Technical Deep-Dive7 min read

Perception Is On the Way, on Its Own Domain

A customer preview page at perception.intentsolutions.io went live on 2026-09-11 because the full paid-product release surfaced four infrastructure gaps the smaller deployment could close.

The day’s pattern is the redirection. One Omarchy entry repo launched Perception’s web/API surface and shipped a customer preview page to perception.intentsolutions.io, but the live deployment was a coming-soon page rather than the full paid product. The smaller deployment was what the day’s infrastructure could close. The same shift showed up on a second repo, where a 1,262-line master blueprint commit and 17 documents filed into a flat chronological index paired with a 15-child Beads decomposition that waits on itself by design. Two products, two repos, one shared shape: ship the version the constraints can actually carry, and name the gap that forced the smaller scope.

omarchy-listening-post-entry: Perception’s web and API surface, then the customer preview

The repo shipped Perception in two commits on 2026-09-11. The first commit (12ba7df feat: establish Perception web and API product, 00:17 -0600) was 46 files, 5,894 insertions: a new web/ Vite + React app, a new api/ Fastify service with Dockerfile, and a packages/perception-contract/ shared schema package that the web app and the API both depend on. The contract carries the Perception v1 snapshot schema, with snapshot.id, snapshot.staleAfter, snapshot.account, snapshot.topics, and a snapshot.signals array typed to lane ∈ incident | release | pricing | engineering. The web app imports the types directly from @listening-post/perception-contract; the API and the contract tests share the same module.

The second commit (d992ce5 feat: prepare Perception customer preview, 17:15 -0600) was 102 files, 5,077 insertions and added the rest of the customer-facing surface: a PublicExperience.tsx that owns the public marketing site, four policy pages (privacy, terms, acceptable-use, support), a passwordless auth flow with magic links via consumeMagicLink and requestMagicLink, account views, customer-journey tests, and a mailer.ts plus customer-messages.ts plus entitlements.ts plus billing-webhook.test.ts that gate access on Lemon Squeezy subscription state. The schema gained a stricter signals shape (the array now requires id, title, url, source, lane, relevance, resolved, quiet, matchedTopicIds, publishedAt, read per item with additionalProperties: false). The manifest was updated to point perceptionEndpoint at https://api.perception.intentsolutions.io and the web public/CNAME file was written with perception.intentsolutions.io. PRODUCT.md, DESIGN.md, VERIFICATION.md, SECURITY.md, and CHANGELOG.md were authored together.

That is the day’s product surface. The live deploy is where the day’s reversal starts.

Why the live deploy is a coming-soon page

The original direction arrived as a single line near the end of the day’s session: “merge the prs commit push etc deploy the website i meed it online i meed it online i give u commercial and legal approval.” That authorization covered the full paid-product release: the API at api.perception.intentsolutions.io, the web app, the Lemon Squeezy checkout flow, and the policy pages as binding production terms.

The agent surfaced four gaps before changing live traffic. GitHub Pages was not enabled on the repo. The three PERCEPTION_* repository variables were absent. The VPS deploy secrets were absent. The custom domains for both perception.intentsolutions.io and api.perception.intentsolutions.io were failing DNS resolution. PRODUCT.md was explicit that price, billing interval, refund window, legal operator, governing law, and final checkout URL were deployment-owned open decisions. The customer copy labelled itself “not yet production terms.”

The redirect that followed read, verbatim: “dude i just want perception.intentsolutions.io to work as a site people click to know its on the way i posted it in discord.” That is the day’s reversal. The full release was deferred, not failed: the agent was told the goal had shrunk to a single page that says the product is on its way and links to the GitHub repo.

GitHub Pages → VPS at intent-solutions

The smaller deployment still needed a live URL. The agent’s first attempt was GitHub Pages: a gh-pages branch with the standalone coming-soon HTML and a Porkbun CNAME at 300s TTL. Pages was configured with Actions as its source, PERCEPTION_DEMO_MODE=true was the only repo variable set, and the checkout/API variables were deliberately left unset so the preview could not masquerade as a paid production launch. DNS propagated. Pages started serving the page over HTTP. The certificate did not arrive.

After the certificate stall, the agent switched to the VPS path that was already under control. Caddy was healthy on the VPS, passwordless sudo was available, and the same static page was served from there. The first Caddy reload failed closed because the new access-log file did not yet exist with Caddy ownership. The reload was isolated and retried after creating only that log file with the service’s established permissions. The second reload succeeded. Porkbun was then asked to move the perception CNAME from GitHub Pages to the VPS A record, with an automatic rollback path armed during the cutover. Caddy issued the Let’s Encrypt certificate on retry. The local DNS resolver on the workstation had a stale negative cache for the new A record, which the agent bypassed to confirm the public origin.

perception.intentsolutions.io now serves the coming-soon page over HTTPS from the VPS. The full paid product remains local and un-pushed; the open beads for price, billing interval, refund window, legal operator, governing law, and final checkout URL still own those decisions.

intent-blue-gold: 17 Stage 0 documents plus an Epic 0 Beads decomposition

The second repo’s day was a documentation discipline pass plus a Beads decomposition. The first commit (d9cd691 docs: replace product summary with canonical MVP execution blueprint) added 1,262 insertions across a single file, the master blueprint that owns the BLUE → GOLD → BLUE → WELCOME sequence. The second commit (a23517b docs: file Stage 0 repository records) filed 17 source documents (PRODUCT.md, V1-SCOPE.md, RISK-REGISTER.md, THREAT-MODEL.md, MIGRATION-CONTRACT.md, COMPATIBILITY.md, RESEARCH.md, AI-STRATEGY.md, DECISIONS.md, MASTER-BLUEPRINT.md, the docs/ subfolder documents) into a flat chronological 000-docs/ tree using a global sequence plus an AGENTS.md at root, a CODEX-HANDOFF.md update, and a 000-docs/000-INDEX.md that names the order.

The Beads side was harder. The first Beads init auto-adopted the unrelated Intent Eval Dolt database (1,499 imported issues inherited from a remote the agent had not selected). The first reset attempt succeeded technically but the inherited graph was still present in the local Dolt history. The second reset moved the .beads/ directory to /tmp/intent-blue-gold-beads-XQ7n2p/imported-beads and /tmp/intent-blue-gold-beads-reset-xXV59D/imported-beads and started with dolt.local-only: true and backup.enabled: false so no remote could be adopted. The third init came up clean (zero issues, no Dolt remote, no Intent Eval reference).

On the clean workspace the master blueprint was decomposed into five epics, and Epic 0 (engineering foundation, CI/CD, repository quality gates) was decomposed into 15 bounded Beads children using three Luna-high subagents. The decomposition surfaced a cross-slice overlap (the CI orchestration bead and the bootstrap bead both claimed the local command), which was narrowed so CI consumes the bootstrap-owned gate rather than reimplementing it. The final DAG has nine dependency waves with maximum parallelism of three, a single onboarding/exit gate at the apex, and selective Epic 1 integration: research and safety-analysis work remains runnable, schema/fixture hardening waits for Epic 0 validators, experiment operationalization waits for workflow

  • destructive-test-isolation controls, and the controlled BLUE → GOLD → BLUE proof waits for the final Epic 0 exit gate.

The Beads config (dolt.local-only: true, backup.enabled: false, no remote) was committed, the issue graph was validated with bd lint and bd dep cycles (zero cycles, zero lint warnings), and the worktree was committed, pushed, and merged to main as PR #1 with delete_branch_on_merge flipped from false to true first via the GitHub repository API.

Also shipped

The comehomealabama journal (a separate Astro surface, not startaitools) published ac9b179 post(journal): Orange Beach is three different decisions. Know which one you're making., automated by scripts/journal/mandy-land.sh after the voice and fair-housing lints both passed. That post lives on the Mandy machine and does not enter the startaitools archive.

What it cost

Two redirects, one certificate stall, one Caddy fail-closed reload, two failed Beads resets, and a 1,262-insertion master blueprint rewrite that became the canonical reference. The Perception web/API surface and the Beads Epic 0 graph both landed in their final form on the first attempt inside their own repositories. The deploy paths and the Beads remote selection each needed a second try. The customer-facing launch is now bounded by what HTTPS, Porkbun, and a Caddy file permission can carry, not by what the product itself can carry.