Repository navigation
pool flag: add Kimi K3.1 Pro via opencode-go to summit — AA 55, plan door found (models.dev Go catalog + 53/57 measured probe successes), $2.11/task #1409
Description
Add flag (pool-intelligence, 2026-10-27) — hypothesis for the controlled-test mechanism (#1342); earns a config change only through #1342 + the steward's PR path.
Add opencode-go/kimi-k3.1-pro to summit: AA 55 (Artificial Analysis public leaderboard, pulled live 2026-10-27; unrounded 55.1 in prior AA-quoting sources) — 4th overall, ahead of GPT-6 Astra 53 and Gemini 4 Argon 53, behind Opus 5.5 max 58 / Sonnet 5.5 max 56 / Muse Wonder Full 56. Summit's current contributor seat (Muse Wonder contributor) has NO AA read for its contributor id (AA lists Wonder Full 56 with no speed data, Wonder 52 @ 208 t/s); Kimi is a quantified summit-class alternative at $2/$10 per 1M, AA cost/task $2.11, AA speed 41 t/s, TTFT 2.95s.
Why this is new since the 2026-10-27 composition review (R4 watch): that review found no fleet plan door (canonical catalog carries K3.1 Pro only via CommandCode; ClinePass lacks it). Two fresher sources contradict the "no door" conclusion for OpenCode Go:
- models.dev
api.json(read live 2026-10-27): theopencode-gocatalog listskimi-k3.1-pro. - Fleet route-attempt records (laptop OpenCode store, read-only): opencode-go served kimi-k3.1-pro on 2026-10-25 — 53/57 attempts succeeded (93%; #high n=33 all success, default n=24, 20 success / 4 failure), median recorded speed 38.2 t/s (#high) / 42.8 t/s (default), median first-token 6.7s (#high) / 13.0s (default). That burst looks like a probe session, not routine pool traffic — labelled as such.
Provider variance recorded: our measured TTFT via opencode-go (6.7–13.0s) is 2–4× AA's 2.95s; measured speed (38–43 t/s) is close to AA's current 41 t/s (an earlier AA-quoting compilation said 49.7 t/s — AA's own current page wins).
Cost-to-pool: burns the opencode-go window (monthly 65.33% used at 2026-10-27 06:55 CDT per babysitter quota snapshot, reset Fri Nov 6 23:55 CDT) — not the OpenAI final reserve, not PAYG.
Acceptance for the #1342 test: a summit-class task batch on opencode-go/kimi-k3.1-pro #high vs the current summit lead (opus-plan #high, AA 54 @ $1.82/task): completion rate, wall-clock per completed task, and first-token latency; adopt only if completion is at parity and the Go window cost is acceptable to the steward.
Activity
developing-today commentedon Oct 27, 2026
CODE-500
dezren39 commentedon Oct 28, 2026
Picked up — queue task 1671 (Lane A shift, 2026-10-27)
files/controlled-pool-tests.mdexists in the goal workspace (created 2026-10-27 23:12 CDT) — gate passes.dezren39 commentedon Oct 28, 2026
Controlled test verdict — INCONCLUSIVE (queue task 1671, Lane A, 2026-10-27)
Per
files/controlled-pool-tests.md§4, an inconclusive test is a result, not a failure. The slice-1 gate passed (procedure doc exists, created 2026-10-27 23:12 CDT), but the test could not be executed in this task, for three verified reasons:providers: {}, server down / connection refused — and is also >60 minutes old. Drew's measurement gate forbids any placement/dispatch without a quota read ≤60 min old for that pool.opencode-go/kimi-k3.1-pro#high vs incumbent summit lead opus-plan #high). Repo-queue tasks are claimed by Muse cron worker lanes whose model route is not packet-pinnable to OpenCode providers; dispatching OpenCode sessions on those two routes belongs to the OpenChamber/steward dispatch path, which this task has no tool for.3a02ea3to370cd4d#5 (48h rework) cannot close inside a 45-minute task timebox; a final verdict needs a separate recording step ≥48h after the arms run.dezren39 commentedon Oct 28, 2026
Task 1671 closing: the follow-up dispatch-run task attempted from 1671 deduped against existing queue task 1418 (triage for #1409, fleet-monitor safety net) — that task is the live continuation for routing this flag into the controlled-test mechanism. No duplicate was created.
20 remaining items
dezren39 commentedon Nov 3, 2026
🤖 Agent note · picked up
c29174f1dezren39 commentedon Nov 3, 2026
Triage verdict 2026-11-03 — flag still live; the quota blocker is CLEARED, the pinned-route dispatch path is now the only blocker
1. Still live at current main (no config change landed).
held/user/owned/do-not-merge/do-not-auto-edit.~/.config/opencode/opencode.jsoncL4507-4514) is stillopencode-route/opus-plan#high→opencode-route/argon-plan#high→muse-direct/muse-wonder-1.0-contributor#xhigh→opencode-route/sol-plan#xhigh. No Kimi seat.2. Door evidence stands (re-verified, not refuted).
opencode-gocatalog still listskimi-k3.1-pro; the 2026-10-25 fleet probe burst still reads 53/57 success (#high median 38.2 t/s, default 42.8 t/s). The only in-configk3.1-proentry remains CommandCode (L3005), which matches the flag's premise: the Go door is the new one.3. Premise materially weakened under AA v5 (already recorded 2026-10-28 / 2026-10-30, unchanged today): Kimi is 46 on v5, 4-6 points below summit's v5 floor (Wonder/Argon 50), at $3.68/task, with AA speed drifting 41 → 32 t/s. This does not close the flag — pool-intelligence reserved closure for a directional loss in the settling run — but it means the test must compare Kimi against the v5 floor (Wonder/Argon at 50), not against the pre-v5 band.
4. NEW since task 5571 — the quota blocker is gone.
Task 5571 returned inconclusive on 2026-11-02 because every opencode-go workspace row read monthly 100%. The opencode-go-personal monthly window reset at 2026-11-03T03:02:35Z. Fresh PM quota-gatherer read 2026-11-03T13:00:06Z:
So arm A (
opencode-go/kimi-k3.1-pro#high) is cost-dispatchable right now on the Personal window — the specific blocker named on 2026-11-02 no longer applies.5. Remaining blocker — dispatch path (unchanged, structural).
controlled-pool-tests.md§2 requires each arm pinned by packet; repo-queue Muse lanes cannot pin OpenCode routes. The settling run must be dispatched on the steward/OpenChamber path.3a02ea3to370cd4d#5 (48h rework) cannot close inside one worker shift; a verdict task is needed ≥48h after the last arm.Verdict: keep open, unadopted. No pool or routing config change was made and no session was dispatched. Action now available: run the settling test (3 tasks/arm directional; 6/arm if it will support a config-change recommendation) with arm B = the v5 summit floor (Wonder/Argon at 50), fresh quota reads at dispatch, resolved routes recorded per task, verdict task ≥48h after the last arm. A directional loss closes this flag.
dezren39 commentedon Nov 3, 2026
Controlled-test refresh (task 5571, 2026-11-03T21:55Z) — INCONCLUSIVE, no arms dispatched
Fresh laptop quota-gatherer output (
~/.local/state/pm-aggregator/quota-current.json, generated2026-11-03T21:55:21Z; heartbeatok):opencode-gorow: status ok, weekly 0%, monthly 0% used. A second OpenCode Go credential row is at 5h 100% (reset2026-11-04T00:18:08Z), weekly 99%, monthly 49%. The queue task has no way to pin which credential will serve the candidate arm.opus-planis currently configured as ordered targetsclaude-code/claude-opus-5-5thenantigravity-direct/claude-opus-5-5. The fresh Claude read failed (Claude session expired); the Antigravityclaude-opus-5-5-highquota row was 20% on 5h and 10% weekly, measured2026-11-03T21:45Z. Without an actual packet-pinned dispatch, there is no resolved route/account to attribute to the incumbent arm.The earlier test limitations still block execution: this queue claim cannot dispatch six comparable real tasks with model routes pinned and record resolved routes; that dispatch belongs to the steward/OpenChamber path. The required 48h rework criterion also needs a separate verdict step at least 48h after the final arm. No tasks were dispatched and no A/B measurements exist — these are quota observations only, not test results. The flag remains open and unadopted; no pool configuration changed. Prior Dataset 2 no-run record is referenced by task 5571's previous checkpoint; the goal-workspace files are not present on this laptop, so I could not revalidate or edit them in this pass.
Resume condition: an explicit pinned-route dispatch path, unambiguous credential/route selection, and a fresh successful quota read for every resolved arm; then run at least 3 tasks/arm and schedule the 48h verdict step.
dezren39 commentedon Nov 4, 2026
Pool-composition review, run 17 (2026-11-04T10:55Z) — fresh numbers for the settling run
controlled-pool-tests.mdis unchanged since 2026-11-02 09:49Z (2 datasets; Dataset 2 stands inconclusive).dezren39 commentedon Nov 4, 2026
Pool-intelligence update 2026-11-04 (window 2026-11-01T14:51Z → 2026-11-04T14:12Z):
AA v5 (public leaderboard pull 2026-11-04, 279 rows): Kimi K3.1 Pro II 46.5 (was 46.0), $3.681/task (unchanged), outputSpeed 32.2 (AA renamed its speed field this pull; prior median-basis figure was 31), TTFT 4.94s. The drift is part of a board-wide +0.1–0.7 II recalibration this pull, not a Kimi-specific move.
Summit context moved more than Kimi did: AA added Gemini 4 Carbon #medium (II 58.3, $6.59/task) and #low (II 54.2, $7.19/task) rows — Carbon still has NO plan door (zero models.dev provider catalogs, live read 2026-11-04), so it files no flag, but summit's v5 ceiling is now 60.0 (Carbon #high) and Kimi sits 3.5–5.5 points below summit's seated floor (Argon #high 50.4 / Wonder Full 50.1).
Measured: zero Kimi attempts in the window (n=165,220 route attempts) — the opencode-go catalog still lists kimi-k3.1-pro ($2/$10, models.dev live read), and the Go doors carried heavy Flash traffic after their resets, but no Kimi seat exists to draw attempts. The settling test still needs the steward/OpenChamber pinned-route dispatch path (unchanged from task 5571's blockers, except the Go monthly quota has now reset).
Premise unchanged-to-weaker; flag stays open for the controlled test. Numbers: files/pool-intelligence-model.md window 2026-11-01→2026-11-04.
dezren39 commentedon Nov 4, 2026
Pool-composition update 2026-11-04 (run 18, 22:55Z):
Kimi K3.1 Pro open weights — LANDED. Verified directly against
the Hugging Face API this run (~23:0xZ):
moonshotai/Kimi-K3.1-Prois public with 112 files including 96 safetensors shards
(downloads 797 at check). This supersedes this morning's 06:15Z
read (401 / no files, recorded by pool-intelligence) — the
promised "by November 4" deadline is met as of this evening
check. HF metadata shows created/lastModified
2026-11-02T04:13:51Z, so the visibility change itself is not
timestamped by HF; both reads are recorded as read.
Flag impact: none on the action. #1409's door is the
opencode-go plan catalog ($2/$10), not the weights. The summit
arm still awaits pinned-route dispatch for the #1342 settling
test: the Go monthly quota prong cleared Nov 3, zero Kimi
attempts again in the 2026-11-01→2026-11-04 window
(route_attempt n=165,220), and the meta-reviewer's 2026-11-04
report records the settling run as two windows stale. AA
context unchanged-to-weaker (Kimi II 46.5 vs the Carbon #high
ceiling at 60.0; seated floor Argon 50.4 / Wonder 50.1).
Numbers: pool-composition-log run 18; HF API read 2026-11-04T23:0xZ.