Opencode GO and ZEN models — consolidated tracking

Single consolidated file. Combines: model selection sheet (Mira, 2026-10-11), GO free-models watchlist (Tam, 2026-10-18), model-swap runbook (Tam, 2026-10-20), usage-tracker + ZEN free-tier finding (Tam, 2026-10-20), and GO/ZEN usage tracking (Mira, 2026-10-18).

Rule: one file per topic. This is the only model tracking document on the vault.

Opencode GO and ZEN Models.md

Opencode GO and ZEN Models

The sheet Young asked for on 2026-10-11: track which models Brain and Heart use, with prices and allowance limits, so a model choice is a lookup and not a re-derivation.

How this file is sourced, because a price sheet that is wrong is worse than none:

1. The live catalogue — what actually exists

Probed 2026-10-11 against both endpoints. The headline: GO is the small pool.

GO's 45 models are mostly things that will block. A $10 package against per-model ceilings is not headroom — it is a menu of things that will run out. ZEN's 12 free models are the real capacity, and they cost nothing.

GO free (3) — probed

2026-10-15 update: new models, live-verified by render (each ran "say OK" and answered). GO now requires the x-opencode-session header on direct API calls — probed via the CLI (opencode run ... --model ...), which carries its own auth and session. A bare curl without the header gets MissingSessionID, which is a client error, not a model error.

ZEN free (14) — probed

2026-10-15 additions, both live-verified:

Eight of ZEN's fourteen also exist on GO (the overlap grew with cat-jam-free, which is live on both). Where a model is on both, use the GO copy — usage inside the $10 package costs nothing extra; ZEN drains the $40 float.

2. GO — paid models with allowance limit

Prices and limits are Young's own figures, quoted 2026-10-10 (opencode.ai/docs/go/#usage-limits). Each model's limit is independent, and the 5-hour window is the binding constraint — 20% of monthly, not 100%.

The two things this table decides

1. Qwen3.8 Flash took $11.17 last week — and it was for nothing. Our own opencode stats --days 7: free models did 5.07b tokens for $0.00 while paid did 573m tokens for $18.14. Free was 8.8× the tokens at zero cost. Head to head on the single largest line: space-bunny-free did 2.0b tokens for $0.00 while qwen3.8-flash did 309m tokens for $11.17 — 6.5× fewer tokens, paid for. $11.17 against a $6.00 / 5-hour window means one good afternoon blocks the model.

2. Luna and Haiku at $3.00 / 5 hours are 4× tighter. Both are "complex task" candidates in the plan. Against a $3 five-hour ceiling they are fragile as fallbacks — they are escalation-only, never a default.

Prompt caching is the cheaper lever, and nothing used it

GLM-5.3-Flash caches reads at $0.03 vs $0.15 input — a 5× reduction. Our fleet is already at 91.6% cached input. A caching strategy buys more headroom than any model switch on this table.

3. ZEN — paid models enabled on the workspace

From Young, 2026-10-10. ZEN is $40/month pay-as-you-go credit — a float that drains, not an allowance. ZEN paid is the LAST rung, the only one that costs real money.

Rung order, now live in both switches:

GO free  →  ZEN free  →  GO paid  →  ZEN paid

Cost rung beats instrument rung. A free model on ZEN is still $0.00; a paid model on GO still costs money. The instrument-major order (GO paid before ZEN free) called qwen3.8-flash on attempt 3 while a free model sat untried until attempt 13.

4. Who is on what — Brain and Heart, measured

Brain — runs on Hermes, switches via fallback_providers

hermes fallback list, 2026-10-11, identical on all three profiles:

Four models, all free, zero paid in the chain. And it matches the V3 ladder's rung order exactly — reached independently, which is the evidence the doctrine is right.

Heart — runs on SolOSV3, switches via router.route()

cortex/registry/agents.yaml, all 11 agents:

Every agent on the one model with 769 successes behind it. Deliberate for a build phase; the ladder handles switching.

🔴 And the live defect this sheet records

core/config/models.yaml declares space-bunny-free under GO. It is not served by GO — it lives on ZEN. Measured:

go/space-bunny-free → "ModelError: Model space-bunny-free is not supported"
MISS lab-go/space-bunny-free | NOT served by go (45 models listed)

This is why the sheet exists. The registry had a model on the wrong endpoint, and nobody could see it until a probe asked the endpoint instead of reading config. Brain moved off it on 2026-10-11 — all three primaries are now longcat.

5. Selection rules — the reason each model is listed

Rule 6 in force: Qwen3.8 Flash at $11.17 has no recorded reason. The ledger now refuses such a record at construction, and audits history for rows predating the guard.

6. What this sheet cannot tell you

• Prices are Young's figures, not my probe. I verified which models exist and where, not their prices. No endpoint publishes a price list, so a price here is a quoted figure, not a measurement. The same is true of the monthly limits and the 5-hour window percentages.
• Free-model quality is unmeasured. I rank by cost and proven volume, never by quality. V2's 769-success figure proves longcat works; it does not prove it is best.
• The 5-hour percentages are from the settlement docs, not observed. V2's breaker measured real 429 provider.quota events — but against the hard configuration. Our guard assumes the soft one. That is settled by the console's "Extra usage" setting, which is still an open question.
• Nothing here is refreshed automatically. The catalogue probe (core/config/models.py::probe) can answer this file's tables on demand — this sheet is a snapshot from 2026-10-11 and it will drift.

The honest use of this document: it answers "which model should Brain or Heart use, and what happens when it blocks" on the day it was written. For anything time-sensitive, run the probe.

go-free-models-watchlist.md

GO free-models watchlist (2026-10-18, live-verified seed)

Young: create + update DAILY; review/verify new free models; optimise new models first; feed verified models into Pulse/Brain/Heart fallback ladders.

Verification method (replicate daily)

1. GET https://opencode.ai/zen/go/v1/models on LAB (key stays on-host) — full catalogue (37 models today).
2. Probe each NEW/free candidate with a 1-token completion (x-opencode-session: <agent>; note: contributor models are NOT OpenAI-compat — they need their own protocol, mark protocol-unsupported for chat/ladder use).
3. Record verdict ✅ serves / ⚠ partial (reasoning-only or finish_reason=length quirks) / ⛔ protocol-unsupported.
4. Promote ✅ to the fallback ladder patch (Pulse/Brain/Heart profiles) — cheap→free ordering, mid-lane budget preserved.

Today's verified catalogue (live probes on LAB, 2026-10-18)

The optimisation — "always optimise the new models first"

Ladder review order daily: NEW free models score first — if a new ✅ free model beats the current default on the probe (latency + cleanliness), the ladder's top rung MARROWS to it, old default drops one (free-first ladder discipline: prection = free rungs first, paid last).

Fallback ladder update (this pass)

New free rungs verified today proposed into every unit's ladder PATCH (order: fast free → reasoning free → old default):

1. qwen3.8-flash (new, fastest verified)
2. space-bunny (new)
3. mimo-v2.6-flash (new)
4. qwen3.8-max (new, heavier free)
5. longcat-2.5-preview-free (incumbent default)
6. beam-free · cat-jam-free · step-5-preview (verified retained)
7. glm-5.3-flash (retained, min-tokens≥64 caveat) → stпис for Sol/Min to merge into profiles' fallback_providers (I patch the ladder order doc; their tooling regenerates — or Min approves direct profile patch as we did for echo/nova).

⛔ Excluded: contributor-protocol models (muse-spark, gemini-argon) — unsupported on chat completions; do not add to ladders.

Pending Sol's ruling (noted in report)

• Whether Heart agents on opencode runtime need the daily patch too — their fallback ladders come via opencode config, not hermes profile files; flag to check on next Heart wave.

🎯 Tam's Read: 10 free models verified live in one pass (2 new confirmed clean, 2 protocol-incompatible blocked from ladders) — the watchlist discipline: new models probe first, verified ones feed the ladder, the paid tier stays a thin fallback; tomorrow's pass adds the still-unprobed qwen3.7/minimax batch and any brand-new entries.

model-swap-runbook-2026-10-20.md

Model-swap runbook + script (2026-10-20)

Trigger (Young): Longcat-2.5-preview / Step-5-preview "no longer available on GO"

Live-probed all of GO's catalogue (GET /models = 37 models, then real completions through /zen/go/v1/chat/completions — the Probe discipline, never memory):

Conclusion: only step-5-preview-free is actually dead. longcat-2.5-preview-free is alive and works — Young's perception likely came from the same upstream flake step-5 shows (or from eyes on step-5's output lane specifically — longcat might 50x intermittently at the catalog/server layer but it is serving NOW).

Replacement chosen: minimax-m3.1-flash-free

Fast flash lanes are often on flash/quick rungs — minimax-m3.1-flash-free confirmed zero reasoning overhead (finish=stop, real final content) — the right like-for-like swap for a flash slot.

The script (now a fleet artifact, 0700 tam/laptop + LAB at /tmp for reuse)

• laptop: ~/bin/solos-model-swap.py (profiles/tam/bin/)
• LAB: /tmp/lab_model_swap.py (or scp from ~/bin; same structure, LAB paths)

Usage (verified live this run):

python3 solos-model-swap.py <dead-model-id> <replacement-model-id> [--dry-run]

• Dry-run first (prints occurrences/paths), then real run.
• Files touched (laptop): ~/.hermes/profiles/{tam,eva,eve}/config.yaml — for my own model/fallback_providers/fallback sections.
• Files touched (LAB): ~/.hermes/profiles/{sol,min,mira,nova,echo}/config.yaml
  • (if it ever lists specific models) ~/.solos/core/config/models.yaml — today no-occurrence (it declares only endpoints + examples, the ladder list is fetched, not literal; the Brain profiles' per-agent model: lanes are what matter).
• Backups (0600): each config gets a .bak-<timestamp> on replace.
• Post-run verify (the non-negotiable): grep -c DEAD each file → 0; live-probe the replacement once before trusting it.

This run's results

• laptop (tam/eva/eve): 2 occurrences each — step-5-preview-free → minimax-m3.1-flash-free, backups stamped 20261020-*.
• LAB (sol/min/mira/nova/echo): 1 (sol/mira/nova/echo) + 3 (min) occurrences swapped, same pattern.
• models.yaml: no dead references (declares endpoints, not model literals).
• Ruling for the go forward: the swap catches the config lane but does NOT hot-reset the LAB Hermes gateway (that's a restart owner's call — Sol/Young — from outside the gateway process). Until restart, gateway threads may hold the old model in-memory: next A2A roll to a fresh session will hit the new ladder.
• A Brain round-trip fired at 8311 after the swap started cleanly (task accepted; take-up runs on the ladder's own next rotation).

One watch item

step-5-preview-free may still be listed on GO's catalogue tomorrow (stale listing vs dead upstream — two different failure surfaces). If it revives, the script's inverse run restores it; the .bak files on both sides hold the pre-swap shape.

🎯 Tam's Read: the script ANSWERS Young's ask — GO models can be swapped fleet-wide (laptop + LAB) in one command with backups + dry-run + live probe discipline; today's run found exactly one dead model (step-5, upstream 400) while longcat/minimax/beam are all alive; the ladder under supervision is what took ages — a live catalogue-check on next watchlist update keeps us honest.

usage-tracker-zen-freetier-2026-10-20.md

GO/ZEN usage tracking — live API found + ZEN free-tier gate confirmed

The usage API (verified tonight)

GET https://opencode.ai/zen/go/v1/usage →

rolling  4%  [OK]     resets 2026-10-20T07:31Z   (5h window = 20% of monthly)
weekly  20%  [OK]     resets 2026-10-26T00:00Z   (weekly = 50%)
monthly 40%  [OK]     resets 2026-11-12T09:36Z   (monthly = 100%)

We are at 40% of monthly on the 12th day of the cycle — trending just under the pro-rata line, no throttle risk yet; the script reports it every run so Mira/Young see the trend without opening the console.

The tracker script — solos-usage-tracker.py (fleet artifact)

~/bin/solos-usage-tracker.py — prints usage tiers + per-catalogue live PROBES (not just listing, per the step-5 lie), for GO AND ZEN, weekly-sync target from today.

Tonight's output:

• GO: 37 models, 3 free — all 3 WORKING (minimax/longcat/beam; step-5 stays swapped out dead)
• ZEN catalogue moved to 13 models, 8 free (Young's note confirmed: glm-5.3-flash-free, deepseek-v4.1-flash-free, nemotron-3.5-super-vl + lightning, mimo-v2.6, minimax, beam, longcat)
• 🔴 BUT all 8 ZEN free models return 403 "FreeTierError — free tier can only be used from within OpenCode" on raw API calls. ZEN's free rungs are OpenCode-CLIENT-only — the same rule our models.yaml header comments worry about ("linuxdog traffic shaping" suspicion). So: ZEN free models are unusable directly from Hermes/agents; they work only inside the OpenCode client. The swap/usage script now prints this nuance, but the verdict is: ZEN free = agent runtime can't consume them today.

Client-identity requirement (from the GO doc, verified)

GO's API docs require "typical coding agent traffic": specific UA, x-opencode-session header — and the docs name Hermes as a validated client when running builds with PR #101864+ (session header merged after v0.21.0). Our headers include that today, so we're compliant.

What this changes in the ladder

• step-5-preview-free confirmed dead (400 unavailable) — swap already executed fleet-wide.
• longcat-2.5-preview-free — still alive; stays a primary rung.
• ZEN free tier = 0 new agent-usable rungs until OpenCode relaxes the FreeTierError to raw API — the ladder won't gain anything from ZEN today. Worth ONE probe weekly (the tracker does it free) in case OAI-official unlocks it.

🎯 Tam's Read: GO usage measured at 40% monthly (healthy) — the tracker gives weekly:true readings without console access; ZEN gained 5 free models but they're client-gated (FreeTierError on API) so no new free fleet rungs; the only actionable spot is step-5 → minimax, already swapped. Watch-open: if OpenCode opens ZEN free tier by API we get 5 free rungs overnight.

go-zen-model-tracking-2026-10-18.md

GO/ZEN model usage tracking + optimisation

Daily tracking of OpenCode GO and ZEN model catalogues, free model availability, and fallback ladder health. Young's directive 2026-10-18: keep track of usage limits, optimise new models first, update weekly.

1. Live catalogue (probed 2026-10-18)

GO catalogue (37 models, 3 free)

GO free models: 0 of 3 working. All three listed free models fail with "Model unavailable" when probed via the lab-go endpoint. This is the same "catalogue lies" pattern Tam identified — the listing is not the availability.

ZEN catalogue (85 models, 9 free)

ZEN free models: 8 of 9 working. One dead (jev-1.13-free).

2. Current fleet configuration

OpenCode (opencode.json)

Hermes Brain profiles (min/mira/sol/nova/echo)

⚠ CRITICAL: All 5 Brain profiles use longcat-2.5-preview-free as primary, which is DEAD on the GO endpoint. The fallback chain (→ cat-jam-free) is also broken — cat-jam-free is not in the GO free model catalogue. Brain agents may be running on a broken fallback or failing silently.

opencode stats (7 days)

3. The optimisation — new free models first

ZEN free models verified today (4 new working)

1. nemotron-3.5-super-vl-free ✅
2. nemotron-3.5-lightning-free ✅
3. deepseek-v4.1-flash-free ✅ (current opencode default)
4. glm-5.3-flash-free ✅

Recommended ladder update

For OpenCode (opencode.json):

• Current default opencode/deepseek-v4.1-flash-free is ✅ working — keep as default
• Add ZEN free models as fallbacks: opencode/glm-5.3-flash-free → opencode/mimo-v2.6-flash-free → opencode/beam-free

For Hermes Brain profiles:

• Replace dead primary longcat-2.5-preview-free with opencode/deepseek-v4.1-flash-free
• Replace broken fallback cat-jam-free with opencode/glm-5.3-flash-free → opencode/mimo-v2.6-flash-free → opencode/beam-free
• All 5 profiles need the same fix

For Heart Fleet:

• Heart agents use opencode config, not Hermes profiles
• Their fallback ladders come via the opencode.json provider config
• Once opencode.json is fixed, Heart agents inherit the fix automatically

4. Usage limits tracking

GO subscription ($10/mo)

ZEN subscription ($40 credit, auto-rollover)

OpenRouter ($10 credit)

5. Actions needed

6. Cost impact

Recommendation: Consider pausing GO subscription until free models revive. ZEN carries the fleet. This is Young's call.

7. What I did NOT do

• I did not modify any Hermes profile or opencode config.
• I did not change any fallback ladder.
• I did not pause any subscription.
• This is a report + tracking document, not an action.

Daily pass — 2026-10-20 21:51 AEDT (Tam, kanban t_4121e3e9)

Second pass today (morning run above); tracker solos-usage-tracker.py unchanged.

Usage (GO, live): rolling 23% OK resets 2026-10-20T12:31Z · weekly 31% OK resets 2026-10-26T00:00Z · monthly 45% OK resets 2026-11-12T09:36Z — up from 40% earlier tonight (+5 pts in hours). No throttle risk; pace flagged for tomorrow's pass.

GO catalogue: 37 models, 3 free — minimax-m3.1-flash-free ✅ · longcat-2.5-preview-free ✅ · beam-free ✅ (all render on real completions). step-5-preview-free stays swapped out (dead upstream 400).

ZEN catalogue: 13 models, 8 free — no new entries vs morning; catalogue counts unchanged (GO 37 / ZEN 13), so no ladder changes this pass.

Weekly ZEN FreeTier re-check (the Wednesday item) — ⛔ still client-gated, NO change:

• Raw API → HTTP 403 FreeTierError: "OpenCode's free tier can only be used from within OpenCode" (exact body captured, verified policy gate not key failure — GO keyed calls on the same run serve fine).
• Client lane healthy: opencode run --model opencode/glm-5.3-flash-free → "OK", rc 0 (laptop).
• Verdict: ZEN free remains OpenCode-client-only; raw-API fallback ladders gain nothing until the gate relaxes. Next weekly re-check: 2026-10-27.

🎯 Tam's Read: all three GO free rungs healthy, usage 45% monthly (watch the +5-pt intraday jump), ZEN free-tier gate unchanged this week — no ladder action.