f0173440e4
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>
164 lines
6.1 KiB
TypeScript
164 lines
6.1 KiB
TypeScript
/**
|
|
* A stored conversation, read back into the shape the transcript renders.
|
|
*
|
|
* The store keeps one row per THING that happened — a question, a tool call, a
|
|
* proposed change, an answer — because that is what an append-only ledger has
|
|
* to do to survive a turn that dies half-way through. The transcript renders one
|
|
* block per TURN, with its tools and its approval cards inside it. Folding the
|
|
* rows back into turns is therefore not a formality; it is the difference
|
|
* between reopening a conversation and reopening a log file.
|
|
*
|
|
* The wire shapes below mirror `PiggyConversationDetail` in
|
|
* apps/api/src/services/piggy-conversations.ts. They are restated rather than
|
|
* imported because the browser cannot import from the API package, and every
|
|
* field is optional-tolerant on read for the same reason: a row written by an
|
|
* older build must reopen as a slightly plainer message, never as a blank pane.
|
|
*
|
|
* Standing caveat, stated where it will be found: NOTHING WRITES THESE ROWS YET.
|
|
* `PiggyConversationService.appendMessage` exists and is tested, and the chat
|
|
* relay does not call it — see the report. So today every stored conversation
|
|
* reopens empty, and this module is the half of the loop that is ready.
|
|
*/
|
|
import type { PiggyChatContext, PiggyMode, PiggyProposedChange } from '@pig/core';
|
|
import type { ApprovalStep, ToolStep, TranscriptMessage } from '@/lib/piggy-chat';
|
|
|
|
export interface StoredPiggyMessage {
|
|
id: string;
|
|
seq: number;
|
|
role: 'user' | 'assistant' | 'tool';
|
|
content: string;
|
|
reasoning: string | null;
|
|
model: string | null;
|
|
mode: PiggyMode | null;
|
|
inputTokens: number | null;
|
|
outputTokens: number | null;
|
|
costMicroCents: number | null;
|
|
finishReason: string | null;
|
|
tool: {
|
|
callId: string;
|
|
name: string;
|
|
arguments: Record<string, unknown> | null;
|
|
result: Record<string, unknown> | null;
|
|
ok: boolean | null;
|
|
} | null;
|
|
approval: {
|
|
id: string;
|
|
change: PiggyProposedChange;
|
|
decision: 'apply' | 'reject' | null;
|
|
decidedAt: string | null;
|
|
} | null;
|
|
error: string | null;
|
|
createdAt: string;
|
|
}
|
|
|
|
export interface StoredPiggyConversation {
|
|
id: string;
|
|
title: string;
|
|
model: string | null;
|
|
mode: PiggyMode | null;
|
|
context: PiggyChatContext | null;
|
|
createdAt: string;
|
|
updatedAt: string;
|
|
messages: StoredPiggyMessage[];
|
|
}
|
|
|
|
/**
|
|
* A change that was proposed and never answered.
|
|
*
|
|
* It is not offered as pending on reopening, and that is deliberate rather than
|
|
* cautious: the agent holds a proposal for the length of its own turn, so by the
|
|
* time a transcript is read back from the database there is nothing left at the
|
|
* other end for an Apply button to reach. Showing the buttons would collect an
|
|
* error; showing the card settled says what happened.
|
|
*/
|
|
const UNANSWERED = 'This change was never answered, and the turn that proposed it has ended.';
|
|
|
|
export function toTranscript(messages: StoredPiggyMessage[]): TranscriptMessage[] {
|
|
const transcript: TranscriptMessage[] = [];
|
|
// The assistant turn currently being assembled. Tool rows and approval rows
|
|
// belong to whichever answer they were streamed alongside, and they arrive
|
|
// BEFORE its text — the answer is written last.
|
|
let open: TranscriptMessage | null = null;
|
|
|
|
for (const row of [...messages].sort((a, b) => a.seq - b.seq)) {
|
|
if (row.role === 'user') {
|
|
if (open) transcript.push(open);
|
|
open = null;
|
|
transcript.push({ id: row.id, role: 'user', content: row.content });
|
|
continue;
|
|
}
|
|
|
|
if (row.tool) {
|
|
const turn: TranscriptMessage = open ?? newTurn(row.id);
|
|
turn.tools = [...(turn.tools ?? []), toToolStep(row.tool)];
|
|
open = turn;
|
|
continue;
|
|
}
|
|
|
|
if (row.approval) {
|
|
const turn: TranscriptMessage = open ?? newTurn(row.id);
|
|
turn.approvals = [...(turn.approvals ?? []), toApprovalStep(row.approval)];
|
|
open = turn;
|
|
continue;
|
|
}
|
|
|
|
// A second answer inside one turn cannot happen on the wire, but a repaired
|
|
// or re-run conversation could hold one; starting a fresh block is the only
|
|
// reading that does not silently concatenate two answers into one.
|
|
if (open && open.content) {
|
|
transcript.push(open);
|
|
open = null;
|
|
}
|
|
const turn: TranscriptMessage = open ?? newTurn(row.id);
|
|
turn.id = row.id;
|
|
turn.content = row.content;
|
|
turn.reasoning = row.reasoning ?? undefined;
|
|
turn.model = row.model ?? undefined;
|
|
turn.mode = row.mode ?? undefined;
|
|
turn.inputTokens = row.inputTokens;
|
|
turn.outputTokens = row.outputTokens;
|
|
turn.costMicroCents = row.costMicroCents;
|
|
turn.finishReason = row.finishReason ?? undefined;
|
|
turn.error = row.error ?? undefined;
|
|
transcript.push(turn);
|
|
open = null;
|
|
}
|
|
|
|
if (open) transcript.push(open);
|
|
return transcript;
|
|
}
|
|
|
|
function newTurn(id: string): TranscriptMessage {
|
|
return { id, role: 'assistant', content: '', tools: [], approvals: [], pending: false };
|
|
}
|
|
|
|
function toToolStep(tool: NonNullable<StoredPiggyMessage['tool']>): ToolStep {
|
|
return {
|
|
id: tool.callId,
|
|
name: tool.name,
|
|
arguments: tool.arguments ?? {},
|
|
// `ok: null` is a call the store never saw finish. It is drawn as succeeded
|
|
// rather than running: a spinner in a transcript read back from disk would
|
|
// never stop, and the payload beside it is the evidence either way.
|
|
state: tool.ok === false ? 'failed' : 'succeeded',
|
|
result: tool.result ?? undefined,
|
|
/*
|
|
* No clock. `startedAt` is `performance.now()` on the live path, which is
|
|
* milliseconds since this document loaded and means nothing for a call made
|
|
* last Tuesday. Zero with no `durationMs` renders as a step with no timing,
|
|
* which is honest; a computed one would be fiction.
|
|
*/
|
|
startedAt: 0,
|
|
};
|
|
}
|
|
|
|
function toApprovalStep(approval: NonNullable<StoredPiggyMessage['approval']>): ApprovalStep {
|
|
if (approval.decision === 'apply') {
|
|
return { change: approval.change, state: 'applied', decision: 'apply' };
|
|
}
|
|
if (approval.decision === 'reject') {
|
|
return { change: approval.change, state: 'rejected', decision: 'reject' };
|
|
}
|
|
return { change: approval.change, state: 'failed', error: UNANSWERED };
|
|
}
|