feat(provider): port upstream model-selection option arrays (#2246) - #71
Merged
Conversation
…tgg#2246 commit A) Stage one of five for the upstream pingdotgg#2246 port. Replaces the per-provider ProviderModelOptions object (`{ effort, fastMode, ... }`) with a provider- agnostic ProviderOptionSelections array (`[{ id, value }, ...]`) on the ModelSelection schema. Adds ProviderOptionDescriptor (tagged union of select/boolean) to describe capabilities, and helpers getModelSelectionStringOptionValue / getModelSelectionBooleanOptionValue / createModelSelection / getProviderOptionDescriptors for downstream use. Preserves MarCode-specific fields: - DEFAULT_MODEL = DEFAULT_MODEL_BY_PROVIDER.claudeAgent (Claude-first default) - claudeAgent: "claude-opus-4-6" (vs upstream "claude-sonnet-4-6") - DEFAULT_PROVIDER_KIND = "claudeAgent" - jiraBoard / additionalDirectories / compacting / CLAUDE_COMPACTING_REASON on orchestration aggregates - TurnNotificationMode / CustomNotificationSound / NotificationSoundMap + sidebarProjectGroupingMode settings additions Also lands three upstream-additive fields on ServerProvider (displayName, badgeLabel, showInteractionModeToggle) to avoid a trailing follow-up. Typecheck: @marcode/contracts + @marcode/shared pass. apps/server and apps/web will fail until commits C-E land the retrofits. Tests: 73 contracts + 117 shared passing.
…stream pingdotgg#2246 commit B) Ports upstream's 026_CanonicalizeModelSelectionOptions, renumbered to 030 because MarCode already uses 026 for AuthSessionLastConnectedAt. MarCode's migration head was 029_CleanupInvalidProjectionPendingApprovals; this adds 030. The migration rewrites stored model-selection options from the legacy object shape (`{ effort: "max", fastMode: true }`) to the canonical array shape (`[{ id: "effort", value: "max" }, { id: "fastMode", value: true }]`) in: - projection_threads.model_selection_json.$.options - projection_projects.default_model_selection_json.$.options - orchestration_events payload for thread.created, thread.meta-updated, thread.turn-start-requested, project.created, project.meta-updated effect_sql_migrations row uses (id=30, name="CanonicalizeModelSelectionOptions") — fresh row, no conflict with existing installs. Test uses MarCode-aware bounds (seeds at migration 29, asserts after 30) and covers legacy object, empty object, non-scalar entry drop, already-array (no-op), null selection, and the five relevant event types. 5/5 migration tests pass.
…pstream pingdotgg#2246 commit C) Stage three of five. Replaces provider-specific option access (modelSelection.options.effort / .fastMode / .thinking / .agent / .variant) with the provider-agnostic helpers introduced in Commit A: getModelSelectionStringOptionValue / getModelSelectionBooleanOptionValue / getProviderOptionDescriptors / resolvePromptInjectedEffort. Adds a new `builtInProviderCatalog.ts` module (upstream-introduced) with MarCode-order `BUILT_IN_PROVIDER_ORDER` starting with claudeAgent (Claude-first branding) rather than upstream's codex-first order. ClaudeProvider exports new helpers (resolveClaudeEffort, normalizeClaudeCliEffort) used by ClaudeAdapter for effort normalization. Each built-in model's capabilities now use createModelCapabilities + buildSelectOptionDescriptor / buildBooleanOptionDescriptor. MarCode-preferred defaults retained: Opus 4.6 and Sonnet 4.6 default to "medium" effort (vs upstream's "high"); model names use bare slugs ("Opus 4.6" vs upstream "Claude Opus 4.6"); "Claude/Cursor/Codex/OpenCode is disabled in MarCode settings" branding kept. Preserves: - MarCode's CursorAdapter session/request_permission tool-call hint logic (terminal command display) and ACP toolCallHints map — only retrofitting the modelOptions -> selections parameter name change - MarCode's OpenCodeAdapter tool-activity classification refactor via @marcode/shared/toolActivity (classifyToolLifecycleItemType, extractPlanStepsFromTodos, isTodoWriteTool) — only retrofitting options.agent / options.variant access - MarCode's ClaudeAdapter progressive error classes (ClaudeStreamError variants), interrupt handling, and todo plan-step extraction — retrofitted effort / fastMode / thinking reads via the new helpers Test adaptation: - AnalyticsService.layerTest -> AnalyticsServiceNoopLive (FEATURES.md PostHog-free requirement, pattern from PR #66's resume-drift fix) - Provider test literals migrated to array-of-{id, value} shape via createModelSelection builder - Sonnet 4.6 "fallback to default" test now asserts "medium" (MarCode's isDefault choice) rather than upstream's "high" - Grep tool classification asserted as "file_read" (via MarCode's local classifyToolItemType) rather than upstream's generic "dynamic_tool_call" - Marcode metric prefix applied (marcode_provider_* vs t3_provider_*) Tests: 244 provider + telemetry tests passing; apps/server provider layer typecheck is clean. Git text-generation and orchestration test breaks are expected and will be resolved by Commit D.
…pe (upstream pingdotgg#2246 commit D) Stage four of five. Retrofits every provider-specific modelSelection.options access site in the git-text-generation layers to the provider-agnostic helpers (getModelSelectionStringOptionValue / getModelSelectionBooleanOptionValue / getProviderOptionDescriptors) introduced in Commit A. ClaudeTextGeneration: - Replaces removed normalizeClaudeModelOptionsWithCapabilities helper with the descriptor-based pattern: getProviderOptionDescriptors(selections) + resolveClaudeEffort + normalizeClaudeCliEffort for effort CLI arg, plus typed fastMode / thinking currentValue lookups for the --settings JSON. - Preserves MarCode's fork-exclusive progressive generation code path, --tools "" (tool-lockdown) spawn arg, and Claude-first branching — only the option access layer changed. CodexTextGeneration: - Pulls reasoningEffort / fastMode via the helpers; falls back to the DEFAULT_CODEX_GIT_TEXT_GENERATION_REASONING_EFFORT constant. CursorTextGeneration: - Renames the applyCursorAcpModelSelection argument from `modelOptions` to `selections` (upstream's new parameter name). OpenCodeTextGeneration: - Agent / variant read via getModelSelectionStringOptionValue. RoutingTextGeneration: untouched — it forwards the opaque modelSelection to its sub-layers and never peeks at `options`, so the Claude→Codex fallback routing (FEATURES.md §"Claude-Powered Text Generation") survives without structural changes. Upstream's object-lookup dispatch rewrite was intentionally NOT ported, per the plan ("do not port upstream's structural rewrites"), because MarCode's if/else branching has different semantics. Test literals migrated from `{ options: { effort: "max" } }` to the new array shape in ClaudeTextGeneration / CodexTextGeneration / CursorTextGeneration specs, plus ProviderCommandReactor.test.ts and decider.projectScripts.test.ts. Exit criterion: `(cd apps/server && bun run test)` all-green — 343/343 in git + orchestration, 5/5 in decider, 244/244 in provider (from Commit C), 5/5 migration (from Commit B). Full apps/server typecheck is now clean (remaining errors live in apps/web, Commit E).
… commit E) Stage five of five — the final stage of the upstream pingdotgg#2246 port. Migrates the web composer from provider-specific option objects (`{ effort, fastMode }`) to the provider-agnostic array shape (`[{ id, value }]`). Structural changes: - Delete composerProviderRegistry.tsx + test (gone upstream) and replace with composerProviderState.tsx + test (upstream's descriptor-driven implementation, rebranded to @marcode/*). - Delete TraitsPicker.browser.tsx (gone upstream; no MarCode-unique coverage was there — styling-only fork divergence). TraitsPicker.tsx absorbs the descriptor-based control rendering. - composerDraftStore.ts: drop CursorModelOptions / CursorReasoningOption / CURSOR_REASONING_OPTIONS / ClaudeAgentEffort / CodexReasoningEffort / ProviderModelOptions imports. Introduce local ProviderOptionSelectionsByProvider alias + coerceProviderOptionSelections helper. All MarCode store additions preserved: stickyModelSelectionByProvider, terminalContexts, jiraTaskContexts, quotedContexts, draft thread/project shims, voice-prompting state. Retrofit only (no structural rewrite) in: - ChatView.tsx: formatOutgoingPrompt now uses resolvePromptInjectedEffort; composerProviderControls uses getProviderInteractionModeToggle(providerStatuses). - ChatComposer.tsx: same wiring, plus modelOptions passed as composerModelOptions?.[selectedProvider] instead of the whole by-provider map. - ProviderModelPicker.browser.tsx: 5 capability blocks converted from {reasoningEffortLevels, supportsFastMode, ...} to {optionDescriptors: [...]}; MarCode's Cmd-K redesign (+578 LoC over merge-base from pingdotgg#2153 port) kept intact. - CompactComposerControlsMenu.browser.tsx: same descriptor conversion + 3 literal `options` object -> array migrations. - ChatView.browser.tsx: test fixtures via createModelSelection + createModelCapabilities; expect.arrayContaining for sticky-option assertions (matchObject compares arrays strictly, so we pin the relevant sticky trait and ignore others). - modelSelection.ts, providerModels.ts: import + shape shift only. providerModels.ts adds getProviderDisplayName / getProviderInteractionModeToggle / formatProviderKindLabel (upstream-added, used downstream). - SettingsPanels.tsx: capability labels derived from descriptors (fastMode / thinking / effort-or-reasoning presence). - useSettings.ts: NonNullable cast on textGenerationModelSelection assignment to satisfy exactOptionalPropertyTypes. MarCode customizations preserved: - DEFAULT_PROVIDER_KIND = "claudeAgent" (not upstream's "codex") - All fork-exclusive store fields (Jira chip, voice, terminal contexts, sticky per-provider model selection) - Cmd-K ProviderModelPicker redesign - Voice-prompting flow, sticky model selection behaviors Exit criterion (plan §Commit E): - bun run typecheck: clean across all 10 packages - (cd apps/web && bun run test): 1084/1084 passing All five stages of the pingdotgg#2246 port are now on branch. Regression guard sweep (Phase 3) and migration fixture smoke (Phase 4) come next, then PR.
…ay shape
Three CompactComposerControlsMenu.browser.tsx assertions and one
ChatView.browser.tsx fixture were still using MarCode's pre-port shape
and wording:
- "Fast mode" -> "Fast Mode" in Opus fixture to match ClaudeProvider label
- `toContain("off")`/`toContain("on")` -> `"On"`/`"Off"` (upstream render casing)
- `"On (default)"` -> `"On"` for Haiku thinking (upstream's BoolTrait doesn't
annotate defaults like the old MarCode picker did)
- `"Remove it to change effort."` -> `"Remove it to change this option."`
(upstream's generic descriptor-driven wording)
- ChatView.browser.tsx "prefers draft state" expectation switched from
legacy object `options` to the canonical `[{id, value}]` array shape
(was the only remaining missed retrofit)
All 146 apps/web browser tests pass; fmt:check clean; typecheck all 10
packages successful.Uh oh!
There was an error while loading. Please reload this page.
tyulyukov added a commit
that referenced
this pull request
Apr 24, 2026
- Record PR #68 (cycle ledger bootstrap), #69 (upstream pingdotgg#1996 sidebar timestamp), and #71 (upstream pingdotgg#2246 option-array refactor) under the current cycle's ported set, with deviation notes for pingdotgg#2246. - Document the post-merge composerDraftStore hotfix (9a8c78f) and its regression guard. - Record the Phase 4 real-DB smoke: 0 legacy `$.options` rows across projection_threads, projection_projects, and orchestration_events; 155 canonical thread rows. - Move pingdotgg#1996 and pingdotgg#2246 out of "Pending real work" — no real work outstanding as of 2026-04-24. - Advance "Baseline after cycle" to ececcdc (the #71 merge SHA).
tyulyukov added a commit
that referenced
this pull request
Apr 24, 2026
Brings upstream SHAs up to ada410b (v0.0.21) into main's ancestry without altering the working tree. Every upstream commit in this range is already accounted for per UPSTREAM_DIVERGENCE.md (2026-04-24 cycle): - Ported under new SHAs via PRs #66-#71 (see "Ported in the current cycle") - Already equivalent under a different SHA (see "Already equivalent") - Intentionally skipped: blacksmith runners, nightly channel, fork-specific release ops, upstream 0.0.x version bumps Purpose: reset GitHub's "commits behind" counter. The counter reflects raw SHA reachability and was not aware of the port-rather-than-merge workflow documented in UPSTREAM_DIVERGENCE.md.
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 freeto 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.
Why
Forward-compatibility with upstream's
provider-instance-registrybranch (ceddb40/7a466f/8a82b53/84b0d74 — direct ancestor of8d1d699f8), which extends the new option-array shape. Any future upstream sync that touches provider identity will assume pingdotgg#2246 is in. Landing it now unblocks future port cycles and keeps MarCode's divergence ledger manageable.The refactor itself replaces
modelSelection.optionsfrom a provider-specific object ({ effort: "max", fastMode: true }) with a provider-agnostic array ([{ id: "effort", value: "max" }, { id: "fastMode", value: true }]). Capabilities move from scattered booleans (supportsFastMode,reasoningEffortLevels) to a singleoptionDescriptorstagged union.Ported commit: upstream
8d1d699f8— Refactor provider model selections to option arrays (pingdotgg#2246).What
Staged across 5 bisectable commits. Total surface: 65 files, +3 219 / −3 932.
feat(contracts,shared)— foundationProviderOptionSelection,ProviderOptionSelections(union of canonical array + legacy object withcoerceLegacyOptionsObjectToArraybackward-compat decoder),ProviderOptionDescriptor(select / boolean), helpers (getModelSelectionStringOptionValue,getModelSelectionBooleanOptionValue,getProviderOptionDescriptors,createModelSelection,resolvePromptInjectedEffort). Also lands three upstream-additiveServerProviderfields (displayName,badgeLabel,showInteractionModeToggle) so a trailing follow-up isn't needed.feat(server)— migration 030026_CanonicalizeModelSelectionOptionsrenumbered to 030 because MarCode already uses 026 forAuthSessionLastConnectedAt. Rewrites stored options inprojection_threads,projection_projects, andorchestration_events(thread.created/thread.meta-updated/thread.turn-start-requested/project.created/project.meta-updatedpayloads) from{k: v}to[{id, value}]. Theeffect_sql_migrationsrow uses(id=30, name="CanonicalizeModelSelectionOptions")— fresh row, no conflict.feat(server)— provider adaptersClaudeAdapter,CodexAdapter,CursorAdapter,OpenCodeAdapter, + theirProviderlayers to use the descriptor-based helpers. Adds upstream's newbuiltInProviderCatalog.ts(new file), withBUILT_IN_PROVIDER_ORDERreordered claudeAgent-first (MarCode branding). NewresolveClaudeEffort/normalizeClaudeCliEffortexports inClaudeProvider. Each built-in model's capabilities migrate tocreateModelCapabilities({ optionDescriptors: [...] }).feat(server)— git text-generationClaudeTextGeneration/CodexTextGeneration/CursorTextGeneration/OpenCodeTextGeneration+ their specs.RoutingTextGenerationis intentionally not restructured — MarCode's Claude→Codex if/else fallback (FEATURES.md §"Claude-Powered Text Generation") shouldn't be collapsed into upstream's object-lookup dispatch.feat(web)— composercomposerProviderRegistry.tsx+ test (removed upstream); addscomposerProviderState.tsx+ test (rebranded from upstream). DeletesTraitsPicker.browser.tsx(absorbed upstream).composerDraftStore.tsswaps per-provider option types forProviderOptionSelectionsByProvider+coerceProviderOptionSelectionshelper.ProviderModelPicker.browser.tsx(MarCode Cmd-K redesign, +578 LoC over merge-base) preserved. Retrofit-only inChatView,ChatComposer,CompactComposerControlsMenu,SettingsPanels,TraitsPicker,modelSelection,providerModels,useSettings.Migration renumber: 026 → 030
MarCode's migration ledger already consumed 026 (
AuthSessionLastConnectedAt) before upstream's pingdotgg#2246 merged. MarCode HEAD at port start was 029 (CleanupInvalidProjectionPendingApprovals), so we register this migration as 030 with id=30 in bothMigrations.tsentries and the030_CanonicalizeModelSelectionOptions.tsfilename. The migration body is pure SQL against the json1 extension and is upstream-identical. Test seed bounds shifted totoMigrationInclusive: 29(pre) /30(post).MarCode preservation map (per FEATURES.md)
RoutingTextGeneration.tsClaudeTextGeneration.tsnormalizeClaudeCliEffort); MarCode's--tools ""spawn arg kept.environments/runtime/service.tsservice.notification-wiring.test.tspasses.apps/desktop/src/main.ts__EMBEDDED_MARCODE_JIRA_REDIRECT_URI__/__EMBEDDED_MARCODE_JIRA_TOKEN_PROXY_URL__declarations +embeddedJiraDefaultsloop verified intact at lines 7-8, 296-304.AnalyticsService.layerTestreferences swapped toAnalyticsServiceNoopLive(Claude-Powered Text Generation test and resume-drift test pattern). Metric prefixmarcode_provider_*(nott3_provider_*).store.ts+ projection SQLstore.test.ts -t "shell events are authoritative"passes.composerDraftStore.tsstickyModelSelectionByProvider,terminalContexts,jiraTaskContexts,quotedContexts, draft thread/project shims, voice-prompting state.ClaudeProvider.tsOpus 4.6/Sonnet 4.6keep MarCode-preferredmediumeffort as isDefault (vs upstream'shigh); model names use bare slugs;DEFAULT_PROVIDER_KIND = "claudeAgent"kept.Test plan
Automated (all ran green pre-push)
bun run typecheckmonorepo — 0 errors across 10 packagesapps/serversuite — 950 passed / 4 skipped / 0 failed (101 files)apps/websuite — 1 084 passed / 0 failed (97 files)apps/desktopsuite — 97 passed / 0 failed (16 files)Explicit regression guards (Phase 3 of the port plan)
apps/web/src/environments/runtime/service.notification-wiring.test.ts— turn-notification wiring intactapps/web/src/store.guard.test.ts— incremental event handling intactapps/web/src/store.test.ts -t "shell events are authoritative"— sidebar flag authority intactapps/desktop/src/windowState.integration-guard.test.ts— window-state IPC wiring intact (8/8)apps/web/src/components/chat/workCards.guard.test.ts— rich tool display cards intactapps/web/src/components/jiraChip.integration-guard.test.ts— Jira composer chip intact030_CanonicalizeModelSelectionOptions.test.tscovers legacy object → array transform across all 5 event types plus both projection tablesData-level smoke test (Phase 4)
Applied migration 030's SQL transform to a real
~/.marcode/userdata/state.sqlitecopy (/tmp/marcode-030-test.db, 60 MB):projection_threadsrows + 182orchestration_eventsrows withjson_type(...) = 'object'[{id, value}]{"provider":"claudeAgent","model":"claude-opus-4-6","options":[{"id":"effort","value":"medium"},{"id":"contextWindow","value":"1m"}]}Manual verification checklist
apps/desktop/src/main.tsstill embeds Jira env vars (lines 7-8, 296-304)@t3tools/*imports reintroduced (monorepo grep clean)AnalyticsService.layerTestreferences (replaced withAnalyticsServiceNoopLive)json_setbehavior; existing rows already in array shape are a no-op due to thejson_type(...) = 'object'WHERE guardBisectability
Each commit passes typecheck + the corresponding test slice when checked out individually:
@marcode/contracts+@marcode/sharedsuites green (73 + 117)Realistic effort
The plan called 3–5 focused hours minimum with the bulk of risk in Commit E. Actual: ~6 hours focused, with the web composer rebuild taking the largest share. No dropped MarCode-specific features; no deferred cleanup.
🤖 Generated with Claude Code