mirror of
https://github.com/SHOGGOTH-SECTOR/sica-fondt.git
synced 2026-08-01 08:30:20 +00:00
24f816b6a3
Place useful parts of the surrounding repos into sica-fondt by layer, per the
body model (Ada = membrane; brain/endocrine/capabilities/knowledge non-Ada):
- brain/ LLM reasoning + providers (dapr, hermes, MoMoA)
- capabilities/ REPRAG sidecars: hermes tools/skills, dapr tools, parallel
dispatch, A51 channels, and the OSINT cluster
- knowledge/ LORAG corpus: 754 cyber-skills, agency personas, secure-coding,
MITRE ATT&CK data
- reference/ defensive threat-reference (C3, shhbruh doc) + AdaYaml parser
License handling: AGPL sources (worldosint, advanced_evolution, mercury,
Reticulum) and GPL DeTTECT are SPEC-only clean-room/port descriptions — no
copyleft code copied. MIT/Apache/data parts copied as working trees.
Safety: shhbruh escape/persistence material and C3 covert-C2 kept as reference
only, not wired into the running organism. See CONSOLIDATION.md.
https://claude.ai/code/session_01UehUqEXXJJCsHoA4voCU5c
70 lines
2.3 KiB
TypeScript
70 lines
2.3 KiB
TypeScript
import type { Api } from "@mariozechner/pi-ai";
|
|
import type { ExtensionAPI } from "@mariozechner/pi-coding-agent";
|
|
|
|
type ProviderPayload = Record<string, unknown>;
|
|
|
|
const OPENAI_PARALLEL_TOOL_CALL_APIS: ReadonlySet<Api> = new Set([
|
|
"openai-completions",
|
|
"openai-responses",
|
|
"openai-codex-responses",
|
|
"azure-openai-responses",
|
|
]);
|
|
|
|
function isRecord(value: unknown): value is ProviderPayload {
|
|
return typeof value === "object" && value !== null;
|
|
}
|
|
|
|
function hasTools(payload: ProviderPayload): boolean {
|
|
return Array.isArray(payload.tools) && payload.tools.length > 0;
|
|
}
|
|
|
|
export function addOpenAIApiParallelToolCallsToPayload(api: Api | undefined, payload: unknown): unknown {
|
|
if (!api || !OPENAI_PARALLEL_TOOL_CALL_APIS.has(api)) {
|
|
return payload;
|
|
}
|
|
|
|
if (!isRecord(payload) || !hasTools(payload) || payload.parallel_tool_calls !== undefined) {
|
|
return payload;
|
|
}
|
|
|
|
return {
|
|
...payload,
|
|
parallel_tool_calls: true,
|
|
};
|
|
}
|
|
|
|
export const PARALLEL_TOOL_CALLS_SECTION = `
|
|
## Execution Strategy
|
|
|
|
### Parallel Tool Calls
|
|
|
|
When multiple tool calls are independent, fire them ALL in the same turn. Sequential tool calls waste round-trips and slow down every task.
|
|
|
|
- Multiple file reads in the same turn when you already know the paths
|
|
- Multiple searches with different patterns fired simultaneously
|
|
- Combining read operations with search operations when they gather independent context
|
|
|
|
Before calling any tool, ask: "What other context will I need alongside this?" Then gather it all in one turn.
|
|
|
|
### Context Breadth Before Changes
|
|
|
|
Before modifying any file, gather enough context to get the change right on the first try:
|
|
|
|
1. Read the target file and any files it directly depends on - understand the dependency chain
|
|
2. Find every reference to symbols you will change - know the blast radius
|
|
3. Check for related tests - know what might break
|
|
4. Read inline comments or docs explaining design choices - understand the why
|
|
|
|
Multiple well-informed edits in one pass beats a cycle of edit-then-fix-then-fix-again.
|
|
`;
|
|
|
|
export default function openaiApiParallelToolCallsExtension(pi: ExtensionAPI): void {
|
|
pi.on("before_provider_request", (event, ctx) => {
|
|
return addOpenAIApiParallelToolCallsToPayload(ctx.model?.api, event.payload);
|
|
});
|
|
|
|
pi.on("before_agent_start", async (event) => ({
|
|
systemPrompt: `${event.systemPrompt}\n${PARALLEL_TOOL_CALLS_SECTION}`,
|
|
}));
|
|
}
|