Comparing changes

…g main branch product.overlay.json set licenseUrl/serverLicenseUrl to github.com/levelcodeai/levelcode/blob/main/LICENSE, but there is no `main` branch — the default is `develop` — so the license link baked into the app's About / license metadata 404'd. Use /blob/HEAD/LICENSE: GitHub resolves HEAD to whatever the default branch is, so it works on develop today and survives a future rename to main. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…bution to NOTICE GitHub's license API reported NOASSERTION (not MIT) for this repo: the derivative and Microsoft-trademark paragraph appended after the MIT text pushed the file below the detector's template-match threshold, so the repo carried no MIT label. Split it the way upstream VS Code does — LICENSE is now verbatim MIT (the grant and the copyright line are unchanged), and the Code-OSS provenance + trademark disclaimer move to a dedicated NOTICE file. README's License section points at both. Nothing is lost; MIT now detects. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

fix(license): repair the 404 license link, and make GitHub detect MIT

…HTML into chat A gateway run that hit an upstream 502 surfaced, verbatim: OpenAI API 502: <html> ... <center><h1>502 Bad Gateway</h1></center> ... </html> Two defects in one line: 1. Mislabeled. The gateway resolves to providerId 'openai' to reuse the OpenAI- compatible adapter, so streamAgentTurn stamped the error with the OpenAI provider row's label — even for Opus 5 over LevelCode Cloud. prepProviderRequest already computes the right label ('LevelCode Cloud'); it was just never threaded past the provider lookup. Thread req.label through runAgent/turnOpts, doStream, and compact, and have the router prefer o.label over p.label. BYOK still falls back to the provider's own label, so an OpenRouter failure still reads "OpenRouter". 2. Raw HTML dumped. The three throw sites appended the raw response body; a proxy 5xx is an HTML page, not JSON, so the whole nginx document landed in the transcript. Two pure helpers in openaiCompat: extractApiError() returns a provider's JSON {error:{message}} when present, '' for an HTML page, and a hard-capped string otherwise; httpError() composes "<label> API <status>: <detail>", falling back to the status reason ("Bad Gateway") when there's no usable message, and sets e.status. The 502 itself is a transient upstream blip we can't fix — this is about surfacing it honestly. The same failure now reads: LevelCode Cloud API 502: Bad Gateway providers.test.js +10 (extractApiError / httpError, incl. the exact nginx body). Verified end-to-end with a stubbed fetch: the label threads through the real index.js -> openaiCompat.js chain and BYOK's p.label fallback is intact. Full gate: 24 suites, 0 failures. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…s, so the run survives The 502 that killed the run in the screenshot was a transient upstream blip — the proxy in front of the model briefly couldn't reach a healthy backend. Reporting it cleanly (the previous commit) is good; recovering from it is better. New postChat() centralizes POST /chat/completions for all three OpenAI-shaped entry points (stream, complete, agent turn) and retries ONCE on 502/503/504. It is safe because it runs BEFORE any SSE line is read: on a transient status the response carries no model output, so nothing has streamed or been metered and a retry cannot duplicate output or double-bill the UI. Deliberately NOT retried: 429 (needs Retry-After), 500 (usually a real error), other 4xx, and thrown network/abort errors. A 401 still throws straight through to the agent's existing token-refresh path. The backoff wakes early on abort so Stop stays instant. The agent surfaces a visible "upstream busy (502) — retrying…" status via a new onRetry hook (threaded through the router) rather than a mystery pause; chat/complete retry silently. providers.test.js +6 (recover / exhaust / no-retry-on-4xx / onRetry / clean-200 / abort-not-retried). Verified end-to-end with a stubbed fetch: an agent turn that gets a 502 then a good SSE stream retries once and streams the text, onRetry carrying 502. Full gate: 24 suites, 0 failures. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

@PARAM

#45 review) streamChat/complete/streamAgentTurn read o.label and o.onRetry, and pass onRetry as an object literal into the openaiCompat adapters — but the @PARAM typedefs for both didn't list them. Under // @ts-check that is 9 real diagnostics: TS2339 (property does not exist) on each o.label / o.onRetry read, plus TS2353 (excess property) on each adapter literal. Added label?:string and onRetry?:(info:{attempt,retries,status})=>void to the three router typedefs, and onRetry? to the three adapter typedefs (they already carried label?). The reviewer flagged label; onRetry has the same defect (added by the retry commit) and is fixed in the same pass. Verified with a real tsc 5.6 checkJs run — and not the vacuous kind: a synthetic probe first confirmed tsc actually enforces TS2339/TS2353 here (an earlier npx form was silently not running tsc at all), then stashing the fix gave BEFORE=9 label/onRetry errors, restoring it gave AFTER=0. JSDoc-only; runtime unchanged (full gate 24 suites, 28 provider tests). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

fix(ai): gateway error handling — relabel, sanitize the body, and retry transient 5xx

Navigation Menu

Uh oh!

Comparing changes

Open a pull request

Commits on Jul 25, 2026

This comparison is taking too long to generate.

Uh oh!