main
10 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 claude
|
f0173440e4 |
Put Piggy on Prime Agent, and let it write to the book
Piggy was a hand-rolled OpenAI tool loop. It is now a Prime Agent session — Prime Intellect's own harness, embedded as a Node library — answering from PIG's tools and, for the first time, able to put information into the CRM rather than only read it out. The harness is a coding agent, so the first job was taking the coding agent away from it. `noTools: 'all'` plus an explicit allowlist leaves the model with PIG's ten `pig_*` tools and no bash, no filesystem, no IPython. That holds under attack: a hostile extension, a skill and a settings file planted in the agent's own directory, then `setActiveToolsByName` called with every built-in, still leaves ten tools, all ours. Both lines are load-bearing — `noTools` alone registers nothing, and the allowlist is what admits our own. Writing is gated rather than assumed. A change is proposed, not made: the tool returns a description, the transcript renders a diff card, and nothing reaches the database until someone presses Apply. Contracts, commitments, allocations and compliance always stop for a human whatever the mode. Every write runs through `executeMutation` as the calling user, so their capabilities and the audit trail apply exactly as they would to a human's. Four things about the SDK are wrong in its own documentation and cost a debugging cycle each: models.json does not resolve an env var name for `apiKey`, it sends the literal string; there is no built-in prime-inference provider in 0.84.1; a ResourceLoader you pass in is never reloaded for you; and the stock system prompt is a coding-assistant prompt that must be replaced — but replacing it also silently removes the tool list, because the harness only renders that section when it owns the prompt. AGENTS.md records all four. The expensive one was thinking level. The harness defaults to `medium`, and nemotron spent an entire 4,096-token budget reasoning and returned an empty answer. `low` was worse; `off` omits the parameter so the endpoint's default wins. An explicit `reasoning_effort: none` via `thinkingLevelMap` took a turn from 6,195 output tokens to 149. And a turn is now bounded. The harness loop is `while (true)` with no iteration cap; a runaway on a frontier model would have eaten the credit it is supposed to report on. Ceilings on model calls and tokens, enforced both through the harness hook and independently from the event stream, plus a per-user daily spend limit — and the ledger now records spend on turns that fail, which it previously discarded. Signing in lands on /piggy, which is a workspace: conversations down one side, the agent in the middle, what it did and what it cost beside it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
claude
|
99d165b5e5 |
Rebuild Piggy's interface, and give the demo book a business to describe
Piggy answered in raw markdown, threw away every tool result it streamed, and fought the reader's scroll on every token. The three surfaces that made it worth having — what it read, how it reasoned, what it cost — were all on the wire and none of them reached the screen. The transcript is now composed of five parts under components/piggy: answers render through streamdown, the container sticks to the bottom without pinning the reader there, tool steps say what they read and link to the record, and each turn carries its model and token count. Three lifecycle bugs went with them: Stop left a permanent spinner, a truncated stream was indistinguishable from thinking, and a failed send destroyed the message it failed to send. Underneath, the inference path grew timeouts, jittered retries on 429 and 5xx, tolerance of the malformed frames a 30B model emits, and an agent_runs row per turn so chat spend is observable. The system prompt now states that a field ending in Cents is cents — without it nemotron renders costPerGpuHourCents: 189 as "$189 per GPU-hour", which is a 100x error on the most scrutinised number in the room. The demo book was arithmetically incoherent: every deal's value contradicted its own allocation revenue by up to 3.6x, nothing had ever closed, no customer had any paper, and the marketplace was empty. Deal value is now derived from the allocation, the book clears 5.3% across five blocks with one deliberately underwater, and the renewal, compliance and agent-provenance machinery finally has rows to act on. A --clear that deleted every obligation, SLA term and capacity request in the database regardless of origin is scoped to the demo's own ids. Around that: accounts have a detail page, ⌘K searches the book, Settings can mint the API keys it always claimed to, and deploy.sh actually ships the agent instead of silently skipping its compose profile. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
 karti
|
45b70b17f0 |
Redesign Learn, and give it five real videos in Karti's voice
THE PAGE. The anonymous route rendered outside Shell, so it sat flush against
the viewport edge and read as a form rather than a product — which is the first
thing anyone at Prime Intellect sees when the link is shared. It now brings its
own chrome and leads with a hero; the platform track is a numbered course, the
concept tracks are a poster grid, and admin add/archive moved behind one Manage
toggle so they stop competing with the content. Verified in Chrome at 1440 and
393, light and dark: horizontal overflow is 0 in all three access states.
THE VIDEOS. Five ~30s walkthroughs, narrated in Karti's cloned voice through
Chatterbox and cut against real screen capture of the seeded demo book. The
audio is rendered FIRST and its measured duration drives the capture, because a
shot list that runs short leaves the narrator talking over a frozen frame and
one that runs long gets cut mid-sentence. Levels are loudness-normalised so
clips do not jump between videos.
Cap cannot take a programmatic upload — video.karti.ai needs an interactive
login — so PIG serves these itself. A native <video> on this origin needs no
iframe and therefore no CSP frame-src at all; Karti's own Cap recordings still
render through the existing iframe path, which is why the resolver is now a
discriminated union.
THREE THINGS THE VERIFIERS CAUGHT, all of which shipped green:
- createMediaRoutes was never mounted. Every layer landed — migration, seed,
both feeds, the bind mount, the docs — except the one that serves the bytes,
so /media/learn/* fell through to the SPA fallback and answered HTTP 200
text/html. The player showed a black box with working controls and no error.
The tests certified the route factory in isolation, which proves the handler
and says nothing about whether it is wired in. There is now an assertion
against the ASSEMBLED app, and it fails loudly on content-type — the failure
mode is a 200, not a 404.
- A symlink in the media directory escaped the root. resolve() is lexical and
stat() follows links, so the containment check this file's own header
promised did not hold. realpath before the check closes it.
- Vite proxied only /api, so self-hosted playback broke for anyone running the
app the documented way — in the same invisible 200-text/html manner.
Also: a duplicate media slug used to throw from the middle of seedDemo() and
take out every later section; it now reports and skips that one entry. And the
player has an onError state, because content-addressed filenames mean a
re-render deliberately leaves the old row pointing at a file that is gone.
The three DEMO platform rows are dropped — five real recordings supersede them,
and placeholders sitting under real ones made the page read as half-finished to
the audience it is meant to convince. The supply and demand concept rows stay:
there are no real recordings for those tracks yet, and an empty track hides the
shape of the page.
Tests 275, typecheck clean.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
||
|
karti
|
13dec6b4b8 |
Rebuild the shell, add Calendar and Learn, and govern reads
Seven parallel agents and an adversarial verification pass. The three things worth knowing before reading the diff: RBAC WAS ALREADY BUILT. docs/build-plan.md marks F2 and F3 outstanding and is stale — packages/core/src/permissions.ts and lib/mutation.ts shipped long ago. So this does not rebuild them; it closes the gaps an audit found. The big one is that reads were entirely ungoverned: every GET was "any authenticated member", so a junior demand rep and a research contractor could both pull per-block supplier cost and break-even prices from /api/capacity/margin, and every contract's negotiated terms. For a company whose margin is the business, that was the hole that mattered. Adds book:read / economics:read / team:read, a readGuard middleware, and a `viewer` role below member. THE BUTTON AND THE 403 DISAGREED — the exact thing F3 said must never happen. Contracts.tsx never called can() at all, so its save button was always enabled against a server requiring contract:sign; Capacity.tsx gated commitment creation on deal:write/demand while the server wanted commitment:write/supply. POST /api/activities was the one write bypassing executeMutation: no capability check, and any member could mutate accounts.lastActivityAt as a side effect. It is now a proper mutation() behind activity:write. The shell becomes three panes — a collapsible shadcn sidebar with an account switcher on the Piggy accent, a header with real search, and Piggy docked to the right, page-aware and persistent across navigation. The phone keeps its bottom tab bar, which is the thing this product already beat trycompai/crm on, and gains the sidebar as a sheet. Calendar is a projection over thirteen dated sources rather than a new table, because a table would duplicate dates that already live on contracts, deals and commitments and would drift — and one ledger answering the question is the whole argument. It surfaces export_authorizations and compliance_artifacts, which had indexed expires_at columns, schema comments saying they must be alerted on, and no read endpoint or UI anywhere. Learn carries two tracks. Concepts are members-only; the platform track can be opened with a share code by someone with no account. The code mints a scoped learn-only token and never a Principal — every route here resolves a principal and then checks capabilities, so a principal-minting code would be one missing check away from leaking the book. "Only platform-track rows may be code-visible" is a database CHECK constraint as well as a write-path rule, and a test asserts a valid learn token still gets 401 on /api/dashboard, /api/accounts and /api/contracts — the same invariant scripts/deploy.sh refuses to ship without. CD becomes tag-to-ship. CI publishes an image to the Gitea registry on a release-* tag and cloud-2 pulls it, so no credential on the shared runner can execute anything on production — by construction rather than by policy. Both halves of deploy.sh's original rule survive: nothing on the runner reaches the host, and a human still decides when it ships. deploy.sh gains a rollback and a public-origin check, and PIG_IMAGE now reaches compose through `sudo env`, without which sudo's env_reset silently resolved every release to pig:local. Tests 141 -> 261. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
karti
|
6cf80747cc |
Keep PIG out of search results until it is meant to be found
CI / verify (push) Successful in 3m11s
The app is pre-launch and shared by link with a handful of people at Prime
Intellect. It should not be accumulating a search footprint yet.
Three layers, because each covers a gap the others leave:
- robots.txt asks well-behaved crawlers not to fetch at all.
- The <meta name="robots"> tag covers the HTML document for anything that
fetched anyway.
- X-Robots-Tag covers everything that is NOT the HTML document — og.png,
the manifest, the built assets — which the meta tag cannot reach.
noarchive and nosnippet are there so a cache or an excerpt cannot outlive
the page once this is reversed.
Deliberately NOT stripped: the og:/twitter: tags. Link unfurlers are not
crawlers — they fetch on behalf of the person pasting the link, and a
rendered card is exactly what we want when this is shared.
The real gate remains authentication: / returns the sign-in screen and every
/api/ route returns 401. This only stops the app being indexed.
To go public: delete robots.txt, drop the meta tag, drop the header.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
||
|
karti
|
a6167629cc |
Move from npm to pnpm across the workspace, CI and the image
CI / verify (push) Successful in 3m23s
The monorepo was on npm workspaces. pnpm gives it a content-addressed store shared between the eight packages, a lockfile that records the whole graph rather than a flattened view of it, and — the reason this mattered in practice — `workspace:*`, which makes an internal dependency unambiguous instead of a version range that npm may satisfy from the registry. Mechanics: - `packageManager: pnpm@11.21.0` pins the version; corepack installs it in CI and in the image, so all three environments resolve identically. - The npm `workspaces` array is replaced by `pnpm-workspace.yaml`. pnpm ignores the former, and keeping both would leave two sources of truth. - All six internal dependencies moved to `workspace:*`. - Root scripts use `pnpm -r --if-present` and `pnpm -F <pkg>`. Two findings worth recording, both from running it rather than reading it: `tsx` was a devDependency, but the server runs TypeScript directly in production — the container's command is `pnpm exec tsx apps/api/src/server.ts`. Under npm this was concealed by the runtime stage re-installing tsx by hand after pruning dev dependencies. Under `pnpm install --prod` that sleight of hand stops working and the image simply fails to start. tsx is now declared in `dependencies`, which is what it has always actually been. The first image build failed with ERR_PNPM_ABORTED_REMOVE_MODULES_DIR_NO_TTY. That is not a pnpm bug: it had decided the modules directory was stale and wanted confirmation before deleting it, which a non-interactive build cannot give. The trigger was the host's `node_modules` reaching the build context — there was no `.dockerignore` at all. pnpm's tree is symlinks into a content-addressed store, so copying it into an image produces dangling links and a directory pnpm rightly considers corrupt. Fixed by adding `.dockerignore` and setting `CI=true`, which is required in any non-interactive pnpm build. `esbuild` is denied install scripts via `allowBuilds`. Its platform binary arrives through the optional dependency `@esbuild/linux-x64` and the postinstall only verifies it; confirmed by running the binary directly, which reports 0.25.12. Verified under pnpm: typecheck clean, 150 tests / 0 failures, e2e passes, web builds. The image was built and booted against a real Postgres — health ok, `/api/dashboard` 401 with an issuer configured, `/` and `/capacity` serve the SPA, `/og.png` serves as image/png, and the migrator runs from the pruned runtime stage. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
karti
|
c821b2ca07 |
Authenticate against any OIDC provider, for on-premises installs
CI / verify (push) Successful in 2m55s
The seam existed with only a Supabase implementation, so an on-prem deployment had no way to authenticate. A customer running PIG inside their own network already has Okta, Entra, Keycloak, Auth0 or Google Workspace; asking them to stand up a second identity system is a serious adoption tax and in a regulated environment usually refused outright. Setting PIG_OIDC_ISSUER is normally the whole configuration — the JWKS is discovered from the issuer's well-known document. PIG_OIDC_JWKS_URI skips discovery entirely for an air-gapped network. OIDC takes precedence over Supabase so an on-prem install can leave the hosted values in its environment file without them quietly taking over. Three decisions worth stating: Discovery is resolved lazily and the FAILURE is not cached. Doing it per request would put the customer's identity provider on the critical path of every API call; doing it eagerly at boot would mean their IdP rebooting takes the CRM down with it. So it happens on first use and retries on the next request. The audience check is optional but warned about loudly. Without it, a token the provider issued for ANY other application in the same tenant verifies here — a token minted for an unrelated internal tool would be accepted as a PIG session. It cannot be mandatory because some providers legitimately issue single-audience tokens. Email falls back through email, preferred_username and upn, because providers disagree, but a preferred_username without an "@" is ignored — PIG keys membership on the address, and a bare username must never become an account identity. Also fixed a warning that claimed "authentication is DISABLED" on a correctly configured OIDC deployment. That is worse than silence: an operator who reads it on a secure install learns to ignore the warnings. The dev bypass itself was already correct — it keys on the resolved provider rather than on Supabase. 18 new tests, most of them about what the provider must REFUSE: a foreign signing key, a foreign issuer, a token for a different application, an expired token, a token with no subject, and a discovery outage that must not become permanent. Keys are generated per test and the JWKS is served locally, so they run offline. Verified: production refuses to start with neither provider, starts with OIDC alone, enforces 401 on an unauthenticated request, and warns only about the genuinely missing admin list. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
karti
|
853bde2265 |
Build the agent-native compute CRM platform
CI / verify (push) Successful in 3m6s
|
||
|
karti
|
93818a2d2c |
Document two proxy traps found by deploying rather than assuming
The deployment came up healthy, served a valid certificate, and returned 200 — and was completely unreachable. Two distinct causes, both invisible from inside the host: 1. Every other site on this proxy binds to a private VNIC address. Caddy groups site blocks into servers BY listen address, so a block without `bind` landed in a separate server on :443. The specific listener wins for traffic arriving on that address, which is all public traffic after NAT, so requests hit the server that had never heard of these hostnames and fell through to an empty 200. Testing from the host with --resolve 127.0.0.1 worked perfectly, which is exactly why this was worth chasing from a third machine instead of trusting a local check. 2. The CSP blocked the inline pre-paint theme script, so dark-mode users would have seen a white flash on every load. Fixed with the script's hash rather than 'unsafe-inline', which would have defeated the policy, and rather than an external file, which would have reintroduced the flash. Editing that script changes its hash and silently breaks it, so that is written down. Verified from an independent host: health returns JSON, the app serves, an unauthenticated API call is refused, the short alias redirects, security headers are present, and the existing sites on the proxy are unaffected. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
karti
|
c747eb2aa7 |
Add deployment: Dockerfile, compose, proxy config, and docs
One container plus a Postgres behind any TLS-terminating proxy. Nothing is specific to a particular host. The app and API are served from a SINGLE origin. This is not tidiness: browser auth sessions live in per-origin storage, so splitting them across two hostnames makes sign-in loop in a way that presents as a server fault. The short alias redirects rather than serving a second origin. Two safety properties verified by running the image, not by reading the code: - With NODE_ENV=production and no SUPABASE_URL, the process refuses to start and says why. Serving the whole CRM unauthenticated is a worse outcome than failing to deploy, so the failure is deliberate and loud. - In production the development auth bypass does not apply: an unauthenticated request to /api/dashboard returns 401 rather than adopting the first user in the table. The Dockerfile typechecks all six packages as a build gate, so a deploy that does not compile fails at build time rather than in front of a user. Runtime runs unprivileged as `node`, and Postgres is not published to the host. Docs cover the ontology and why it is shaped this way, agent connection for Claude Code / Codex / prime-agent / Buzz, and the provenance rules governing seed data about real people — including how to have your record removed. Verified: image builds, container reports healthy, serves the SPA, enforces auth, and the production guard exits non-zero. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> |