[baron]

Summary

Fixes #1019.

Custom Claude-compatible backends could not declare a request timeout or context window: environment generation only emitted API_TIMEOUT_MS and CLAUDE_CODE_AUTO_COMPACT_WINDOW inside the Z.ai special case, and persisted backend configuration had no override fields.

• Add optional per-backend requestTimeoutMilliseconds and contextWindowTokens fields to persisted app configuration and provider DTOs, preserving legacy decoding.
• Retain overrides through normalization, slot overrides, persistence, and bridge conversion; validate shared numeric bounds and avoid emitting invalid values.
• Expose both fields in custom-backend settings, with inline invalid-input feedback. Blank values retain Claude Code defaults; existing Z.ai model-specific defaults and Moonshot behavior remain unchanged.
• Emit CLAUDE_CODE_MAX_CONTEXT_TOKENS for explicit context-window overrides (not the auto-compaction threshold), and use the configured window as the app accounting fallback when no provider-reported window is available.
• Keep credentials in the existing secure-storage path; no new secret logging.

Relevance / overlap

Issue #1019 is open with no comments and remains reproducible on base ab84d47e6. PR #356 concerns custom-backend launch isolation/model selection and an execution-location popover crash; it does not add these configuration fields or overrides.

Validation

All builds/tests used the worktree's coordinated daemon.

• Fail first: make dev-provider-test FILTER=ClaudeCompatibleRuntimeSupportTests/testCustomBackendPersistsAndAppliesRequestLimits — 1 test, 4 expected assertion failures before implementation.
• make dev-provider-test — 21 tests passed.
• make dev-test FILTER=ClaudeCompatiblePluginBridgeTests — 10 tests passed, including the new store/bridge/legacy-decode regression.
• make dev-swift-build PRODUCT=RepoPrompt — passed (existing unrelated warnings remain).
• Changed-files-only SwiftFormat; pinned SwiftFormat 0.61.1 check passed.
• make dev-lint — passed with pinned SwiftFormat 0.61.1 and strict SwiftLint. Initial attempt stopped because only system SwiftFormat 0.63.1 was available; installed the repository-pinned tool and reran successfully.
• .agents/skills/rpce-contribution-check/scripts/preflight.sh commit and push — passed, including redacted secret scans and repository guardrails.
• Static Oracle review — no material findings.

Not run: full root test suite, pr-ready, release/package validation, or live UI/provider/MCP smoke. No app was launched, relaunched, or stopped. Draft pending independent adversarial review.

Adversarial review follow-up

• Verified the official model configuration docs: CLAUDE_CODE_AUTO_COMPACT_WINDOW changes the compaction threshold and cannot raise the assumed model window. Explicit overrides now emit CLAUDE_CODE_MAX_CONTEXT_TOKENS; existing GLM preset auto-compaction behavior stays unchanged. The UI warns about [1m], bare claude- names, and recognized-model semantics; no compaction-disabling flags are added.
• Added adapter/sidebar and estimator coverage: a 262,144-token configured fallback accepts 250K usage, while a provider-reported window wins and survives later window-less events. Provider environment tests also cover a 32K window without an auto-compact threshold.
• Shared validation bounds: RepoPrompt timeouts are 1000–2147483647 ms (one-second product minimum to catch seconds-as-ms mistakes; upstream maximum documented in environment variables). Context overrides accept 1–2147483647 tokens as a local numeric bound, with advisory UI warnings outside the typical 4K–1M range, not a claimed upstream MAX_CONTEXT cap.
• UI states the default timeout (600000 ms / 10 min), default for unknown model IDs (200K), units, and that one backend window applies to all three slots. 600 is rejected with guidance to enter 600000 for ten minutes.
• Removed default nil arguments from the public provider config initializer so construction sites must explicitly account for both overrides.
• Expected integration conflict: open PR Fix custom Claude-compatible backend isolation and execution-location popover crash #356 edits the same environment(...) builder in ClaudeCompatibleRuntimeSupport.swift and the same runtime test file. Expect a rebase conflict/adjacent-hunk reconciliation; preserve both launch-isolation behavior and these explicit limits when integrating.

Follow-up validation (conductor-only):

• make dev-provider-test: 21 passed (ticket f517ee3a).
• make dev-test FILTER=ClaudeCompatiblePluginBridgeTests: 10 passed (final ticket 176b4190). An intermediate run failed linking with framework Sparkle not found; the subsequent app build and test retry passed without source changes.
• make dev-swift-build PRODUCT=RepoPrompt: passed (ticket d6d1a3b6).
• make dev-lint: passed (final ticket 870b2d70); changed files formatted with pinned SwiftFormat 0.61.1.
• Commit and push preflights passed, including secret scans and guardrails.
• Full root suite, live smoke, packaging/release validation, and visible app lifecycle operations remain not run.

[baron]

Reviewed 39355a06a275137f580bf00b4e593b21bd950bc5 against its merge base.

No actionable correctness defect identified in the limit plumbing. I checked optional Codable compatibility, host/provider bounds, persistence through normalization and slot overrides, launch-environment construction, and the usage estimator's precedence for provider-reported windows. The controls retain invalid edits locally rather than persisting them, and the tests distinguish the configured context capacity from an auto-compaction threshold.

This branch is currently reported non-mergeable against main, and its observed CI run remains queued (https://github.com/repoprompt/repoprompt-ce/actions/runs/39879288536). Resolve the conflict and revalidate the resulting head; a green run of the pre-resolution tree would not cover that integration. Review was static source/test inspection only; I did not launch Claude Code or run native macOS tests.