ltmoerdani · Oct 20, 2026 · Oct 20, 2026 · Oct 20, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -16,6 +16,10 @@ All notable changes to the **OpenCode Go BYOK Provider** extension are documente
 
 - **`[Models]` Gemini requests no longer send `temperature` to models that declare it unsupported.** The Google body builder sent `temperature` unconditionally even when models.dev marks the model `temperature: false` (`gemini-4-argon-contributor`) — the OpenAI and Anthropic builders already guarded this. Guarded the same way, and the builder now receives the model metadata it needs for both this and the thinking payload.
 
+- **`[Usage]` CLI history reader now reads the opencode CLI 2.x schema — Today shows 0 tokens fixed (issue #262).** The Go usage dashboard read per-message usage only from the CLI 1.x `message` table, which CLI ≥ 2.0 no longer populates, so "Today" showed 0 despite active CLI usage. The reader now probes `sqlite_master` and reads the CLI 2.x `session_message` table (role in the `type` column, provider/model under `$.model.*`, cwd joined from the parent `session_v2.directory`) alongside the legacy table, combined with `UNION` so rows the 2.x migration copied into both schemas are counted once. The issue's suggested `UNION ALL` against `session_v2` was rejected on real data: `session_v2` holds one AGGREGATED row per session (its Go total equals the per-message total), so that patch doubles every dollar/token; and its `providerID='opencode'` filter imports Zen free-model usage into the Go tracker (`opencode` is the Zen provider id — Go is `opencode-go`). A fresh CLI 2.x install without the legacy `message` table is covered too. Regression tests build real SQLite DBs (`src/test/cli2HistoryReader.test.ts`) and prove the dedup: 3 shared + 1 CLI2-only row → 4 rows at the exact total, never 2×.
+
+- **`[Provider]` Test Connection uses live models again (issue #261).** Both `testModelId`s pointed at models that no longer exist on the gateways — Zen's `deepseek-v4-flash-free` is dead upstream (issue #204, already in `KNOWN_UNAVAILABLE_MODEL_IDS`) so every Zen Test Connection failed with "Model is unavailable", and Go's `deepseek-v4-flash` was replaced by `v4.1` in the live catalog. Both now use `deepseek-v4.1-flash` (live on both gateways, cheapest paid tier — `max_tokens: 10` costs a fraction of a cent). A paid model is deliberate for the Zen test: the gateway answers `403 FreeTierError` to free-model requests from non-OpenCode clients regardless of headers or key validity, so a free test model can never distinguish a valid key from an invalid one. Note for #261 reporters: "OpenCode Zen missing from + Add Models" could not be reproduced from the contribution/registration side (vendor entries and `when` clauses verified correct, live Zen catalog returns 85 models incl. 10 free) — if Zen is still absent after this fix, run `OpenCode: Model Picker Diagnostics` and include the output.
+
 ## [0.7.9] — 2026-10-13
 
 ### Fixed

diff --git a/docs/issues/100-20261020-issue262-cli2-history-issue261-zen-test-model.md b/docs/issues/100-20261020-issue262-cli2-history-issue261-zen-test-model.md
@@ -0,0 +1,68 @@
+# Issue #262 (+ #261) — CLI 2.x usage history unreadable; Zen/Go Test Connection dead models
+
+**Status:** ✅ Solved (branch `fix/issue262-cli2-history-and-issue261-zen-test-model`)
+**Topic:** usage / provider / models
+**Updated:** 2026-10-20
+**Tags:** #usage #sqlite #cli2 #zen #test-connection #byok
+**GitHub Issues:** [#262](https://github.com/ltmoerdani/opencode-copilot-chat/issues/262), [#261](https://github.com/ltmoerdani/opencode-copilot-chat/issues/261)
+**Related:** [88-20260903-issue204-block-unavailable-zen-free-models.md](88-20260903-issue204-block-unavailable-zen-free-models.md) · repo memory `opencode-go-usage-accuracy.md`
+
+---
+
+## Problem
+
+### #262 — usage dashboard shows 0 tokens for today (opencode CLI ≥ 2.0)
+
+`src/usage/history.ts` read Go per-message usage only from the CLI 1.x `message` table. The CLI 2.x schema moved that data to `session_message` (and stopped writing `message`), so "Today"/"Yesterday"/the charts showed 0 despite active CLI usage.
+
+### #261 — "OpenCode Zen does not show in add models"
+
+Reporter gave no environment/repro detail. Investigation found two **verified broken** Zen surfaces adjacent to the report (both fixed here), but could NOT reproduce the missing "+ Add Models" entry itself.
+
+## Root Cause (evidence-based)
+
+### #262 — two schemas, and the reporter's patch was wrong twice
+
+Verified against a live `~/.local/share/opencode/opencode.db` that went through the 1.x → 2.x migration:
+
+| fact                                                                           | evidence                                                                                                                                                                  |
+| ------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `message` and `session_message` hold the same per-message rows after migration | 927/929 `message` ids exist in `session_message` with identical cost/tokens/modelId/`time.created` (the CLI's `v1-migration.bun.ts` copies legacy rows preserving ids)    |
+| `session_v2` is a per-session **aggregate**                                    | its Go total (57 rows, $2.671938) exactly equals the per-message total (191 rows, $2.671938)                                                                              |
+| reporter's `UNION ALL message + session_v2` **double-counts**                  | measured: $5.343876 = exactly 2× the true total                                                                                                                           |
+| reporter's `providerID='opencode'` filter imports **Zen** usage                | `opencode` = Zen free models (243 rows, 17.1M tokens); Go is `opencode-go` — confirmed by model-id sets and the CLI's own models.dev snapshot keys                        |
+| CLI 2.x column moves                                                           | role → `type` column; provider/model → `$.model.providerID`/`$.model.id`; `$.path.cwd` gone → joined from `session_v2.directory` (matches legacy cwd for all 191 Go rows) |
+| fresh CLI 2.x DB may lack `message`                                            | caught by the unit tests — the reader probes `sqlite_master` and falls back to a v2-only query                                                                            |
+
+### #261 — verified broken surfaces
+
+1. **Zen `testModelId` was `deepseek-v4-flash-free`** — dead upstream (issue #204, listed in our own `KNOWN_UNAVAILABLE_MODEL_IDS`), so **every** Zen "Test Connection" failed with "Model is unavailable" before validating anything.
+2. **Go `testModelId` was `deepseek-v4-flash`** — no longer in the live Go catalog (replaced by `deepseek-v4.1-flash`).
+3. **Zen free tier is client-gated**: live probes show `403 FreeTierError "OpenCode's free tier can only be used from within OpenCode"` for free models from any non-OpenCode client, regardless of `User-Agent` / `x-opencode-client` / key presence. So a free test model can never validate a credential.
+4. **Not reproducible**: the `opencodezen` vendor contribution, `when` clause (`config.opencodezen.enabled`, registered, default `true`), activation event, and registration in `extension.ts` are all symmetric with Go and correct; the live Zen `/models` returns 85 models (10 free). Most plausible remaining causes are user-side: "+ Add Models" clicked in the **Agents window** (agent vendors hidden by default via `showAgentModelsInManagePanel=false` + BYOK bridge settings) or an outdated VS Code.
+
+## Fix
+
+| file                                 | change                                                                                                                                                                                                                                                                                                                                                                             |
+| ------------------------------------ | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `src/usage/history.ts`               | Added `HISTORY_ROWS_SQL_V2` (CLI 2.x projection incl. `session_v2.directory` cwd join) and `HISTORY_ROWS_SQL_COMBINED` (`UNION` → dedups rows shared by both schemas). A `sqlite_master` probe (`parseHistorySchemasProbeRow` / `selectHistoryRowsSql`) picks legacy-only / combined / v2-only per database — used by BOTH readers (node:sqlite and the `sqlite3` binary fallback) |
+| `src/provider/definitions.ts`        | `testModelId` → `deepseek-v4.1-flash` for Go and Zen (live on both gateways, cheapest paid tier; `max_tokens: 10` ≈ negligible cost)                                                                                                                                                                                                                                               |
+| `src/test/cli2HistoryReader.test.ts` | new — builds real SQLite DBs via `node:sqlite` and proves: provider filtering, dedup (3 shared + 1 v2-only → 4 rows, exact total, never 2×), cwd join, and the fresh-2.x-no-legacy-table case                                                                                                                                                                                      |
+
+## Verification
+
+- `npm run lint` — 7/7 (editorconfig, eslint, markdown, prettier, shell, typecheck, tests incl. the new suite).
+- SQL variants validated against the real local `opencode.db` before implementation (combined UNION returns 191 rows / $2.671938 — identical to the per-message truth).
+- Live gateway probes: `deepseek-v4.1-flash` present on both `/zen/go/v1/models` and `/zen/v1/models`; old ids absent.
+
+## Lessons Learned
+
+1. **Read the actual schema before patching a usage query** — the reporter's locally-"verified" patch double-counted every dollar because `session_v2` is an aggregate table, not a per-message one.
+2. **Provider-id assumptions are dangerous**: `opencode` (Zen) vs `opencode-go` (Go) vs legacy `opencodego` alias — always verify against the model-id sets actually present per provider.
+3. **A missing table is a prepare-time error** — schema-optional SQL needs a probe statement, not a guard inside the SELECT. The unit tests caught the fresh-CLI-2.x-without-`message` case that the first implementation missed.
+4. **Connection-test models must be (a) live and (b) paid on Zen** — free-tier models 403 for any non-OpenCode client regardless of key validity, so they cannot validate credentials.
+
+## Open follow-up (not in this PR)
+
+- #261 needs reporter confirmation (VS Code version + `OpenCode: Model Picker Diagnostics` output) if Zen is still absent from "+ Add Models" after this fix.
+- Zen free-tier `403 FreeTierError` on real chat requests (not just the test) deserves its own investigation — it may require OpenCode-side allowance for this client.
diff --git a/src/provider/definitions.ts b/src/provider/definitions.ts
@@ -141,7 +141,10 @@ export const PROVIDERS: Record<ProviderDefinition["vendor"], ProviderDefinition>
     messagesUrl: goRoutes.messagesUrl,
     responsesUrl: goRoutes.responsesUrl,
     usageUrl: goRoutes.usageUrl,
-    testModelId: "deepseek-v4-flash",
+    // Issue #261/#262 batch: `deepseek-v4-flash` is gone from the live Go
+    // catalog (replaced by v4.1). Test Connection must use a live model or it
+    // fails with "Model is unavailable" before ever validating the key.
+    testModelId: "deepseek-v4.1-flash",
     fallbackModels: [
       // Curated active set from models.dev (2026-09-08) — deprecated/legacy
       // models are excluded, mirroring the live catalog's own filtering.
@@ -182,7 +185,13 @@ export const PROVIDERS: Record<ProviderDefinition["vendor"], ProviderDefinition>
     chatCompletionsUrl: zenRoutes.chatCompletionsUrl,
     messagesUrl: zenRoutes.messagesUrl,
     responsesUrl: zenRoutes.responsesUrl,
-    testModelId: "deepseek-v4-flash-free",
+    // Issue #261: `deepseek-v4-flash-free` is dead upstream (issue #204, in
+    // KNOWN_UNAVAILABLE_MODEL_IDS) — every Test Connection failed with
+    // "Model is unavailable". A live PAID model is used instead: the free
+    // tier also answers 403 FreeTierError to non-OpenCode clients, which
+    // cannot distinguish a valid key from an invalid one. v4.1-flash is the
+    // cheapest live model (~$0.30/1M in; max_tokens 10 ≈ negligible).
+    testModelId: "deepseek-v4.1-flash",
     fallbackModels: [
       "claude-opus-4-7",
       "claude-opus-4-6",