Command

PePersonal JEJust Electrolyte RTRelease Timer TtTacTile CSClearSource AtAthletic

TacTile›Launch TacTile›MCP surface·47 in this wave

Approve: MCP surfaceTo do · blocker 13 production secrets overdue for rotation — the gate cannot be green until this is resolvedTo do · blocker Deploy step 2 — verify/rotate production MCP Worker secretsTo do Deploy step 4 — fix the stale "NOT DEPLOYED" banner and deploy docsDone Add action.* and bluetooth.* op groups to the MCP catalogDone Build real handlers for group.* (9 ops)Done Build real handlers for family.* (7 ops)Done Build settings.* MCP tools including custom domain managementTo do Verify/build the Connected Apps management UI for MCP OAuth clientsDone Add rate limiting, quota, and cost controls to the hosted MCP serverDone Honesty check — reconcile investor deck MCP claims against what shipsTo do Verify "everything a human can do, an agent can do over MCP" end to endTo do staging-api is a day behind main — and writes to the production SupabaseDone MCP GAP INVENTORY — every unwired op, with the decision attached to eachDone Demo account for screenshots + QA — seeder is live, shape is frozenTo do SECURITY: hosted MCP tool agent.approval.resolve may let the model approve its own proposalsDone Approvals: resume re-executes approved rows, REST resolveApproval is non-atomic, inbox.ts selects a missing expires_atTo do Build the full medialink.* MCP tool set (3 ops)Done Build the payment.* connector MCP tool set (3 ops)To do Build the full webhook.* MCP tool set (4 ops)Done Build zap.* MCP tools or explicitly cut Zapier from v1.0 MCP scopeTo do Build link.schedule.repoint, link.compound, wizard.suggest or cut themTo do QA: the public MCP tools page opens without sign-in and lists the tools (desktop web and phone)Blocked Deploy step 1 — verify DB migration + RLS on OAuth tablesDone Deploy step 3 — confirm CF route/DNS for mcp.tactile.to is production-gradeTo do Deploy step 5 — formalize the MCP smoke testTo do Build nfc.lock and nfc.unlock MCP toolsTo do Build contact.vcard.export and contact.delete MCP toolsDone Build linkhub.create and linkhub.delete MCP toolsDone Build videoaudio.create and videoaudio.replace MCP toolsTo do Build document.create and document.replace MCP toolsTo do Build the full social.* MCP tool set (6 ops)To do Build the full paylink.* MCP tool set (4 ops)Done Build wifi.edit MCP toolDone Build the full form.* MCP tool set (7 ops)Done Build the analytics automation layer MCP tools (6 ops)Blocked Wire inbox.list and inbox.mcp_approval to real handlersDone Add architecture test guarding against a second tool-handler implementationDone Publish a public MCP tool manifest / discovery docsDone Webhook signing secret: add the one surface that can reveal itDone A Website destination's URL cannot be edited through the agent surface at allDone SPEC → backend: webhook secret reveal/rotate endpoints + secret_fingerprint on the listDone NEEDS MATTHEW: decide whether resolve_trigger_destination keeps anon EXECUTEDone 5 proposal-only MCP ops are advertised with no task row building themTo do contact.delete MCP tool has no destination-type guard: any hash the caller owns is hard-deletedDone Add rule E to the single-tool-handler-path test: a handler map keyed by tool name must go red (follow-up to #698, t_27dbb775)Done notes.delete, document.delete, menu.delete and videoaudio.delete soft-delete any destination type the caller owns, and report success for a hash that matched nothingDone

Wave

MCP surface

Thesis pillar 3 — "Agentic AI + MCP." The investor deck claims first-in-category MCP-native. Audit (2026-07-26) found the hosted server IS live at mcp.tactile.to with real OAuth production traffic (14 clients, 11 tokens) — but the code's own deploy banner still says "NOT DEPLOYED" (stale), and tool coverage is genuinely incomplete: roughly half the 147-op catalog are proposer stubs that never mutate state, and 2 of 14 destination types (Action, Bluetooth) have zero MCP presence at all. Not cuttable: this is a stated structural moat.

TacTile › Launch TacTile › MCP surface

27/47 done20 open2 open blockers

Selected taskclear

notes.delete, document.delete, menu.delete and videoaudio.delete soft-delete any destination type the caller owns, and report success for a hash that matched nothing

DoneOwner · Vivt_35cf60f9

Found 16 Oct while auditing contact.delete's siblings (t_ec1ab373). packages/mcp-server/src/tools/destination-content.ts:56-63 softDeleteDestination PATCHes `destinations?hash_id=eq.<hash>` with deleted_at and NO type filter and NO return=representation read-back; makeSoftDeleteTool (:65-81) is notes.delete (:144), document.delete (:186), menu.delete (:434) and videoaudio.delete (:511). So notes.delete with a Wi-Fi, payment or website hash the caller owns sets deleted_at on it (the public link stops resolving), and any of the four returns ok:true for a hash that matched no row. Soft, not a hard delete, so recoverable by clearing deleted_at, but no MCP tool clears it. form.delete (form.ts:653) and linkhub.delete (link-hub.ts:351) already do it right: `&type=eq.form` / `&type=eq.link_hub` in the PATCH filter plus `return=representation` and an error on zero rows. destinations.delete (destinations.ts:413) is the intentionally generic soft delete and is not this bug.

No dependencies in this wave — nothing gates it, and nothing waits on it.

Show the prompt

You are Viv, the agent for the TacTile lane of Command.

Open the lane and work there:

```bash
cd ~/Command/tactile && claude
```

Read that lane's CLAUDE.md and lane.json before you touch anything — they
carry your identity, your scope, and the hazards specific to this lane.
Read state from the hub (https://internal.tactile.to) or ask. Never guess at it.

## The task

**notes.delete, document.delete, menu.delete and videoaudio.delete soft-delete any destination type the caller owns, and report success for a hash that matched nothing** · done

Found 16 Oct while auditing contact.delete's siblings (t_ec1ab373). packages/mcp-server/src/tools/destination-content.ts:56-63 softDeleteDestination PATCHes `destinations?hash_id=eq.<hash>` with deleted_at and NO type filter and NO return=representation read-back; makeSoftDeleteTool (:65-81) is notes.delete (:144), document.delete (:186), menu.delete (:434) and videoaudio.delete (:511). So notes.delete with a Wi-Fi, payment or website hash the caller owns sets deleted_at on it (the public link stops resolving), and any of the four returns ok:true for a hash that matched no row. Soft, not a hard delete, so recoverable by clearing deleted_at, but no MCP tool clears it. form.delete (form.ts:653) and linkhub.delete (link-hub.ts:351) already do it right: `&type=eq.form` / `&type=eq.link_hub` in the PATCH filter plus `return=representation` and an error on zero rows. destinations.delete (destinations.ts:413) is the intentionally generic soft delete and is not this bug.

Where it sits: TacTile › Launch TacTile › MCP surface

## The brief on file

[2026-10-17 · Viv] r35 CLOSE-OUT 19:20 PDT, CI re-read, stays DONE: draft PR #753 (https://github.com/TacTile-dev/TacTile/pull/753, head 5c35ed73 live, MERGEABLE, not merged, not deployed) now shows its CI finished. verify fails at exactly one step, "Secrets rotation audit" (the sanctioned red, t_3e28c5b5, a calendar check on the whole repo, not a function of the 3 files in this commit). playwright, changes, dep-scan, semgrep and static-firewall pass; swift-parity is skipped. Nothing red under the diff. The closer's earlier "re-read CI before anyone treats this as green" is now answered.

[2026-10-17 · Viv] DONE on draft PR #753 (https://github.com/TacTile-dev/TacTile/pull/753), commit 5c35ed73 (head of r30/api-backend-r2-ship, pushed 27dcddbc..5c35ed73, no force). notes.delete, document.delete, menu.delete and videoaudio.delete now PATCH destinations with type=eq.<type> (videoaudio: type=in.(video,audio)), select=hash_id and Prefer return=representation, and throw on zero rows, so a wrong-type hash, an unknown hash and another account's hash all refuse and none reports success; blank hash refused before any request (packages/mcp-server/src/tools/destination-content.ts, copying form.ts:670-677). Evidence: new tests/unit/mcp-destination-content-soft-delete-guard.test.ts 101 green incl. 51 wrong-type cases; mutation 1 (drop type filter) 64 red, mutation 2 (drop zero-row throw) 71 red, both restored; acceptance reader also red on 3 more mutations (videoaudio narrowed 16, hash unencoded 4, return=minimal 9), restored. Related files 276 green, verify:unit 240 files / 5023 tests green. Fresh closer run 17 Oct 13:47 at load1 14: guard + mcp-contact-delete-guard + mcp-destub-r18 = 3 files / 157 tests passed. Gate run 1 of 3 on 5c35ed73: exit 0, 55 of 56 passed, only the known-red verify:secrets-rotation. Readers: acceptance CONFIRMED, safety CONFIRMED (3 files only, no hub or internal-hub, no AGENTS.md/CLAUDE.md root, no migration, no secrets, fetch stubbed so no network). CI on #753 at 13:47: semgrep, changes and dep-scan pass, swift-parity skipped, static-firewall, verify and playwright still pending, none red: re-read CI before anyone treats this as green (verify:ts-source-coverage has a known 1-in-3 node SIGSEGV flake, re-run that check alone if it is the only red). NOT MERGED, NOT DEPLOYED, no production call, no simulator; the fix reaches users only when Matthew merges #753 and the MCP server deploys. Wording note: the refusal text uses an em dash (U+2014) like form.ts:676 and link-hub.ts:374, not the ASCII hyphen the old spec printed; the test pins it. ui_check none, so no QA card.

[2026-10-17 · Viv] SHIPPED to #753 (draft): 5c35ed73 pushed 27dcddbc..5c35ed73 (fast-forward, no force), gate run 1 of 3 passed 55 of 56 with only verify:secrets-rotation red, exit 0, started at load1 6.77 (CI idle, no xcodebuild or Gradle). Both readers CONFIRMED (acceptance and safety). PR body, title and a comment updated: https://github.com/TacTile-dev/TacTile/pull/753#issuecomment-6240764089. Not merged, not deployed, no production call. Status left as is for the closer; supersedes the BUILT-not-yet-gated line below.

[2026-10-17 · Viv] BUILT, not yet gated or pushed: local commit 5c35ed73 on r30/api-backend-r2-ship (#753's branch, was 27dcddbc). softDeleteDestination now takes the type(s) and copies form.delete: PATCH destinations?hash_id=eq.X&type=eq.<type>&select=hash_id (videoaudio: type=in.(video,audio)), Prefer return=representation, throws "No <type> <hash> found on your account — nothing was deleted." on zero rows; blank hash refused. Types confirmed from tools/destinations.ts:49-50 and notes.ts:102, menu.ts:145, destination-content.ts:484. New tests/unit/mcp-destination-content-soft-delete-guard.test.ts: 101 green; mutation 1 (drop type filter) 64 red incl. all 51 wrong-type cases, mutation 2 (drop zero-row throw) 71 red; both restored. Related files 276 green; verify:unit 240 files / 5023 tests rc 0; tools/AGENTS.md updated (it stated the old behaviour). Next: readers (acceptance, safety), gate, push, #753 body. destination-content.ts is in no other open PR. No production call.

FIX: in packages/mcp-server/src/tools/destination-content.ts give softDeleteDestination a `type` parameter (notes -> notes, document -> document, menu -> menu, videoaudio -> video and audio, as an `in.(video,audio)` filter) and copy form.delete's shape (form.ts:653-678): PATCH with `&type=eq.<type>&select=hash_id`, `prefer: "return=representation"`, throw `No <type> <hash> found on your account - nothing was deleted.` when zero rows come back. makeSoftDeleteTool takes the type per tool. TEST: new tests/unit/mcp-destination-content-soft-delete-guard.test.ts with a stubbed fetch (599 for anything unstubbed), the in-memory destinations table pattern from tests/unit/mcp-contact-delete-guard.test.ts (t_ec1ab373, on #753): each of the four deletes its own type with one PATCH carrying the type filter; a wifi hash passed to notes.delete throws and the PATCH matches no row; an unknown hash throws; mutation (drop the type filter) turns the wrong-type cases red. destination-content.ts and tests/unit/mcp-destub-r18.test.ts are in no open PR at 16 Oct (re-check); tests that pin the old PATCH URL may need the filter added. DONE WHEN: the four tools soft-delete only their own type, throw on zero rows, mutation red then restored, verify:unit green. Draft PR, no deploy, no production call.

## How to work this

- Do the work, don't just plan it. If the fix is obvious and safe, make it.
- Surface anything needing Matthew's sign-off instead of deciding it yourself.
- Short replies. Lead with the answer, then stop.
- Finish by saying what changed, what you verified, and what you'd do next.

To doIn progressDoneBlockedPick a wave on the left · tap a task to expand it

Lanes

MCP surface

notes.delete, document.delete, menu.delete and videoaudio.delete soft-delete any destination type the caller owns, and report success for a hash that matched nothing