[viztor]

Phase 0 of the architecture review. It does three things:

• makes the suite deterministic: right engine, offline, independent of test order;
• fixes twelve confirmed defects, each with a test that fails without its fix;
• adds docs/improvement-plan.md, the reviewed plan for everything after this PR.

The suite

Before	After
Below Node 24, three cases failed for engine reasons (node:sqlite, Intl.DurationFormat, ICU strings) that looked like regressions	the run refuses to start below engines.node and names both versions; .node-version pins Node 26
engines promises Node 24, but no job ran it	a new CI job runs the unit suite shuffled, plus the mock suite, on Node 24
the "offline" unit suite made 54 requests per run to models.opencode.ai (the catalog's background refresh), so a networked runner could swap the catalog mid-test	test/support/offline.ts refuses every non-loopback request in the unit and mock suites; e2e keeps the network
shuffled runs failed on seeds 42 and 1234 (captured keys, session spend and the live catalog leaked between tests)	test/support/reset-process-state.ts resets all three after every test; pnpm run test:shuffle
the client-bundle boundary was guarded only by a size tripwire	architecture.test.ts walks the bundle's import graph and fails on node:*, cordis, pi or process.env (checked by planting a violation)
the host/client parity test grepped for a string literal, so a namespace passed as a constant slipped through	it resolves constants and checks every declared Remote method

Defects fixed

Every fix has a regression test, and each test was run against the old code and failed.

• Host crash risk. A rejected internal-route mount escaped as an unhandled rejection. mountInternalRoutes now owns the mount and reports the failure.
• Leaked registrations. Three early exits leaked the Gemini bridge registrations, so the next reload failed with DUPLICATE_ADAPTER. A mount that failed to start kept its fiber and its routes.
• Reloads ignored. Model discovery answered from the first activation's settings for the life of the process. Phase P3 deletes that registration anyway, because Settings never queries it.
• Pricing. Zen turns were priced at the Go rate. deepseek-v4-pro was billed at 0.66 instead of 1.74 per million input tokens.
• Foreign providers. The listing and model-info patches rewrote providers this plugin does not claim, such as a direct anthropic or openai route.
• Relay listing. A relay's Go /models was answered with the Zen catalog.
• Free models. The free-model check never read a Request body, which is how DSH sends a turn, so every keyless free turn was reported as a missing credential.
• Meter session. The meter's session fallback picked the newest session, not the one that just ran.
• Key tiers. A key's tier was read from its prefix. OpenCode issues oc_sk_ for Go too, so a working Go key was filed under Zen, and an OpenAI sk- key on a claimed relay was filed under Go.
• Store inspection. The store-inspection row called opencodeStore, which the client never mounted. The descriptor is now added, with a whitelist codec.
• Store reader. The credential reader's legacy-row fallback was not per integration, so one flagged Go row hid the Zen key.

Reviewer notes

• tierForRequest no longer takes the key. Three tests that pinned the prefix rule were rewritten to the rule docs/engineering-notes.md records as measured: the observed URL decides, then the route.
• CLIENT_BUDGET is raised from 128 to 132 KB for the store descriptor and its parser. That is about 1.2 KB of source, not a dependency, and is the gate's own "raise it" case.
• The no-credential warning no longer says "sign in through the settings card", which this build does not have.
• AGENTS.md no longer describes nine login modules that do not exist. CONTRIBUTING.md documents the new test rules.

The plan

docs/improvement-plan.md is the plan of record. A nine-lens review, with two independent skeptics per finding (78 of 82 upheld), is folded into it. It contains:

• the owner's decisions (D1–D22);
• what the DeepSeek Harness host supports, read from its source (F1–F20);
• the defects this audit found that are not fixed yet;
• the routing design, measured on the real host with pi 1.1.0;
• 20 phases, landing as stacked pull requests on this one, with a status table each phase updates.

The plugin owns opencode and opencode-go.

• An outermost internal/config filter keeps pi-ai from claiming the two routes, and the plugin registers its own adapter under the same ids. It never deletes another plugin's registration.
• That adapter serves every protocol per model, any catalog model, and any custom id.
• The user's file is never written. A read-only wrapper on settings.describe keeps the Models page showing both providers.

Decisions:

• Protocol. The catalog map decides it, never a guess at the model's name.
• Precedence. Catalog first: the user's row, then the catalog, then pi.
• Reasoning. A model whose reasoning is only on/off is offered as Off / On. Old Minimal/Xhigh/Max selections on such models break, and the release notes say so.
• Releases. Everything ships as 1.x minors. internalRoutes, resolveRoutedKey, three plugin settings and the deprecated @viztor/dsh-opencode wrapper go.

Before routing changes, two harnesses pin today's behaviour:

• a real-host test harness (a bare cordis/llm/llm-pi-ai tier, and a boot tier with ConfigEditor and Settings);
• a CI job that boots the published dsh offline, for both pi 1.1.0 and whatever npm installs.

Verification

All run locally:

Node	Step	Result
26	pnpm run check	0 errors; 23 warnings, unchanged from main
26	unit suite	1186 tests pass (1155 before)
26	pnpm run test:mock	51 tests pass
26	pnpm run test:shuffle	passes on 8+ seeds
26	build and scripts/check.ts	pass
26	coverage	96.4% lines, above the ratchet
24	test:shuffle and test:mock	pass

🤖 Generated with Claude Code

https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD

[viztor]

813e6f3 ports the nightly catalog refresh (#111) into this branch, so this diff now touches src/catalog-data.ts and src/go-limits-data.ts. Both files are exactly what the generator wrote; neither was edited by hand.

The vendor added Go kimi-k3.1-pro, which turned catalog freshness red on every branch of the stack. On main, #111's own test job is red as well, because three assertions and two docs hard-code the catalog's size. The commit updates those counts here:

• the Go shim test: 28 → 29;
• docs/redirect-planes.md: Go served 28 → 29, and the no-SDK Go cell 17 → 18;
• the Go completions list in both READMEs.

I merged the commit up through #106 → #107 → #108 → #109 → #110 → #112. On each branch, the unit suite has the same test count as before and passes.

The port becomes a no-op once main carries the refresh. If #105 merges first, #111 has nothing left to add and can be closed. P-T2 turns these counts into invariants, so later refreshes stop needing this.

---

Generated by Claude Code

[viztor]

2ebaffc ports today's regenerated vendor tables into this branch, the same way 813e6f3 did, so catalog freshness is green again. In the Go plane, gemini-4-argon-contributor became gemini-4-argon and is now priced, and iceman-contributor-free was added. The fixtures, counts and README lists that named the old id moved with it.

e2e · live gateway failed on the same commit because of the vendor's live state. Three cases failed:

• Two get 400 Upstream request failed: Model is unavailable for the free nemotron-3-ultra-free on completions.
• One gets a 429 rate limit from the free beam-free.

That is the gateway's own "model is unavailable / rate limited" answer for free models. The commit touches neither model; it changes only the generated tables, test fixtures and docs. The same suite passed on this branch's previous head.

I re-ran the failed job once to tell a transient outage from a free model being retired upstream. If it fails the same way, the suite's free-model cases need to pick a model the gateway still serves rather than a fixed id. That work belongs to the stack's later phases (P-T2 removes hard-coded vendor ids from the suites), and I'll report it there rather than widen this PR.

---

Generated by Claude Code