in the next full run
• claude plugin validate on plugins/codex, plugins/grok, plugins/cursor, plugins/fusion, and the marketplace root -> all passed
• codex-cli 0.162.1 through the repo companion: setup --json -> ready: true, compatibility: tested; a luna task with --output-schema -> done with {"ok": true}; --resume <thread-id> -> done with RESUMED; native review -> done; the 9 gpt-6-luna write runs in the lane snapshot all passed their task verifiers
• installed grok 1.0.52: grok inspect --json with all eighteen bridge pins set -> every compatibility cell false@env; a managed consult reached the model call and failed as quota (http 402), so the strict handshake matched; /grok:setup reports every required headless flag
• lane snapshot: composer-2.5 and composer-2.5-fast 3 of 3 on T05 and T09 at 60 to 75s medians and 0 of 3 on T06 (doubled quotes inside quoted csv fields), gpt-6-luna at xhigh 9 of 9 at 105 to 131s; on the consults muse-spark-1.3 and kimi-k3 matched deep-reasoner with no false claims, gemini-3.8-flash made one false claim, glm-5.2 missed a latent defect, and the research brief separated nobody
• the re-score is applied to both fusion hosts, and their model-routing.json files match row for row

reviewer should verify

• node bench/lanes/run.mjs --task T05-mechanical-codemod --engine cursor --mode write --model composer-2.5 --effort default --repetition 1 --results /tmp/lanes appends one record with verdict: pass and billingPool: first-party, and node bench/lanes/summarize.mjs --results /tmp/lanes prints its row
• /codex:setup on codex-cli 0.162.1 reports compatibility: tested

notes

• correction to re-verify the codex and grok companions on codex 0.162 and grok 1.0.5x #72: its "identical cells with all eighteen pins set" observation came from a zsh word splitting mistake in the probe (env $vars passed one malformed variable); set correctly, every cell reports false@env
• the grok live pass (consult, write, and review on 1.0.52), the interval bump, a runtime check of claude plugin startup, and the grok burst arm of the lane snapshot move to live pass for grok 1.0.52 and the grok burst arm of the lane snapshot #75, because the grok balance is exhausted (http 402) and OrbStack's ~/.orbstack/run/docker.sock makes the companion refuse grok writes on both hosts
• the consults showed that the T01 reference solution keeps a latent double removal its hidden test misses; bench T01 reference solution keeps a latent double removal #74 tracks the bench fix
• rules edits, per CONTRIBUTING's rules release checklist: the bridge variable sentence in orchestration.md and troubleshooting.md is corrected rather than softened (it adds the claude code plugin path and its two suppression keys), the model_usage row sentence gains cache creation, troubleshooting's write roots move from the stale "entire grok home" to ~/.grok/sessions, the cursor table rows are re-scored, rules-manifest.json is regenerated, and tests/rules-sync.test.mjs pins the new wording. nothing is deleted or softened