Path to impressive

Pickup: Handoff.md to execute; this file is the strategy, not the checklist.

Strategic companion to Pre-Public Hardening. That doc is reputation polish. This doc is product ambition: how Maidan becomes the tool people reach for when agents need a shared workplace, without becoming a mediocre Slack clone.

Snapshot date: 2026-08-25. Program D closed at v266 (2026-08-24). The optional-deferrals sweep (267–272) has since shipped in full (tags through v273.0.0); the current backlog is Open Work.md. Expansion bets in Expansion Bets.md.

North star (decide this once)

Maidan wins as the best multi-agent collaboration substrate (durable shared state, capability-scoped tools, task orchestration, HITL), not by matching Slack chrome.

Recommended: agents-first + thin HITL console + bridges. /ui is operator surface, not the product. If you flip that, browser automation becomes P0.

1. UI testing and browser assurance

What you have

• Static shell markers: ui_static_e2e, ui_search_e2e, ui_tokens_e2e, ui_channels_e2e
• Session /ui/api e2e (~1.6k lines across ui_*.rs): channels, threads, messages, admin, collab, edits
• WS via session: ui_ws_tail_e2e, parts of ui_v2_e2e
• ui_js_contract.rs: static analysis that bare JS calls resolve (Cluster 133 class). No browser.
• openapi_e2e lists /ui/api paths
• Ten ui_* test files. None of them is Playwright.

Implementation: one vanilla static/index.html (~2.4k lines). No SPA framework.

What you do not have

• No headless browser job in CI
• No real DOM click-path coverage
• No screenshot / a11y / visual diff gates
• No proof client-side WS handlers work under a real browser event loop

What to do

If /ui stays operator-thin: keep API + JS contract; optionally add one headless smoke later (load UI, post message, assert row). Document that it is an operator console. UI browser e2e is still not the product this quarter.

If humans live in /ui: budget months for real browser e2e — string checks are not that.

2. Ecosystem gaps that limit adoption

Already good

High bounce for target users

Do not chase

External vector DBs as primary (Pinecone/Qdrant) — keeping vectors beside RBAC messages is a feature. Native GCS/Azure blob only if demanded. Search replica routing (271–272) is the other agent's job (PostgresSearch has its own PgPool); not an expansion bet.

Provider matrix (the "whatever they already run" job)

Users will not adopt a workspace that forces one cloud's Postgres and OpenAI. They also will not wait for a third database engine.

Dialects we keep (two, not N): Postgres and SQLite. Production HA, LISTEN bus, pgvector, LSN replicas = Postgres. Laptop / Pi / tests = SQLite. A MySQL/Mongo backend would duplicate the 228-method Store and still lack NOTIFY/pgvector. That is a no.

Hosts we owe a tested page for (same code, different URLs):

Execution checklist: Hardening I (docs/Providers.md, then recipes). Do not add embedding SDKs or a third Store impl.

Protocol matrix (the "whatever they already speak" job)

Users will not adopt a workspace that invents a fourth agent protocol. They also will not wait for GraphQL, gRPC-for-REST, or IBM ACP.

Layers we keep (industry 2026, AAIF): MCP = tools, A2A = peers, REST/OpenAPI + WS + webhooks = existing IT. AG-UI is a frontend protocol we do not native-ize until /ui is the product. Zed ACP is an optional worker adapter (OpenTag-shaped), not the workspace.

Wires we owe honesty on (same code, clearer contract):

Execution checklist: Hardening J (docs/Protocols.md). Do not add a Maidan-native agent protocol.

Do

Publish both matrices (hosts + wires). Ship interop packs (MCP JSON, thin SDK, framework recipes) after documenting the MCP protocol freeze.

3. Slack gaps (triage)

Skip: huddles/voice, native mobile, emoji packs, Workflow Builder UI, org hierarchy, SAML-in-Maidan, SCIM.

Worth for HITL: notification polish, rich unfurls/task cards for supervisors.

Already ahead of Slack (market these): DAG, leases, skill claim, claim_next_thread, wait_for_*, structured results, tool transcripts, capability tokens, context export, A2A, at-least-once, read-your-writes tokens.

4. Steal from other platforms

Highest-ROI for us (Claude agent owns import + search routing): Slack projector, client SDKs + MCP pack (protocol honesty then examples/), hero multi-agent demo (offline DAG, no LLM), durable mail retry. Workspace import is 269–270 — not our bet. /ui SPA is not the product.

5. Performance, load, and optimization

This is a code-improvement track (Hardening H), not an expansion bet and not more Program D.

Already shipped: sharded fan-out, filtered ANN, concurrent context, transactional bus outbox (maidan_outbox, 205/84), NOTIFY self-heal, LSN read replicas (265–266 closed at v266), scripts/loadgen.sh + #[ignore] load_baseline (198), criterion search_hot / store_hot (109 / 120), scale-out smoke + replica-harness.sh.

Do not "finish 266." It shipped. Search token-aware replica routing is 271–272 (other agent; PostgresSearch owns its pool).

The remaining job is in Pre-Public Hardening.md section H:

1. Record a Postgres loadgen baseline (today the default is SQLite in-process, REST post/read/search only).
2. Extend the mix to MCP / WS / claim_next_thread (the agent path).
3. Optimize only what those numbers move: context filter-before-build, cheaper search deny-set, maybe DM-in-SQL. Mail retry is Bet 4, not perf.
4. Fix Production.md, which still says load/throughput is "not covered."

Do not SPA-rewrite /ui for speed. Do not add Redis until Evidence shows a bus/DB bottleneck. Do not chase external vector DBs for vanity benches. Do not reopen batched pg_notify (declined in Open Work).

6. Most useful tool possible

useful ≈ agent outcomes × ease of adopt × trust

90-day sequencing

The Claude agent shipped 269 and is finishing 270–272 (import REST+remap+409 / search token-aware replica routing). That is not our quarter. Do not duplicate it. Hardening P0 can overlap.

Our quarter is pack-and-prove, then Slack projector. /ui browser e2e is still not the product.

• Now (parallel with 270–272): Hardening P0 (tone, README first command, mail.rs lie) and H1/H5 (Postgres loadgen baseline + Production.md honesty) and I1/J1 (docs/Providers.md, docs/Protocols.md). Measurement and docs do not collide with import PRs.
• Then MCP 2026, then pack: Hardening J3 / Bet 2 M.0 is the required 2026-07-28 upgrade (not a 2024 freeze). M.1 examples/ + 2026 snippets. M.2 offline DAG seed using existing hero tools (claim_next_thread, wait_for_result, post_message, set_thread_result, request_approval, search_messages) — no new runtime, no LLM. Optional Bet 3 TS client, pinned to a 7-method OpenAPI freeze, ≤15 methods, REST+WS, not Crew.kickoff.
• Then (if claiming the front door): Bet 1 Slack projector (HTTP Events, mention-only, final-message first). Bet 4 mail retry only if we claim reliable notifications (new mail_outbox).
• Star-tax (GIF / topics / homepage) stays parked.

Impressive to each audience

Decision checklist

1. Primary user: agent or human? (Recommend agent.)
2. Humans in Slack via bridge, or in /ui? (Recommend bridge.)
3. Next quarter: package-and-prove after D — D is done; optional-deferrals (269–272) are the other agent's in-flight work; then package (examples + MCP honesty + optional SDK). Not more substrate. Not /ui SPA.
4. First interop bet: MCP pack, SDK, or Slack bridge? (Recommend pack M.0→M.2, then SDK, then bridge.)

Record answers in Decisions.md when picked.

See also

• Pre-Public Hardening.md
• Expansion Bets.md — researched Slack/MCP/SDK/mail bets after 270-272
• Protocols.md — 2026 integration wires vs what we speak
• Providers.md — host matrix
• Launch.md — public cut, production-ready extras, announce
• Handoff.md — pickup page for a later agent session
• Open Work.md / Remaining Work.md
• Embeddings.md / Production.md / Threat-Model.md
• AGENTS.md