Skip to content

Latest commit

 

History

History
1288 lines (754 loc) · 87.8 KB

TODOS.md

File metadata and controls

1288 lines (754 loc) · 87.8 KB

TODOS

Source links on cards (from the CEO review of docs/designs/ask-me-source-links.md, 2026-11-08)

Links to connector records that come with only an id

What: Build a record's web link for each connector whose tools return only an id, HubSpot first: from the account and the record's id and type, make the link to its page.

Why: Until then a record with no web address shows as a Named only source ("Masa Global · HubSpot company 1234") the owner can't click. Owner deferred this (SRC-3) because link formats differ by service and account region, and a wrong link opens the wrong record.

Context: sourcesOf in src/shared/askSources.ts reads Named only entries as {"label", "id"}. Check each service's documented record URL against a live account before building, and keep the Named only fallback when the account details are unknown.

Effort: M (human: M / CC: S) Priority: P2 Depends on: the Sources row (docs/designs/ask-me-source-links.md)

A source may carry the line it backs up

What: Let a source carry one short quoted line (up to about 200 characters), shown on hover and in See full context.

Why: A link opens a long page, and the owner still has to hunt for the sentence the card rests on. Owner deferred this (SRC-6) until real Sources rows show which kinds need it.

Context: It would be an optional quote field on a source, rendered as plain text. Members must quote exactly, never paraphrase.

Effort: S (human: S / CC: S) Priority: P3 Depends on: the Sources row

Ask me notices and leftovers (from the ship review of mit/ceo-review-workflow-clarity, 2026-11-05)

A skipped kind still surfaces when it has real stakes

What: Word the "Don't show me these" memory lines (the member's and Michael's) so a notice of that kind that carries real stakes (money, a deadline, a rejection or suspension, security or legal) still gets a card.

Why: The kind label is the member's own words, so a later App Store rejection filed under a kind the owner turned off for routine results is held back by both the member and Michael. Red team, pass 2.

Effort: S Priority: P2

Wider phrasings for the two card checks

What: Teach the email link check possessive and verb forms ("Will's email", "Dana emailed us", "the message Will sent") and to leave Slack and chat threads alone; teach the change option check "Approve with my changes", "Send with my changes, then show me again", and to apply a negation only to the words just before "show me again".

Why: Both checks are a backstop to Michael's rules and only read English words; these phrasings slip past or, for one compliant label, flag it until he uses the canonical wording. Ship review pass 2, 2026-11-05.

Effort: S Priority: P3

The Ask me count can run one ahead for a moment

What: Count the case study and update cards from the same check the cards render with (the case study screen open, a restart in progress).

Why: While the case study tab is open, or while a restart is under way, the pill and the strip count a card Ask me hides. It clears when main hears the case study was seen, or when the app restarts. Red team, pass 1.

Effort: S Priority: P3

Memory and voice trust gaps (from the 0.2.7 ship review, 2026-10-26; owner chose TODOS)

Owner standing per note, not per tidy pass

What: Have each owner sourced op name the owner note it comes from, and accept source owner only for an op that cites a note under FROM THE OWNER in this pass.

Why: validateOps gets ownerWords for the whole pass (src/main/memoryTidy.ts). When any owner note is pending, an op the model built from an inbox line can claim owner, and only the prompt keeps the two apart. Michael gets an owner note for every dock message, so his passes almost always run that way.

Go with it: keep only the owner's own words (the answer, the edit, the note) in owner_notes.text and the quoted context (the member's question, a card title, an email subject, pasted text) in a field of its own that the tidy prompt labels as not the owner's; today one line in the instructions asks the model to tell them apart (D4). And file an owner note only when a pass used it, or keep the ones it set aside for a later pass, so a planted inbox note can't talk the tidy-up into dropping the owner's words (adversarial review, DeepSeek, 2026-10-26).

Effort: M Priority: P2

Owner entries in memory.md can be planted

What: Keep the ids (or a hash) of the owner entries the tidy-up writes in harness.db, and read any other entry marked owner as the member's own.

Why: memory.md sits in the member's folder, which its shell can write; parseIndex trusts any | owner | line, and the new rules then protect it. Older than this branch.

Effort: M Priority: P2

A correction typed in a member's own terminal can't replace an owner entry

What: Record what the owner types to a member as an owner note, the way the dock does for Michael, or give the member an app path that writes one.

Why: Members still write the owner's corrections to their inbox, which never counts as the owner's, so an older owner entry (from an Ask me answer) can't be replaced that way.

Effort: M Priority: P2

Run comparison reads transcripts it finds by name

What: Record each member session's transcript path (and size) when the app starts the session, and have run comparison read only that file, instead of finding <session id>.jsonl anywhere under Claude's projects folder.

Why: On a Mac members' shells are walled off from that folder, but on Windows a member's shell could write a file under its own session id. The member's own transcript already holds whatever it chose to say, and every suggestion still waits for the owner's Review and Apply, so this is defense in depth. The injection screen on a suggestion is a pattern list, not a boundary; the owner's Apply is (adversarial review, DeepSeek, 2026-10-26).

Effort: S Priority: P3

Voice confirms for schedules and settings carry no target

What: Carry the staged schedule id or setting key on the pending confirm, and cancel when the confirm names another one, as namesSomeoneElse does for members.

Why: "confirm, delete the Invoice check schedule" while a delete of Morning report waits deletes Morning report. Older than this branch. Also: a member whose name is an ordinary word ("Will") can cancel a valid confirm; it fails safe.

Effort: S Priority: P2

Overlap split (deferred from the CEO review, 2026-10-24)

Explain how they split it in the hire wizard

What: Give the hire wizard's overlap step the same "Explain how they split it" that Edit agent has, writing the new hire's handoffs.

Why: A new hire who works with a teammate (sends them part of the work) hits the same wall a changed job did: cut the duty or guess at wording.

Pros: One way of describing teamwork everywhere a job is written.

Cons: The wizard's overlap box has mailbox, topic and connection chips, and its topic chip writes the boundary the opposite way ("Only X." on the new hire, "Not for X" on the teammate); its Next gate needs the explanation to count. About M.

Context: Edit agent's version: JobChangeReview's explain state, the revise call's explanation and teammates, checkDistinct mode 'change', and the handoff list in src/shared/jobChange.ts (docs/designs/work-style-change-request.md, Overlap split). Deferred by the owner (CEO D8) to prove it in Edit agent first.

Effort: M Priority: P3 Depends on: the Edit agent version shipping

Release notes accuracy (from the eng review of the site's 0.2.5 and 0.2.6 plan, 2026-10-24)

The 0.2.5 changelog says engine records are closed to every member's shell

What: Reword CHANGELOG.md's 0.2.5 line "closed to every member's tools and shell, Michael included" to what the app does on every platform.

Why: On Windows Claude Code has no sandbox, so shell commands are not limited; only Claude's own file tools are walled there (src/shared/folderAccess.ts:16-20). As written, the line is a privacy claim that is false on Windows, in public release history.

Context: Commit 8f6b9038 (#387) walls the engine records: folderAccess.ts:97-101 adds them to the sandbox deny lists and to the Read and Edit deny rules for every agent, Michael included, and test/folder-access.test.cjs covers it. The site's wording is the model: "The app keeps each team member's conversation records out of every other member's file tools, Michael's included." The published v0.2.5 GitHub release body carries the same words and needs a hand edit too; release bodies are built from RELEASE.md on the tag run, so edit the published body directly rather than re-running the tag workflow.

Effort: S Priority: P2 Depends on: None

Connections

Step cards for iCloud, Yahoo and Zoho app passwords

What: Give iCloud Mail, Yahoo and Zoho the same numbered step cards Google gets in Add a mailbox (eng review 2026-10-26, D3 B): each step with an Open button to that service's own two step and app password pages, and a plain line on what to copy.

Why: Each still shows one paragraph with an address to type; the owner saw a business user struggle with the same paragraph for Google.

Pros: One look across Add a mailbox; owners on those services set up without help.

Cons: Three more providers' pages to find, link and keep current when they move.

Context: The Google cards (AddMailboxDialog.tsx, Google steps) are built to take a per provider list of steps; start from mailboxes.help.icloud/yahoo/zoho in the locales and add each provider's steps and page addresses.

Effort: human S / CC S Priority: P3 Depends on: the Google step cards landing.

One call at a time per connected app, and a throttle notice

What: The broker now backs off per connected app after a 429 or a 503 with Retry-After (src/main/brokerThrottle.ts, PR #399): a wait of 20 s or less is held, a longer one is answered at once with when to try again. Two parts are left. Send one call at a time per app, so the calls held during a wait don't all go out together when it ends and hit the limit again. And show the owner a throttle that lasts, in plain words, instead of it showing only in agents' answers.

Why: Gmail, Google Business Profile, Meta and QuickBooks all throttle. Held calls released together can trip the limit straight away, and a long throttle is invisible to the owner today.

Context: Decision 22 (business-mode-office-packs.md) named a serialized executor with at most one call in flight per connection as the home for this; Decision 25 named the attention strip, which does not exist yet.

Effort: M Priority: P2 Depends on: Decision 2 (per-agent grants), Decision 22 (serialized executor); only relevant from Phase 2, when agents make real third-party calls.

Research agent (deferred from the third ship of mit/named-links, 2026-10-13)

Catch money amounts in the privacy check's pattern recheck

What: Add a pattern for money amounts to the recheck in src/shared/dataClasses.ts: digits with a currency mark or code before or after them (Rs, INR, ₹, $, USD, €, £ and the like, with the groupings people write, such as 1,35,000 or 2,000.50). Today the only money pattern is the card number check.

Why: Found by PR 411 (Tejashree0405), confirmed on main 2026-10-18: when the model's cleaned rewrite of a message to Pete keeps "Rs 1,35,000", "₹5,000" or "$2,000", the recheck lets it through, so business figures (private by default) rely on the model alone. In the contributor's live runs, one of 30 private messages kept a payroll figure.

Behavior change: Pete gets more messages refused or trimmed: a message to Pete that names an amount is cleaned or refused even when the model passed it. Tests should keep years, counts and percentages passing.

Effort: S Priority: P2

Give Pete mail folders it can only read

What: Keep the Research team member's inbox and the outbox the app reads in folders only the app writes, and give Pete a way to mark a message handled and to hand in an answer that does not need its shell to write there.

Why: Three review passes and two Codex passes each found a new way for a steered Pete to trick the app through its own folders: links, pipes, hard links, forged messages, swapped folders. Each is fixed and tested (src/main/hive.ts researchDir, researchDelivered, dropResearchEntry; src/main/researchNotes.ts ownResearchDir), and its shell now writes only inbox, outbox, memory and its desk (verified live). The review still ended with converged false (owner, third ship D6): the class of attack stays as long as the app acts in folders Pete writes.

Context: The app delivers stamped copies already (researchStamp), so an app owned inbox Pete reads is a small step; marking handled could be a message to the app instead of a move into .done. The outbox needs a drop folder the app moves out of at once, or a hook on Pete's Write tool that hands the file to the app.

Effort: M Priority: P1 Depends on: None

Refuse the web tools when a member's set up fails

What: When hive set up fails for a team member other than Research (src/main/index.ts, [hive] ensureAgent failed), the member still starts, with no hooks and no deny rules, so its web page and web search tools are not refused. Either refuse the spawn, as Research already does, or pass --disallowedTools WebFetch WebSearch on that path.

Why: The site and /security say the app turns off the web page tool and the web search tool for every team member but Pete. On this path it does not, on Mac or Windows. Found in the website's 2.8.1.0 ship review (2026-10-20); moved here from the site's TODOS when its guardrail item closed (site CEO review, owner D10, 2026-11-04).

Effort: S Priority: P2 Depends on: None

Take the internet from team members' shells on Windows

What: Hold team members' shell commands off the internet on Windows, where the engine has no sandbox. Until then, the Research screen says so on Windows (done 2026-10-18).

Why: On Windows only the web tools are refused; a shell command can still reach the internet, so the rule that only Pete uses the internet is not enforced there. The CHANGELOG names this.

Effort: M Priority: P2 Depends on: A Windows shell sandbox, or a firewall rule per team member.

Wall members' notes from every engine's shell, and on Windows

What: Keep a team member's shell from reading another member's hive folder (agents//) and the hive's git history on every engine, not only Claude on a Mac and Linux, and on Windows.

Why: Member memory is private (owner, 2026-10-16), and the walls are Claude Code's sandbox Read rules plus the hook on file tools. Codex, Gemini and the other engines get no such shell rule, and Windows has no sandbox at all, so a shell command there can still read a teammate's notes; on Windows that includes Research, which can reach the internet from its shell. Found by the ship review and the Codex adversarial review of the running branch (2026-10-17, owner D3.3).

Context: The prompts tell every member to ask instead; the file tool hook refuses everywhere. Each engine's own sandbox config, where one exists, is the place to add the rule; Windows needs the same shell sandbox as the item above.

Effort: M Priority: P2 Depends on: A Windows shell sandbox for the Windows part.

Keep a teammate's question to Pete across a quit while its check waits

What: Save a teammate's own message to Research while it waits for the privacy check (up to 30 minutes), as the owner's and the app's notes are saved, and tell the sender and Michael after a restart.

Why: queueForResearch writes only the owner's and the app's notes to research-pending.json; a teammate's question that waits through a quit or a crash is lost with no word to anyone. The wait used to be seconds; it is now up to 30 minutes, mostly while the team starts. Found by the ship review red team (2026-10-17, owner D3.3).

Effort: S Priority: P3 Depends on: None

Team performance (from the CEO review of docs/designs/team-performance.md, 2026-10-15)

Compare runs for members on other engines

What: Read Codex (and other engines') session files for a run's window, so "Went wrong" can compare their runs too.

Why: Run comparison (src/main/runCompare.ts) reads only Claude Code's conversations, so members on other engines get no suggestions; the jobs list shows an info icon for them.

Context: From the eng review of "Compare runs we already have", 2026-10-24 (D8). The slicer finds a member's sessions through the private session record and reads <claude projects>/<folder>/<session>.jsonl; pulseBackfill.ts has the same limit. Each engine needs its own folder and line format, and its own runEntry reading.

Effort: M (human) / S (CC) Priority: P3 Depends on: Run comparison (this feature)

Give Fable a tier and a price

What: Add Fable's tier (top, middle or small) to modelCatalog.json and its price rows to pricing.ts, from Anthropic's published pricing.

Why: Model advice in the Team view needs a tier, so members on Fable, the default model (config.ts:525), get no cheaper or stronger model advice until it has one.

Context: The team performance build adds tier to Claude catalog entries and knownPrice(model); a model without a tier gets no model advice. The remote catalog (modelCatalog.ts) must carry the field too. Owner decision D18.

Effort: S (human) / S (CC) Priority: P2 Depends on: The team performance build

What: Read transcripts for the backfill line by line with a yield every few hundred lines, skip lines before the window before parsing tool results, and let the 30 s tick drain backfilled points instead of folding up to 100,000 at once; prune telemetry_runs with the events.

Why: On the first launch after an update the backfill and its fold can hold the main thread for seconds while every agent starts and sends hook frames; a very large office log or transcript is read whole into memory (Codex review).

Context: From the /ship performance review, 2026-10-16. Migration 4 added the indexes the hot reads needed; this is the remaining part. pulseBackfill.ts readTranscript and pulseBackfillOnce in index.ts.

Effort: M (human) / S (CC) Priority: P2 Depends on: None

Pulse keeps a card finished just before a quit

What: Record card status changes where Michael's board write lands (or persist the last seen board) instead of only diffing the 8 s poll in memory.

Why: A card that reaches Done in the seconds before the app quits or crashes is taken as the new baseline on the next launch and never counted in Tasks done.

Context: From the Codex adversarial review, 2026-10-16. telemetryStore.ts tasks(): the first read of an office only remembers.

Effort: S (human) / S (CC) Priority: P3 Depends on: None

Key Pulse by a stable office id, not its folder path

What: Store an office id in the office and key telemetry rows by it, or re-key rows when Settings moves the office; resolve the path with realpath meanwhile.

Why: Moving the office, or reaching it by another spelling (a symlink, a different case), splits Pulse history and runs the backfill again under the new key.

Context: From the /ship red team review, 2026-10-16. pulseOffice() in index.ts returns expandTilde(harnessHome).

Effort: M (human) / S (CC) Priority: P2 Depends on: None

Hand the hook secret over where another process cannot read it

What: On engines and platforms where a sibling process can read another's environment (Linux /proc, Windows), pass the hook secret through a 0600 file or descriptor only that agent's sandbox can read, or check the peer on the socket.

Why: All agents run as the same OS user, and a sibling's environment is readable: on a Mac, ps eww showed a test HIVE_HOOK_SEAL of a node process (checked 2026-10-16). A member that reads a teammate's secret can send hook frames as them (Pulse numbers, a stuck restart), the forgery #61 closes.

Context: From the /ship security and Codex adversarial reviews, 2026-10-16; the owner chose to ship 0.2.2 with this named limit. Today the secret is in each agent's env (hive.ts HIVE_HOOK_SEAL), read by every bridge shim. A per member file outside the hive, with a sandbox read deny for every other member, or a socket peer check (LOCAL_PEERPID on macOS), would close it; Windows has no sandbox, so it stays weaker there.

Effort: M (human) / M (CC) Priority: P1 Depends on: None

Fill in Pulse's earlier checks off the main process

What: Run the one-time backfill of Pulse's questions and answers lists, and the parse of the 62 day list, off the main process (a worker, or rows in the telemetry database queried by time and way).

Why: The first open of the list scans every member's inbox and sent folders and then parses the whole checked.jsonl on the main process; a large office stalls the window and every timer for that time. The live file is now read without blocking, but the scan and the parse are not. Found by the Codex adversarial review (2026-10-17).

Effort: M Priority: P3 Depends on: None

Retry the backfill when it skipped files it could not read

What: Write the backfill's done mark only after a scan that read every file it meant to, and record how many it skipped, so a scan cut short by files being written or a slow volume runs again.

Why: ensureCheckedBackfill writes checked-backfill.json whatever it read, so a scan that met unreadable files never fills in their details. Links and pipes an agent left are skipped on purpose and must not count as a failure. Found by the Codex adversarial review (2026-10-17).

Effort: S Priority: P4 Depends on: None

Company foundations (deferred from the ship of mit/file-tab-no-bottom-bar, 2026-10-20)

See and put back a foundation answer's past versions

What: A screen, beside the owner's foundations form when it comes, listing each question's past answers (text, coverage, from when, recorded when, by whom), with a way to put one back.

Why: The foundation answers design keeps every past answer (part 6, owner 2026-11-03), but nothing shows them, so a wrong change to a list such as the bank and card accounts cannot be spotted or undone from the app.

Pros: The owner can check what changed, when and who changed it, and undo a bad change without asking Michael.

Cons: A new screen with its own design and tests; it belongs with the owner's foundations form, which is itself later work.

Context: docs/designs/foundation-answers-stay-open.md (part 6 and the eng review, D8). The history lives in foundations.json, written only by src/main/foundations.ts; the renderer reads view() today, which should not carry history to members.

Effort: M (human) / S (CC) Priority: P3 Depends on: The foundation answers build (docs/designs/foundation-answers-stay-open.md)

Name a missing foundation's proof in Early look and Recommendations

What: Once the kept as is evidence exists for a question, add a line to Early look and Recommendations such as "Answer who we sell to: Ryan's posts are kept as is 45% without it, 85% with", built from the same proof the member detail box shows.

Why: The strip in Recommendations names the jobs whose work hit an empty answer, but not what answering it would win; the proof today is only in a member's detail box, one click away.

Context: Deferred from plan: docs/designs/office-foundations.md (section 7, "Early look and Recommendations"), owner choice D2 at /ship, 2026-10-20. The proof is computed in src/main/pulse.ts foundations() and shown by FoundationsBox in PulseView.tsx; advice lines come from src/shared/teamAdvice.ts.

Effort: S (human) / S (CC) Priority: P1 Depends on: None

Read an edited document in again off the main thread

What: Move kg-core reindex (and the index rewrite it does) to a worker or child process, or keep the index in one file per document, so a save's work is bounded to the edited document.

Why: Every save of a document opened from Settings rewrites the whole index.jsonl synchronously on the main process; with a large store one save can stall routing and the window for a moment (Codex adversarial review).

Context: From the Codex adversarial pass at /ship, 2026-10-20 (owner choice D11). Saves now wait 3 s of quiet before a read in (knowledgeEdits.ts), and the index is replaced whole by rename (kg-core.cjs writeAtomic). A copy kept from a broken foundations.json (foundations.json.broken-* in the private folder) is not yet erased by a factory reset.

Effort: M (human) / S (CC) Priority: P2 Depends on: None

Keep the other hidden checks' text out of Claude's history

What: Turn on isolateConfig (src/main/hiddenClaude.ts) for the other hidden checks that paste private text: the Research privacy check (src/main/index.ts, the trust layer, which already sets dropTranscript) and the mail standing fit check. Measure the cost first: each isolated run is about 1.4 s slower and writes about 1 MB to a temporary folder, and the trust layer can run a dozen checks at app launch.

Why: Ship review 2026-10-23 found Claude Code keeps every pasted prompt in its shared ~/.claude/history.jsonl and paste-cache, which dropping the transcript does not touch. The file fit check is isolated now (verified live: nothing written under ~/.claude); a message to Pete and an email's text still land there.

Context: Isolation relies on CLAUDE_SECURESTORAGE_CONFIG_DIR, found in the Claude Code binary and not documented. If a later Claude drops it the session stops at sign in and the check fails closed, which for the trust layer means every message to Pete is held, so check that before widening it. The API key sign in and Windows are untested. The file fit check is isolated on every platform (owner, 2026-10-23), so check a foundation card file upload on Windows first: Claude keeps its sign in in .credentials.json inside the settings folder there, and if the variable doesn't reach it every file is refused as "couldn't check".

Effort: S Priority: P2 Depends on: None

Give the other job helpers the job change's settings

What: convertWorkStyle (plain words) and checkFocusArea in src/main/workStyleConvert.ts still start their hidden session with the office folder's settings and hooks and keep the transcript. Give them what reviseWorkStyle has: settingSources: 'user', disableAllHooks and dropTranscript.

Why: Codex adversarial pass 2 (2026-10-23): the job text reaches any hook of the owner's or the office folder's. The job change itself keeps no isolated folder by the owner's choice (D9: speed; Claude's history is already closed to members), so its prompt stays in the owner's own Claude history.

Effort: S Priority: P2 Depends on: None

A foundation file the card did not take is always removed

What: In saveFileAnswer (src/renderer/src/shell/foundationSave.ts), when the card refuses the answer and kgRemove also fails, keep the document id, say so on the card, and remove it again at the next launch. Also ingest the bytes that were checked, not the path again (src/main/foundations.ts addFile, knowledge.ts ingest), so a file replaced in the seconds between check and save can't slip in.

Why: Codex adversarial passes 1 and 2 (2026-10-23). Both need two things to go wrong at once, so they were left out of the ship.

Effort: S Priority: P3 Depends on: None

Terminal links (from the ship review of mit/named-links, 2026-10-13)

A quoted phrase that starts with a slash links as a path

What: Quoted prose that starts with a real folder still links as one path: "/tmp is full, see notes.md" underlines the whole phrase and ⌘-click finds nothing. PR 378 (2026-10-17) fixed the case with a space next to a slash ("/ see notes.md" now links only notes.md); a phrase like this one has no such space, so the pattern cannot tell it from a real file named with spaces.

Why: A link that leads nowhere reads as broken.

Context: PATH_TOKEN_RE (quotedPathRe) in src/shared/terminalPaths.ts. Fix: check a quoted match exists on disk before underlining it (the chip path already does), and fall back to the file name inside it when it does not.

Effort: S Priority: P3 Depends on: None

Reset (deferred from the ship of mit/named-links, 2026-10-14)

Clear what a reset could not remove on the next launch

What: Reset & start over retries each delete for about a second, then logs what is left. Save what still could not be removed and erase it early on the next launch, telling the owner if it still cannot go.

Why: The reset erases right after it stops the team without waiting for them to exit; a slow agent can write a file back, and on Windows a file still open can't be deleted, so a few leftovers can stay while setup looks fresh. A saved list of paths to delete at startup needs checks that it only names the app's own folders.

Effort: M Priority: P3

A memory reset clears each other engine's own history before that engine ships

What: Before Codex, Grok, Antigravity, Crush, OpenCode or Pi joins BUILD_ENGINES, Reset memory deletes that engine's own conversation history for each member (Codex sessions and archived_sessions and its state database; .gemini-hive, .pi-agent, .opencode, .crush-data), never the engine's sign in or settings.

Why: Reset memory clears Claude members fully; a member on another engine gets a new conversation but its old history stays in its own folder, where it can look back (owner, 2026-10-22, D8).

Context: docs/designs/reset-agent-memory.md row R4. History locations: src/main/hive.ts:5204, :5336, :5455, :5492, :5540. Add it to each engine's ship checklist with a test that the history goes and the sign in stays.

Effort: S (human) / S (CC) per engine Priority: P3 Depends on: The memory reset; the engine joining BUILD_ENGINES

A memory reset also clears a member's conversations from earlier working folders

What: Reset memory deletes a member's recorded Claude Code conversations and auto memory only in the working folders the roster names today (current, restorable and archived, worktrees included). A member moved to another folder earlier keeps its old conversations and auto memory in the old folder's Claude Code project.

Why: If the owner points that member back at the old folder, Claude Code's auto memory for it loads again, so the member is not on its first day there (Codex adversarial pass at ship, 2026-10-22). Its old conversations can't be resumed (the session ids are cleared), so the risk is the auto memory.

Context: src/main/memoryReset.ts foldersFromRoster and memoryResetTargets. Options: record each member's past working folders in the private record beside its sessions, or delete <sid>.jsonl under every project folder for the trusted session ids. Either way, nothing a member can write may name a folder (R5).

Effort: S (human) / S (CC) Priority: P2 Depends on: The memory reset

Attachments (from the Codex adversarial review of design/studio-v2, 2026-10-01)

Keep each agent's mail pass out of other agents' reach

What: Move each agent's md-mail.mcp.json (it holds that agent's broker token) out of the shared hive folder, or block other agents from reading <hive>/agents/<other>/md-mail.mcp.json, and delete it when the agent's terminal closes.

Why: Sandboxed team agents cannot reach the broker, but Michael runs unsandboxed and could read a Can send agent's token and send from its mailbox. Found in the pre-landing security review (src/main/index.ts, the md-mail block before ptyManager.spawn).

Effort: S Priority: P1

Read one email without downloading its attachments

What: MailService.read fetches the whole message, attachments included, before trimming the text to 50K. Fetch the body structure plus the text part only, and take attachment names and sizes from the structure.

Why: A 25 MB message is held in memory in the main process for one read.

Effort: M Priority: P2

Original email view (deferred from the CEO review of docs/designs/original-email-view.md, 2026-10-29)

Show the whole conversation in the email tab

What: Under the one email the tab opens, list the earlier messages of the same conversation, collapsed, newest at the top, found from its In-Reply-To and References headers.

Why: A reply often only makes sense next to what it answers (our nudge, the sender's first message); today the owner still needs the mailbox for those.

Pros: Full context in one place; reuses the email tab, the owner fetch and the header lookups.

Cons: More folder searches per open (each earlier Message-ID), slower open, more failure paths; Gmail threads and other providers' references differ.

Context: Owner deferred it (CEO review D5, 2026-10-29) so the single email view ships first and is used. Start from MailService.original and the thread headers threadOf already reads (src/main/mail.ts:762).

Effort: human M / CC S Priority: P3 Depends on: the email tab shipping.

Open or save an email's attachments from the email tab

Status (2026-11-05): Open is done (docs/designs/source-attachments.md): a click opens an attachment in its own app, held to save_attachment's kind check and size cap, marked as from outside, from the app's own folder cleared at quit. Save to a folder the owner picks is still open.

What: The email tab lists attachment names and sizes; add Save to a folder the owner picks, and Open in the office for safe kinds (PDF, pictures).

Why: The owner still needs the mailbox to open what was attached to the email an ask is built on.

Pros: Everything about the email in one place.

Cons: Brings files from outside senders onto the Mac; needs its own checks (kind from content not name, size cap, no launching executables, quarantine flag on macOS).

Context: Owner deferred it (CEO review D6, 2026-10-29). Pair with "Read one email without downloading its attachments" under Attachments: download only the one part asked for. Team members can already save attachments themselves (save_attachment, docs/designs/agent-save-attachment.md): reuse its kind check and safe write here.

Effort: human M / CC S Priority: P3 Depends on: the email tab shipping.

Message times and upkeep (deferred from the ship review of mit/file-tab-no-bottom-bar, 2026-10-18)

Read a reply's time from the app's record, not the agent's file

What: The dock and Michael's open questions read Michael's replies back from his outbox and its .sent folder, with the time he wrote. Use the time the app routed it instead: keep it in an app private record (or stamp the .sent copy even when it can't be rewritten in place), and never let a time earlier than the question it answers count as an answer.

Why: The router now keeps an agent's time no later than now, but the files it reads back are Michael's own, and a past time (a timezone slip) can sort a reply before your question or leave a done question open.

Effort: S Priority: P2

Compare To do reminder times by the instant

What: dueReminders and guardCards in src/shared/cardGuards.ts compare nextAt as text. Compare by Date.parse, as waitingChecks now does.

Why: Rewriting a past due time in another spelling (Z or +00:00) sends the Next step due reminder again and resets Michael's backstop.

Effort: S Priority: P3

Keep a schedule decision for Pete while he isn't reporting

What: A schedule decision for Research while it isn't reporting is now dropped quietly with the other upkeep. Keep it (without Michael) and deliver it, with its upkeep kind, once Research reports again.

Why: A decline carries Michael's or the owner's reason; dropped, Research may ask for the same change again.

Effort: S Priority: P3

Michael reply readability (deferred from the eng review, 2026-10-17)

Check the figures in Michael's answer against the team's report

What: In code, check that every amount in Michael's answer to the owner matches, by label and value, an amount in the team report it drew on; a mismatch goes back to Michael, never the owner.

Why: Michael now copies figures exactly by rule (hive.ts owner answer rule), but a retyped or rounded figure still reaches the owner unseen. Research on multi agent systems found the lead agent's rewrite of worker output is where most errors come from, and models are weak at checking numbers themselves.

Pros: A wrong figure is caught before the owner reads it.

Cons: Needs a link from an answer to the reports it used, which the hive does not record today; a check that matches any recent number would flag dates, counts and Michael's own totals.

Context: Owner D9 A in the readability eng review (decision 383ee82c). Start by recording, on Michael's answer, the ids of the team messages it used. The marker gstack-shortcut(dec-383ee82c) sits at routeMessage's toOwner branch in src/main/hive.ts.

Effort: M Priority: P2 Depends on: linking an answer to its source reports

An eval for agent prompt changes

What: A small eval that runs Michael (and one team member) on a few fixed questions, such as a figures question and a compare two options question, and checks the reply's shape in code: answer first, a list or small table, figures copied exactly, currency present, no dashes.

Why: Prompt changes ship with source pins only, which prove the text is in the app, not that agents follow it; the owner is the first to see a regression.

Pros: Catches a prompt regression before an office sees it; makes later wording changes safe.

Cons: Model calls per run; answers vary, so checks must be about shape, not exact text.

Context: Owner D12 A in the readability eng review, 2026-10-17. Prompt rules: src/main/hive.ts (owner answer rule, team report rule), src/shared/ownerRequests.ts ownerQuestionsContext. Start with a script under tools/ that runs claude -p with the built prompt and checks the reply with plain code.

Effort: M Priority: P2 Depends on: None

Mail approvals and Michael replies (deferred from ship of feat/mail-approvals-michael-replies, 2026-10-07)

Rebuild Michael's owner requests from open items only

What: Bound the 8 second rebuild of Michael's owner requests to messages newer than the oldest open item, and keep message bodies out of the never evicted cache.

Why: Since the fixed floor replaced the 30 day window, the work and the cached bodies grow with all conversation history.

Context: Deferred by the owner (ship D5). src/main/hive.ts refreshOwnerRequests and ownerMessages.

Effort: M Priority: P2 Depends on: None

Update notices (deferred from ship of feat/claude-code-update-notice, 2026-09-30)

Secret store (deferred from ship of fix/atomic-secret-store, 2026-09-30)

QuickBooks through the Claude account (deferred from ship of feat/qbo-claude-channel, 2026-09-30)

Stop a nested Claude from reaching the account's connectors

What: Keep agents from starting their own claude (for example claude -p) outside the app's hooks, or turn off the Claude account's connectors for such runs.

Update 2026-10-02 (docs/designs/claude-connectors.md): once that plan ships, an agent with no connector grants starts with ENABLE_CLAUDEAI_MCP_SERVERS=false, which a nested claude inherits, so this is closed for them. It stays open for agents with at least one grant (explicit or the QuickBooks role default): a nested claude loads every connector with no deny rules. Options: a separate Claude config folder per agent, or blocking claude in agent shells.

Update 2026-10-14 (#325): a claude started from an agent's bash shell now gets ENABLE_CLAUDEAI_MCP_SERVERS=false through CLAUDE_ENV_FILE, so the everyday case is closed on Mac, Linux and Git Bash. Still open: PowerShell agents on Windows, and an agent that unsets the variable on purpose; the hard guarantee is a separate Claude config folder per agent.

Why: The QuickBooks switch and Capabilities are enforced by the app's PreToolUse hook, which reaches an agent only through --settings (src/main/hive.ts). A nested Claude started from Bash has no hook but inherits the account's QuickBooks (and email), so the switch off is not a hard guarantee. Stated in CHANGELOG 0.0.11.

Effort: M Priority: P1

Closing time (deferred from ship of feat/closing-time-progress, 2026-09-29)

Rows in a second window

What: Give the closing-time rows names and detail in every window, not only the one that saw the hook events.

Why: Detail lives in the renderer store of the window that received hive:hookEvent; another window shows names but no detail.

Effort: S Priority: P3

Parallel tool calls clear each other's detail

What: Track detail per tool call so one call's PostToolUse does not clear another's.

Why: With parallel calls, the row goes blank while a call is still running.

Effort: S Priority: P3

Replace the source-regex dialog tests with pure helpers

What: Move the checks in test/closing-time-dialog.test.cjs that grep component source into tested helpers.

Why: They break on harmless edits (comments, added props) and passed review cycles several times only after loosening.

Effort: M Priority: P3

Archiving a live agent during closing time

What: Archiving through voice or hive:setArchived drops the agent from the wait list but does not call closingTime.refresh or tell Michael.

Why: Michael can keep waiting for an ACK that never comes until the 6-minute timeout.

Effort: S Priority: P3

Mark the detail line as the agent's own words

What: The detail is agent-written text shown where the owner decides to close without someone.

Why: An agent can write "safe to close without me"; the owner's choice is steered by untrusted text.

Effort: S Priority: P3

Task board (deferred from ship of mit/local-app-version-stuck-ee4e87, 2026-10-12)

A triage filing the app creates itself, so none is ever overwritten

What: Team members send a filing as an outbox message to "triage", and the router creates the file in triage/ with an exclusive create, instead of each agent writing there itself.

Why: The write guard refuses an edit, another agent's name and a file that already exists (src/main/harnessGuard.ts, isPlumbing), but the check runs before the write, so two writes of the very same filing name at the same instant could still lose the first. Codex adversarial pass 2 found it; the owner accepted the risk for this release (ship D11), since agents do not write the same name twice at once.

Context: The router already moves every outbox file (HiveManager.routeOnce); open the new file with wx there and archive the outbox copy, as messages do. Then the triage folder can be the app's alone in the write guard.

Effort: S Priority: P3 Depends on: None

Schedules

Michael turns repeated work into a routine (issue #41)

What: When the same kind of work keeps getting done (for example chasing overdue invoices three weeks running), Michael sets it up as a scheduled job for the team member who does it, with its how in that member's Work style.

Why: Owners rarely notice which work repeats, so it stays manual. Starter jobs only cover what a business type usually needs, not how this office actually works.

Context: The app works out the repeats and Michael decides, so the duty never lives only in his prompt (card-lifecycle rule): group done cards that look like the same job (title, assignee, what was done) and keep each group in front of Michael until he acts or passes. Approval follows the office management rule (owner, 2026-10-10, issue #41): Michael sets up routine jobs himself through the schedule path and adds the how to the Work style; a job that touches money, customers, legal or security goes to Ask me first with the proposed job filled in. Nothing starts without one of those two.

Effort: M Priority: P3 Depends on: nothing.

Agents and 1:1

End a hidden Claude call early when it can't answer

What: A hidden call whose session is stuck on a trust or sign-in screen now waits for its full timeout (60 to 180 seconds). Detect a known blocking screen, or no transcript at all, and give up early so the rules fallback answers quickly.

Why: The owner waits on "checking..." for a minute when Claude is signed out.

Context: From the /ship pre-landing review (2026-09-27): captureAndFinish in src/main/hiddenClaude.ts re-arms until its own answer appears.

Effort: S Priority: P3 Depends on: nothing.

Test Michael deciding schedule requests end to end

What: Runtime tests for receiveScheduleRequest, michaelDecides, decideScheduleRequest and sweepScheduleRequests in src/main/index.ts (unknown id, stale, escalated, pending list, the 12 hour escalation), which today are checked by source patterns only.

Why: These decide whether a request reaches Michael or the owner at all.

Context: From the /ship coverage audit (2026-09-27). Needs the functions moved into a module that tests can load with fake config and hive.

Effort: M Priority: P2 Depends on: nothing.

Hire through Michael

What: The owner tells Michael "hire another admin for catering" and Michael asks the app to open the hire wizard on Review, prefilled with a job he adjusted. The owner still presses Hire; Michael can never hire on his own.

Why: The owner runs the office by talking to Michael, so hiring from chat fits how the app is used.

Context: Deferred in /plan-ceo-review of the hire wizard redesign (2026-09-27, D5.5) because it needs the wizard first (docs/designs/hire-redesign.md). To build: a Michael to app message (an MCP tool, since sandboxed Bash cannot reach the app), a prefill path into the Review step that runs the same distinct job check, and a guard that no agent message can press Hire.

Effort: L Priority: P2 Depends on: the hire wizard redesign.

One rule for "stuck at a prompt" in main and the panel

What: Put the terminal prompt check (isTerminalPrompt in src/main/hooks.ts) in shared code and use it in the renderer's status parsing (src/renderer/src/hooks/useHive.ts), passing notification_type through the hook event.

Why: Michael is told an agent is stuck from the prompt's type; the panel decides from words in the message. For a dialog without those words, Michael hears "stuck" while the panel shows the agent idle, with no coral bar and Talk 1:1 not leading. Found by the red team in the pre-landing review of feat/per-agent-schedules (2026-09-25); the owner chose a follow-up.

Priority: P2 Completed: isTerminalPrompt moved to src/shared/hookEvents.ts and consumed in useHive.ts's Notification hook handler alongside notificationType.

Block mouse and focus reports in a watch-only terminal

What: While a team member's terminal is locked (outside 1:1), drop mouse report and focus report sequences in term.onData (src/renderer/src/components/terminalPool.ts), letting the terminal's own query replies through.

Why: The lock covers keys, paste, drop and composition. If the agent's program turns on mouse tracking, a click in the watch-only terminal still reaches it. Low confidence in the pre-landing review (2026-09-25); needs a check with a mouse-aware prompt.

Priority: P3

Business mode (deferred from plan, v0.0.1 ship)

Deferred from plan: docs/designs/business-mode-office-packs.md (owner chose "ship, defer as P1 TODOs" on 2026-09-24).

Enforcement compiler and the default-deny hook

What: Compile each agent's levels into real permissions: dontAsk, compiled allow/deny rules, a default-deny PreToolUse hook (Decision 4, tests T2/T3).

Why: Today only the prompt layer (teamMemberGoal) limits agents, bypassPermissions is still the default, and the hook fails open. "Asks before sending" is instructed, not enforced.

Context: src/shared/agentDefinition.ts already carries levels and outward capabilities; src/main/harnessGuard.ts shows the PreToolUse pattern. Start by compiling one agent's levels into its per-session settings.

Effort: L Priority: P1 Depends on: None

Approvals inbox (with the two-zone panel and draft-to-post flow)

What: Approval action payloads, an executor, and the approvals UI (Decisions 31 and Ryan's "copy to clipboard and mark posted").

Why: Outward actions (emails, posts, spending) need a place for the owner to say yes; Ask me covers questions only.

Context: Ask me is the first tab on Michael's panel now; approvals likely sit beside it.

Effort: XL Priority: P1 Depends on: Enforcement compiler

Office schedule (officeHours)

What: Honour each pack's officeHours: soft pause outside hours, office state, wired to weeklySchedule.

Why: Validated in the pack schema but never used, so the office runs around the clock regardless.

Context: src/shared/officePack.ts (officeWindow), src/shared/weeklySchedule.ts.

Effort: M Priority: P1 Depends on: None

Attention strip and owner floor layout (Decisions 25, 27)

What: The minimal attention strip, plain-language agent cards, approvals in the right panel.

Why: Owners need one place that says what needs them now.

Context: The floor now has OFFICE, TASKS and GRAPH views and a blocked-count badge; the strip is the next layer.

Effort: M Priority: P1 Depends on: None

Close dialog and tray (Decisions 1, 39)

What: A close dialog with a safe default and keyboard contract, and a tray so the office keeps running.

Why: Closing the window ends the day for the whole office; there is no way to keep it running in the background.

Context: Since the Studio redesign, closing the main window with terminals running starts closing time on the floor (ClosingTimeBar, via app:closeRequested in src/main/index.ts), the same as quitting. What is left is the tray, and a choice between closing the window and quitting.

Effort: M Priority: P1 Depends on: None

Per-agent levels screen (Decision 33)

What: A screen per agent in the order job, limits, connect.

Why: Autonomy is one global checkbox; owners cannot give Oscar less freedom than Pam.

Context: Pairs with the enforcement compiler.

Effort: L Priority: P1 Depends on: Enforcement compiler

First-success confirmation and morning report (Decision 32)

What: Confirm the first finished job to the owner, and a morning summary, plus a weekly digest for owners who open the app once a week (issue #81): emails handled and drafted, tasks closed by each team member, scheduled jobs run, and what still waits on the owner.

Why: New owners need proof the office works while they were away.

Context: Packs declare firstAction (Decision 29), used only in agent goals today. Every count comes from what the app recorded (the task ledger, the mail send records, the schedule runs), never from Michael's memory, and a week with nothing in it says so plainly. Deliver it the way reports already work: Michael posts it on Ask me as a report (kind: "report"), from a schedule the owner can see and change in Schedules. No estimated time saved: the owner declined it (2026-10-10, #81), because minutes per task set by the owner would end up wrong for everyone.

Effort: M Priority: P1 Depends on: None

Owner floor narrow layout (Decision 38)

What: A layout for narrow windows.

Why: The floor and side panel crowd each other on small laptop screens.

Context: src/renderer/src/App.tsx floor area and SidebarSplitter.

Effort: M Priority: P1 Depends on: None

14px text floor on the remaining setup screens (Decision 36)

What: Raise the last setup text under 14px to 14px or more.

Why: The owner's standing rule. The Business and Team steps (with TeamCard and PackTile) were raised in PR #138.

Context: Still under 14px: the company profile labels in src/renderer/src/components/CompanyProfileFields.tsx (shared with other screens; the PR #138 author offered to do these), and the Welcome and Permissions steps in src/renderer/src/components/OnboardingWizard.tsx (12 and 13px titles and notes).

Effort: S Priority: P1 Depends on: None

Documents and security (review items skipped for v0.0.1)

Gate web tools for members on other engines (issue #230 follow-up)

What: Members that run on Codex, OpenCode, Crush or another bridged engine keep their own web tools and shell network; the Research rule (only Research uses the internet) reaches Claude members only.

Why: An owner who puts a member on another engine may believe that member is fenced in like the rest of the office.

Context: Ship review 2026-10-13, decision D4: Settings, Research & privacy, names those members with an info icon. Fix per engine: map each bridged provider's web tool names into the hook's gate where its bridge posts PreToolUse, and give its shell a network rule where its sandbox allows one; otherwise refuse web for that engine.

Effort: M Priority: P1 Depends on: nothing.

Research agent and data privacy guard (issue #230)

What: No agent reaches the internet except an optional Research agent (off by default, turned on in Settings). Members ask Research through the hive; a guard checks every message to Research against the owner's data classes (patterns plus a quick no-tools AI check, fails closed) and returns refused ones to the sender only. A Data privacy page in Settings lists the classes, seeded with safe classes Allowed and sensitive ones Private.

Why: Owners want research to stay possible without any agent that holds private data (HubSpot, mailboxes, finance, the board) sending it to the web.

Context: Design and eng review in docs/designs/research-agent-privacy.md (owner direction and decisions D1 to D7, 2026-10-13). Research is walled off from the hive, business folder, palace and company knowledge (sandbox deny rules), excluded from broadcasts; other members cannot write its inbox, and the app sweeps any message it did not deliver; its replies are screened as outside text. Known limit: on Windows (no sandbox, #187) Research can be turned on with a warning, but its shell is not confined there. Michael's shell reaches no host in any office (D9.1, every office since ship 2 D16). Replaces the per member Sites code applied uncommitted on mit/named-links (keep the hook gate, sandbox.network plumbing and restart reason; remove the Sites group). Upgrade impact: every office loses web research until the owner turns Research on.

Status (2026-10-13): built on mit/named-links; at /ship the refuse only guard became a two way trust layer (pass, clean or deny in; clean and flag out), see the design doc's "Trust layer" section and its review pass 2 decisions (D6.1 to D6.7). Live sandbox checks on a Mac passed: Research cannot read tasks.json, the home folder, other sessions' temp output or teammates' moved session files; a member cannot read or write Research's folders. Left: Windows (no sandbox, #187) and members on other engines (P1 above); Research on engines other than Claude is refused until they have equal walls. Named Pete by default; the owner may rename it in Edit agent, the name is kept in the app settings (research.name) and nobody else may take it (ship 2 D17); the second ship's three review passes made every file tool call judged on its real path (trimmed and resolved, kept when a big call is cut), walled each agent's session scratch folder from the others, and made a path from agent text only ever shown in Finder (D12).

Effort: L Priority: P1 Depends on: nothing.

Engines

Run Gemini members in Gemini's own sandbox

What: Start Gemini team members with Gemini's own sandbox (-s) where it works, so a Gemini member has a guard besides its reports to the app.

Why: The app starts Gemini with --yolo and no sandbox (src/shared/agentProvider.ts), so a Gemini member whose reports stop has no guard at all. The #326 design therefore stops a silent Gemini member on every platform; with a sandbox it could keep working the way a Claude member on a Mac does.

Pros: Gemini offices stop far less often; parity with Claude and Codex on a Mac.

Cons: Gemini's sandbox needs Docker or Podman on some platforms, may block tools the team uses, and needs its own writable folder setup (like Codex --add-dir).

Context: From the #326 eng review, 2026-10-16 (decision R2). Start in hive.ts where Gemini's --include-directories are added, and check what -s needs on each platform.

Effort: M Priority: P2 Depends on: none; independent of #326.

Make a second engine (Codex first) ready for an office

What: Close the gaps that keep engines other than Claude Code out of BUILD_ENGINES, starting with Codex, then add it to the list.

Why: An audit on 2026-09-24 found that the code for Codex, Gemini CLI and Antigravity exists, but none is ready for an office without Claude. Only Codex has live history (hooks, idle, messaging and resume were debugged on real workers). Every picker and voice hiring now offer only BUILD_ENGINES, so nothing half-working can be chosen until this is done.

Context: In order of impact:

  1. The team always starts on the default engine. Done 2026-09-26: setup now writes defaultCommand from Michael's engine (teamDefaultsFromMichael, test/team-engine-from-michael.test.cjs).
  2. Cost and token data: Codex is done. The telemetry fallback reads a Codex agent's rollout in its CODEX_HOME (src/main/codexUsage.ts, hive.codexHome), and src/main/pricing.ts prices GPT-5 models. Still open: Gemini and the other engines (no usage source yet), and the lifetime cost ledger, which only records live OTel samples. For Gemini, issue #113 has the plan: read the session .jsonl at the hook payload's transcript_path (already kept as hookServer.transcriptPath(id)), skip $set and $rewindTo lines, keep the last copy of each message id (late token updates re-append it), keep rewound turns (they were billed), map cacheRead = cached, input = input - cached + tool, output = output + thoughts; fall back to the project slug in ~/.gemini/projects.json when no hook has fired; add Gemini Pro and Flash rows to pricing.ts; prove it live with one Gemini worker under a low token cap.
  3. MCP servers go only into Claude's --settings (src/main/hive.ts:1163). Codex now gets bundled skills in its private CODEX_HOME/skills; other engines still need native skill discovery. The md-mail server is the same gap (issue #109): index.ts writes it and passes --mcp-config only for Claude, so a Codex member with email starts with no mail tools. Fix: write an [mcp_servers.md-mail] entry in the agent's private CODEX_HOME/config.toml (same command, args and env, a fresh broker token each spawn), removed at teardown like the Claude file. The mail gate in hooks.ts matches only Claude's mcp__md-mail__ tool names, so the same change makes it match Codex's names too: a mail tool on any engine passes the same check, and a test proves a Codex mail call is gated. md-integrations has the same gap: it attaches only for Claude, and the Access tab says so to a member on another engine. The same entry fixes it.
  4. Antigravity gets no extra writable directories, so hive file writes may be refused. Gemini CLI done 2026-10-11: the folders Codex gets with --add-dir go into its per-agent settings file as context.includeDirectories and onto argv as --include-directories (Gemini CLI 0.60 and later skip a system settings file in a folder the user owns; geminiIncludeDirArg, test/auto-mode-sandbox.test.cjs).
  5. Antigravity agents never get their standing goal, and an Antigravity Michael never gets the roster (no SessionStart/UserPromptSubmit event).
  6. The app's own background calls run only on Claude (issue #85). Done 2026-10-17 (PR 354): runHiddenModel in src/main/hiddenModel.ts picks the runner from Michael's engine for all six (memory tidy up, the standing approval check, the hire check and the three Work style helpers). Claude keeps its runner and models; Codex runs codex exec read only in an empty folder with every tool off and the prompt on stdin, using Michael's model or the CLI default; any other engine answers unavailable and the caller keeps its fallback (test/hidden-model.test.cjs, test/hidden-model-callers.test.cjs). Still open: a runner for Gemini.
  7. Michael's first prompt is still written for one engine (INITIAL_GOD_PROMPT in src/renderer/src/hooks/useHive.ts:77). The spawn prompt and PROTOCOL.md now say "AI agents".
  8. No permission-prompt detection for these engines (no Notification event), which matters only with auto mode off.
  9. The access policy reaches Claude only (issue #48). Folder privacy (denyRead/denyWrite), connector deny and the MCP list are written into Claude's --settings (hookSettings in src/main/hive.ts); Codex gets its own workspace sandbox and --add-dir, but no folder read privacy or connector deny. Before any engine joins BUILD_ENGINES: build the one policy into each engine's own settings, record per engine which rules it can enforce, and where it can't, limit that agent (mail Draft only, connectors off) with the reason on its Access tab. A test per engine tries a read of another member's folder, a write outside its own and a denied connector, and passes only when each is blocked or the agent is limited. test/engine-readiness.test.cjs checks the preset half today. Issue #139 adds: make it a typed isolation capability on each engine preset (OS sandbox plus deny rules, a PreToolUse check only, or none), so BUILD_ENGINES refuses an engine that has not proved its kind, and an engine with none never hosts a member who holds a private folder; put the app's private mail token folder (agentPrivatePaths.ts, #63) in every engine's denied set; and note the hook gaps: pi has only PreToolUse, and the proxy tier (qwen) has no hooks at all, so its shell is unguarded.
  10. Hook events carry a per start secret (issue #61, built with the team performance work): a new engine's bridge must send HIVE_HOOK_SEAL in every frame (never a name like a token or key: Claude Code's subprocess scrub strips those) and pass the forged frame checks in test/hook-auth.test.cjs before it joins BUILD_ENGINES.
  11. No test of a run on each engine (issue #83). With a fake CLI process standing in for the engine (no paid calls), one shared test per engine checks: it starts and finishes a task, reports completion, and handles a sign-in failure, a plan or quota limit, a cancel and its cleanup, each with a plain reason to the owner. It runs in CI on every platform the engine ships for, beside test/office-privacy-conformance.test.cjs and test/engine-readiness.test.cjs.

A live end-to-end run is required before adding an engine: Gemini has never been run for real, and Codex needs an OpenAI account with credits.

Effort: L Priority: P3 Depends on: A decision to offer a second engine.

A backup engine through a plan limit (issue #129)

What: An optional backup engine per team member, set by the owner (default none). When the office break starts (issue #124), a member with a backup restarts on it with a handoff note as its first message (its task, recent context, files touched) and keeps the same folder, memory and mailbox. At the reset it goes back to its own engine at its next idle point. The floor shows it ("Pam is on Codex until 3:40 PM"); members without a backup keep the break.

Why: On a Pro plan the break can last hours of a business day. An owner who also has a ChatGPT plan could keep work moving.

Context: The trigger is the office break (src/main/officeBreak.ts); the handoff can reuse hive.takeHandoff. Two facts shape the design. The usage limit is per Claude account, so every Claude member, Michael included, reaches it at once: there is no member with headroom left, and failover has to cover Michael (a Codex Michael needs gaps 5 and 7 of the second engine item), or the moved members work with nobody to report to until the reset. The backup is the owner's own choice per member, never picked by the app (the model choice rule). Behind a buildFeatures.ts flag until it has run live.

Effort: M Priority: P3 Depends on: Make a second engine (Codex first) ready for an office, in full: the access policy (#48) and hook authentication (#61) above all, so a member that fails over keeps the owner's folder and connector limits.

Onboarding

Release

Repo

Remove or rewrite the old project's website files in docs/

What: What is left of the old project's website: docs/blog/ with its source in blog/, docs/sitemap.xml, .github/workflows/blog.yml, and the upstream promo media that was in docs/media/ (removed on 2026-10-01 with the v2 redesign). (The site pages, docs/CNAME, robots.txt, the Google verification file, the research, drops, hires gallery, evidence and deck were removed on 2026-09-28. docs/llms.txt and docs/llms-full.txt went in the 0.0.1 land: GitHub Pages is off for this repo, so nothing served them.)

Why: They describe the other product.

Context: The website for this app lives in its own repo; keep only files the app reads at runtime (docs/model-catalog.json, docs/hero.json).

Effort: S Priority: P3 Depends on: None

Record demo videos of Don't Be Michael

What: New demo videos of this app: a short hero clip for the top of the README, and an agents-at-work clip with a poster image for the README's "Watch the office work" section, which now shows a still.

Why: The README shows stills only. The old project's recordings (docs/media/hero.mp4, docs/media/demo/agents.mp4) were removed on 2026-10-01 with the v2 redesign, since they showed the Munder Difflin pixel UI.

Context: The studio lab (npm run lab, docs/demo/studio-lab.html?clean&autoplay) plays the office on a neutral example office and is made for recording. Keep files small (under 5 MB), commit an mp4 plus a jpg poster under docs/media/, and point the README at them. Delete whatever of landing-remotion/ is no longer used. The website repo may want the same clips.

Effort: S Priority: P2 Depends on: None

Completed

Sign and notarize the Mac build with an Apple Developer ID

What: Add the Developer ID Application cert and notarization credentials as repo secrets (APPLE_CERTIFICATE_P12, APPLE_CERTIFICATE_PASSWORD, APPLE_ID, APPLE_APP_SPECIFIC_PASSWORD, APPLE_TEAM_ID). release.yml and build/notarize.cjs already use them when present.

Why: 0.0.1 ships ad-hoc signed. Owners must clear a "could not verify" dialog through Privacy & Security on first open, and Squirrel.Mac cannot update an ad-hoc app in place, so every update falls back to "download the new version" instead of restart to install. macOS also forgets folder grants between versions without a stable signature.

Context: Needs an Apple Developer Program membership (paid, per organization). Once the secrets exist, the ad-hoc fallback in release.yml switches off by itself; update RELEASE.md's install steps then.

Effort: S (after the Apple account exists) Priority: P1 Depends on: Apple Developer Program membership Completed: v0.2.6 (2026-10-23) is the first Mac release signed with Developer ID team B5EHD4JBD2, notarized and stapled (spctl: source=Notarized Developer ID). RELEASE.md, README.md and dontbemichael.com now give the signed steps, and a tagged release without the certificate fails instead of shipping ad hoc (REQUIRE_DEVELOPER_ID, test/release-mac-only.test.cjs).

A write to a connected app is never sent twice (issue #54)

What: The send once rule mail already has, for writes the broker forwards to a connected app (anything other than GET or HEAD): an identical write is recorded before it goes upstream and answered from that record if it comes again, across a restart too.

Why: The broker forwards a write with no dedupe behind a 30 second timeout (src/main/integrationBroker.ts, UPSTREAM_TIMEOUT_MS). If the provider accepts it but the answer is lost, or the app stops before it comes back, the agent's retry creates a second invoice, booking or contact. Plain mail sends got this in PR #126; connected app writes did not.

Context: Reuse PR #126's pattern inside the broker, not a new store: an app private journal like mail-send-journal.json, keyed on the agent, integration, method, path and a hash of the body, written as pending before the upstream call and settled after. Pass the key upstream where the provider takes one (an Idempotency-Key header, QuickBooks requestid). A write left pending after a stop that the app cannot check is refused with a plain reason telling the agent to tell Michael, as PR #126 does for mail; the app never makes or moves a card for it (Michael owns every card). Issue #54 also proposed a per task work record for engine hand off and proof of done; those stay with "Make a second engine (Codex first) ready for an office", #36, #48 and #51.

Effort: M Priority: P2 Depends on: None Completed: Connected app writes are recorded in broker-write-journal.json before they go upstream (src/main/brokerWriteOnce.ts). A repeat of a write that worked gets the first answer; a failed one may go again with the same key. A write still pending after a restart is refused, and the agent is told to tell Michael. No card is opened. QuickBooks also gets a stable requestid. A damaged journal is set aside and a fresh one started (PR #212).

Message when the OS cannot encrypt stored credentials

What: A plain-language path for when Electron's safeStorage reports encryption unavailable, so connecting an account fails visibly instead of silently.

Why: The secret store refuses to write without OS encryption — "a secret is never written unless safeStorage.isEncryptionAvailable()" (src/main/integrations.ts:12, enforced at :111). That is the correct security behavior, but no screen explains it. On Linux without a keyring, or a Mac with a damaged keychain, the owner clicks Connect and nothing happens.

Effort: S Priority: P2 Completed: merged 2026-10-17. Mailboxes (PR #136), connected apps (#211, by DavidFHCh) and saved AI keys (#384, by fayzan-cbr) each check the secret store first and say what to do on each platform.

Read the registry once per router tick

What: In routeOnce, read the registry and Pete's name once per tick, not once per outbox file.

Why: Each outbox file asks addressedToResearch with a fresh registry() read, so a busy tick reads registry.json once per message (performance review, third ship).

Effort: S Priority: P3 Completed: merged 2026-10-17. routeOnce reads the registry and Research's name once per tick, the first time a message needs them; delivery still reads per message (#385, by amartya1523).

End a credit or sign in stop only on a finished turn

What: When the office stopped because the Claude account ran out of credit or its sign in stopped working, end the stop when a turn finishes, never because a check heard nothing for ten minutes.

Why: After a restart, or with no Claude session alive to take Michael's check, silence ends the stop (CHECK_TIMEOUT_MS in src/main/officeBreak.ts) and the team is given work on an account that still cannot run it, then stops again.

Effort: S Priority: P3 Completed: merged 2026-10-17. A credit or sign in check that hears nothing goes back to waiting and checks again after RETRY_MS; only a finished turn ends the stop. The plan limit keeps its old ending and a key stop still waits for the owner (#383, by amartya1523).

Copy a link's address from a card

What: A right-click "Copy link address" on links in the app's rendered markdown.

Why: Named links (docs/designs/named-links.md) show "HubSpot" instead of the raw URL, so selecting the text copies the name, not the address. The app has no context menu (src/main/index.ts sets setWindowOpenHandler and will-navigate handlers, no context-menu), so an owner who forwards a record link has to open it and copy it from the browser.

Effort: S Priority: P3 Completed: merged 2026-10-17. Right-clicking a web or email link anywhere in the app shows one item, Copy link address, which copies the real address; the label follows the app language (src/shared/linkMenu.ts, #372, by Yash-Nimkar0, closes #370).

A refused tool call still shows its detail

What: PreToolUse that the hook denies (paused or gated) still sends detail.

Why: The row names a step that never ran.

Effort: S Priority: P3 Completed: merged 2026-10-13. A refused PreToolUse call omits detail via emitRefused() so the row stays clear (#261, by AnonymousSurya).

Log the webhook format editor's parse error

What: console.error the raw JSON parse message in src/renderer/src/components/triggers/WebhookSchemaEditor.tsx, and make error a boolean.

Why: The owner now sees plain words, but nobody can see where the mistake was.

Context: Third review pass, 2026-09-24.

Effort: S Priority: P2 Depends on: None Completed: merged 2026-10-12, ships in 0.1.9. A failed save logs the parser's message once with console.error, error is a yes or no flag, and the owner sees the same words; the check is the pure schemaParseError helper with its own test (#141, by piyush12kunjilwar, closes #131).

Cap the total size of an Office file's zip entries

What: In src/main/docText.ts (unzip filter, ~line 108) cap total uncompressed bytes and entry count, not only each entry.

Why: A crafted .docx/.xlsx/.pptx can declare many large entries and exhaust memory in the main process, taking every agent down. Owner skipped this at ship review 2026-09-24.

Context: Count accepted entries and sum originalSize in the filter; reject past e.g. 200 MB / 5,000 entries.

Effort: S Priority: P1 Depends on: None Completed: 0.1.7 (2026-10-09). Before inflating anything, fromOoxml adds up the declared sizes of the XML parts it would read and refuses the file past 200 MB or 5,000 parts; media still never counts (#58).

Decide the Git for Windows check

What: On the clean PC, install Claude Code and start one session. If it needs Git Bash, add a Git for Windows check to onboarding (E3).

Why: If Git Bash is missing, a Windows beta user would get stuck at their first agent start.

Context: E3 in docs/designs/windows-11-installer.md depends on that smoke test, which has not run yet.

Effort: human S / CC S Priority: P1 Depends on / blocked by: the PC smoke test. Completed: 0.1.6 (2026-10-09). A Windows office without Git for Windows runs Claude Code, which then runs hooks in PowerShell; the app now writes hooks for that shell (src/main/hookShell.ts), so no Git for Windows step is needed in onboarding.

Clean up a "Not for" clause when its teammate leaves

What: When a hire is bound to a mailbox or topic, overlapping teammates' routing lines gain "Not for ; that goes to ." If that teammate is later renamed or removed, rewrite or drop the clause.

Why: A stale clause sends Michael to a name that no longer exists, and the work bounces.

Context: Found by the spec review of the hire wizard redesign (2026-09-27). The clause format is bindingLines in src/shared/hireTemplates.ts; hook the rename path in hive.ts and agent removal in the renderer.

Effort: S Priority: P2 Completed: 0.1.0 (2026-10-04). A rename rewrites the clause on every teammate (renamePatches); closing or forgetting the hire takes it off and keeps it on the hire, and it goes back when they return (src/renderer/src/shell/releaseBindings.ts).

Recognise the connector behind an opaque server name

What: When the connector's server has a UUID-like name, 16 of its tools (including money_onboarding_application_submit) are not recognised as QuickBooks and are not gated.

Why: Claude Code names claude.ai connectors claude_ai_Intuit_QuickBooks, which matches; a differently named server would not. Match on the connector URL from claude mcp list or on the full tool catalog.

Effort: S Priority: P2 Completed: 0.0.15 (2026-10-02). Any MCP server that is not a discovered, granted connector is now refused by default (docs/designs/claude-connectors.md), so an opaque name can no longer slip past.

Subagents keep their parent's QuickBooks access

What: Tag hook payloads with the harness agent id so a subagent's own agent_id does not replace it.

Why: If Claude Code sends a subagent id, the hook finds no capability and refuses Oscar's delegated QuickBooks reads (fails closed).

Effort: S Priority: P3 Completed: 0.0.15 (2026-10-02). The hook shim always sends the agent's own id and keeps Claude Code's subagent id as subagent_id, so delegated calls follow the agent's QuickBooks, connector and folder rules.

Back the Claude account switch with more than the hook

What: The Your Claude account switch was enforced only by the PreToolUse hook, which let calls through when the hook could not reach the app, and matched connectors by server name.

Completed: 0.0.15 (2026-10-02). Replaced by per-connector grants (docs/designs/claude-connectors.md): connectors are kept out at spawn, the hook shim fails closed for MCP calls, and connectors are matched by tool name. Left open: local-scope and project .mcp.json servers are not in the start-up deny list (the hook still refuses them).

Point the contributors workflow at this fork's credit, or turn it off

What: .github/workflows/contributors.yml regenerates CONTRIBUTORS.md from this repo's merged pull requests and opens a bot pull request.

Why: On this fork that list would drop the original project's contributors, which the README credits.

Context: Found in the 0.0.1 land audit. A bot pull request is harmless until someone merges it.

Effort: S Priority: P3 Depends on: None Completed: 2026-09-24 (the workflow stays on; CONTRIBUTORS.md now lists this repo's contributors only, and the README no longer credits the upstream list there)

Claude account connectors (deferred from /plan-ceo-review, 2026-10-02)

Per-connector activity log

What: Record each connector call (agent, connector, tool, allowed or refused) and show recent use on each row of Settings > Connections > Claude connectors.

Why: Once connectors are granted per agent, the owner will ask what an agent did in HubSpot or Drive; refusals are otherwise visible only in the agent's own terminal.

Pros: A clear record of who touched which system; builds on the hook's refusal log lines from the connectors plan.

Cons: A new screen, log rotation and privacy wording.

Context: Deferred as E5 (D7) in docs/designs/claude-connectors.md. The hook already sees every MCP call (src/main/hooks.ts).

Effort: M (human) / S (CC) Priority: P2 Depends on: the Claude connectors plan.

Disable the owner's Claude Code plugins in agent sessions

What: Start agents with the owner's Claude Code plugins (skills, hooks, agent definitions) disabled, not only their MCP servers.

Why: E1 in docs/designs/claude-connectors.md blocks personal and plugin MCP servers, but plugin skills, hooks and agents still load in every agent session; a plugin hook runs code on agent events.

Pros: Closes the rest of the leak E1 targets; agents run only what the app provides.

Cons: Needs a probe of which flag or setting disables plugins for one session without touching the owner's own Claude Code.

Context: Seen 2026-10-02: plugins enterprise-search and finance load in agent sessions (stream-json init).

Effort: S (human) / S (CC) Priority: P2 Depends on: nothing.

Watch list (no issue seen yet; act only if the symptom shows up)

Agents skip their folder or company knowledge

Watch for: an agent answers a policy, price or "how we do it" question from memory or the internet when the answer is in the company knowledge store; an agent asks the owner for a document already in its folder; this happens more late in a long session or after a compaction.

How it works today: agents are only told, never reminded. At spawn, injectedPrompt (src/main/hive.ts) puts the folder line ("read it for context before searching the internet"), companyKnowledgeLine ("when a task touches any of that, search it") and memoryRule into --append-system-prompt. At session start the hook (src/main/hooks.ts) adds the company profile, which ends with a pointer to the knowledge store, and the memory index. knowledgeUpdate speaks only when the owner turns the store on or off. Nothing repeats per task, nothing searches for the agent, nothing lists the folder's files, and nothing records whether kg search ran.

Smaller gaps: Michael's folder line does not say to read it for context; Michael's hand-off rules say "where to look" but never name company knowledge; Codex, Grok and agy get the text only as their first message.

Fixes, strongest first:

  1. On a new task, have the hook run kg search on the task text and attach the top 2 or 3 matches (capped, skipped when nothing matches).
  2. A one-line reminder through the hook when a new inbox task arrives.
  3. A short list of the agent's folder files at session start, sent again when it changes (like the roster).
  4. Log kg calls from activity events so agents that never search are visible. Do this first if the symptom appears, to confirm it.

Context: Investigated 2026-10-02 on the owner's question; owner chose to watch rather than build, since no issue has been seen.

Effort: S to M (human) / S (CC) Priority: P3 (watch) Depends on: nothing.

Windows

Sign the Windows installer

What: Choose a Windows signing service (Microsoft Trusted Signing, or a vendor cloud key such as DigiCert KeyLocker or SSL.com eSigner), then wire release.yml so the Windows leg signs when its secrets exist and builds unsigned, without failing, when they don't.

Why: Unsigned installers show "Windows protected your PC"; people must click More info, then Run anyway. Some small business owners stop there.

Pros: Removes or shortens the warning; the publisher name shows in the install prompt.

Cons: A paid account and a one-time identity check; a new signature can still warn until Windows' reputation system trusts it.

Context: Accepted as E1 in docs/designs/windows-11-installer.md (CEO review 2026-10-04), with wiring deferred by R5 until the owner picks a service at sign up. New certificates must keep their key in hardware or a cloud vault, so a certificate file in CI secrets is not an option. Mirror the Mac notarize step: absent secrets mean an unsigned build and no failure.

Effort: human M / CC S Priority: P2 Depends on / blocked by: the owner choosing a service and passing its identity check.

Go live steps for the Windows beta

What: In order:

  1. Push an rc tag (vX.Y.Z-rc.1).
  2. Read the first "Tests (Windows)" CI run and fix each POSIX-only test, or skip it on win32 with a reason. Done 2026-10-09: tests are platform scoped (test/platform.cjs, CONTRIBUTING.md). Text files check out with LF (.gitattributes), Windows test path assumptions were made portable, and Mac or POSIX only tests skip on Windows with a reason.
  3. Run the rc.1 and rc.2 checklist on the clean x64 Windows 11 PC: install through SmartScreen, onboard, hire, task, mail archive search, schedule, rc.1 updates to rc.2, a user name with a space and an accent, an office folder in OneDrive, and agent start time under Defender. Also check that rc.2 is offered the clean release.
  4. Once it is green, remove continue-on-error: true from the test-windows job Done 2026-10-09 (ci.yml now also runs "Tests (macOS)", and every job blocks). Done 2026-10-10: Typecheck, Build, Tests (macOS) and Tests (Windows) are required checks on main in the "protect main" ruleset, and admins can bypass them only by merging a pull request.
  5. Set the repository variable WINDOWS_RELEASE=on. Done 2026-10-05: the owner chose to ship Windows on every release before the PC check. v0.1.1 got its Windows installer added after release.
  6. After the first clean release that carries Windows, add the Windows (beta) download to the website, and confirm with the owner before that push.

Why: Windows ships on every release before the PC check (owner, 2026-10-05); these steps confirm it works. The website push is a production deploy.

Context: Deferred from the ship of feat/windows-beta-askme-fixes (2026-10-05). These are post-merge steps from docs/designs/windows-11-installer.md (T5, T10, R9, R10, E5).

Effort: human M / CC S Priority: P1 Depends on / blocked by: this branch merged; the physical PC.

Ask me (deferred from ship of feat/windows-beta-askme-fixes, 2026-10-05)

FEATURES.md entries at land

What: Add the Ask me file rows and the Windows beta to docs/FEATURES.md when this branch lands.

Why: FEATURES.md is updated at each land.

Context: Deferred from plan: docs/designs/ask-me-open-file.md and docs/designs/windows-11-installer.md.

Effort: human S / CC S Priority: P1

A React test harness

What: Add a small render harness (jsdom with react-dom) so these renderer paths get behavior tests instead of source regex pins:

  • Ask me rows
  • the per-index answer
  • AskFileRows states
  • Task detail withdrawn text
  • the floor's flight rules

Context: The coverage audit for this ship found 7 paths that are only pinned by source regex (73% value-weighted). The same applies to the IPC handlers in src/main/index.ts, which tests cannot load.

Effort: human M / CC M Priority: P2

Trust boundaries Codex raised on Ask me (2026-10-05)

What: Codex adversarial pass 2 raised these items. Each one either already exists on main or needs a design decision:

  1. Question text in Michael's request. Done 2026-10-18: the question goes in as one line labelled as card text, through cardText, which also folds C1 line breaks and drops format characters.
  2. Who closed the question. Open or closed is decided by whether a string is present, so a worker that writes tasks.json can set a or dismissedAt and hide a question. Give answers and withdrawals provenance: the recorded owner answer keys for answers, plus a trusted withdrawal marker.
  3. The 30-day window. Done 2026-10-06 (michael-replies.md): a fixed floor replaces the rolling window.
  4. Reply alias. Done 2026-10-06 (michael-replies.md): any name that reaches the owner closes a request.
  5. File swap race. fs:openAskFile checks the file and then opens it by path. The swap window is milliseconds, because the check runs at click time. Open a verified copy or a file handle instead.

Effort: human M / CC S Priority: P2

Hire wizard offers Send only from an existing mailbox (deferred 2026-10-06)

  • What: In the hire wizard's mailbox step, also list "Send only from ()" for roles whose pack card sends mail.
  • Why: A new sales hire could send outreach from the owner's address from day one instead of needing a second step on the Access tab.
  • Pros: Teaches Send only where it matters; no extra trip to the Access tab.
  • Cons: One more choice in an already full wizard step.
  • Context: docs/designs/shared-mailboxes.md (E4, owner deferred in the CEO review). Send only grants live on the Access tab first.
  • Depends on / blocked by: Send only grants (shared-mailboxes.md accepted scope) shipping first.

Translate the update surfaces

Completed: next release after 0.2.0 (2026-10-14), contributed in #290.

Row still names a finished tool

Completed: next release after 0.2.0 (2026-10-14), contributed in #288.

Parse documents off the main thread

Completed: next release after 0.2.0 (2026-10-14), documents are read in a process of their own, src/main/docTextProcess.ts.