lucas/omp-deck

Watch

Fork 0

issue_label dispatch crashes on null entries in the Forgejo issue timeline #262

New issue

Closed

opened 2026-11-12 10:43:08 +01:00 by omp-deck · 0 comments

omp-deck commented

2026-11-12 10:43:08 +01:00

Collaborator

Problem

An issue_label / label_updated webhook delivery fails with:

null is not an object (evaluating 'c.label')

The trigger is recorded as failed with sessionId: null — no agent session is ever started, so label-driven intake silently does not happen for the affected issue.

Cause. apps/server/src/forgejo/dispatch-api.ts:535-549, listLabelEvents:

for (const c of res.data) {
    const name = c.label?.name;
    if (c.type !== "label" || !name) continue;

The element access c.label is unguarded; only the property below it is optional-chained. RawTimelineComment (same file, line 121) declares every field optional but the element as non-nullable, so TypeScript does not flag it.

Forgejo's GET /repos/{owner}/{repo}/issues/{n}/timeline can return null array elements: the serializer maps timeline rows it cannot convert into an API comment to nil and emits them as JSON null instead of dropping them (unverified — Forgejo's source is not reachable from this repo; the observed crash is the evidence). Typical sources are rows whose referenced object no longer exists (deleted comment/review, removed label, dependency on a deleted issue). One such element makes the loop throw.

Path to the failure. The issue_label payload carries no diff of what changed, so Dispatcher.latestAddedLabels (apps/server/src/forgejo/dispatcher.ts:1079-1091) asks the API which labels the newest change added; listLabelEvents pages the timeline and throws on the first null entry; the dispatcher aborts before resolving the label trigger (resolveLabelTrigger, dispatcher.ts:1115).

Blast radius. Not a one-off: every issue_label trigger on an issue whose timeline contains a null entry fails the same way, permanently, until the loop skips null entries. Long-lived issues with edited/deleted history are the likely population.

The same unguarded-element assumption exists across the whole Forgejo module, in three shapes, all verified in the current tree:

fetchAllPages<T> (apps/server/src/forgejo/tools/shared.ts:236-255) declares client.json<T[]> and accumulates elements unchanged — 13 call sites.
Ten direct, unpaged client.json<X[]> fetches that never touch fetchAllPages and map over their elements immediately.
Two nested arrays, ApiCombinedStatus.statuses in dispatch-api.ts:554 and tools/ci.ts:162, whose elements are dereferenced right away.

The fix must therefore be one shared mechanism, not a guard at the crash site. (ApiActionRunList.workflow_runs is already element-safe at ci.ts:186 and needs no change.)

Plan

Approach

Step 1 — add the null-dropping primitive and the list fetch helper in apps/server/src/forgejo/tools/shared.ts.

Insert both next to fetchAllPages (after line 229, before the fetchAllPages doc comment). No equivalent exists: the module repeats the inline idiom Array.isArray(x.data) ? x.data : [], which handles a non-array body but not null elements.

/**
 * A JSON array body as a usable list: `[]` when it is not an array, and
 * without the `null` elements Forgejo emits for rows it cannot serialise
 * (deleted comments, reviews, labels and the like). Only `null`/`undefined`
 * are dropped, so a legitimately falsy element survives.
 */
export function compactList<T>(value: unknown): T[] {
	if (!Array.isArray(value)) return [];
	return value.filter((v): v is T => v != null);
}

/**
 * One request to an endpoint whose body is a JSON array, with the elements
 * compacted by `compactList`. Error results pass through unchanged; `meta`
 * (the paging headers) is preserved.
 */
export async function jsonList<T>(
	client: ForgejoClient,
	method: string,
	path: string,
	opts: RequestOpts = {},
): Promise<ApiResult<T[]>> {
	const res = await client.json<(T | null)[]>(method, path, opts);
	if (!res.ok) return res;
	return { ok: true, meta: res.meta, data: compactList<T>(res.data) };
}

RequestOpts is exported from ../client.ts (it is the third parameter type of ForgejoClient.json, client.ts:203); extend the existing type-only import on shared.ts:26 to include it. If RequestOpts turns out not to be exported, export it from client.ts rather than restating its shape.

Step 2 — route fetchAllPages through compactList, keeping the raw page length for paging.

In fetchAllPages (shared.ts:244-253), fetch each page as (T | null)[] and compact only what is accumulated. The short-page break must keep using the raw server page length: compacting first would make a full page that contains nulls look short and silently truncate paging.

const res = await client.json<(T | null)[]>("GET", path, {
    query: { ...opts.query, page, limit: pageSize },
    signal: opts.signal,
});
if (!res.ok) return res;
const items = Array.isArray(res.data) ? res.data : [];
all.push(...compactList<T>(items));
if (items.length < pageSize || !hasMorePages(res.meta, page, pageSize)) break;

Do not call jsonList here — it would discard the raw length. This step alone fixes the reported crash for all 13 fetchAllPages consumers; no call site changes.

Step 3 — convert every direct array fetch to jsonList. Mechanical and independent of each other; each one replaces client.json<X[]>("GET"|"PUT", path, opts) with jsonList<X>(client, "GET"|"PUT", path, opts) (note the element type loses its []), and the Array.isArray(<res>.data) ? <res>.data : [] line that immediately follows becomes a plain <res>.data. Error handling at each site is unchanged — jsonList returns the same error variant, so existing if (!res.ok), res.status === 404 and apiError(res) branches keep working.

File	Line	Current element type	Follow-up `Array.isArray` guard to drop
`forgejo/dispatch-api.ts`	299	`ApiReviewComment[]`	line 303 `if (!Array.isArray(inline.data)) continue;` — delete the line
`forgejo/dispatch-api.ts`	320	`ApiReviewComment[]`	line 326
`forgejo/dispatch-api.ts`	402	`RawIssue[]`	line 407
`forgejo/tools/issues.ts`	76	`ApiComment[]`	line 78
`forgejo/tools/issues.ts`	104	`ApiIssue[]`	line 116
`forgejo/tools/issues.ts`	263 (`PUT`)	`ApiLabel[]`	line 272
`forgejo/tools/prs.ts`	75	`ApiReviewComment[]`	whichever guard follows in `fetchInlineComments`
`forgejo/tools/prs.ts`	308	`ApiComment[]`	line 319 `Array.isArray(comments.data) ? comments.data : []`
`forgejo/tools/prs.ts`	609	`ApiPull[]`	line 614
`forgejo/tools/prs.ts`	624	`ApiIssue[]`	line 629
`forgejo/tools/ci.ts`	381	`ApiActionRunJob[]`	line 386
`forgejo/tools/ci.ts`	431	`ApiActionRunJob[]`	line 435
`forgejo/tools/pr-diff.ts`	221	`ApiCommit[]`	line 226 (`Array.isArray(res.data) ? res.data[0]?.sha : undefined` → `res.data[0]?.sha`)

In dispatch-api.ts the helper is called as jsonList<X>(client, …); in tools/*.ts as jsonList<X>(ctx.client, …). Add jsonList to each file's existing import from ./shared.ts / ./tools/shared.ts.

Leave single-object fetches alone (client.json<ApiPull>, client.json<ApiActionRun>, client.json<ApiCombinedStatus>, …): they are not arrays.

Step 4 — apply compactList to the two nested statuses arrays.

forgejo/dispatch-api.ts:554: let statuses: ApiCommitStatus[] = compactList<ApiCommitStatus>(combined.data.statuses);
forgejo/tools/ci.ts:162: the same replacement for combined.data.statuses ?? [].

Both then dereference elements (s.context, s.status, s.target_url) safely. Import compactList alongside the existing shared.ts imports in both files.

Step 5 — make createDispatchApi accept an injected client so listLabelEvents is testable.

createDispatchApi() (dispatch-api.ts:262-265) builds its own ForgejoClient from readForgejoConfig(), which reads the DB; nothing can reach the real listLabelEvents loop today. Change the signature to take an optional client and skip the config read entirely when one is given — that keeps the test free of any DB (getDb() throws unless openDb() ran, and db/index.ts opens nothing at import time):

/**
 * A `DispatchApi` for the configured instance, or null when Forgejo is not
 * configured. `injected` overrides the configured connection (tests); with
 * it, the stored config is not read and the result is never null.
 */
export function createDispatchApi(injected?: ForgejoClient): DispatchApi | null {
	let client = injected;
	if (!client) {
		const cfg = readForgejoConfig();
		if (!cfg) return null;
		client = new ForgejoClient({ url: cfg.url, token: cfg.token });
	}
	return {
		…
	};
}

Naming the parameter injected keeps client as the identifier the returned object's methods already close over, so no line inside the returned object changes. client is ForgejoClient | undefined after the let, but narrowed by the if block before the return; if the narrowing does not survive into the closures, bind const api = client; after the block and rename the closure uses in one pass. Existing callers (repo-system-prompt.ts:2 and the dispatcher) pass nothing and are unaffected — the parameter is optional.

Step 6 — mechanism tests in apps/server/src/forgejo/tools/shared.test.ts.

Add one describe("list compaction") block. Stub the client with the pattern already used at bridge/git-tools/repo-info.test.ts:40:

const stub = (pages: unknown[][]) => {
	let i = 0;
	return {
		json: async () => ({ ok: true, data: pages[Math.min(i++, pages.length - 1)], meta: { totalCount: null, hasNext: i < pages.length } }),
	} as unknown as ForgejoClient;
};

Tests:

compactList drops null/undefined, keeps order, keeps 0 and "", and returns [] for a non-array (undefined, an object).
jsonList over a body [{id:1}, null, {id:2}] resolves {ok: true} with data [{id:1},{id:2}] and the stub's meta intact; an {ok:false, status:404} body passes through unchanged.
fetchAllPages paging is driven by the raw page length: with pageSize: 2 and pages [[{id:1}, null], [{id:2}]], the result is [{id:1},{id:2}] — proving the null-bearing first page did not stop paging (if the filter leaked into the length check the result would be [{id:1}]).

Step 7 — end-to-end regression test, new file apps/server/src/forgejo/dispatch-api.test.ts.

Build the API with an injected stub client and assert the reported failure is gone:

const timeline = [
	{ type: "label", body: "1", created_at: "2026-11-01T10:00:00Z", user: { login: "alice" }, label: { name: "Ready" } },
	null,
	{ type: "label", body: "0", created_at: "2026-11-01T10:00:01Z", user: { login: "bob" }, label: { name: "Draft" } },
];

The stub returns that body for the timeline path. Assert listLabelEvents("o", "r", 7) resolves {ok: true} (not a throw) with exactly two events, in page order, [{label: "Ready", added: true, author: "alice", …}, {label: "Draft", added: false, author: "bob", …}]. Add a second case where the null sits on a full first page (pageSize elements, stub serving a second page) to show the null did not cut the timeline short. Document in the file header that the client is a stub and the API is built via the injected-client path.

Critical files & anchors

apps/server/src/forgejo/tools/shared.ts:236-255 — fetchAllPages; the raw-vs-compacted page length in the break condition is the one subtle line of the change.
apps/server/src/forgejo/dispatch-api.ts:535-549 — listLabelEvents, the reported crash; after Step 2 it needs no edit of its own, which is the point of the central fix.
apps/server/src/forgejo/dispatch-api.ts:262-266 — createDispatchApi; the config read must stay on the non-injected path only.
apps/server/src/forgejo/tools/ci.ts:162 and apps/server/src/forgejo/dispatch-api.ts:554 — the nested statuses arrays, the only element dereferences a list helper cannot reach.
apps/server/src/forgejo/client.ts:25-32,203 — PageMeta, ApiResult, ForgejoClient.json; the shapes jsonList must preserve.

Verification

From apps/server:

bun test src/forgejo/tools/shared.test.ts src/forgejo/dispatch-api.test.ts — the new tests. The decisive assertion is Step 7's: a timeline page containing null yields both surrounding label events in order instead of throwing null is not an object (evaluating 'c.label').
bun test — the full server suite, including dispatcher.test.ts (which fakes DispatchApi wholesale and must be untouched by the signature change) and tools/*.test.ts for the converted call sites.
bun run typecheck from the repo root — proves every converted call site passes jsonList the element type rather than the array type, and that no caller of createDispatchApi broke.

Assumptions & contingencies

The null elements are assumed to carry no information worth recovering (they are rows Forgejo itself could not serialise), so they are dropped silently, with no logging added to tools/shared.ts (which currently imports no logger). If a later investigation needs to see them, the single place to add a counter is compactList.
If RequestOpts is not exported from client.ts, export it there rather than duplicating the option shape in shared.ts.
If any direct call site in Step 3 turns out to use res.data in a way that depends on positional alignment with another array (none does in the current tree — the mapLimit sites align result arrays, not elements within a body), leave that site on client.json and note it, rather than changing its semantics.

## Problem An `issue_label` / `label_updated` webhook delivery fails with: ``` null is not an object (evaluating 'c.label') ``` The trigger is recorded as `failed` with `sessionId: null` — no agent session is ever started, so label-driven intake silently does not happen for the affected issue. **Cause.** `apps/server/src/forgejo/dispatch-api.ts:535-549`, `listLabelEvents`: ```ts for (const c of res.data) { const name = c.label?.name; if (c.type !== "label" || !name) continue; ``` The element access `c.label` is unguarded; only the property below it is optional-chained. `RawTimelineComment` (same file, line 121) declares every *field* optional but the *element* as non-nullable, so TypeScript does not flag it. Forgejo's `GET /repos/{owner}/{repo}/issues/{n}/timeline` can return `null` array elements: the serializer maps timeline rows it cannot convert into an API comment to nil and emits them as JSON `null` instead of dropping them (*unverified — Forgejo's source is not reachable from this repo; the observed crash is the evidence*). Typical sources are rows whose referenced object no longer exists (deleted comment/review, removed label, dependency on a deleted issue). One such element makes the loop throw. **Path to the failure.** The `issue_label` payload carries no diff of what changed, so `Dispatcher.latestAddedLabels` (`apps/server/src/forgejo/dispatcher.ts:1079-1091`) asks the API which labels the newest change added; `listLabelEvents` pages the timeline and throws on the first `null` entry; the dispatcher aborts before resolving the label trigger (`resolveLabelTrigger`, `dispatcher.ts:1115`). **Blast radius.** Not a one-off: *every* `issue_label` trigger on an issue whose timeline contains a null entry fails the same way, permanently, until the loop skips null entries. Long-lived issues with edited/deleted history are the likely population. The same unguarded-element assumption exists across the whole Forgejo module, in three shapes, all verified in the current tree: - `fetchAllPages<T>` (`apps/server/src/forgejo/tools/shared.ts:236-255`) declares `client.json<T[]>` and accumulates elements unchanged — 13 call sites. - Ten direct, unpaged `client.json<X[]>` fetches that never touch `fetchAllPages` and map over their elements immediately. - Two nested arrays, `ApiCombinedStatus.statuses` in `dispatch-api.ts:554` and `tools/ci.ts:162`, whose elements are dereferenced right away. The fix must therefore be one shared mechanism, not a guard at the crash site. (`ApiActionRunList.workflow_runs` is already element-safe at `ci.ts:186` and needs no change.) ## Plan ### Approach **Step 1 — add the null-dropping primitive and the list fetch helper in `apps/server/src/forgejo/tools/shared.ts`.** Insert both next to `fetchAllPages` (after line 229, before the `fetchAllPages` doc comment). No equivalent exists: the module repeats the inline idiom `Array.isArray(x.data) ? x.data : []`, which handles a non-array body but not null elements. ```ts /** * A JSON array body as a usable list: `[]` when it is not an array, and * without the `null` elements Forgejo emits for rows it cannot serialise * (deleted comments, reviews, labels and the like). Only `null`/`undefined` * are dropped, so a legitimately falsy element survives. */ export function compactList<T>(value: unknown): T[] { if (!Array.isArray(value)) return []; return value.filter((v): v is T => v != null); } /** * One request to an endpoint whose body is a JSON array, with the elements * compacted by `compactList`. Error results pass through unchanged; `meta` * (the paging headers) is preserved. */ export async function jsonList<T>( client: ForgejoClient, method: string, path: string, opts: RequestOpts = {}, ): Promise<ApiResult<T[]>> { const res = await client.json<(T | null)[]>(method, path, opts); if (!res.ok) return res; return { ok: true, meta: res.meta, data: compactList<T>(res.data) }; } ``` `RequestOpts` is exported from `../client.ts` (it is the third parameter type of `ForgejoClient.json`, `client.ts:203`); extend the existing type-only import on `shared.ts:26` to include it. If `RequestOpts` turns out not to be exported, export it from `client.ts` rather than restating its shape. **Step 2 — route `fetchAllPages` through `compactList`, keeping the raw page length for paging.** In `fetchAllPages` (`shared.ts:244-253`), fetch each page as `(T | null)[]` and compact only what is accumulated. The short-page break must keep using the **raw** server page length: compacting first would make a full page that contains nulls look short and silently truncate paging. ```ts const res = await client.json<(T | null)[]>("GET", path, { query: { ...opts.query, page, limit: pageSize }, signal: opts.signal, }); if (!res.ok) return res; const items = Array.isArray(res.data) ? res.data : []; all.push(...compactList<T>(items)); if (items.length < pageSize || !hasMorePages(res.meta, page, pageSize)) break; ``` Do not call `jsonList` here — it would discard the raw length. This step alone fixes the reported crash for all 13 `fetchAllPages` consumers; no call site changes. **Step 3 — convert every direct array fetch to `jsonList`.** Mechanical and independent of each other; each one replaces `client.json<X[]>("GET"|"PUT", path, opts)` with `jsonList<X>(client, "GET"|"PUT", path, opts)` (note the element type loses its `[]`), and the `Array.isArray(<res>.data) ? <res>.data : []` line that immediately follows becomes a plain `<res>.data`. Error handling at each site is unchanged — `jsonList` returns the same error variant, so existing `if (!res.ok)`, `res.status === 404` and `apiError(res)` branches keep working. | File | Line | Current element type | Follow-up `Array.isArray` guard to drop | | --- | --- | --- | --- | | `forgejo/dispatch-api.ts` | 299 | `ApiReviewComment[]` | line 303 `if (!Array.isArray(inline.data)) continue;` — delete the line | | `forgejo/dispatch-api.ts` | 320 | `ApiReviewComment[]` | line 326 | | `forgejo/dispatch-api.ts` | 402 | `RawIssue[]` | line 407 | | `forgejo/tools/issues.ts` | 76 | `ApiComment[]` | line 78 | | `forgejo/tools/issues.ts` | 104 | `ApiIssue[]` | line 116 | | `forgejo/tools/issues.ts` | 263 (`PUT`) | `ApiLabel[]` | line 272 | | `forgejo/tools/prs.ts` | 75 | `ApiReviewComment[]` | whichever guard follows in `fetchInlineComments` | | `forgejo/tools/prs.ts` | 308 | `ApiComment[]` | line 319 `Array.isArray(comments.data) ? comments.data : []` | | `forgejo/tools/prs.ts` | 609 | `ApiPull[]` | line 614 | | `forgejo/tools/prs.ts` | 624 | `ApiIssue[]` | line 629 | | `forgejo/tools/ci.ts` | 381 | `ApiActionRunJob[]` | line 386 | | `forgejo/tools/ci.ts` | 431 | `ApiActionRunJob[]` | line 435 | | `forgejo/tools/pr-diff.ts` | 221 | `ApiCommit[]` | line 226 (`Array.isArray(res.data) ? res.data[0]?.sha : undefined` → `res.data[0]?.sha`) | In `dispatch-api.ts` the helper is called as `jsonList<X>(client, …)`; in `tools/*.ts` as `jsonList<X>(ctx.client, …)`. Add `jsonList` to each file's existing import from `./shared.ts` / `./tools/shared.ts`. Leave single-object fetches alone (`client.json<ApiPull>`, `client.json<ApiActionRun>`, `client.json<ApiCombinedStatus>`, …): they are not arrays. **Step 4 — apply `compactList` to the two nested `statuses` arrays.** - `forgejo/dispatch-api.ts:554`: `let statuses: ApiCommitStatus[] = compactList<ApiCommitStatus>(combined.data.statuses);` - `forgejo/tools/ci.ts:162`: the same replacement for `combined.data.statuses ?? []`. Both then dereference elements (`s.context`, `s.status`, `s.target_url`) safely. Import `compactList` alongside the existing `shared.ts` imports in both files. **Step 5 — make `createDispatchApi` accept an injected client so `listLabelEvents` is testable.** `createDispatchApi()` (`dispatch-api.ts:262-265`) builds its own `ForgejoClient` from `readForgejoConfig()`, which reads the DB; nothing can reach the real `listLabelEvents` loop today. Change the signature to take an optional client and skip the config read entirely when one is given — that keeps the test free of any DB (`getDb()` throws unless `openDb()` ran, and `db/index.ts` opens nothing at import time): ```ts /** * A `DispatchApi` for the configured instance, or null when Forgejo is not * configured. `injected` overrides the configured connection (tests); with * it, the stored config is not read and the result is never null. */ export function createDispatchApi(injected?: ForgejoClient): DispatchApi | null { let client = injected; if (!client) { const cfg = readForgejoConfig(); if (!cfg) return null; client = new ForgejoClient({ url: cfg.url, token: cfg.token }); } return { … }; } ``` Naming the parameter `injected` keeps `client` as the identifier the returned object's methods already close over, so no line inside the returned object changes. `client` is `ForgejoClient | undefined` after the `let`, but narrowed by the `if` block before the `return`; if the narrowing does not survive into the closures, bind `const api = client;` after the block and rename the closure uses in one pass. Existing callers (`repo-system-prompt.ts:2` and the dispatcher) pass nothing and are unaffected — the parameter is optional. **Step 6 — mechanism tests in `apps/server/src/forgejo/tools/shared.test.ts`.** Add one `describe("list compaction")` block. Stub the client with the pattern already used at `bridge/git-tools/repo-info.test.ts:40`: ```ts const stub = (pages: unknown[][]) => { let i = 0; return { json: async () => ({ ok: true, data: pages[Math.min(i++, pages.length - 1)], meta: { totalCount: null, hasNext: i < pages.length } }), } as unknown as ForgejoClient; }; ``` Tests: 1. `compactList` drops `null`/`undefined`, keeps order, keeps `0` and `""`, and returns `[]` for a non-array (`undefined`, an object). 2. `jsonList` over a body `[{id:1}, null, {id:2}]` resolves `{ok: true}` with `data` `[{id:1},{id:2}]` and the stub's `meta` intact; an `{ok:false, status:404}` body passes through unchanged. 3. `fetchAllPages` paging is driven by the raw page length: with `pageSize: 2` and pages `[[{id:1}, null], [{id:2}]]`, the result is `[{id:1},{id:2}]` — proving the null-bearing first page did not stop paging (if the filter leaked into the length check the result would be `[{id:1}]`). **Step 7 — end-to-end regression test, new file `apps/server/src/forgejo/dispatch-api.test.ts`.** Build the API with an injected stub client and assert the reported failure is gone: ```ts const timeline = [ { type: "label", body: "1", created_at: "2026-11-01T10:00:00Z", user: { login: "alice" }, label: { name: "Ready" } }, null, { type: "label", body: "0", created_at: "2026-11-01T10:00:01Z", user: { login: "bob" }, label: { name: "Draft" } }, ]; ``` The stub returns that body for the timeline path. Assert `listLabelEvents("o", "r", 7)` resolves `{ok: true}` (not a throw) with exactly two events, in page order, `[{label: "Ready", added: true, author: "alice", …}, {label: "Draft", added: false, author: "bob", …}]`. Add a second case where the null sits on a full first page (`pageSize` elements, stub serving a second page) to show the null did not cut the timeline short. Document in the file header that the client is a stub and the API is built via the injected-client path. ### Critical files & anchors - `apps/server/src/forgejo/tools/shared.ts:236-255` — `fetchAllPages`; the raw-vs-compacted page length in the break condition is the one subtle line of the change. - `apps/server/src/forgejo/dispatch-api.ts:535-549` — `listLabelEvents`, the reported crash; after Step 2 it needs no edit of its own, which is the point of the central fix. - `apps/server/src/forgejo/dispatch-api.ts:262-266` — `createDispatchApi`; the config read must stay on the non-injected path only. - `apps/server/src/forgejo/tools/ci.ts:162` and `apps/server/src/forgejo/dispatch-api.ts:554` — the nested `statuses` arrays, the only element dereferences a list helper cannot reach. - `apps/server/src/forgejo/client.ts:25-32,203` — `PageMeta`, `ApiResult`, `ForgejoClient.json`; the shapes `jsonList` must preserve. ### Verification From `apps/server`: 1. `bun test src/forgejo/tools/shared.test.ts src/forgejo/dispatch-api.test.ts` — the new tests. The decisive assertion is Step 7's: a timeline page containing `null` yields both surrounding label events in order instead of throwing `null is not an object (evaluating 'c.label')`. 2. `bun test` — the full server suite, including `dispatcher.test.ts` (which fakes `DispatchApi` wholesale and must be untouched by the signature change) and `tools/*.test.ts` for the converted call sites. 3. `bun run typecheck` from the repo root — proves every converted call site passes `jsonList` the element type rather than the array type, and that no caller of `createDispatchApi` broke. ### Assumptions & contingencies - The null elements are assumed to carry no information worth recovering (they are rows Forgejo itself could not serialise), so they are dropped silently, with no logging added to `tools/shared.ts` (which currently imports no logger). If a later investigation needs to see them, the single place to add a counter is `compactList`. - If `RequestOpts` is not exported from `client.ts`, export it there rather than duplicating the option shape in `shared.ts`. - If any direct call site in Step 3 turns out to use `res.data` in a way that depends on positional alignment with another array (none does in the current tree — the `mapLimit` sites align *result* arrays, not elements within a body), leave that site on `client.json` and note it, rather than changing its semantics.

omp-deck added the

labels

2026-11-12 10:43:08 +01:00