A working manual for solix.ibex.now/user/app/. It describes capability, workflow, and — for each surface — what Foundation is actually doing for you when you press a button.
Foundation is an intelligence and execution platform. The user app exposes five working surfaces and three operational ones:
Everything is a single-page app. Each tab is independently deep-linkable.
You don't need to understand any of this to use Foundation, but reading it once gives you a mental model that makes the rest of the manual obvious.
Tenants and identity. Everything you do in Foundation happens inside a tenant — your organization's private data boundary. Your account belongs to exactly one tenant. Other tenants can't see your data; your data lives only in yours. The role on your account (Owner, Admin, Operator, User) decides which surfaces you can change, not what you can read.
The knowledge store. Foundation maintains a layered knowledge store. One layer is shared across your organization — documents your team has ingested, public sources Foundation has crawled, finished work products produced by everyone's Scouts. A second layer is private to you — notes, drafts, private uploads. Both layers are searchable not just by keyword but by meaning, which is why a search for "renewal risk" can surface a memo titled "Q4 churn signals" even though neither phrase contains the other's words.
Connectors. When you connect a third-party data source (Google Workspace, Dropbox, HubSpot, LinkedIn, NotebookLM, tl;dv, Notion), Foundation receives a delegated-access token — never your password. The token has a scope you approved (read this folder, write that channel) and an expiry. Foundation refreshes tokens silently in the background. A connector marked "needs attention" is one where refresh failed — usually because you revoked it on the provider's side, your password changed, or the provider invalidated the app.
Scouts run on a schedule. A Scout is a stored mission. Foundation has a scheduler that wakes Scouts at the right time. On each wake, a planner reads the mission, decides which tools to use, runs the tools, composes a draft, runs it through a quality check, and writes the output to wherever you said it should go (Workspace, your inbox, a Notion database, etc.). Every step is recorded.
Triton runs in turns. When you send a message in Triton, Foundation routes it through a chat engine that has your conversation history, your preferences, and a list of tools it can use. It thinks, it streams text back to your browser as it goes, and — whenever it wants to take an action that touches the outside world — it stops at an approval gate so you can confirm. The conversation history is preserved between turns; tool calls are recorded separately for the evidence drawer but don't clutter the chat.
Approvals are contracts. Every time Foundation does something with a side effect (sends an email, books a meeting, writes to a Notion database, publishes a file), it records the action as a PACT — a small contract that says what was done, on whose behalf, with what permission, and what the result was. The Inbox surfaces PACTs that need your approval. The Workspace's Evidence tab shows you the PACT chain for each finished work product.
The Inbox is a stream of events. Anything Foundation considers noteworthy — a Scout completed, an Alexandria signal worth your attention, a connector failed, a Triton action wants approval — gets written to a notification log. The bell in the top bar polls that log every 30 seconds. Marking an item read or dismissing it appends a marker to the log; nothing is ever deleted.
Autonomy is a slider, not a switch. Settings let you tell Foundation how much it may do without confirming. At the conservative end, Foundation only suggests; at the bold end, Foundation acts on high confidence without asking. The autonomy slider, the confidence threshold, the review-time window, and the per-surface overrides are all consulted at runtime — they are real constraints on the engine, not cosmetic preferences.
Audit trails are append-only. Anything Foundation publishes to a public URL, every PACT it records, every version of a Workspace document — these live in append-only logs with cryptographic chaining. A later entry can't quietly rewrite an earlier one without breaking the chain. If you ever need to prove what Foundation knew, did, or said at a particular moment, the audit chain is the source of truth.
You sign in at the app root with email and password. A successful sign-in establishes a session; every subsequent request carries that session, and the app re-checks it on every load.
OAuth consent prompts (Google, Dropbox, HubSpot, etc.) that appear after sign-in come from connecting third-party data sources to Foundation — not from the app's own auth. They originate when you, or a Scout you launch, asks for access to a corpus you haven't authorised yet.
Sign-out is immediate and server-confirmed. It does not affect Scouts already running on your behalf — those continue under your stored autonomy settings until they reach their own terminal state.
On first sign-in (or any visit with ?onboarding=1 appended), a four-step wizard runs. None of the steps is required; the wizard tracks readiness as a percentage but lets you finish, skip, or defer to Settings at any point.
| Step | What it sets | Why it matters |
|---|---|---|
| Role | Founder, Investor, Operator, Researcher, Product, Sales, Legal, or Custom | Tunes Alexandria's surfaced signals and Triton's default tone |
| Trusted source | One of the connector providers, or none | Tie-breaker when sources disagree later |
| Watch target | Free-text — a company, market, regulation, or product | Creates a paused Scout you can launch when ready |
| Brief style | Output style + notification posture | Defaults for how finished work is delivered and how chatty the Inbox is |
"Seed Foundation" applies the choices and drops you into Workspace. "Tune later" opens Settings instead. Every choice the wizard makes is editable afterwards.
The top bar carries seven tabs.
| Tab | Scope | URL |
|---|---|---|
| Alexandria | Public intelligence — research, startups, OSS, patents, decision desk | /user/app/alexandria |
| Scouts | Delegated missions, recurring or one-off | /user/app/scouts |
| Workspace | Private library of completed work | /user/app/workspace |
| Triton | Conversational driver — can run live work and request approvals | /user/app/triton |
| Inbox | Notifications and pending approvals | /user/app/inbox |
| Settings | Identity, security, connectors, preferences, autonomy, posture | /user/app/settings |
| Bugs/Features | Structured feedback intake | /user/app/bugs |
Default landing is Alexandria. Change it under Settings → Preferences → Default landing.
Alexandria is Foundation's curated read of the outside world. Five sub-surfaces cover different bodies of evidence:
Foundation continuously ingests public sources — news APIs, arXiv, patent registries, GitHub trending, financial filings, social feeds — and normalises every item into a signal. Each signal is scored against your interests (your role from onboarding, the topics you've engaged with, the watch targets you've set), your trust order (who you've told Foundation to prefer when sources disagree), and a freshness curve (newer beats older, subject to a tail for evergreen items). The Alexandria feed is the top of that ranked list.
Auto-pull every 60 seconds re-fetches the top slice of the feed without disturbing your scroll position. "Pull latest" is the manual one-shot version. The "New Alexandria arrivals" strip surfaces unread items as a triage queue.
Cards show the artifact, its provenance, a one-line "why surfaced" rationale, and modality (Video, Audio, PDF, Image, Brief, Article, Doc, Tweet). Open a card for the spotlight panel — full preview, related entities, evidence list, source update timeline. Pin to mark for later; dismiss to suppress.
Sort with Balanced, Newest first, or Most used. Group with Domain stacks (one card per source) or Flat feed (one card per artifact). The since-window narrows the date range.
The search bar runs semantic retrieval. Some shapes that work:
tier-1 LP letters published in the last two weeks mentioning AI infrastructure
patent filings this quarter naming "diffusion" and "GPU memory"
startups in Singapore raising bridge rounds with a Korean lead investor
OSS projects with a Triton or vLLM dependency that crossed 1k stars this month
anything that contradicts our case "generative AI in retail commerce is ready for production"
The last form targets Decision Desk and returns counter-evidence to a stored thesis — Foundation will surface what disagrees with you, not what agrees.
A Scout is a delegated mission. You describe the work; Foundation runs it on the schedule you choose; the output lands where you say.
When a Scout fires, Foundation does roughly this:
A Scout moves through states:
| State | Meaning |
|---|---|
| Draft | Saved but not launched. Preflight may or may not have run. |
| Ready | Configuration is complete and preflight passed; not currently executing. |
| Running | A run is in flight. |
| Paused | Halted by you or by Foundation (e.g., budget exhausted). Can be resumed. |
| Completed | The most recent run finished successfully. For on-demand Scouts this is a terminal state; for recurring Scouts it just means the last cycle finished. |
| Needs attention | Last run hit an error that Foundation cannot auto-recover. Usually a connector or budget problem. |
| Archived | Hidden from the dashboard but not deleted. |
The wizard's starter grid carries 40 templates plus a Custom Scout card. The grid is split into two named categories: Guided Templates (broad, general-purpose) and Power Templates (deeper, scenario-specific). Each template card carries a readiness pill — Available if you can launch immediately, Needs setup if the template requires a connector you haven't authorised, Coming soon if Foundation has staged the template but hasn't certified the runtime path.
Custom Scout — Build pill. Start from a blank mission with sources, outputs, budget, and approvals under your control. Use it when no template matches what you want.
These are general-purpose missions. The Available pill means launch is unblocked for any account; the only thing you need to fill in is the mission prompt and the target.
| Template | What it does |
|---|---|
| Watch Anything | Keeps an eye on one company, person, topic, website, or source and tells you when something important changes. |
| Track A List | Watches a group of companies, accounts, vendors, products, or topics and shows which ones changed, which stayed quiet, and what needs attention. |
| Research Deep Dive | Answers a hard question with evidence, citations, open questions, and a clear recommendation. |
| Summarize And Synthesize | Turns a pile of documents, Scout results, notes, or feeds into a short decision-ready summary. |
| Compare Options | Compares choices side by side — strengths, weaknesses, evidence, tradeoffs. |
| Score And Rank | Ranks candidates using your criteria, explains the score, flags where evidence is weak. |
| Challenge A Claim | Stress-tests a claim, thesis, forecast, or memo; shows what supports it, what contradicts it, what is still unknown. |
| Connect The Dots | Finds useful relationships between people, companies, products, papers, patents, events, or signals. |
| Process Data Or Media | Takes a file, transcript, recording, feed, or dataset and turns it into structured notes, evidence, clips, or rows. |
| Source-Bound Research | Researches only inside the sources you approve, so the answer stays grounded in your documents, corpus, or workspace. |
| Refresh Or Repair A Web Source | Re-checks a website, feed, or stale citation and refreshes the evidence in your knowledge base. |
| Generate Context Or Test Data | Builds examples, context packs, or test sets that help another agent, Scout, or workflow perform better. |
| Run A Model, Algorithm, Or Simulation | Runs a calculation, model, simulation, or code-assisted analysis and returns the result with assumptions and proof. |
| Consumer Task Scout | Handles a recurring personal task (a trip, renewal, appointment, deadline) and asks before taking action. |
| Fill An Evidence Gap | Finds the missing proof behind a claim, report, entity, or prior Scout result. |
| Verify Evidence Or Program | Checks whether a claim, application, supplier file, compliance pack, or evidence bundle is complete and defensible. |
| Provenance Chain Of Custody | Traces where a document, claim, file, or result came from and whether its evidence chain is trustworthy. |
| Cost And Economics Scout | Models costs, tradeoffs, budgets, unit economics, or financial scenarios so you can compare choices. |
| Anomaly Pattern-Break Scout | Looks for unusual spikes, silence, drift, or pattern breaks and explains why they may matter. |
| External Feed Or Webhook Scout | Connects to a feed, API, webhook, or watched folder and only reports when new records arrive. |
| Multi-Source Fusion Scout | Combines documents, feeds, prior Scouts, media, and live sources into one clean answer — without mixing up what came from where. |
| Media Forensics And Passport Scout | Checks a document, image, audio, or video for origin, key claims, useful clips, and signs that the material needs closer review. |
| Knowledge Pack Builder | Turns a set of documents, links, files, or source lists into an organised knowledge pack future Scouts can rely on. |
These are deeper, scenario-specific missions. Several require a connector before you can launch.
| Template | What it does | Readiness |
|---|---|---|
| OAuth Document Analysis Scout | Reads documents from apps you connect during launch, answers a question with evidence, and writes the finished analysis to an approved destination. | Needs setup — Dropbox |
| M&A Diligence Scout | Reviews a target, shortlist, or diligence packet and returns evidence-backed signals, red flags, open questions, and next actions. Does not provide legal, tax, audit, investment, or transaction-safety opinions. | Available |
| Market Entry War Room Scout | Evaluates a market-entry decision with live evidence, competitor signals, operating risks, and scenario options you can act on. | Available |
| Regulatory Impact Scout | Monitors a regulation or policy area and maps what changed to your products, customers, policies, and operating model. | Available |
| Supply Chain Resilience Scout | Watches vendors, regions, facilities, logistics, sanctions, outages, and financial-stress signals so supply-chain risk shows up before it becomes a failure. | Available |
| IP / Patent White-Space Scout | Finds technical white space, patent and publication patterns, collision concerns, and research directions with source-backed evidence. | Available |
| Customer Churn Early-Warning Scout | Finds customers that may be at risk before renewal or disengagement and returns evidence, intervention options, and approval-gated follow-up drafts. | Available |
| Competitive Countermove Scout | Turns a competitor move into response options with evidence, likely customer impact, risks, confidence, and next experiments. | Available |
| Board / Investor Readiness Scout | Prepares a board, investor, or executive review with live evidence, likely questions, risk register, gaps, and follow-up work items. | Available |
| Scout Mission Commander | Breaks a complex objective into child Scouts, routes work between them, verifies the evidence, and returns one closeout with receipts and next moves. | Available |
| Twin Gap-Fill Scout | Turns missing or weak Digital Twin evidence into focused Scout work, then writes back a source-backed gap report for human review. | Available |
| Twin Scenario Drill Scout | Tests a Digital Twin scenario or hypothesis against live evidence and returns what held up, what broke, and what to test next. | Available |
| Expert Review Relay Scout | Packages a Scout result and routes it to the right expert, peer Scout, or Triton review stream with PACT receipts and approval controls. | Needs setup — Peer Router, Triton Review |
| Rich Media Evidence Room Scout | Turns videos, audio, documents, screenshots, or source bundles into a reviewable evidence room with media passports, clips, provenance, and a workspace artifact. | Needs setup — Media Evidence |
| Scientific / Quant Experiment Scout | Runs a bounded analysis, model, algorithm, or simulation and returns assumptions, code or compute receipts, results, and verification notes. | Available |
| Live Feed Fusion Scout | Watches APIs, webhooks, feeds, folders, or source streams, deduplicates changes, fuses signals, and reports only what matters. | Needs setup — External Source, Webhook |
| Payment-Gated Agent Scout Preview | Previews a paid agent, premium tool, wallet, or HTTP 402 / x402 handoff. Stops safely until a live payment provider and approval are available. | Needs setup — Wallet Payment, X402, Budget |
You can either start from a template (which pre-fills sensible defaults) or pick Custom Scout to build from scratch. The wizard fields:
| Field | What it does |
|---|---|
| Name | Display name for the Scout |
| Target / scope | Company, domain, topic, URL, document, account, or corpus |
| Mission prompt | The objective in your own words |
| Triton Assist | Click to have Triton rewrite the mission prompt more sharply. It returns a suggested rewrite as a card; you choose Use rewrite or Keep original. |
| Connected data sources | Per-connector read/write scopes. Tighten these — a Scout you give read-only access to Drive cannot accidentally write back. |
| Cadence | How often the Scout fires (next section) |
| Boundaries | Free-text guardrails. Foundation parses these and applies them at runtime. |
Picking Custom Scout also exposes:
A weak Scout prompt:
Watch competitors
A strong one:
Every weekday at 06:00 Tokyo time, scan news, blog posts, and earnings calls from the previous 24 hours for these competitors: NVIDIA, AWS, Google, Microsoft, OpenAI, Anthropic, Meta. Focus on agentic infrastructure announcements. Produce a one-page brief — five signals max, one-line so-what per signal. Output to Workspace and an Inbox notification. Boundary: never browse paywalled content.
The strong version names: the cadence, the time window, the entity list, the topic filter, the output shape, the destination, and one explicit boundary. Foundation can be vague-tolerant when you must, but it produces better work when the brief is sharp.
More worked examples are in the Triton prompt library below.
| Cadence | Behaviour |
|---|---|
| On-demand | Fires only when you click Run or another Scout/Triton triggers it. |
| Daily | Fires once per calendar day. Foundation defaults to off-peak; you can override. |
| Weekly | Fires once per week on a day/time you pick. |
| Monthly | Fires once per month on a day you pick. |
| Continuous | Polls a bound external source (RSS, JSON API, webhook, watched folder) on a fixed inner cadence — default once an hour — and emits output only when the source changed. Empty polls do not produce a notification. |
| Recurring | Custom interval — every N hours or N days. Useful for "twice a day" or "every six hours" cases. |
Each Scout writes its output to one or more destinations. The router enforces order: your Workspace always gets the canonical copy first; external destinations follow. Available destinations:
| Destination | Where it lands |
|---|---|
| User Workspace | Your private library — the default. |
| Public Workspace | A public URL under solix.ibex.now/public/, gated by the entity-metadata contract. |
| Dashboard | Pinned widget in Foundation's operator dashboard. |
| Chat | Replies into the Triton thread that started the Scout, if origin was chat. |
| Formatted report to addresses you configured on the Scout. | |
| Slack / Discord / Telegram / Teams | Posts to a configured channel — requires the workspace auth. |
| SMS | Short summary to a configured number. |
| Google Drive / Dropbox / OneDrive | Saved file in the folder you authorise. |
| Notion | Adds or updates rows in a Notion database. |
| Airtable | Appends rows to a configured table. |
| Structured data | Stored in Foundation's structured data lake for downstream analytics. |
| File export | A downloadable CSV / JSON / PDF artifact. |
| Alexandria | Contributes findings back into the corpus so future Scouts can retrieve them. |
| Webhook / API callback | POSTs JSON to a URL you registered. Retries with backoff. |
| Calendar | Creates an event with the finding as the description. |
| CRM | Creates or updates company / contact / deal records. Approval gate enforced. |
| Scout | Feeds findings into another Scout's input — the "pipeline" pattern. |
| Peer Router | Routes to another agent (Cortex etc.) for peer review. |
| Triton Review | Routes a live PACT stream to Triton for refinement. |
| FMS Workspace | Publishes rendered media to Foundation's Media Suite. |
Failed deliveries retry three times with exponential backoff; persistent failures land in a dead-letter log and surface in the Inbox as "Needs attention".
Preflight runs before the first launch and on every config change. It verifies, in order:
A fail blocks launch. A warn allows launch but surfaces the warning in the wizard. Power templates set a stricter mode where warnings also block.
Boundaries you type in plain English are parsed and applied, not just left as hints:
If a boundary is violated mid-run, the Scout pauses and a "Needs attention" Inbox item appears with the violation reason.
Every Scout has four budget knobs:
| Knob | Default | What it caps |
|---|---|---|
max_iterations |
10 | How many tool-call cycles the planner may run. |
mcp_tool_budget |
20 | How many tool calls total the Scout may issue per run. |
planner_budget_ms |
0 (no cap) | Wall-clock budget for the planner step. |
| Run timeout | 15 min | Hard ceiling on total run time. |
Hitting any of these pauses the Scout at the next clean boundary and writes a partial output explaining what was achieved and what was left.
Workspace is the private library of finished work products. Drafts and uploads-in-progress are kept separate.
Each Workspace document is a record. The record carries:
When you click a card, Foundation loads the latest version and shows the four reader tabs:
The reader's top bar exposes Public link (mints a shareable URL via the public publish gate), Open public page (follows it), and Close. Inside the Evidence tab you can reprocess media intelligence or soft-delete the document (recoverable from elsewhere).
Triton is the only surface where you have a conversation with Foundation. It is also the only surface that runs live work in real time. Triton can:
Send a message. Foundation does roughly this:
Triton runs at the conservative PROPOSE autonomy tier by default. You can broaden that per surface in Settings → Preferences → Autonomous work.
Four practical ways to send a turn:
While a turn is in flight, four controls light up: Pause reply, Resume, Stop, Try again. Stop cancels the turn — it does not roll back work already committed (a saved Workspace doc stays saved). Try again re-runs the original prompt from a clean state.
Tools that touch the outside world surface an inline approval card with four actions:
The decision is recorded in the turn's evidence drawer.
Every turn produces an activity rail with nine rungs: Stage, Understanding, Planning, Actions, Approvals, Saved work, Artifacts, Evidence, Recovery. The Context drawer shows the identity, role, and autonomy mode Triton operated under. The Verification drawer shows token usage, model, and the turn's terminal state. You don't need to read these for routine work, but they are authoritative when something goes wrong.
Examples by job to be done. Copy and adapt.
One-page brief on the state of sovereign-AI infrastructure in Japan as of this week. Cap at eight sources. Evidence-table output style.
Meeting brief on NTT Data for Tuesday: shipped product in the last six months, hiring signal, public stance on agentic AI, three sharp questions to ask them.
Synthesise — not list — the three best articles published this month on long-context model serving. Under 300 words.
Read the latest ten papers tagged "diffusion + GPU memory" on arXiv. Tell me which two are most likely to matter for inference cost in 2026, and why.
Weekly recurring Scout. Watch these competitors: [list]. Output destination: Workspace. Notify only on pricing changes or senior hires. Boundary: no paywalled content.
Continuous: watch the Japan FSA announcements feed. Notify within an hour on any "generative AI" or "large language model" mention.
Subscribe to LinkedIn job postings of these 12 Japanese AI startups. Weekly hiring-velocity table. Quiet posture.
Working thesis: "Edge inference will outgrow centralised inference for consumer AI by 2027." Find the three strongest counter-arguments. One piece of evidence per counter.
Steel-man the case for and against partnering with NotebookLM Enterprise as our default doc-corpus tool. Board-ready output style.
Across every Workspace doc tagged "Tokyo POC", what is the single most likely thing to go wrong in Q3? Cite at least two pieces of evidence.
Draft a one-pager titled "Why Foundation should ship the Triton Aide tier first." Action-list output style. Save to Workspace tagged internal/strategy.
First-draft LP update for Q4 2026. Two wins, one challenge, an honest note on burn. Pull wins and challenge from the last three Workspace docs tagged portco-updates.
Compose the agenda for Tuesday's product review meeting. Themes: Triton autonomy rollout, Alexandria refresh policy, Scout time-budgets. Each item is a decision question, not a topic.
Read every Workspace document tagged "fundraise-pack". Identify deck sentences not backed by the data room.
Cross-check the figures in this brief [paste] against the underlying Scout outputs. Flag any number I cannot defend from a primary source.
Open the last Triton output on Tokyo POC. Find any claim that disagrees with the customer's last NotebookLM brief.
Read the last 30 tl;dv calls with enterprise-infrastructure prospects. Top three objections, ranked by frequency.
Pull LinkedIn updates from everyone who attended the July product council. Flag job changes, fundings, or public posts about us.
Draft the renewal email to
. Include the two product changes they cared about most (last call transcript). Avoid the topic of seat-count.
Daily 07:30: a 15-minute synopsis of overnight Alexandria, Inbox, and any Scout that ran while I was asleep. Quiet posture. One notification.
Friday 16:00, block 30 minutes for "Read Triton briefs". Reuse the same room. No confirmation needed.
Whenever a new Workspace doc is saved tagged for-board, summarise in 150 words and email me at maximusclaudius2026@icloud.com.
M&A Diligence Scout
Target: [target company]. Produce an evidence-backed diligence pack — material risks, customer concentration, key-person risk, regulatory exposure, integration debt. Sources: public filings, last four quarters of earnings, LinkedIn org changes, the eight most-cited industry analyst notes. No legal or tax opinions. Cite every claim. Output: Workspace + Inbox notification. Cadence: on-demand.
Customer Churn Early-Warning Scout
Watch the 40 enterprise customers in HubSpot pipeline stage "Renewal next 90 days". Weekly. For each, score churn risk from: tl;dv calls in the last 30 days, support ticket volume, product usage drop-off, recent LinkedIn departures of champions. Output an Airtable row per customer with risk score, evidence summary, and a draft outreach email. Boundary: do not send anything externally; approval gate on every draft.
Regulatory Impact Scout
Watch every announcement from the EU AI Office and the UK AI Safety Institute. Continuous. Map each new announcement to: which of our products it affects, what change is required, who in the org owns the response, what the deadline is. Output: Workspace + Slack #regulatory.
Scout Mission Commander
Mission: "Decide whether to enter the German enterprise CRM market in 2026." Break into child Scouts — market sizing, top 5 incumbents deep-dive, regulatory landscape, hiring landscape, two reference-customer interviews via the connected meeting transcripts. Return one closeout brief with a recommendation and named open questions. Budget: 50 USD. Approval gate on every child Scout's external action.
Worked Refine-request example
The Inbox aggregates anything that needs a human eye: Scout completions, Alexandria arrivals, Triton approvals, connector status changes.
Three filter axes:
Each card carries a severity icon, source label, relative timestamp, title, body, and the right CTA: Open Workspace, Open Alexandria, or Open Inbox. Approval notifications expose Approve and Reject inline. "Mark all read" applies to whatever the current filter shows.
The Inbox polls in the background; new items appear without a refresh.
Settings is split into six panels.
Read-only summary: display name, email, role, workspace.
Change password. Minimum eight characters; new password must differ from current.
Per-provider authorisation: Google Workspace, Dropbox, HubSpot CRM, LinkedIn, NotebookLM Enterprise, tl;dv, Notion. Each card shows status and a connect/disconnect action. Google Workspace separates Ibex (primary) and Antler accounts; tl;dv accepts an API key inline.
Connectors are the source of authority for what Scouts and Triton can read and write. Check this panel first when something says "needs attention".
| Control | Effect |
|---|---|
| Notifications | Master switch for Inbox + bell |
| Compact view | Higher information density in lists |
| Show sources | Inline citations in Triton replies |
| Default landing | Which tab loads on sign-in |
| Control | Effect |
|---|---|
| Autonomy level | Observe, Suggest only, Act on high confidence, Act broadly |
| Confidence threshold | 0.70–0.99 — minimum confidence for autonomous action |
| Review time | 0–60 seconds — how long Foundation waits for your "stop" before proceeding |
| Quick confirmations | Single-tap approvals for common, low-risk actions |
| Per-surface autonomy | Independent autonomy for Scouts, Workspace, Alexandria |
These settings are read at runtime. Observe stages every action but never executes; Suggest only shows you proposals; Act on high confidence runs things above your threshold and pauses below; Act broadly approves more aggressively. Raise the tier deliberately, surface by surface, as you learn what Foundation does well.
Operating posture:
Advanced controls for fine-tuning Triton's memory, interests, and behavioural weights. Defaults are sensible; visit when Triton is consistently misjudging tone or priority.
A single form. Required: Title, Details. Optional: Type (Bug, Feature, Confusing UX, Data gap), Surface (which tab), Impact (High, Medium, Low), a screenshot up to 5 MB (PNG, JPEG, WebP).
Good reports name the surface, the action, the expected result, the observed result, and the frequency. Foundation triages these to the responsible team.
The bell polls every 30 seconds. Its dropdown lists the eight most recent notifications, each linking to the right surface. The full Inbox is one click away from the dropdown.
The account control shows your initials when no profile photo is configured. Sign out terminates the session immediately.
Every tab is deep-linkable. Useful parameters:
| Parameter | Effect | Example |
|---|---|---|
?onboarding=1 |
Re-trigger onboarding | /user/app/alexandria?onboarding=1 |
?tab=<name> |
Open Settings on a specific panel | /user/app/settings?tab=connectors |
?document_id=<id> |
Open Workspace at a specific document | /user/app/workspace?document_id=... |
?action_id=<id> |
Open Workspace at a specific action | /user/app/workspace?action_id=... |
Unknown paths under /user/app/ redirect to Alexandria.
| Symptom | First action |
|---|---|
| Stuck on "Loading Foundation App…" longer than 10 seconds | Refresh once; if it persists, sign in again |
| Login screen appears mid-session | Session expired; sign in again |
| Triton header shows "Connection lost" | Start a new chat; the WebSocket reconnects on the next turn |
| Triton input is disabled | A turn is still in flight; wait or press Stop |
| Triton reply ends with "Needs your review" | A tool was gate-denied; address the reason in a follow-up |
| Scout stuck on Running past its time limit | Check Settings → Connectors for any that need attention |
| Scout shows Needs attention | Open it; the reason is in the run log — usually budget, connector, or boundary violation |
| Workspace document fails to render | Pull latest; if still unavailable, it's in Preparing state |
| Notification bell count looks wrong | Wait 30 seconds (poll interval) or refresh |
| Bugs/Features won't submit | Title or details empty, or screenshot exceeds 5 MB / unsupported format |
| Need to redo onboarding | Append ?onboarding=1 to any user-app URL |
/user/app/.If something on screen isn't covered here, ask Triton in the app. It can describe its own surface and capabilities.