fix(selfhost): align Docker + smoke CI with flat repo layout by seungdo-keum · Pull Request #8 · cliwant/practiq-oss

seungdo-keum · 2026-05-27T05:21:05Z

Why

The docker/ skeleton and selfhost-smoke.yml were drafted assuming an apps/web monorepo layout. The extracted repo is flat — the Next.js web app is at the root, and packages/mcp is the only workspace. That mismatch made the self-host smoke matrix structurally unrunnable (every docker compose / build path was wrong), which blocks DoD #3 — self-host 3-OS PASS evidence.

What

docker/Dockerfile.web — build at repo root (drop apps/web + nonexistent packages/core); npm install (repo ships no lockfile, so npm ci can't work); npx prisma generate; full-copy next start runner (robust vs. output: standalone tracing, which isn't enabled and is finicky with Prisma). Copies prisma.config.ts so runtime db push resolves schema + datasource.
docker/entrypoint.sh — schema path → root; migrate deploy → db push (no prisma/migrations/ folder ships); POSIX while loop; no --accept-data-loss (a destructive diff fails loudly instead of dropping a self-hoster's data; fresh DBs push cleanly).
src/app/api/healthz/route.ts (new) — liveness probe: 200 iff server up + Postgres reachable. Distinct from /api/health (readiness — 503s when optional Stripe/Resend/OpenRouter are unconfigured, which is wrong for docker compose up --wait on a minimal install). Both the compose healthcheck and the CI probe already point at /api/healthz.
.github/workflows/selfhost-smoke.yml — flat-repo path filters (src/**, prisma/**, public/**, root configs); COMPOSE_FILE=docker/docker-compose.yml job-env so every docker compose step finds the file without -f.
.gitattributes — pin Dockerfile* to LF (Alpine build safety on Windows checkouts).

Test plan

selfhost-smoke matrix green on ubuntu-22.04
selfhost-smoke matrix green on macos-14
selfhost-smoke matrix green on macos-13
ci (type-check / lint / build / test / mcp-pack) green

Closes the structural blocker on #5.

🤖 Generated with Claude Code

The docker/ skeleton and selfhost-smoke.yml were drafted assuming an apps/web monorepo layout, but the extracted repo is flat (web app at the root, packages/mcp the only workspace). That mismatch made the self-host smoke matrix unrunnable. This fixes the whole path end-to-end: - Dockerfile.web: build at repo root (no apps/web / packages/core), use `npm install` (repo ships no lockfile), `npx prisma generate`, and a full-copy `next start` runner (robust vs. standalone tracing). Copy prisma.config.ts so `db push` resolves schema + datasource at runtime. - entrypoint.sh: schema path -> root; switch `migrate deploy` -> `db push` (no migrations/ folder ships); POSIX while-loop; no --accept-data-loss so a destructive diff fails loudly instead of dropping self-hoster data. - New GET /api/healthz liveness probe (server up + Postgres reachable), distinct from /api/health readiness which 503s on optional integrations. Both the compose healthcheck and the CI probe already point at it. - selfhost-smoke.yml: flat-repo path filters (src/prisma/public/configs), and COMPOSE_FILE=docker/docker-compose.yml so every `docker compose` step finds the file without -f. - .gitattributes: pin Dockerfile* to LF (Alpine build safety). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

actions/setup-node@v4 with `cache: 'npm'` requires a lockfile to compute its cache key. The repo intentionally ships no package-lock.json, so every CI job was failing at the setup-node step before install even ran. Drop the cache directive across all five jobs (type-check / lint / build / test / mcp-pack) so installs run via `npm install`. Follow-up (not this PR): commit a package-lock.json to restore cacheable, reproducible installs. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The Prisma client uses the new prisma-client generator with a custom output (src/generated/prisma), imported as @/generated/prisma/client. The type-check and test jobs never generated it, so prisma.* resolved to `any` — the root of the bulk of the strict-mode TS7006/TS2339 cascade. The build job + Docker already run generate; add it to these two jobs. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The gmail/outreach + google-sc modules import googleapis, but the dep was dropped during the OSS package.json restructuring, causing 7 TS2307 "Cannot find module 'googleapis'" errors that cascade. Restore the exact version the app uses (^144.0.0). Note: this operator growth/outreach tooling (outreach crons, Search Console SEO automation, Slack summaries) is surplus to the self-host product and is tracked for removal as a follow-up; restoring the dep unblocks the build now. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…rift) The last 2 of 149 type errors: stripe@^22 caret floats (no lockfile) and its pinned-version literal type drifted to 2026-04-22.dahlia, rejecting our deliberately-pinned 2026-03-25.dahlia. Cast to Stripe.LatestApiVersion to keep the intended runtime value type-stable across SDK minors. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The cast to Stripe.LatestApiVersion failed (TS2694 — not an exported member in stripe@22). Omitting apiVersion entirely is always type-valid and uses the Stripe account's default pinned version — the correct default for a self-host deployment, and immune to the lockfile-free SDK caret drift. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…Docker) GitHub arm64 macOS runners lack nested virtualization, so Colima/docker compose fails at VM start (exit 125) — macos-14 could never pass the compose smoke. Split the matrix: docker-compose smoke runs on ubuntu-22.04 + macos-13 (full boot + DB + healthz), and a new native-build job proves Apple-Silicon parity by running prisma generate + next build on arm64 macOS (exercises the arch-sensitive Prisma engine / bcryptjs / lightningcss). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

tsc includes tests/, so the e2e spec's hardcoded apiVersion literal was the last type error (the bootstrap script mirrors it for consistency). Same fix as the client: omit apiVersion to stay type-stable across stripe@22 minors. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The class extends the built-in Error; transpilation breaks `instanceof` for genuine instances unless the prototype is reset. Add Object.setPrototypeOf in the constructor — fixes the lone failing vitest assertion (token-budget.test.ts > assertBudget — trial plan). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

`next build` of this heavy app (pdfjs/exceljs/pptxgenjs/tiptap/react-pdf) exceeded the default V8 old-space heap on the 7 GB ubuntu runner — "Ineffective mark-compacts near heap limit". Set NODE_OPTIONS --max-old-space-size=6144 in the CI build job and the Docker builder stage. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Colima on the Intel macos-13 runner is too memory-starved (4 GB VM) to build this app and is slow/flaky; arm64 macos-14 can't run Docker at all. Drop Docker-on-macOS entirely: keep the full docker-compose smoke on ubuntu-22.04 (the canonical Linux self-host target), and prove macOS self-host parity (both Intel + Apple Silicon) via a native `next build` matrix [macos-13, macos-14] with the same heap bump. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The real cause of the failing assertion: FREE_TRIAL.trialTotalTokens was bumped 200K → 700K on 2026-05-15, but the test still fed 200K usage and expected a trial_exceeded throw. 200K < 700K → no throw → expect.fail() threw the AssertionError we saw. Reference FREE_TRIAL.trialTotalTokens directly so the test tracks the constant and can't go stale again. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…protocol/sdk The Docker builder copied only the root package.json before npm install, so the packages/mcp workspace deps were never installed. `next build` type-checks the whole project (incl. packages/mcp, which imports @modelcontextprotocol/sdk) and failed: "Cannot find module '@modelcontextprotocol/sdk/server/stdio.js'". Copy packages/mcp/package.json before install so npm (workspaces) resolves it. (CI Build + native macOS passed because their npm install ran against the full checkout.) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…hip client Two runtime issues that kept the web container from going healthy: 1. entrypoint ran `prisma db push --skip-generate`, but Prisma 7 removed that flag — it errored, printed help, and crashed the container (set -e). Use a plain `prisma db push`. 2. The Prisma client uses a custom output (src/generated/prisma, not node_modules/.prisma) that the runner stage never copied, so the server + db push couldn't resolve @/generated/prisma/client. Copy it explicitly. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…macos-13 Two things kept the ubuntu docker-compose smoke from going healthy / the run from completing: 1. The schema has a vector(1024) column but docker-compose used postgres:16-alpine (no pgvector) and nothing created the extension, so `prisma db push` failed: ERROR: type "vector" does not exist. Switch to pgvector/pgvector:pg16 + a first-boot init that runs CREATE EXTENSION vector. This is a real self-host correctness fix, not just CI. 2. macos-13 native build sat queued 45+ min (GitHub Intel-mac runner unavailability) and wedged the whole workflow from ever completing. Drop it from the matrix; macos-14 (Apple Silicon) covers native build parity. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

The container went "unhealthy" (compose --wait failed) even though the app boots fine (Postgres reachable, db push synced, Next.js Ready). Decouple: `docker compose up -d --build` (no --wait), and let the explicit probe step gate — now printing the HTTP code + body each attempt so a 503 reveals the actual healthz error instead of an opaque "unhealthy". Either goes green or surfaces the precise runtime cause. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Operator decision: merge now, track the Docker HTTP-serving gap. The ubuntu job now asserts the core self-host proof as fatal — image builds, container boots, Postgres reachable, schema synced via prisma db push — and treats the in-container HTTP route probe as non-blocking, because the hand-rolled `next start` runtime doesn't serve built routes (404; needs output:standalone, tracked as a follow-up issue). Makes selfhost-smoke green on what genuinely works instead of red on the one packaging gap. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…, skill promotion Three new autonomous agents that close the AI-Native Agent loop's "proactive preparation + intelligent intervention" principles: - P2-02 src/lib/agents/anomaly-detector.ts — scans each client's ClientContext + recent ConversationMessages for out-of-pattern signals (threshold breaks, contradicting facts, unusual cadence). Writes ApprovalItem of type "anomaly_alert" with severity high/medium/low → priority 80/50/25. Confidence < 0.65 outputs drop silently so the queue doesn't fill with noise. Pulls active pattern-learner rules into the system prompt so previously- dismissed pattern shapes get downweighted automatically. - P2-03 src/lib/agents/comms-drafter.ts — once-daily (17:00 local), drafts client-facing reminder emails for each pending action item that needs a recipient nudge. Tone-aware via client preferences + pattern-learner. Each draft becomes an ApprovalItem of type "email_draft" with priority weighted by nudgeWithinDays so soonest- needed nudges sort to the top of the queue. - P2-05 src/lib/skill-generator.ts — nightly batch that promotes AgentRule rows from "candidate pattern" to "promoted skill" once appliedCount ≥ 5, confidence ≥ 0.85, and recent verdictWeight average ≥ 0.85. Promoted rules surface in the system prompt with stronger language ("APPLY this template" vs "consider this pattern"). Demotion is just a flag flip — no row deletes if the operator loses confidence later. Implemented inside the existing AgentRule schema rather than introducing a new Skill table; the promoted flag lives in the action JSON. Lovable-mark gate touched: #8 (background agent producing real value — nightly briefing + anomaly detector + comms drafter now stack multiple ApprovalItems per active client per day). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…03) + self-paging wired (P1-07) Closes lovable-mark gate #2 (memory loop) for the *whole product*, not just daily-briefing. Replaces the flat `prisma.clientContext. findMany({ take: 50 })` block scattered across runner/daily-briefing/ anomaly-detector/comms-drafter/chat with a single 5-tier composer that aggregates: - **T0 ClientProfile** (250 tokens) — always-inline core memory a-la Letta. Identity + industry + relationship length + operator-set tone / formats / note. Built from Client.preferences. - **T1 RollingDigest** (500 tokens) — nightly-compacted 30-day rolling summary stored in ClientContext.category="digest". Falls back to top-3 pinned facts when the compactor hasn't run yet. Zep "Community subgraph" plays this role. - **T2 VectorHits** (600 tokens) — hybridSearchKnowledgeBase (0.4×trigram + 0.6×cosine over content_embedding) + loadActiveFacts (bitemporal ClientFact rows). Only fires when the caller passes `query`. Mem0g-style structured beliefs + paragraph hits for grounding. - **T3 EpisodicTimeline** (300 tokens) — last 5 AgentTask summaries + last 3 ApprovalItem decisions, newest-first. Zep "Episode subgraph" — the model finally sees "we already sent that reminder Tuesday" as a tier rather than reconstructing from titles. - **T4 FirmPatterns** (250 tokens) — promoted vs candidate AgentRule entries. Surfaces the operator's repeated decisions to ALL agents now, not just daily-briefing. Promoted rules get "APPLY this template" framing per Hermes/skill-promotion lit. Composer (`src/lib/memory/loader.ts`) does proportional scaling when budget < 2000, greedy fit in T0→T1→T4→T3→T2 order, emit in stable T0→T1→T4→T3→T2 sequence for prompt-cache stability. Returns full per-tier observability map. Migrations: - daily-briefing.ts → 1800-token budget, no query (full scan). - anomaly-detector.ts → 1700-token budget, query="anomaly out-of- pattern unusual transaction threshold". - comms-drafter.ts → 1700-token budget, query="pending action items requiring outbound nudge reminder". - chat/route.ts → 1500-token budget (tighter — tool-use loops compound the system prompt across rounds), query=last user message. renderSystemPrompt drops the legacy raw-contexts rendering and embeds `memory.prompt` directly under "Quick reference". P1-03 Digest Compactor: `src/lib/memory/digest-compactor.ts` + nightly cron `/api/cron/digest-compactor`. Every active client (any ConversationMessage / AgentTask / ApprovalItem in last 30d) gets one 400-word past-tense plain-English digest written by Haiku 4 (~$0.30 per night for an entire firm). Idempotent — prior digest rows flip to `category: "digest_archive"` in the same tx. SLACK_WEBHOOK_URL receives an error ping when ≥1 client digest fails. CRON_SECRET-gated; `DIGEST_COMPACTOR_DISABLED=1` is the emergency kill. P1-07 Self-paging: `recall_archival` tool was already wired into chat tools by sub-agent A. Confirmed runner-managed agents see the same tool dispatch path so background agents can self-page mid-run. Hard 800-token cap injected into next-turn user message prevents runaway recall. Operator probe: `GET /api/dev-test/memory-snapshot?clientId=...& query=...&budget=...` — auth-gated, returns the rendered prompt + per-tier breakdown so an operator can see exactly what their agents are loading. Token economics: median chat turn dropped from ~3.5K input tokens to ~1.7K (-52%) on seeded fixture client. With 200 active clients × 12 chat turns/day × Anthropic Sonnet 4.5 input pricing, that's ~$1.50/day saved per active firm — or ~$45/month per firm of useless context cost. Validation: - `npm run type-check` clean. - `npm run lint` clean (only pre-existing warnings in unrelated files). - `vitest src/lib/memory/loader.test.ts` 19/19 pass — token counter, T0 profile renderer, composer skip + budget bounds + preloadedClient short-circuit + missing-client fallback. - New `tests/e2e/memory-tiering.spec.ts` covers prod auth-gated probe (skips when E2E_TEST_EMAIL unset). - All 4 migrated agents + chat type-check + lint clean. Research artifact: `.cycle/research/2026-04-28-memory-deep-dive.md` documents the external pattern survey (Letta MemGPT, Zep/Graphiti, Mem0/Mem0g, Cognee, MemKraft), brutal evaluation of the prior flat-list path, and the design rationale for each tier choice. Lovable-mark gates touched: #2 (memory loop closes — pattern learner + temporal facts + vector + digest + episodic ALL demonstrably influence every agent + chat output now), #8 (background agents producing real value — composer makes their prompts coherent rather than wall-of-text). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…efresh crons Two new daily/weekly crons that close audit gap #8 (FactEdge writes absent) and AEO research priority #6 (freshness refresh boosts ~2.5× citation rate). 1. **FactEdge inference** (`/api/cron/factedge-inference`, 04:00 UTC daily) - `inferEdgesForClient(clientId)` walks up to 30 active ClientFact rows per client, asks Claude (tool_use schema) to propose directed edges with one of 5 relations: causes / contradicts / refines / depends_on / follows. - Cost-bounded: skips clients with < 5 active facts. One Claude call per client. ~$0.50/day across ~10 firms × 5 clients. - Idempotent: writes via the existing `@@unique([fromFactId, toFactId, relation])` constraint; on conflict refreshes weight if higher. Re-running is safe. - Edge weight floor 0.55 — proposed edges below that drop. Match to the corpus's exact fact ids (no invented ids). - Slack agent_cron_summary / _warning emits aggregate counters at the end. Failure-per-client doesn't kill the cron. 2. **30-day freshness refresh** (`/api/cron/freshness-refresh`, 09:00 UTC Tuesday) - Picks up to 6 oldest blog posts whose `dateModified` is > 30 days ago. Per post, asks Claude to spot ONE stale claim + propose a replacement (verbatim phrase to update + 1-3 sentence proposed update + rationale + confidence). - Persists proposals to `AuditLog` with `action="content_freshness_proposal"`. Operator reviews via the audit table or a future admin page. **No automated PR creation** — the operator must validate + manually apply, which keeps the cron Vercel-runtime-safe (no gh credentials required). - 14-day cooldown per slug so a previously-proposed post doesn't get a duplicate proposal next Tuesday. - Confidence < 0.6 → skip. Most posts on most weeks have nothing to refresh; empty answer is correct most of the time. - Cost: ~$0.12/week (6 posts × ~$0.02 each at Sonnet 4.5). Safety guarantees: - Both crons are READ-mostly. Only writes are FactEdge rows (existing schema + unique constraint) and AuditLog rows (existing model). - Both use the existing CRON_SECRET / SEO_DEPLOY_SECRET / x-vercel-cron auth path that all other crons share. - Both fail closed on per-item errors so one bad client / post doesn't cascade. - vercel.json gets 2 new entries (Hobby-compliant: daily + weekly, not hourly). 233/233 vitest pass · type-check ✅. New crons additive — zero behavior change to existing daily-briefing / anomaly-detector / comms-drafter / digest-compactor flows. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix(selfhost): align Docker + smoke CI with flat repo layout The docker/ skeleton and selfhost-smoke.yml were drafted assuming an apps/web monorepo layout, but the extracted repo is flat (web app at the root, packages/mcp the only workspace). That mismatch made the self-host smoke matrix unrunnable. This fixes the whole path end-to-end: - Dockerfile.web: build at repo root (no apps/web / packages/core), use `npm install` (repo ships no lockfile), `npx prisma generate`, and a full-copy `next start` runner (robust vs. standalone tracing). Copy prisma.config.ts so `db push` resolves schema + datasource at runtime. - entrypoint.sh: schema path -> root; switch `migrate deploy` -> `db push` (no migrations/ folder ships); POSIX while-loop; no --accept-data-loss so a destructive diff fails loudly instead of dropping self-hoster data. - New GET /api/healthz liveness probe (server up + Postgres reachable), distinct from /api/health readiness which 503s on optional integrations. Both the compose healthcheck and the CI probe already point at it. - selfhost-smoke.yml: flat-repo path filters (src/prisma/public/configs), and COMPOSE_FILE=docker/docker-compose.yml so every `docker compose` step finds the file without -f. - .gitattributes: pin Dockerfile* to LF (Alpine build safety). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(ci): drop npm cache from setup-node (repo ships no lockfile) actions/setup-node@v4 with `cache: 'npm'` requires a lockfile to compute its cache key. The repo intentionally ships no package-lock.json, so every CI job was failing at the setup-node step before install even ran. Drop the cache directive across all five jobs (type-check / lint / build / test / mcp-pack) so installs run via `npm install`. Follow-up (not this PR): commit a package-lock.json to restore cacheable, reproducible installs. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(ci): run prisma generate before type-check and test The Prisma client uses the new prisma-client generator with a custom output (src/generated/prisma), imported as @/generated/prisma/client. The type-check and test jobs never generated it, so prisma.* resolved to `any` — the root of the bulk of the strict-mode TS7006/TS2339 cascade. The build job + Docker already run generate; add it to these two jobs. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(deps): restore googleapis (dropped in extraction, still imported) The gmail/outreach + google-sc modules import googleapis, but the dep was dropped during the OSS package.json restructuring, causing 7 TS2307 "Cannot find module 'googleapis'" errors that cascade. Restore the exact version the app uses (^144.0.0). Note: this operator growth/outreach tooling (outreach crons, Search Console SEO automation, Slack summaries) is surplus to the self-host product and is tracked for removal as a follow-up; restoring the dep unblocks the build now. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(stripe): cast apiVersion to LatestApiVersion (lockfile-free SDK drift) The last 2 of 149 type errors: stripe@^22 caret floats (no lockfile) and its pinned-version literal type drifted to 2026-04-22.dahlia, rejecting our deliberately-pinned 2026-03-25.dahlia. Cast to Stripe.LatestApiVersion to keep the intended runtime value type-stable across SDK minors. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(stripe): omit apiVersion (Stripe.LatestApiVersion type not exported) The cast to Stripe.LatestApiVersion failed (TS2694 — not an exported member in stripe@22). Omitting apiVersion entirely is always type-valid and uses the Stripe account's default pinned version — the correct default for a self-host deployment, and immune to the lockfile-free SDK caret drift. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * ci(selfhost): native build job for macos-14 (Apple Silicon can't run Docker) GitHub arm64 macOS runners lack nested virtualization, so Colima/docker compose fails at VM start (exit 125) — macos-14 could never pass the compose smoke. Split the matrix: docker-compose smoke runs on ubuntu-22.04 + macos-13 (full boot + DB + healthz), and a new native-build job proves Apple-Silicon parity by running prisma generate + next build on arm64 macOS (exercises the arch-sensitive Prisma engine / bcryptjs / lightningcss). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(stripe): omit apiVersion in e2e test + bootstrap script too tsc includes tests/, so the e2e spec's hardcoded apiVersion literal was the last type error (the bootstrap script mirrors it for consistency). Same fix as the client: omit apiVersion to stay type-stable across stripe@22 minors. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(token-budget): restore prototype chain on BudgetExceededError The class extends the built-in Error; transpilation breaks `instanceof` for genuine instances unless the prototype is reset. Add Object.setPrototypeOf in the constructor — fixes the lone failing vitest assertion (token-budget.test.ts > assertBudget — trial plan). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(build): raise V8 heap for next build (was OOMing on CI runner) `next build` of this heavy app (pdfjs/exceljs/pptxgenjs/tiptap/react-pdf) exceeded the default V8 old-space heap on the 7 GB ubuntu runner — "Ineffective mark-compacts near heap limit". Set NODE_OPTIONS --max-old-space-size=6144 in the CI build job and the Docker builder stage. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * ci(selfhost): ubuntu-only docker + native build matrix for both Macs Colima on the Intel macos-13 runner is too memory-starved (4 GB VM) to build this app and is slow/flaky; arm64 macos-14 can't run Docker at all. Drop Docker-on-macOS entirely: keep the full docker-compose smoke on ubuntu-22.04 (the canonical Linux self-host target), and prove macOS self-host parity (both Intel + Apple Silicon) via a native `next build` matrix [macos-13, macos-14] with the same heap bump. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(token-budget): update stale trial-allowance test (200K → constant) The real cause of the failing assertion: FREE_TRIAL.trialTotalTokens was bumped 200K → 700K on 2026-05-15, but the test still fed 200K usage and expected a trial_exceeded throw. 200K < 700K → no throw → expect.fail() threw the AssertionError we saw. Reference FREE_TRIAL.trialTotalTokens directly so the test tracks the constant and can't go stale again. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(docker): install workspace deps so next build finds @modelcontextprotocol/sdk The Docker builder copied only the root package.json before npm install, so the packages/mcp workspace deps were never installed. `next build` type-checks the whole project (incl. packages/mcp, which imports @modelcontextprotocol/sdk) and failed: "Cannot find module '@modelcontextprotocol/sdk/server/stdio.js'". Copy packages/mcp/package.json before install so npm (workspaces) resolves it. (CI Build + native macOS passed because their npm install ran against the full checkout.) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(docker): runtime boot — drop Prisma 7-invalid --skip-generate + ship client Two runtime issues that kept the web container from going healthy: 1. entrypoint ran `prisma db push --skip-generate`, but Prisma 7 removed that flag — it errored, printed help, and crashed the container (set -e). Use a plain `prisma db push`. 2. The Prisma client uses a custom output (src/generated/prisma, not node_modules/.prisma) that the runner stage never copied, so the server + db push couldn't resolve @/generated/prisma/client. Copy it explicitly. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(selfhost): pgvector Postgres image + create extension; drop dead macos-13 Two things kept the ubuntu docker-compose smoke from going healthy / the run from completing: 1. The schema has a vector(1024) column but docker-compose used postgres:16-alpine (no pgvector) and nothing created the extension, so `prisma db push` failed: ERROR: type "vector" does not exist. Switch to pgvector/pgvector:pg16 + a first-boot init that runs CREATE EXTENSION vector. This is a real self-host correctness fix, not just CI. 2. macos-13 native build sat queued 45+ min (GitHub Intel-mac runner unavailability) and wedged the whole workflow from ever completing. Drop it from the matrix; macos-14 (Apple Silicon) covers native build parity. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * ci(selfhost): gate on explicit verbose healthz probe, not compose --wait The container went "unhealthy" (compose --wait failed) even though the app boots fine (Postgres reachable, db push synced, Next.js Ready). Decouple: `docker compose up -d --build` (no --wait), and let the explicit probe step gate — now printing the HTTP code + body each attempt so a 503 reveals the actual healthz error instead of an opaque "unhealthy". Either goes green or surfaces the precise runtime cause. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * ci(selfhost): assert build+boot+db-sync (fatal); HTTP serve = known-gap Operator decision: merge now, track the Docker HTTP-serving gap. The ubuntu job now asserts the core self-host proof as fatal — image builds, container boots, Postgres reachable, schema synced via prisma db push — and treats the in-container HTTP route probe as non-blocking, because the hand-rolled `next start` runtime doesn't serve built routes (404; needs output:standalone, tracked as a follow-up issue). Makes selfhost-smoke green on what genuinely works instead of red on the one packaging gap. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…, skill promotion Three new autonomous agents that close the AI-Native Agent loop's "proactive preparation + intelligent intervention" principles: - P2-02 src/lib/agents/anomaly-detector.ts — scans each client's ClientContext + recent ConversationMessages for out-of-pattern signals (threshold breaks, contradicting facts, unusual cadence). Writes ApprovalItem of type "anomaly_alert" with severity high/medium/low → priority 80/50/25. Confidence < 0.65 outputs drop silently so the queue doesn't fill with noise. Pulls active pattern-learner rules into the system prompt so previously- dismissed pattern shapes get downweighted automatically. - P2-03 src/lib/agents/comms-drafter.ts — once-daily (17:00 local), drafts client-facing reminder emails for each pending action item that needs a recipient nudge. Tone-aware via client preferences + pattern-learner. Each draft becomes an ApprovalItem of type "email_draft" with priority weighted by nudgeWithinDays so soonest- needed nudges sort to the top of the queue. - P2-05 src/lib/skill-generator.ts — nightly batch that promotes AgentRule rows from "candidate pattern" to "promoted skill" once appliedCount ≥ 5, confidence ≥ 0.85, and recent verdictWeight average ≥ 0.85. Promoted rules surface in the system prompt with stronger language ("APPLY this template" vs "consider this pattern"). Demotion is just a flag flip — no row deletes if the operator loses confidence later. Implemented inside the existing AgentRule schema rather than introducing a new Skill table; the promoted flag lives in the action JSON. Lovable-mark gate touched: #8 (background agent producing real value — nightly briefing + anomaly detector + comms drafter now stack multiple ApprovalItems per active client per day). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…03) + self-paging wired (P1-07) Closes lovable-mark gate #2 (memory loop) for the *whole product*, not just daily-briefing. Replaces the flat `prisma.clientContext. findMany({ take: 50 })` block scattered across runner/daily-briefing/ anomaly-detector/comms-drafter/chat with a single 5-tier composer that aggregates: - **T0 ClientProfile** (250 tokens) — always-inline core memory a-la Letta. Identity + industry + relationship length + operator-set tone / formats / note. Built from Client.preferences. - **T1 RollingDigest** (500 tokens) — nightly-compacted 30-day rolling summary stored in ClientContext.category="digest". Falls back to top-3 pinned facts when the compactor hasn't run yet. Zep "Community subgraph" plays this role. - **T2 VectorHits** (600 tokens) — hybridSearchKnowledgeBase (0.4×trigram + 0.6×cosine over content_embedding) + loadActiveFacts (bitemporal ClientFact rows). Only fires when the caller passes `query`. Mem0g-style structured beliefs + paragraph hits for grounding. - **T3 EpisodicTimeline** (300 tokens) — last 5 AgentTask summaries + last 3 ApprovalItem decisions, newest-first. Zep "Episode subgraph" — the model finally sees "we already sent that reminder Tuesday" as a tier rather than reconstructing from titles. - **T4 FirmPatterns** (250 tokens) — promoted vs candidate AgentRule entries. Surfaces the operator's repeated decisions to ALL agents now, not just daily-briefing. Promoted rules get "APPLY this template" framing per Hermes/skill-promotion lit. Composer (`src/lib/memory/loader.ts`) does proportional scaling when budget < 2000, greedy fit in T0→T1→T4→T3→T2 order, emit in stable T0→T1→T4→T3→T2 sequence for prompt-cache stability. Returns full per-tier observability map. Migrations: - daily-briefing.ts → 1800-token budget, no query (full scan). - anomaly-detector.ts → 1700-token budget, query="anomaly out-of- pattern unusual transaction threshold". - comms-drafter.ts → 1700-token budget, query="pending action items requiring outbound nudge reminder". - chat/route.ts → 1500-token budget (tighter — tool-use loops compound the system prompt across rounds), query=last user message. renderSystemPrompt drops the legacy raw-contexts rendering and embeds `memory.prompt` directly under "Quick reference". P1-03 Digest Compactor: `src/lib/memory/digest-compactor.ts` + nightly cron `/api/cron/digest-compactor`. Every active client (any ConversationMessage / AgentTask / ApprovalItem in last 30d) gets one 400-word past-tense plain-English digest written by Haiku 4 (~$0.30 per night for an entire firm). Idempotent — prior digest rows flip to `category: "digest_archive"` in the same tx. SLACK_WEBHOOK_URL receives an error ping when ≥1 client digest fails. CRON_SECRET-gated; `DIGEST_COMPACTOR_DISABLED=1` is the emergency kill. P1-07 Self-paging: `recall_archival` tool was already wired into chat tools by sub-agent A. Confirmed runner-managed agents see the same tool dispatch path so background agents can self-page mid-run. Hard 800-token cap injected into next-turn user message prevents runaway recall. Operator probe: `GET /api/dev-test/memory-snapshot?clientId=...& query=...&budget=...` — auth-gated, returns the rendered prompt + per-tier breakdown so an operator can see exactly what their agents are loading. Token economics: median chat turn dropped from ~3.5K input tokens to ~1.7K (-52%) on seeded fixture client. With 200 active clients × 12 chat turns/day × Anthropic Sonnet 4.5 input pricing, that's ~$1.50/day saved per active firm — or ~$45/month per firm of useless context cost. Validation: - `npm run type-check` clean. - `npm run lint` clean (only pre-existing warnings in unrelated files). - `vitest src/lib/memory/loader.test.ts` 19/19 pass — token counter, T0 profile renderer, composer skip + budget bounds + preloadedClient short-circuit + missing-client fallback. - New `tests/e2e/memory-tiering.spec.ts` covers prod auth-gated probe (skips when E2E_TEST_EMAIL unset). - All 4 migrated agents + chat type-check + lint clean. Research artifact: `.cycle/research/2026-04-28-memory-deep-dive.md` documents the external pattern survey (Letta MemGPT, Zep/Graphiti, Mem0/Mem0g, Cognee, MemKraft), brutal evaluation of the prior flat-list path, and the design rationale for each tier choice. Lovable-mark gates touched: #2 (memory loop closes — pattern learner + temporal facts + vector + digest + episodic ALL demonstrably influence every agent + chat output now), #8 (background agents producing real value — composer makes their prompts coherent rather than wall-of-text). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…efresh crons Two new daily/weekly crons that close audit gap #8 (FactEdge writes absent) and AEO research priority #6 (freshness refresh boosts ~2.5× citation rate). 1. **FactEdge inference** (`/api/cron/factedge-inference`, 04:00 UTC daily) - `inferEdgesForClient(clientId)` walks up to 30 active ClientFact rows per client, asks Claude (tool_use schema) to propose directed edges with one of 5 relations: causes / contradicts / refines / depends_on / follows. - Cost-bounded: skips clients with < 5 active facts. One Claude call per client. ~$0.50/day across ~10 firms × 5 clients. - Idempotent: writes via the existing `@@unique([fromFactId, toFactId, relation])` constraint; on conflict refreshes weight if higher. Re-running is safe. - Edge weight floor 0.55 — proposed edges below that drop. Match to the corpus's exact fact ids (no invented ids). - Slack agent_cron_summary / _warning emits aggregate counters at the end. Failure-per-client doesn't kill the cron. 2. **30-day freshness refresh** (`/api/cron/freshness-refresh`, 09:00 UTC Tuesday) - Picks up to 6 oldest blog posts whose `dateModified` is > 30 days ago. Per post, asks Claude to spot ONE stale claim + propose a replacement (verbatim phrase to update + 1-3 sentence proposed update + rationale + confidence). - Persists proposals to `AuditLog` with `action="content_freshness_proposal"`. Operator reviews via the audit table or a future admin page. **No automated PR creation** — the operator must validate + manually apply, which keeps the cron Vercel-runtime-safe (no gh credentials required). - 14-day cooldown per slug so a previously-proposed post doesn't get a duplicate proposal next Tuesday. - Confidence < 0.6 → skip. Most posts on most weeks have nothing to refresh; empty answer is correct most of the time. - Cost: ~$0.12/week (6 posts × ~$0.02 each at Sonnet 4.5). Safety guarantees: - Both crons are READ-mostly. Only writes are FactEdge rows (existing schema + unique constraint) and AuditLog rows (existing model). - Both use the existing CRON_SECRET / SEO_DEPLOY_SECRET / x-vercel-cron auth path that all other crons share. - Both fail closed on per-item errors so one bad client / post doesn't cascade. - vercel.json gets 2 new entries (Hobby-compliant: daily + weekly, not hourly). 233/233 vitest pass · type-check ✅. New crons additive — zero behavior change to existing daily-briefing / anomaly-detector / comms-drafter / digest-compactor flows. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix(selfhost): align Docker + smoke CI with flat repo layout The docker/ skeleton and selfhost-smoke.yml were drafted assuming an apps/web monorepo layout, but the extracted repo is flat (web app at the root, packages/mcp the only workspace). That mismatch made the self-host smoke matrix unrunnable. This fixes the whole path end-to-end: - Dockerfile.web: build at repo root (no apps/web / packages/core), use `npm install` (repo ships no lockfile), `npx prisma generate`, and a full-copy `next start` runner (robust vs. standalone tracing). Copy prisma.config.ts so `db push` resolves schema + datasource at runtime. - entrypoint.sh: schema path -> root; switch `migrate deploy` -> `db push` (no migrations/ folder ships); POSIX while-loop; no --accept-data-loss so a destructive diff fails loudly instead of dropping self-hoster data. - New GET /api/healthz liveness probe (server up + Postgres reachable), distinct from /api/health readiness which 503s on optional integrations. Both the compose healthcheck and the CI probe already point at it. - selfhost-smoke.yml: flat-repo path filters (src/prisma/public/configs), and COMPOSE_FILE=docker/docker-compose.yml so every `docker compose` step finds the file without -f. - .gitattributes: pin Dockerfile* to LF (Alpine build safety). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(ci): drop npm cache from setup-node (repo ships no lockfile) actions/setup-node@v4 with `cache: 'npm'` requires a lockfile to compute its cache key. The repo intentionally ships no package-lock.json, so every CI job was failing at the setup-node step before install even ran. Drop the cache directive across all five jobs (type-check / lint / build / test / mcp-pack) so installs run via `npm install`. Follow-up (not this PR): commit a package-lock.json to restore cacheable, reproducible installs. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(ci): run prisma generate before type-check and test The Prisma client uses the new prisma-client generator with a custom output (src/generated/prisma), imported as @/generated/prisma/client. The type-check and test jobs never generated it, so prisma.* resolved to `any` — the root of the bulk of the strict-mode TS7006/TS2339 cascade. The build job + Docker already run generate; add it to these two jobs. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(deps): restore googleapis (dropped in extraction, still imported) The gmail/outreach + google-sc modules import googleapis, but the dep was dropped during the OSS package.json restructuring, causing 7 TS2307 "Cannot find module 'googleapis'" errors that cascade. Restore the exact version the app uses (^144.0.0). Note: this operator growth/outreach tooling (outreach crons, Search Console SEO automation, Slack summaries) is surplus to the self-host product and is tracked for removal as a follow-up; restoring the dep unblocks the build now. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(stripe): cast apiVersion to LatestApiVersion (lockfile-free SDK drift) The last 2 of 149 type errors: stripe@^22 caret floats (no lockfile) and its pinned-version literal type drifted to 2026-04-22.dahlia, rejecting our deliberately-pinned 2026-03-25.dahlia. Cast to Stripe.LatestApiVersion to keep the intended runtime value type-stable across SDK minors. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(stripe): omit apiVersion (Stripe.LatestApiVersion type not exported) The cast to Stripe.LatestApiVersion failed (TS2694 — not an exported member in stripe@22). Omitting apiVersion entirely is always type-valid and uses the Stripe account's default pinned version — the correct default for a self-host deployment, and immune to the lockfile-free SDK caret drift. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * ci(selfhost): native build job for macos-14 (Apple Silicon can't run Docker) GitHub arm64 macOS runners lack nested virtualization, so Colima/docker compose fails at VM start (exit 125) — macos-14 could never pass the compose smoke. Split the matrix: docker-compose smoke runs on ubuntu-22.04 + macos-13 (full boot + DB + healthz), and a new native-build job proves Apple-Silicon parity by running prisma generate + next build on arm64 macOS (exercises the arch-sensitive Prisma engine / bcryptjs / lightningcss). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(stripe): omit apiVersion in e2e test + bootstrap script too tsc includes tests/, so the e2e spec's hardcoded apiVersion literal was the last type error (the bootstrap script mirrors it for consistency). Same fix as the client: omit apiVersion to stay type-stable across stripe@22 minors. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(token-budget): restore prototype chain on BudgetExceededError The class extends the built-in Error; transpilation breaks `instanceof` for genuine instances unless the prototype is reset. Add Object.setPrototypeOf in the constructor — fixes the lone failing vitest assertion (token-budget.test.ts > assertBudget — trial plan). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(build): raise V8 heap for next build (was OOMing on CI runner) `next build` of this heavy app (pdfjs/exceljs/pptxgenjs/tiptap/react-pdf) exceeded the default V8 old-space heap on the 7 GB ubuntu runner — "Ineffective mark-compacts near heap limit". Set NODE_OPTIONS --max-old-space-size=6144 in the CI build job and the Docker builder stage. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * ci(selfhost): ubuntu-only docker + native build matrix for both Macs Colima on the Intel macos-13 runner is too memory-starved (4 GB VM) to build this app and is slow/flaky; arm64 macos-14 can't run Docker at all. Drop Docker-on-macOS entirely: keep the full docker-compose smoke on ubuntu-22.04 (the canonical Linux self-host target), and prove macOS self-host parity (both Intel + Apple Silicon) via a native `next build` matrix [macos-13, macos-14] with the same heap bump. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(token-budget): update stale trial-allowance test (200K → constant) The real cause of the failing assertion: FREE_TRIAL.trialTotalTokens was bumped 200K → 700K on 2026-05-15, but the test still fed 200K usage and expected a trial_exceeded throw. 200K < 700K → no throw → expect.fail() threw the AssertionError we saw. Reference FREE_TRIAL.trialTotalTokens directly so the test tracks the constant and can't go stale again. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(docker): install workspace deps so next build finds @modelcontextprotocol/sdk The Docker builder copied only the root package.json before npm install, so the packages/mcp workspace deps were never installed. `next build` type-checks the whole project (incl. packages/mcp, which imports @modelcontextprotocol/sdk) and failed: "Cannot find module '@modelcontextprotocol/sdk/server/stdio.js'". Copy packages/mcp/package.json before install so npm (workspaces) resolves it. (CI Build + native macOS passed because their npm install ran against the full checkout.) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(docker): runtime boot — drop Prisma 7-invalid --skip-generate + ship client Two runtime issues that kept the web container from going healthy: 1. entrypoint ran `prisma db push --skip-generate`, but Prisma 7 removed that flag — it errored, printed help, and crashed the container (set -e). Use a plain `prisma db push`. 2. The Prisma client uses a custom output (src/generated/prisma, not node_modules/.prisma) that the runner stage never copied, so the server + db push couldn't resolve @/generated/prisma/client. Copy it explicitly. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * fix(selfhost): pgvector Postgres image + create extension; drop dead macos-13 Two things kept the ubuntu docker-compose smoke from going healthy / the run from completing: 1. The schema has a vector(1024) column but docker-compose used postgres:16-alpine (no pgvector) and nothing created the extension, so `prisma db push` failed: ERROR: type "vector" does not exist. Switch to pgvector/pgvector:pg16 + a first-boot init that runs CREATE EXTENSION vector. This is a real self-host correctness fix, not just CI. 2. macos-13 native build sat queued 45+ min (GitHub Intel-mac runner unavailability) and wedged the whole workflow from ever completing. Drop it from the matrix; macos-14 (Apple Silicon) covers native build parity. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * ci(selfhost): gate on explicit verbose healthz probe, not compose --wait The container went "unhealthy" (compose --wait failed) even though the app boots fine (Postgres reachable, db push synced, Next.js Ready). Decouple: `docker compose up -d --build` (no --wait), and let the explicit probe step gate — now printing the HTTP code + body each attempt so a 503 reveals the actual healthz error instead of an opaque "unhealthy". Either goes green or surfaces the precise runtime cause. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> * ci(selfhost): assert build+boot+db-sync (fatal); HTTP serve = known-gap Operator decision: merge now, track the Docker HTTP-serving gap. The ubuntu job now asserts the core self-host proof as fatal — image builds, container boots, Postgres reachable, schema synced via prisma db push — and treats the in-container HTTP route probe as non-blocking, because the hand-rolled `next start` runtime doesn't serve built routes (404; needs output:standalone, tracked as a follow-up issue). Makes selfhost-smoke green on what genuinely works instead of red on the one packaging gap. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

seungdo-keum and others added 17 commits May 27, 2026 14:19

seungdo-keum merged commit 02bfa76 into main May 27, 2026
7 of 8 checks passed

seungdo-keum deleted the phase-4/selfhost-smoke-flat-repo branch May 27, 2026 08:17

Navigation Menu

Uh oh!

fix(selfhost): align Docker + smoke CI with flat repo layout - #8

fix(selfhost): align Docker + smoke CI with flat repo layout#8
seungdo-keum merged 17 commits into
mainfrom
phase-4/selfhost-smoke-flat-repo

seungdo-keum commented May 27, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant