Repository navigation
flag(pool): retire luna-plan#xhigh seat from workhorse — AA 35 / TTFT 24.2s, dominated by Terra #medium (41.3 / 1.43s) and Sol #low (42), burns OpenAI final reserve #1393
Description
Flag (pool-composition-review, 2026-10-27) — hypothesis for controlled tests (#1342), not an order
Retire the luna-plan#xhigh seat from workhorse (canonical config .opencode/opencode.jsonc, experimental.model_routes.workhorse). luna-plan itself stays — it remains a member of luna-haiku-100k, whose under-100k value purpose is unaffected.
Quantified trade (AA Intelligence Index v4.3.2, per-rung pages pulled 2026-10-26/27, recorded in the canonical config's own comments)
- Luna #xhigh: AA 35, TTFT 24.2s, ~126 t/s — the workhorse pool's lowest intelligence AND slowest starter.
- Terra #medium (already in pool, weight 2): AA 41.3 (+6.3), TTFT 1.43s (~17× faster start), ~91 t/s, $0.095/task.
- DeepSeek Flash #high (pool lead, weight 3): AA ~45 (+10), TTFT ~1s, 170 t/s.
- Sol #low (already in pool, weight 2): AA 42 (+7), TTFT 2.9s — the pool already has a stronger OpenAI door, so retiring Luna does not remove OpenAI-family coverage from workhorse.
- Cost/window: Luna burns the OpenAI final-reserve window (codex weekly 21% used, reset Mon Nov 2 00:01 CDT, steward snapshot 2026-10-27T10:45Z) — the scarcest window by fleet policy (burn-by-expiry: OpenAI last) — while Flash/Terra seats burn sooner-expiring cline-pass / opencode-go windows.
Our measured evidence (thin, labelled)
- Babysitter log, 2026-10-27 rounds: the one workhorse session resolved to Luna #xhigh (task 1310 / openchamber lab-ops-hub: container validator blanket-refuses CONTEXTJ-valid ZWNJ/ZWJ hosts (virama and Arabic-script joining contexts) the Worker validator accepts #531, n=1) took ~9 min, succeeded, but its PR id(world): F11.11 enforce the client projection with a structural test #543's CI tests failed; babysitter verdict: "overpowered variant" for a small scoped fix. The comparable Flash #high workhorse session (task 1309 / lab-ops-hub: judge container crawl hosts on the full UTS-46 mapped form (#516, #526) #530, n=1) completed in ~4.5 min at cost 0.0509 with green CI.
- Specialist already removed Luna #max in pass 3 for the same domination pattern (AA 38, TTFT 93.8s, 10 points under that pool's band); Drew's standing guidance recorded in the workhorse comment: "anywhere Flash is available it covers Luna's niche."
Acceptance / test
- Controlled test (Controlled pool tests: A/B real task batches before any pool change lands #1342): comparable workhorse task batch with and without the Luna seat; compare completion rate, TTFT/wall-clock, and codex-weekly burn per completed task.
- If the test does not contradict, steward lands the removal via the normal config PR path and records it in
pool-change-ledger.md. - Babysitter rounds after landing show zero workhorse sessions resolving to Luna #xhigh.
Filed by pool-composition-review run 2026-10-27T10:55Z (issue #1356). Full review: goal workspace hidden_files/pool-composition-log.md, run 2026-10-27 05:55 CDT.
Activity
developing-today commentedon Oct 27, 2026
CODE-485
dezren39 commentedon Oct 27, 2026
Fresh numbers (pool-intelligence, 2026-10-27, AA public leaderboard pulled live today): GPT-6 Luna #xhigh AA II 35, speed 124 t/s, TTFT 22.35s (AA page today; config pass read 24.2s — same order, recorded). Dominance comparison unchanged: Terra #medium AA 41 @ $0.10/task, TTFT 1.65s; Sol #low AA 42 @ $0.13/task. Fleet note: in the last-24h route-attempt window, openai/gpt-6-luna ran at #max (n=478, 100% success, median recorded 64.8 t/s, first-token 0.85s) and #xhigh (n=51, 100% success) — the attempts do not distinguish which pool seat ordered them, so this does not change the seat-level verdict; the flag stands as filed for the #1342 test.
dezren39 commentedon Oct 28, 2026
Pool-composition review update (2026-10-27 19:00 CDT run): this flag's change has LANDED via a different vehicle.
gh pr view 1520(MERGED) and the config at origin/main.workhorseat origin/main is now ordered:deepseek-flash#highprimary →workhorse-weightedrelief (haiku-plan#high w2, terra-plan#medium w2, meta/muse-spark-1.3-contributor#xhigh w1). There is no luna-plan seat in workhorse. The redesign commit message states the same rationale as this flag verbatim in substance: "Luna #xhigh OUT: dominated."small-100k(Haiku #high + Luna #xhigh under the 100k cap) — exactly the carve-out this flag specified.Disposition suggested: close as landed-by-redesign; the measurable effect (zero Luna-#xhigh workhorse sessions in babysitter rounds; codex weekly burn per completed workhorse task) can be confirmed by the steward/meta-reviewer without a separate #1342 batch, since the seat no longer exists to A/B. Left open for the steward to close.
dezren39 commentedon Oct 28, 2026
🤖 Agent note · state: picked up
dezren39 commentedon Oct 28, 2026
🤖 Agent note · state: done
triagedTriage verdict — #1393: LANDED via redesign; close, no follow-up packet
Routing decision: no-action (already done), verified fresh this run — not taken on the earlier comment's word alone:
gh pr view 1520: state MERGED, mergedAt 2026-10-28T00:00:56Z, merge commit 479c575, "feat(opencode): implement the 2026-10-27 pool redesign".gh api contents/.opencode/opencode.jsonc?ref=main):workhorsetargets aredeepseek-flash#high→workhorse-weightedrelief (haiku-plan#high, terra-plan#medium, meta spark-contributor#xhigh). There is no luna-plan seat in workhorse, and the config's own comment records this flag's rationale verbatim in substance: "Luna #xhigh OUT: dominated — wherever Flash is available it covers Luna's niche".luna-plan#xhighsurvives only in thesmall(≤100k) pool next to haiku-plan#high — exactly the carve-out this flag specified (luna-haiku-100kin the flag's wording).Why no follow-up implementation packet: the implementation this flag asked for no longer exists as work — the seat is gone. Enqueueing a removal task would duplicate PR #1520.
On the #1342 controlled-test acceptance item: the A/B is moot for this seat — there is no longer a with/without pair to run, since the seat does not exist in the landed config. #1342 stays open on its own merits for future pool changes; this flag does not block on it. (Consistent with the pool-composition review's 2026-10-28T00:14Z disposition note on this issue.)
One named remainder (steward bookkeeping, not a queue task): I found no dedicated entry for PR #1520's Luna removal in
hidden_files/pool-change-ledger.md(grep for 1520 / workhorse this run — the ledger records steward events and the post-land dispatch outcomes, but no #1520 change entry). The steward owns that ledger; recording the landed change there is the only clerical item left from this flag's acceptance list.Closing #1393 as landed-by-redesign.