Skip to content

fix: report the gateway's refusals honestly - #175

Merged
viztor merged 256 commits into
mainfrom
claude/vigilant-mendel-bctz9n-p-r3
Nov 16, 2026
Merged

viztor merged 256 commits into
mainfrom
claude/vigilant-mendel-bctz9n-p-r3

Conversation

@viztor

@viztor viztor commented Nov 14, 2026 •

Copy link
Copy Markdown
Owner

Phase P-R3 of docs/improvement-plan.md, on main (P-T3, #173, is merged).

What changed

Item Change
L2: honest refusal codes New pure module src/utils/gateway-refusal.ts. It reads the gateway's typed refusal from pi's failure text and maps it to one of two codes. FREE_TIER covers FreeTierError and FreeUsageLimitError. MODEL_UNAVAILABLE covers RegionError, ModelDeprecated and ModelProtocolUnsupported. Neither code is AUTH, QUOTA or ACCOUNT_QUOTA, so the chat UI shows the gateway's own words instead of replacing them with "API key is invalid" or "Request quota exhausted.". Neither code is in the host's default retryable set. pi's message is kept unchanged, and status is read only from the head of the text.
Plane adapter PlaneAdapter.prepareCall and stream() wrap the prepared stream and rewrite the final finish failure. pi-ai's own routes and stepped-aside planes (D27) keep the bridge's codes.
L7: free-tier tool fallback src/tool-fallback.ts now treats a read or bash tool without non-empty properties as missing, and replaces it rather than adding a duplicate. A free chat-completions body that had no tools gets tool_choice: "none". Responses bodies are never touched.
Probe rows (recorded, never asserted) test/support/keyed-probes.ts adds four kinds of row: OC9 (a raw, non-ses_ session id, keyed and free), OC8 (session-only headers, keyed), OC6 (the harness User-Agent, keyed and free), and a free baseline. A keyless 429 from the usage cap now reads as free-usage-limited, not rate-limited, and is left out of the Retry-After record. They are formatted into the variants table in docs/redirect-planes.md §13.
OC10 Two remaining bullets in the README's paired section "With OpenCode Console installed".
L9, L10 No code. Run 43542273415 answered no keyed 429, so there was no Retry-After to act on. L10's reasoning_effort: "none" was never measured: the reasoning model was refused both with and without the field.
Docs and tests The plan, AGENTS.md, CONTRIBUTING.md, docs/testing.md and the CI comments now describe the above. New tests: test/gateway-refusal.test.ts, test/mock/plane-refusals.mock.ts, and new cases in the plane-provider, tool-fallback, keyed-probes and endpoint-record tests.

Behaviour changes

  • Refusal codes. A plane turn the gateway refuses on its own terms (free tier, usage cap, region, retired model, unsupported protocol) used to finish as AUTH, QUOTA, INVALID_REQUEST or PI_AI_ERROR, depending on how the bridge read the text. It now finishes as FREE_TIER or MODEL_UNAVAILABLE, and is not retried by default. The conversation shows the gateway's words and its status. Google-protocol models are the exception: their failure text carries no status, so none is shown.
  • Tool fallback. The free-tier fallback replaces a read or bash that has no properties instead of adding a second copy. It sends tool_choice: "none" on a free chat-completions turn that had no tools.
  • Probe record. The live keyed-probes record's empty Retry-After line now reads "Retry-After on a keyed 429: no keyed 429 was answered." (it was "no 429 was answered."). The old wording appeared to contradict keyless free rows that answer free-usage-limited (429) in the variants table.

Review

The chain ran three rounds, each with correctness, tests and rules lenses, and a verify pass after each fix round.

Round Found Upheld Fixed
1 24 22 22
2 10 10 10
3 6 5 5

Those rounds ran on the general model, because the dedicated reviewer was unavailable. The dedicated reviewer then re-reviewed the phase: round 1 upheld one finding, which is fixed, and round 2 was clean. Each merge of main (#177, #178, #181, #182 and #173) was reviewed for faithfulness, and found nothing.

Verification

At the current head (main through #173), on Node 26.12.1:

Step Result
pnpm run check 0 errors, 12 warnings
unit (alpha) 132 files; 2163 passed, 1 skipped
mock 9 files; 120 passed
rc channel: unit and mock same counts; switching back to alpha left no residue
coverage floors met
build + scripts/check.ts passed

CI is green on every check, catalog freshness included.

🤖 Generated with Claude Code

https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD

claude added 30 commits October 22, 2026 23:54
Replace docs/improvement-plan.md with the plan of record. It now holds:
- the owner's decisions (D1-D11);
- what the DeepSeek Harness host supports, with references into
  packages/llm, llm-pi-ai, session-controller, gateway and
  ui-settings-models;
- the defects this audit found that are not yet fixed;
- the target architecture and the phases, one PR each, with a status
  table that each PR updates.

The decisions that change the code:
- Settings offers the whole catalog, and the composer offers only the
  declared models.
- The plugin never writes the user's profile and follows the user's
  configuration at dispatch.
- The adapter patch exists to merge provider modes and nothing else. It
  is generalized to every mode, and its catalog fill is removed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
Investigation of the host's source and of the pi catalog it pins (1.1.0)
changed the routing design.

- llm-pi-ai resolves `api` and `baseUrl` per model, and pi's installed
  `opencode` catalog already spans Anthropic, Google, Completions and
  Responses with per-protocol endpoints. One route serves every mode for
  every model pi knows.
- The plugin fills only three gaps: models newer than pi's catalog
  (including all Gemini on Go), endpoint drift, and profile-level
  `api`/`baseURL` collapsing the modes.
- The new design is one unregistered catalog adapter per plane, keyed by
  the user's route id, and the adapter patch (merge-modes) delegates only
  gap models to it. That deletes the hidden routes, the listing hide,
  the stream redirect, the child mount and the Gemini bridge
  registration.
- A real-host test harness (P-T1) comes before any of it.
- One upstream change (per-entry `api`/`baseURL` in llm-pi-ai, U1)
  would remove the routing layer entirely. It would also let Settings
  save a model pi doesn't know, which pi-ai's strict write validation
  rejects today.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
pi-ai registers exactly the providers in its resolved config, and the
fiber keeps whatever the internal/config waterfall returns, while the
config editor writes the raw value. A prepended listener that resolves
pi-ai's config without opencode/opencode-go therefore keeps pi-ai from
claiming them, without touching the user's file, and the plugin registers
its own PiAiAdapter under the same ids through the public API.

Measured on the real host with pi 1.1.0 and the harness's patch: both
routes held by the plugin, pi-ai's other routes unchanged, every protocol
on both planes (Gemini on Go included) sent to its own endpoint, a
catalog-unknown model id saved, listed and dispatched, and both routes
handed back to pi-ai on release.

Replaces the upstream asks (no upstream changes) and the save-gate
design, which is kept only as the fallback.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
A nine-lens review of the routing design, with two independent skeptics
per finding (78 of 82 upheld), changed five things before any code moves:

- Settings reads pi-ai's resolved config, so a read-only wrapper on
  settings.describe puts the plane entries back into the displayed value;
  writes were never affected, the Models page edits by path.
- The route claim gets an explicit lifecycle: commits are told from
  validations by identity, reconcile runs on commit, mount and teardown
  are ordered, and pi-ai restarts at most once per transition.
- The plugin's adapter is built from public pieces only, with pi-ai's
  credential rule and no captured keys on the adapter path.
- The protocol comes from the catalog map, never from a guess at the
  model's name; metadata and levels are catalog first.
- Releases stay on 1.x, phases land as stacked pull requests, and a real
  host harness plus a binary e2e job pin the behaviour first.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
- The deprecated @viztor/dsh-opencode wrapper re-exported a `default` the
  barrel does not have, an ESM link error on every host load. It now
  forwards `export *` only, written by one shared helper that a test links
  against the built barrel in a real Node process.
- isZenProvider("opencode-google") answered false: "go" inside "google"
  read as the Go plane. It now uses the plane rule the catalog already has.
- enrichModelsResponse answered a 401, 403, 429 or 5xx with a 200 catalog.
  Those statuses now pass through; a 404 is still filled from the catalog,
  because the Go plane's inference base has no /models of its own.
- The warning colour named a token the host does not define, so warnings
  rendered as caption text. They now use --dsw-alias-state-warn-label.
- The settings spinner only animated while the meter was mounted; the
  settings surface now carries its own keyframes.
- The stream hook wrote debugFile with debug off. debug is the switch and
  debugFile only where it writes, the same gate the fetch patch uses.
- The purge row reported the host's own { ok: false } failure as "not a
  purge result" and cast the answer; it now shows the host's message and
  parses through the contract's whitelisting parser.
- The endpoint copy described modern as /inference/v1, which the plugin
  never sends. It now names the per-protocol inference family, and auto as
  its alias, which is what the code already did.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
- Delete adapter-replacement.ts and its test. Nothing imported it, and
  deleting another plugin's registry entry is not a host API; the
  engineering notes record why, and what the plugin does instead.
  A gate fails the build if src/ writes the llm registry map directly.
- Delete the stream hook's `prepared` handling: the host passes the
  prepared dispatch as its own argument and never reads options.prepared.
- Untrack .freebuff/ and ignore it.
- The settings partition test walks Config.dict, which is how debugFile
  went missing from it; a manifest test checks cordis.patch.yml against
  the schema and the field registry, and the manifest's phantom
  showUsagePrice, wrong marker text and incomplete lists are fixed.
- utils/retry.ts takes its timer and logger as parameters, and the utils
  purity gate now also rejects console and global timers, walking the
  folder recursively.
- tool-fallback reads free-tier ids from the request's own plane (Go iff
  opencode-go or opencode-go-*) and loses its unreachable Anthropic
  branch. A Zen-only free model is no longer treated as free on Go.
- test/support/src.ts is the one door tests use into src/, so later
  module moves touch one file; CONTRIBUTING.md states the rule.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
The barrel is the host's contract and nothing more: apply, Config,
inject, name and the config types. Everything else is internal, and
test/public-api.test.ts pins the list, so widening it is a reviewed
change. The 1.x line keeps its version scheme: nobody consumes the
removed exports (D6).

- Every test reaches src through test/support/src.ts, importing each
  name from the module that defines it; a gate rejects a static src
  import anywhere else outside its named exemptions.
- package.json no longer exports ./src/*, and check.ts fails when an
  exports target is outside the published files.
- resolveRoutedKey is deleted: it was exported and never called.
- The internalRoutes setting is deleted; the per-protocol routes always
  mount until the plugin owns the plane routes. A row that still carries
  it validates and the key is ignored.
- @viztor/dsh-opencode is no longer published. The deprecation message
  already on npm stays as its final word, and check.ts asserts that the
  release workflow verifies exactly what publish-scoped.ts publishes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
The test install now runs the host users run.

- pi is pinned to 1.1.0, the version the harness's own lockfile pins,
  with the harness's patch applied verbatim (patches/, hash recorded by
  pnpm). A pnpm override holds every copy in the tree to it, not just the
  direct devDependency.
- The @deepseek-ai packages move from 0.2.1-alpha.5 to 0.2.1-rc.2, the
  version a registry install of the published dsh resolves (cordis 4.0.5,
  loader 1.0.6, schemastery 3.18.5). The boot-tier packages the real-host
  harness needs next (dsh-llm, dsh-app-boot, dsh-config-editor,
  dsh-settings, cordis-plugin-loader) are added as devDependencies.
- test/support/host-versions.ts names those versions once.
  test/host-versions.test.ts checks that the pi the plugin loads at
  runtime is 1.1.0 and the same copy the tests import, that the patch is
  applied, and that the installed cluster matches.
- mount-protocols now checks that every catalog SDK maps to a protocol
  pi's exported api modules implement, and records as a characterization
  row that pi 1.1.0's opencode-go provider has no Google API.

No src/ behaviour changes: the rc.2 builds of llm-pi-ai, cordis and
schemastery are code-identical to alpha.5.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
Add an in-process harness that runs the plugin inside the real host,
so the routing work that follows can be measured against today's
behaviour rather than against fakes. Additive only: no src/ change.

test/support/host.ts gives two tiers. Tier A is a bare cordis Context
with LlmRuntime and llm-pi-ai mounted with { providers }, for the
config waterfall and pi-ai's credential paths. Tier B is dsh-app-boot's
boot() with a temporary HOME, a test bundle, a fake credentials service
and a launch-environment snapshot, for the config editor,
settings.describe and the volatile commit. Tier B's rows carry package
names through a stub module loader, and it fixes the apply order,
because the loader starts rows in parallel.

On top of it:

- contract tests for the waterfall, the profile and Config shapes,
  credentials, and the config editor (host-waterfall, host-shapes,
  host-credentials, host-editor-contracts), plus a tier A smoke test;
- a characterization of today's behaviour on pi 1.1.0, each row tagged
  where P-R1 is expected to flip it, with S0_TODAY shared by the
  binary e2e;
- the keyless 401-vs-404 matrix: a pure classifier, its unit test and
  a gated live run;
- a binary e2e gated on DSH_HOST_E2E=1 with its probe bundle, the
  e2e-host CI job (a pinned, patched cell on PRs that touch the host
  seams or vite.config.ts, and a floating cell nightly and on
  dispatch), and the test:e2e:host script.

The binary e2e now reports why a run failed: a timeout only when its
own deadline killed the host, the signal for a crash, and the probe's
fatal message whenever there is one.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
Add src/host/patch-method.ts (patchMethod), now the only way the
plugin wraps a method on a host object, and move the four instance
patch sites in src/models-discovery.ts onto it: the three listings
hidden by hideResponsesRoute, decorateModelDiscovery,
decorateModelInfo and decoratePlaneAdapters.

On the real host the three llm-level disposers never restored:
ctx.llm is cordis's traceable proxy, every method read returns a new
shadow proxy, so the `===` identity checks were never true and each
reload stacked a wrapper over the last. The helper reads identity
from own descriptors instead. A dispose restores only while its
wrapper is still current (otherwise it becomes a pass-through),
deletes the own property when the original was inherited, and a
same-site reinstall wraps the unpatched method rather than the
earlier install, so either disposal order ends at the exact
original.

decoratePlaneAdapters becomes one patch per (adapter, method) that
composes the adapter's plane routes in registry order, which keeps
today's nesting on pi-ai's shared adapter. rewrapOnAdaptersUpdated is
built and tested but wired to no site yet.

test/reload.test.ts runs apply A, apply B, dispose A, dispose B (and
the reverse order) for every site on own and inherited originals, on
the real host (tier A) and through apply(); test/patch-method.test.ts
covers the helper's contract. The existing per-site pins in
test/models-discovery.test.ts stay unchanged; five new cases there
record the deliberate edge changes on hosts that refuse the patch.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
The vendor added kimi-k3.1-pro to the Go plane, so `catalog freshness`
went red on every branch of this stack. This takes the two generated
files exactly as the nightly refresh produced them (#111), with no
hand edits, and updates the counts that hard-code the catalog's size:
the Go shim test (28 -> 29), the plane table in redirect-planes.md
(Go served 28 -> 29, the no-SDK Go cell 17 -> 18), and the Go
completions list in both READMEs. The listed half of "29 served / 33
listed" is models.dev's count, inferred from the one added entry,
because models.dev cannot be reached from here.

It no-ops once main carries #111. P-T2 turns these counts into
invariants so a refresh stops needing this.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
The plugin now serves the opencode and opencode-go routes itself,
through supported host paths only, and the hidden-route workaround is
gone.

A prepended internal/config listener hands pi-ai a copy of its config
without the two plane entries, so pi-ai keeps every other route. On a
commit the plugin reconciles: it builds one plane adapter, a subclass
of the host's PiAiAdapter, over pure profiles (utils/plane-profile,
plane-model, model-metadata and protocols) and registers only the
planes that are declared. A Settings candidate has the plane entries
it changes validated strictly. A read-only settings.describe graft
puts the committed plane entries back, so the host still sees them.

Deleted: responses-provider, responses-routes, google-bridge,
pi-provider and utils/google-replay; the listing hide and the
discovery decoration; the stream redirect; the D19 settings; and the
hidden ids in DEFAULT_PROVIDERS and cordis.patch.yml.

What users see: any saved id is accepted and served, and an id no
catalog knows dispatches as completions. The protocol comes from the
catalog and model info follows the user's own row first. A model
whose reasoning is only on or off offers Off/On. The default endpoint
is the per-protocol inference family. Keys resolve by pi-ai's own
rule, and a plane lists its declared ids even when no key resolves.

When several pi-ai rows declare planes, the row the plugin serves
stays the same while it is live and still declares one, so no row is
left stripped with nobody serving it. If the host's pi-ai module
cannot be read, the plugin steps aside even before any plane is
declared, so a later Settings save goes to pi-ai's own check.

The characterization rows this changes are flipped to new baselines;
every other row stays green unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
# Conflicts:
#	README.md
#	README.zh-CN.md
#	docs/redirect-planes.md
A measurement phase: the default endpoint family stays `modern`, so no
plugin behaviour changes and no characterization row flips.

- Name the default in one constant, AUTO_ENDPOINT_FAMILY in
  src/utils/plane-model.ts, which is what `auto` resolves to, and pin
  it in test/endpoint-family.test.ts.
- Extend the keyless matrix to three URL shapes (legacy, root, modern)
  over four protocols and two planes, add LEGACY_DISPATCH, and add one
  keyless free-model probe with its own pure classifier.
- Add minimal keyed probes (pure test/support/keyed-probes.ts and its
  e2e): one tiny request per plane, protocol and shape, sent with the
  plane's own key and the plugin's identity from one shared constant,
  PROBE_IDENTITY. Only the cells the default sends to are asserted;
  the rest are recorded.
- Gate the live job by tier with OPENCODE_E2E_KEYED (none, probes,
  full; unset means full). CI arms `probes` on a pull request that
  touches what the probes send, `full` nightly and on dispatch, and
  passes the secrets only when the tier is not `none`.
- Write a dated endpoint-family record from both live suites to the
  run summary and an `endpoint-families` artifact, and add §13 to
  docs/redirect-planes.md with every cell pending the first CI run.

The live part runs in CI only; no live result is invented here. A
failing keyed modern cell in CI is the evidence for a later `fix:`
that flips the default.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
For the opencode and opencode-go routes, the Settings model list now
shows pi-ai's own rows first, in pi's order, and then every model the
plane's catalog can serve that pi does not list and the catalog has
not retired. Ids that pi's registry has retired are dropped, a field a
row lacks is filled from the catalog, and every row is cut down to the
five fields the host's discovery contract accepts, with absent values
omitted rather than set to undefined.

While pi-ai restarts and its discovery is withdrawn (NO_DISCOVERY),
the list answers the catalog rows alone instead of the error.

This rewrites the existing discoverModels wrapper rather than adding a
second one. Other routes keep today's enrich-only merge, because pi-ai
serves them and cannot serve an id it does not list. The plugin's own
model-discovery namespaces, which Settings never asked, are deleted
with their lifecycle tests, and RouteClaim gains a claiming() seam.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
Phase P-T2 of docs/improvement-plan.md. Test-only: no source file
changes and no behaviour changes.

- Every socket-bound case is in the mock suite. The nine unit files
  that booted the mock gateway through testing-kit.ts are merged into
  test/mock/request-shapes.mock.ts and a new protocol-endpoints.mock.ts
  table. The two offline e2e files that bind a port
  (patched-fetch-headers, refusal-classification) moved there too. The
  device-login routes left the gateway mock.
- createMockContext, HarnessSimulator and the rest of testing-kit.ts
  are gone. Services are built on a real cordis Context, and
  lifecycle.test.ts mounts the plugin the way the Loader does
  (test/support/composition.ts). degradedHost remains only for the
  degraded-composition diagnostics.
- 43 cases that tested only the harness or the mock gateway are
  deleted. 33 duplicated cases are merged and 16 restatements are
  deleted; the plan names each one.
- Hard-coded catalog counts and vendor ids are now invariants, or a
  row picked by shape, so a vendor catalog refresh no longer breaks
  the suite by construction.
- The seven largest test files are split by subject, the live suites
  read their environment through test/support/e2e-env.ts, and the
  duplicated helpers and hard-coded ses_ ids are removed.
- src/utils/session-id.ts, src/opencode-store-contract.ts and the e2e
  environment reader have their own tests, and src/lifecycle.ts,
  src/stream-hook.ts and src/fetch-patch.ts have per-file coverage
  floors.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
The vendor renamed the Go plane's one Gemini model (dropping its
contributor suffix, halving its price and publishing an allowance for
it where the limits table had none) and added a free model with no
SDK, so `catalog freshness` went red on every branch of this stack.
This takes the two generated files exactly as the nightly refresh
produced them (#111), with no hand edits, and follows them:

- the Go Gemini fixtures in six test files (attachment-services,
  catalog, go-google, go-google-prepare, models-discovery,
  responses-routes) now name the renamed id, since each means "the
  Go Gemini model the catalog serves";
- the usage panel test built its unknown allowance from the table's
  one unpriced row, which the vendor has now priced, so it builds the
  `Number.NaN` allowance the generator emits itself;
- the Go shim count in catalog.test.ts (29 -> 30);
- redirect-planes.md: Go 29 -> 30 served, the no-SDK Go cell
  18 -> 19, and "0 of 29" -> "0 of 30" in the prose;
- both READMEs: the renamed Gemini id, and the new model in the Go
  completions list, since it declares no SDK.

The listed half of "30 served / 34 listed" is models.dev's count,
inferred from one rename (no change) and one addition (+1), because
models.dev cannot be reached from here. The dated S0 measurement row
in improvement-plan.md keeps the id it was measured with.

It no-ops once main carries #111.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
The live routing suite picks its free-tier subjects from the catalog by
asking the gateway, but it counted "400 Upstream request failed: Model
is unavailable" as an observable answer. The upstream withdrew
nemotron-3-ultra-free while the catalog still lists it, so the picker
chose it and two cases asserted on a model the gateway no longer
serves; the reachability case also failed on a 429 throttle.

A withdrawn model (that 400, or a 410) is now skipped like a throttled
one. The free-tier routing case only accepts a candidate whose answer
is the free-tier gate it asserts, so an unmetered model cannot become
its subject, and the reachability walk moves past gated, withdrawn and
throttled candidates before it reports.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
The review ran three rounds through the correctness, rules and matrix
lenses. What it found, and what changed:

- The binary e2e's beforeAll budget was smaller than the blocking
  subprocess deadlines it serializes. Each plugin install now has its
  own deadline, the hook's budget is derived from the installs and the
  boot, and afterAll stops a host a failed hook left running.
- providerWithApiKey claimed to be a line-for-line restatement but
  skipped upstream's credentialRef(ref) check. It now throws the same
  TypeError for a ref outside the grammar (a host-editor-contracts
  case pins it), and its comment names the deliberate differences.
- The pinned e2e-host cell let cordis and the loader float at patch
  level. The overrides now force both, the cache key includes them,
  and the job asserts them against HOST_VERSIONS.
- The pi-ai-first characterization asserted every facet under one
  KNOWN_DEFECT title. Each apply order now generates one test per row
  and facet, titled with that order's own tags.
- CONTRIBUTING said the contract tests pin F11-F20, and the plan said
  P-T1 pins every fact. Both now give the same accounting: which facts
  are pinned by number, which P-R1 exercises first, and why F9 and F17
  are not pinned.
- CONTRIBUTING put the e2e's dummy keys in the profile's credentials
  file; they are in the temporary home's credentials document.
- F5 had no real-host pin: host-shapes now pins that a route-level api
  and baseURL collapse pi's protocols onto one. F6 is recorded as first
  exercised by P-R1.
- The credential-mapping row pinned only the api-key branch of
  toPiCredential. host-credentials now covers all three branches, and
  the env branch is asserted through pi's Models.getAuth, so the case
  fails without it.
- The plan's reason for restating hasProviderApiKey was inaccurate; it
  now says no installed copy reaches the function.
- The new credential cases cited toPiCredential at the wrong lines;
  they now cite auth.ts:74-84.
- F4's listing and dispatch halves were pinned only by KNOWN_DEFECT
  rows P-R1 flips. A plugin-less tier-A case in host-shapes now pins
  that an id pi cannot place is not listed and that a turn on it is
  INVALID_CONFIG with no request; the save and deferred-read halves
  are tagged F4 too.
- The one facet that depends on the apply order lost its tag when the
  pi-ai-first block was split. The fixture case now checks only
  structure (the facets whose value moves are the facets whose tag
  moves), and the exact pin is its own KNOWN_DEFECT case that boots no
  host. CONTRIBUTING and the plan say P-R1 drops todayRowsFor's
  override and turns that case into a BASELINE check in the same diff.

Nothing under src/ changed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
…claude/vigilant-mendel-bctz9n-p3

Brings in the nightly catalog refresh (#171: Zen 87 -> 88, adds
gemini-4-hydrogen; Go unchanged at 32) as P-R2 carries it, with P-R1's
follow-up of the docs and tests.

Resolution: no textual conflicts. git auto-merged README.md,
README.zh-CN.md, docs/protocol-routing-and-merge.md,
docs/redirect-planes.md and test/catalog.test.ts, where P3's own edits
and the refresh's count and list edits touch different lines. The merged
tree is exactly P3's own change (d563282..8b2fa44) applied on top of
65b8443: the two diffs differ only in blob ids and in the one context
line of docs/protocol-routing-and-merge.md that carries the Zen count
(87 -> 88).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
…aude/vigilant-mendel-bctz9n-p-t2

Brings the nightly catalog refresh (#171: Zen 87 -> 88, adds
gemini-4-hydrogen on @ai-sdk/google; Go unchanged at 32) through P3 into
P-T2's test layout.

Resolution:

- test/catalog.test.ts (the one conflict): P3 re-pinned the Zen shim
  length from 87 to 88; P-T2 had replaced that pinned count with an
  invariant (non-empty, ids unique), and the per-SDK counts are compared
  with the docs by catalog-docs.test.ts. Kept P-T2's invariant; the file
  is P-T2's text unchanged.
- docs/protocol-routing-and-merge.md, docs/redirect-planes.md: merged
  cleanly. P3's count refresh (88 Zen, Google 10, 2026-11-13, #171) sits
  beside P-T2's renamed guard references (catalog-docs.test.ts,
  catalog-routing.test.ts, the host-planes*.test.ts files).
- README.md, README.zh-CN.md, src/catalog-data.ts: P3's side only (the
  generated file and both READMEs in step).
- test/thinking-levels.test.ts: P3's side only (the gemini-4-hydrogen
  xhigh exception); P-T2 did not move or touch this file, so it stays
  where it is in P-T2's layout.

The parent added no new test titles (git diff 8b2fa44 285aeaf -- test/
adds no it/test/describe line), so nothing needed relocating.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
Phase P-T3 of docs/improvement-plan.md. CI now runs the unit suite once
per install (the coverage run, in the required check job), builds lib/
before any test that reads it, type-checks test/**/*.tsx, and sets up
every job through one composite action (checkout, pnpm, node, install).
The catalog job's fork condition and the stale pnpm comments are fixed,
and the paid keyed e2e runs only on a schedule or a dispatch, apart from
the minimal keyed probes on pull requests that touch src/host/**.

scripts/check.ts keeps the lib freshness check (once), the consumer
install (now loading the client half too) and the size tripwire; its
pure file checks moved to vitest. The generator scripts share helpers
and one staleness rule, every verify-* script exits non-zero on a
mismatch, scripts/capture-ui.sh carries no machine paths, and the
no-local-paths test also covers .github/ and *.sh files.

The CLI-version script reads the registry's dist-tags and the 2.x
User-Agent format, and the nightly host-drift report now proposes a
bump pull request from a tokenless planning step; a test pins every
build approval the pinned pnpm acts on so that step cannot run a
dependency's install script unseen.

The plan gains a "Lessons from @opencode/console" section (OC1-OC13,
each with its verdict, evidence and phase).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
Brings in #172 (the live e2e job signs in with the runner's OIDC
identity). Merged against 9be88d6, the tree this branch was cut from;
ci.yml keeps the composite setup action and adds the sign-in step
after it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
The chat UI replaces an AUTH failure with "API key is invalid" and a
QUOTA one with "Request quota exhausted.", so a plane turn the gateway
refused for its own reasons (free tier, usage cap, region, retired
model, unsupported protocol) read as a bad key or an empty quota. A new
pure module, src/utils/gateway-refusal.ts, reads the gateway's typed
refusal from pi's failure text and finishes the turn as FREE_TIER or
MODEL_UNAVAILABLE instead: neither code is replaced by the UI and
neither is retried by default, so the user sees the gateway's own
words and, except on Google-protocol models, its status. The plane
adapter applies it in prepareCall and stream; pi-ai's own routes and
stepped-aside planes keep the bridge's codes.

The free-tier tool fallback now treats a read or bash tool without
properties as missing, replacing it rather than adding a duplicate,
and sends tool_choice "none" on a free chat-completions body that
carried no tools.

The keyed probes gain recorded rows for a raw session id, session-only
headers and the harness User-Agent, plus a free baseline; a keyless
usage-cap 429 now reads as free-usage-limited and is left out of the
Retry-After record. The README gains the remaining bullets of its
"With OpenCode Console installed" section. The rate-limit and
reasoning-effort lessons land no code: the recorded run answered no
keyed 429 and never measured the field.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
Takes #174 (the Console client id from a repository variable), #167 (the
Go limits page read by its tab labels) and #160 (pnpm/action-setup 6.1.1).
The workflows keep the composite setup action, which now carries the 6.1.1
pin, the one site that holds it; generate-go-limits.ts reads its sources
through #167's firstDoc/docFrom/chooseTables inside generator-kit's
arguments, staleness rule and outcome, with go-limits-render's renderer;
AGENTS.md and docs/testing.md carry both sides' text.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
…s it was

withReadmePolicy rewrote one field at a time over text earlier fields had
already edited, so a field's new value could be read as a later field's old
one: a bridge raising its pi floor to the pi the record holds while pi moves
on wrote the new pi where the floor belongs, and named nothing for review.
Every plan's replacements are now decided against the original paragraphs
and applied in one pass; two fields that would rewrite one value apart are
both left for review.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
…ep guards

A pnpmfile's updateConfig hook can set the allow-all, and the pinned pnpm
loads one from the root (.pnpmfile.cjs or .pnpmfile.mjs) or wherever the
workspace's pnpmfile names, so adding one turned every dependency build on
in the plan step's installs while the guard stayed green. The guard now
fails on a root pnpmfile, the workspace keys that load one (pnpmfile,
globalPnpmfile, configDependencies), .npmrc's and a --pnpmfile flag, each a
known answer; ci.yml, CONTRIBUTING, AGENTS.md and the plan name it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
…verify-session-band failing

The spawned catalog-generator and verify-catalog cases fed only made-up ids,
which the old rules called stale too, and neither script had a spawned
current answer. A payload that reproduces the shipped shim (the inverse of
the parser, held by the pure case) now runs both: current, then the same
ids with only the Go plane's endpoint or one Zen family endpoint moved,
stale, and each mutant that drops the endpoints from the comparison fails.

verify-session-band was spawned only in its passing state; a scratch copy
of it beside a src/session.ts whose derivation leaves the band now prints
the MISMATCH line, no CONFIRMED, and exits 1, which fails once the script
stops setting its exit code. The docs drop the "no seam" and "no payload
reproduces the shim" reasons, which did not hold.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
…request

The parity plan still credited scripts/check.ts with keeping lib/client.js
free of host-only imports, which it never scanned; the import-graph walk in
test/architecture.test.ts is what enforces it, as AGENTS.md and
docs/credentials.md already say. Its coverage floors, four numbers behind
the ratchet, now point at vite.config.ts. The plan's phase table names #173.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
…ings

The tokenless plan step's build-approval guard named a root pnpmfile and
three global spellings that no case exercised: the known answer handed
buildApprovals a ready-made pnpmfile list, and it carried none of
globalPnpmfile, .npmrc's global-pnpmfile or a --global-pnpmfile flag, so
emptying the probe or dropping a spelling left the file green.

The probe is now pnpmfilesIn(root), read for the repository and for a
scratch root (empty, one file planted, both planted, a missing directory),
and the known answer carries each global spelling with its expected line.
Emptying PNPMFILES, either global key or the regex's global group now
turns the file red.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD

# Conflicts:
#	docs/improvement-plan.md
The keyless-label assertion read only the first row picked by its own
"a free model" prefix, so it held by construction and tied nothing to
`keyed: false`. Walk every printed variant row instead: it contains
"no key" exactly when its variant is keyless, and the keyless rows are
the known six (three free variants on each plane).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
No file conflicted. Resolutions, file by file:

- test/freshness-scope.test.ts: auto-merged; main's new es-module-lexer 3
  case ("reads a file off every loading form...") sits beside this
  branch's rewritten cases, both kept.
- scripts/freshness-scope.ts: main's es-module-lexer 3 port taken as is
  (this branch does not touch the file).
- package.json, pnpm-lock.yaml, channels/rc/pnpm-lock.yaml,
  pnpm-workspace.yaml: main's, unchanged (#177).
- src/catalog-data.ts, src/go-limits-data.ts: main's generated tables,
  unchanged (#178).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
Brings in p-t3's merge of main: the #177 toolchain bump (vite-plus 1.5.0
with vitest 5.1.0 and the dev-dependency group, the es-module-lexer 3
port of scripts/freshness-scope.ts and test/freshness-scope.test.ts)
and the #178 refresh of the generated catalog and limits tables.

No file conflicted: p-r3's own change touches none of the files p-t3
changed, so every file merged automatically.

- package.json, pnpm-lock.yaml, channels/rc/pnpm-lock.yaml,
  pnpm-workspace.yaml: taken from p-t3 (main's #177) unchanged; p-r3
  does not touch them.
- scripts/freshness-scope.ts, test/freshness-scope.test.ts: taken from
  p-t3 unchanged (#177's es-module-lexer 3 port).
- src/catalog-data.ts, src/go-limits-data.ts: taken from p-t3 (main's
  #178) exactly; generated, never hand-edited.

The merged tree is p-r3's own change (its diff against 9911b30) on top
of the new parent.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD

viztor commented Nov 16, 2026

Copy link
Copy Markdown
Owner Author

This phase's earlier reviews ran without the dedicated reviewer, so it has now had a second review on that reviewer. The branch moved fast-forward to 0f97615.

Re-review:

  • Round 1: 1 finding, upheld and fixed in 46175e3. The "a keyless row says so in its label" assertion checked a label constant instead of the keyless switch. Each variant row's label is now tied to its keyed switch.
  • Round 2: nothing found.
  • The round-3 doc items left unreviewed earlier were confirmed fixed: the Retry-After and 429 wording, the variant counts, and the 403 comment.

Merges: the re-reviewed P-T3 (ef74a26, which carries main's #177 and #178) is merged in, with no conflicts. P-R3's own change is byte-identical to before.

Gate:

  • oxlint 1.91: 0 errors and 12 warnings.
  • Unit suite: 2155 passed, 1 skipped, on both channels, plain and shuffled.
  • Mock suite: 120 passed. Coverage floors met. check.ts passed.
  • Binary e2e in both npm cells: 15 passed and the 2 HOST_LIMITATION skips each.

Generated by Claude Code

github-actions Bot and others added 2 commits November 16, 2026 13:04
The nightly refresh (#182) added hy4-flash-free to both planes (Zen 88 -> 89, Go 32 -> 34 with qwen4-max on @ai-sdk/anthropic) and qwen4-max to the Go limits (29 -> 30). test/catalog-docs.test.ts held five doc counts to the old catalog; the docs follow it:

- README.md, README.zh-CN.md: the per-SDK routing table's no-SDK row 27 -> 28, the stated total 88 -> 89 (twice) and its date 2026-11-16; the Go plane's Chat Completions list gains hy4-flash-free and its Anthropic Messages list gains qwen4-max, so the lists stay the whole Go catalog. Kept in step for the docs-pairing gate.
- docs/protocol-routing-and-merge.md: every "N of 88" becomes "of 89", the no-SDK row 27 -> 28, the count date 2026-11-16, and the refresh history names #182.
- docs/redirect-planes.md: the served halves 88 -> 89 (Zen) and 32 -> 34 (Go); the shapes table's no-SDK row 27/21 -> 28/22 and Go Anthropic 5 -> 6; "0 of 89 on Zen, 0 of 34 on Go"; the Go non-completions sentence 11 -> 12; both count dates and histories name #182. The listed halves stay the ones #156 recorded, since the registry was not re-read.

No test pins a count any more (the catalog tests assert invariants), the Go limits count appears in no document, and no fixture names a moved row, so no test changed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
The catalog refresh added hy4-flash-free and qwen4-max to Go, and the
newest host cell's pi lacks both, so Settings' fetch on Go now appends
them there too. Three test comments still described an empty Go gap on
the newest cell; they now say what holds on every refresh. The phase
record in docs/improvement-plan.md keeps its statement, marked as true
at the time.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
The parenthetical beside the per-plane gap assertion said the newest
cell's pi lacks the catalog's newest ids. Nothing asserts that, and it
moves with pi's releases, so it now says only that which ids a cell's pi
lacks moves with the catalog and with pi.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
…claude/vigilant-mendel-bctz9n-p-t3

No file conflicted. Twelve files came in from the parent; P-T3 had touched
two of them, in hunks git merged on its own:

- docs/redirect-planes.md: P-T3 rewrote the section 7 paragraph on the
  OPENCODE_E2E_KEYED tiers (a dispatch's `keyed` input, the composite setup
  action); the refresh moved the section 1 and 2 counts (Zen 89, Go 34, the
  no-SDK row 28/22, Go Anthropic 6, Go non-completions 12) and their #182
  stamp. Both kept: P-T3's paragraph, the parent's counts.
- docs/improvement-plan.md: P-T3 rewrote the phase table and its own P-T3
  entries; the refresh reworded one P-PD1 restack bullet (the newest cell's
  pi then listed every Go catalog model). Both kept.

Taken from the parent unchanged, P-T3 did not touch them: README.md,
README.zh-CN.md, docs/protocol-routing-and-merge.md, pnpm-workspace.yaml
(vite-plus and vite-plus-core 1.6.0), pnpm-lock.yaml and
channels/rc/pnpm-lock.yaml, src/catalog-data.ts and src/go-limits-data.ts
(generated), test/host-discovery.test.ts, test/host-planes-foreign.test.ts.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
…to claude/vigilant-mendel-bctz9n-p-r3

Brings in the parent's merge of main's toolchain bump (#181: vite-plus
and vite-plus-core 1.6.0, oxlint 1.93, oxfmt 0.78, both lockfiles
relocked, two test helpers moved to .then().catch()) and the nightly
refresh (#182: Zen 89 and Go 34 catalog models with hy4-flash-free and
qwen4-max, Go limits 30), followed through the README pair, the
protocol, redirect-planes and improvement-plan docs and two host tests.

No conflicts. README.md, README.zh-CN.md, docs/improvement-plan.md and
docs/redirect-planes.md were changed on both sides and merged cleanly in
separate hunks: this branch's refusal reporting rows and plan entries
beside the parent's catalog counts. The merged tree is this branch's own
change (ef74a26..0f97615) applied unchanged on top of the parent
(f36e849); the lockfiles and pnpm-workspace.yaml are the parent's.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
History only: main's tree equals the p-t3 tree (f36e849) this branch already contains.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
@viztor
viztor changed the base branch from claude/vigilant-mendel-bctz9n-p-t3 to main November 16, 2026 17:43
@viztor
viztor merged commit 630da90 into main Nov 16, 2026
9 checks passed
viztor pushed a commit that referenced this pull request Nov 16, 2026
History only: main's tree equals the p-r3 tree (71316be) this branch already contains.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0141fKp5k6hiaJcP5ovhGtSD
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants