fix(tools): repair SDK call mismatches that broke five tools

[abhay-codes07]

What

Five @supermemory/tools operations were written against a different SDK surface than the supermemory@^3 the package pins, so they failed on every invocation. Because every tool wraps its body in try/catch and returns {success: false, error}, these never threw loudly — the model just gets a failure (or worse, a plausible-looking success with undefined data) and moves on.

Tool	Bug	Effect
documentDelete (ai-sdk)	passed {docId} to documents.delete(id: string)	no document could ever be deleted
memoryForget (ai-sdk + openai)	called client.memories.forget, which doesn't exist in the v3 SDK	TypeError on every call, surfaced as success: false
documentList (ai-sdk + openai)	read response.documents; the list response keys items as memories	always returned documents: undefined — models conclude the user has no documents
ClaudeMemoryTool str_replace / insert	falsy checks rejected new_str: "" and insert_text: ""	deleting text or inserting a blank line — standard memory-tool operations — were impossible
ClaudeMemoryTool delete / rename	delete returned success without calling any API; rename never removed the old document	"deleted" files kept appearing in listings and search — the model believes forgotten data is gone when it isn't

How

• documents.delete(documentId) — plain string, matching the SDK signature (the OpenAI variant already did this correctly).
• New src/shared/forget-memory.ts helper that calls DELETE /v4/memories directly with {containerTag, id?, content?, reason?} — the v3 SDK has no memories.forget, and the package already raw-fetches /v4/profile and /v4/conversations the same way, so this follows the established pattern rather than forcing a risky SDK major bump.
• documentList returns response.memories (still exposed as documents in the tool result, so consumers are unaffected) and its schema now exposes the API's real page-based pagination. The old offset/status params were silently ignored by the API — passing offset: 10 returned the same first page — so they were removed rather than kept as dead knobs.
• str_replace rejects only a missing new_str (empty string is the documented way to delete text); same for insert_text. old_str still must be non-empty since replacing "" would prepend.
• delete and rename now call documents.delete on the backing document (rename skips the delete when both paths normalize to the same customId, since the add already replaced the content).

Testing

• New offline unit suite src/tool-operations.test.ts (13 tests) mocks the SDK the same way the existing claude-memory.test.ts does, plus stubs global fetch for the forget endpoint, and asserts the exact call shapes: delete receives a string, list forwards page and returns the memories array, forget issues DELETE /v4/memories with the right body/auth header, empty-string new_str/insert_text work, delete/rename actually delete. 9 of the 13 fail against the previous code (the 4 that pass are negative-path guards that were already correct).
• Existing claude-memory.test.ts still passes; the stale getToolDefinitions().length assertion in tools.test.ts is updated from 2 to the actual 7.
• tsc --noEmit error count drops from 155 to 148: the diff removes all eight SDK-mismatch errors (TS2345 on delete, TS2339 Property 'documents'… ×3, TS2339 Property 'forget'… ×2, plus two implicit-anys) and introduces none. The remaining 148 are the pre-existing zod/ai schema-typing incompatibility that hits every tool() call in the package — separate issue, not touched here.
• biome check clean on all touched files (two pre-existing noExplicitAny warnings on lines I didn't change).

The live integration tests in tools.test.ts need SUPERMEMORY_API_KEY/OPENAI_API_KEY, which I don't have — happy to iterate if a staging run surfaces anything, but the call shapes are verified against the SDK's type definitions and the repo's own API docs (list-memories/overview.mdx shows the response keyed "memories").

cc @MaheshtheDev

---

Session Details

• Session: View Session
• Requested by: Unknown
• Address comments on this PR. Add (aside) to your comment to have me ignore it.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[graphite-app]

The rule 'Include any supporting data as the cause argument instead of inlining into the string' is violated here. The error is constructed by inlining the status code, status text, and error body directly into the error message string. Instead, the supporting data (status, statusText, errorText) should be passed as the cause option to new Error(). For example:

throw new Error('Supermemory forget memory failed', {
  cause: { status: response.status, statusText: response.statusText, body: errorText },
})

Suggested change

	throw new Error(
	`Supermemory forget memory failed: ${response.status} ${response.statusText}. ${errorText}`,
	)
	throw new Error('Supermemory forget memory failed', {
	cause: { status: response.status, statusText: response.statusText, body: errorText },
	})

Spotted by Graphite (based on custom rule: TypeScript style guide (Google))



Is this helpful? React 👍 or 👎 to let us know.

[vorflux]

Testing

Targeted validation passed for the internal packages/tools SDK wrapper fixes. The testing subagent classified this as an internal tooling/library bug fix with no UI or mobile testing required.

Commands run:

cd packages/tools && bun run test src/tool-operations.test.ts src/claude-memory.test.ts
cd packages/tools && bun run build
bunx biome ci packages/tools/src/ai-sdk.ts packages/tools/src/claude-memory.ts packages/tools/src/openai/tools.ts packages/tools/src/shared/forget-memory.ts packages/tools/src/tool-operations.test.ts packages/tools/src/tools-shared.ts packages/tools/src/tools.test.ts

Result:

Targeted tests: PASS — 16 tests passed.
Package build: PASS — `tsdown` built JS and `.d.ts` outputs, including `dist/shared/forget-memory.js`.
Changed-file Biome gate: PASS exit 0 with 2 warnings in `claude-memory.ts` about existing `any` usage.
Worktree after build: clean.

Verdict

✅ Passed. Targeted tests, package build, and changed-file lint/format checks all completed successfully for the PR-relevant files.

[vorflux]

Summary

Reviewed the SDK call-shape fixes across packages/tools, including the AI SDK, Claude memory, OpenAI tool wrappers, shared forgetMemoryRequest, and updated tool operation coverage. The review checked the changed diff against the installed supermemory SDK types and documented DELETE /v4/memories behavior.

Verdict

✅ Reviewed — no issues found. The changed wrappers align with the SDK/API contracts and no bugs, breaking changes, security issues, or data-loss risks were identified in the reviewed diff.