fix: batch 216-223 — model-list cache, output channel, tool-call pairing, zero-part parsing, 429 retry, session headers - #224
Merged
Conversation
VS Code polls provideLanguageModelChatInformation every few hundred ms; ModelListFetcher.fetch() performed a live GET /models on every poll because MODEL_LIST_CACHE_TTL_MS was only consulted on the failure path. - consult the fresh cached snapshot first; stale snapshots fall through to the live fetch + retry path - add ModelListFetcher.invalidate() and wire it into Refresh Models so a manual refresh still performs a real upstream fetch - log the 'Models registered' summary only when its signature changes
prepareChatRequest() called vscode.window.createOutputChannel("OpenCode")
once per request; every call registers a new Output-tab entry, so long
sessions accumulated dozens of duplicate channels and leaked the old ones.
- drop the per-request channel from chatPrep; transports now receive the
provider's lazy singleton channel (already disposed via subscriptions)
The Console Go upstream rejects the whole request with 'No tool output found for function call <id>' when a function_call item has no matching function_call_output — e.g. after history trimming dropped the tool result, or when VS Code replays a tool message whose tool_call_id was lost. The old converter even fabricated a 'tool-<timestamp>' call_id in that case, which could never match its call. - never fabricate a function_call_output.call_id; drop the item instead - add pairResponsesFunctionCallItems() enforcing strict 1:1 pairing: orphaned outputs are dropped, calls without an output are dropped, duplicated outputs keep only the first - wire the pairing pass into buildResponsesRequestBody + unit tests
…ors (closes #217) gpt-5.6-luna streams could end with zero extractable parts while the gateway billed completion tokens; VS Code then surfaced its empty-response-loop guard. Verified against the published 0.7.4 VSIX: neither error string exists in our build, so the loop guard is Copilot Chat reacting to a zero-part stream from our provider. - accept record-shaped payloads on response.output_text.delta (delta.text / delta.content / text.value) and nested reasoning deltas so new gateway payload shapes still produce parts - throw a dedicated zero-part error that names the token count, event stats, and the [diag-sse-event-*] diagnostic path after retries are exhausted
Console Go upstream rate limits ('Upstream request failed:
[rate_limit_exceeded]') come from the model provider, not the gateway —
the extension cannot lift them, but it can stop turning them into hard
failures when the upstream names a wait.
- parse Retry-After (delta-seconds and HTTP-date) via parseRetryAfterMs
- on a 429 with a Retry-After within RATE_LIMIT_MAX_RETRY_AFTER_WAIT_MS
(30s), wait once and transparently retry — nothing has been streamed
at that point, so no content can be duplicated
- longer waits still surface the existing rate-limit error untouched
…loses #218, closes #223) - docs/issues/94-98: full root-cause records for #222, #220, #216, #217, #221 - docs/issues/99: triage for #218 (VS Code dropdown limitation, use Configure Utility Models) and #223 (stack trace belongs to vizards.deepseek-v4-for-copilot; our requests always send x-opencode-session) - devlog batch entry + CHANGELOG 0.7.5
…thy tool-call turns [diag-empty-response] plus the raw [diag-sse-event-*] dump fired on every tool-call-only Responses turn (gpt-5.6-luna): tool calls are flushed in the transport's finally block AFTER the diagnostic runs, so extractedPartCount is legitimately 0 there while the stream is healthy. Skip the dump when a real finish_reason was extracted — genuine format mismatches (issue #217 class) never produce one.
…ent shapes)
Automates the manual verification checklist against the exact event shapes
captured from a live gpt-5.6-luna /v1/responses session:
- a 40-tool-call-group session that triggers history trim must still yield
strictly paired function_call/function_call_output items after the pairing
pass (the gateway 400s on any orphan)
- the captured luna tool-call event sequence (response.created → reasoning
items → function_call added/delta/done → response.completed) must extract
into a complete tool call despite zero text parts
- output_text.delta must extract for flat, delta:{text}, and text:{value}
payload shapes
OpenCode's Go docs (updated 2026-09-07) now require a stable x-opencode-session on every request, and requests missing it may start erroring from 2026-09-06. Only the main chat path sent the header; four auxiliary call sites hit the gateway without it: - GET /models (ModelListFetcher) - inline completions POST /chat/completions (ChatCompletionEngine) - GET /usage (fetchGoUsage via GoUsageTracker server sync) - Manage Provider test-connection POST These have no conversation, so they share one persisted per-installation session id (auxiliarySessionId(), globalState-backed) instead of the per-conversation hash the chat path uses.
Header-capture tests stubbing global fetch prove the session header is sent by ModelListFetcher (GET /models), ChatCompletionEngine (inline completions), and fetchGoUsage (GET /usage); plus persistence semantics of auxiliarySessionId (stable per installation, distinct across installs). Also teaches the shared vscode mock about extensions.getExtension so getUserAgent resolves in tests.
…s in devlog/CHANGELOG
…y download CI failed on the Editorconfig lint step: editorconfig-checker@6.1.1 could not download the 'ec-linux-amd64' binary on the Linux runner. 6.2.0 (2026-08-27) adds 'support editorconfig-checker binary name for v4 with legacy ec fallback' (#435), which fixes the binary-name resolution. Node >=20.11 requirement stays compatible with the CI Node 20 matrix.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Batch of fixes for issues #216, #217, #218, #220, #221, #222, #223. Each commit carries
closes #N, so merging this PR auto-closes all seven.Changes
GET /modelsserved from cache (was refetching every UI poll);Refresh Modelsstill forces a real fetch;Models registeredlogs only on change9f0eba95d2b525function_call/function_call_output; no more fabricatedcall_ided53ecb573f757Retry-Afteron upstream 429 (single transparent retry, capped 30s)8d8e9f06c0f184[diag-empty-response]false positive on healthy tool-call turnsac7a9a1b99f449,25f0301x-opencode-sessionnow sent on all auxiliary gateway requests (models, inline completions, usage, test-connection)418c4c4Verification
Notes
docs/issues/99).x-opencode-sessionfollow-up addresses the gateway enforcement announced in the Go docs (updated 2026-09-07).