|
| 1 | +# Refactoring sweep — July 2026 |
| 2 | + |
| 3 | +Full-codebase sweep (src, tests, tooling, configs) for defects and refactoring |
| 4 | +opportunities. Line numbers are as of 2026-07-08 — verify before use. |
| 5 | + |
| 6 | +Filed as issues (do not duplicate here): |
| 7 | + |
| 8 | +- **#1064** — defects: helper-tool `inputSchema` required/default bug, `smithery.yaml` |
| 9 | + broken entry path, Node-floor drift, payment-client bypass, hardcoded store URL. |
| 10 | +- **#1065** — `defineHelperTool` factory + `mockApifyClient()` + context-stub merge. |
| 11 | +- **#1066** — quick wins: `getTaskOrThrow`, loader mode-gate dedup, `resolveWidgets` |
| 12 | + memoization, `catchNotFound`, search-actors strings, `injectMcpSessionId`, widget |
| 13 | + boolean split, dead Python tooling, vitest timeout split. |
| 14 | +- **#658** (umbrella) — sync/task tool-call dispatch dedup; sub-issues #1061, #684, |
| 15 | + #1062, #1063, #974. |
| 16 | + |
| 17 | +## Backlog — not filed, pull when there's appetite |
| 18 | + |
| 19 | +### `types.ts` split (M) |
| 20 | + |
| 21 | +746 lines mixing six concerns. Move types next to their owners (pattern already used by |
| 22 | +`WidgetActor` in `actor_card.ts`, `PricingTier` in `pricing_info.ts`): telemetry types |
| 23 | +(`ToolCallTelemetryProperties`, `CallDiagnostics`, `AjvErrorDetails`) → beside |
| 24 | +`tool_status.ts`/`telemetry.ts`; `StructuredActorCard`/`ActorCardOptions`/ |
| 25 | +`ConsoleLinkContext` → `actor_card.ts`/`console_link.ts`; `ServerCard` → |
| 26 | +`server_card.ts`. Keep the tool model (`ToolBase`, `ToolEntry`, `TOOL_TYPE`) + `Input` as |
| 27 | +the lean core. Do as pure moves + re-exports first. Also: `types.ts` holds runtime values |
| 28 | +(`TOOL_TYPE`, `SERVER_MODE`) despite the name. |
| 29 | + |
| 30 | +### Payment seam closure (M) — sequence after #1062 |
| 31 | + |
| 32 | +`PAYMENT_REQUIRED_HEADER` + its base64→JSON decode exist twice (`payments/x402.ts` and |
| 33 | +`utils/payment_errors.ts`); `server.ts` branches on `isX402PaymentRequiredError` directly |
| 34 | +in both catch blocks instead of the provider owning its error path; `prepareToolCallContext` |
| 35 | +(`payments/helpers.ts`) carries a split-me TODO and registers the x402 axios interceptor |
| 36 | +for every client regardless of provider. Fix: unify header consts/decode in |
| 37 | +`payments/const.ts`; add a `buildPaymentRequiredResult(error)` hook to `PaymentProvider` |
| 38 | +so a third provider needs no `server.ts` edits. The catch blocks are exactly what #1062's |
| 39 | +`mapToolCallError` extracts — do this as its follow-up, not in parallel. |
| 40 | + |
| 41 | +### `RunResponse` assembly dedup (S/M) |
| 42 | + |
| 43 | +The canonical structuredContent + `respondOk` mirror is hand-assembled in |
| 44 | +`buildStartRunResponse`, `fetchActorRunData`, and re-implemented (~20 lines) in |
| 45 | +`abort_actor_run.ts`. Extract `buildRunResponseContent(run, storages)` + |
| 46 | +`respondRunResponse(...)` so abort stops drifting from get-actor-run. |
| 47 | + |
| 48 | +### `call_actor.ts` core extraction (M) |
| 49 | + |
| 50 | +Lines 47–543 are helpers exported mainly for `widgets/call_actor_widget.ts` |
| 51 | +(`callActorPreExecute`, `resolveAndValidateActor`, `buildCallActorErrorResponse`, |
| 52 | +`callOptionsSchema`); the tool entry itself is ~30 lines at the tail. Move the shared |
| 53 | +engine to `actors/call_actor_core.ts`; the widget then depends on an explicit core module |
| 54 | +instead of reaching into a sibling tool file. |
| 55 | + |
| 56 | +### `actor_run_response.ts` split (M/L) |
| 57 | + |
| 58 | +875 lines, four banner-separated jobs: field normalization + response types; |
| 59 | +status→summary/nextStep templates (~210 lines, the most-edited part); storage |
| 60 | +fetch/enrichment; wait/orchestration (`raceAbort`, `waitForRunWithProgress`). Mechanical |
| 61 | +4-file split; watch storage tools importing `normalizeDatasetFields`. |
| 62 | + |
| 63 | +### `internals.js` export narrowing (M/L, cross-repo) |
| 64 | + |
| 65 | +`index_internals.ts` exports raw tool-catalog functions the hosted repo consumes directly |
| 66 | +(`getDefaultTools`, `getCategoryTools`, `getActorsAsTools`, `processParamsGetTools`, |
| 67 | +`getToolPublicFieldOnly`, …) against the stated "expose methods on `ActorsMcpServer`" |
| 68 | +policy, plus deprecated aliases awaiting internal migration (`addActor as addTool`, |
| 69 | +`redactSkyfirePayId` — see #604, `HelperTools`). Convert one export at a time to a server |
| 70 | +method; needs internal-repo coordination per export. |
| 71 | + |
| 72 | +### Server-mode lifecycle encapsulation (M) |
| 73 | + |
| 74 | +One lifecycle ("auto → resolved on initialize, buffer tool loads until then") tracked by |
| 75 | +five mutable fields (`serverModeOption`, `serverMode`, `serverModeResolved`, |
| 76 | +`pendingToolsAfterModeResolved`, `clientSupportsUi`) mutated from four places; invariants |
| 77 | +comment-enforced. Extract a `ServerModeResolver` owning option/resolved/buffer. Land the |
| 78 | +#1066 loader-dedup first — it shrinks this. |
| 79 | + |
| 80 | +### `pricing_info.ts` formatter consolidation (M, deliberately deprioritized) |
| 81 | + |
| 82 | +Five pricing models × (complete|simplified) × (text|structured) run in near-parallel |
| 83 | +function pairs; tier logic re-walked four ways. Fix: one resolved intermediate per model, |
| 84 | +thin renderers on top. Well-tested and correctness-sensitive (see |
| 85 | +`pricing_output_contract.md`) — risk offsets payoff; only touch with the E1–E8 oracle |
| 86 | +green. |
| 87 | + |
| 88 | +### Test import time (M, measure first) |
| 89 | + |
| 90 | +`test:unit`: 7s running tests, ~30s importing modules (vitest transform+import stats). |
| 91 | +Likely heavy barrel imports (`tools/index.js`, server construction) pulled by most of the |
| 92 | +72 files. Measure the dep graph before changing anything; do NOT blindly flip |
| 93 | +`isolate: false` — the suite leans on `vi.mock`. |
| 94 | + |
| 95 | +### Smaller items |
| 96 | + |
| 97 | +- **Evals type-check policy**: CI type-checks only `src`+`tests`; standalone eval scripts |
| 98 | + (`run_evaluation.ts`, `create_dataset.ts`, `eval_single.ts`) are never compiled, and |
| 99 | + oxlint ignores `evals/*.ts` but lints `evals/*/**.ts`. Pick one policy for the tree. |
| 100 | +- **Log field vocabulary**: `statusCode` vs `failureHttpStatus` vs `failure_http_status` |
| 101 | + for the same concept; mixed `[HandlerName]` tag conventions. Standardize when it next |
| 102 | + bites an alert query. |
| 103 | +- **Web mock typing**: `MOCK_ACTOR_DETAILS_RESPONSE` (`web/src/utils/mock-actor-details.ts`) |
| 104 | + is an untyped literal with fields absent from the `Actor` type it mocks; annotate it so |
| 105 | + drift becomes a type error. Longer term, generate `web/src/types.ts` from server schemas. |
| 106 | +- **`TTLLRUCache.set()`** does a redundant get-then-remove before add; module-global cache |
| 107 | + singletons in `state.ts` have no reset hook (stale ~30 min across hot-reloads). |
| 108 | +- **`structured_output_schemas.ts`**: `apifyConsoleUrl` property object hand-written 4×, |
| 109 | + `userTier` enum 3× — extract consts (or fold into the #1065 factory work). |
| 110 | + |
| 111 | +## Checked and healthy — no action |
| 112 | + |
| 113 | +- Zod+AJV double validation is consistent by design (every tool compiles the same schema). |
| 114 | +- `legacyToolNameToNew` shim is small, documented, load-bearing. |
| 115 | +- content/structuredContent mirroring is intentional for mixed MCP clients. |
| 116 | +- Lint suppressions are not piling up (16 across 9 files, scattered). |
| 117 | +- `scripts/` (check-agents-links, check_widgets, dev_standby) are solid. |
| 118 | +- `evals/workflows` TS code is live and unit-tested (unlike the Python side — #1066). |
0 commit comments