ThothII/AGENTS.md

AGENTS.md

This file provides guidance to Codex (Codex.ai/code) when working with code in this repository.

Start here

Read PROJECT_STATE.md for the current-state snapshot: what was last built, pending manual gates, workspace/secret layout, and design-doc locations. This file holds the stable commands + architecture mental model; PROJECT_STATE.md holds the evolving detail. Design history lives in docs/superpowers/specs/ and docs/superpowers/plans/.

Commands

The repo has three independently-built layers. Run the local Docker stack with ./scripts/run-stack.sh after creating deploy/env/local.env; it starts the base+local Compose profile with frontend, core, qdrant, embedding, and the one-shot embedding-model-init. The core image contains Pi. Qdrant and Ollama are internal Compose services; DWH and LLM remain external configuration endpoints.

Native host CLI tht (tools/tht/)

• Operator surface: setup, start, stop, status, doctor, auth, workspace, and pi.
• Use tht --installation <absolute-path>/thothii-installation.yaml <command> for installation, authentication, diagnostics, lifecycle, and workspace operations.

harness/ (Python workflow tht CLI + Pi gate extension)

• Install: cd harness && python -m venv .venv && pip install -e ".[dev]" (puts tht on PATH)
• Test: .venv/bin/pytest -q — l2 (real GLM + remote DB) is opt-in via addopts = -m 'not l2'; l0 (testcontainers) needs Docker
• Single test: .venv/bin/pytest tests/test_session_mutations.py::test_set_name -v (or -k <pattern>); include e2e with -m l2
• Lint: .venv/bin/ruff check . (line-length 100)

backend/ (Fastify + TypeScript, vitest)

• Dev: npm run dev (tsx watch src/server.ts) · Build: npm run build (tsc → dist/)
• Test: npx vitest run · Single: npx vitest run test/routes-sessions.test.ts -t "rename"
• Typecheck: npx tsc --noEmit -p . (vitest does NOT type-check — run this before committing)

frontend/ (React 18 + Vite + vitest)

• Dev: npm run dev (Vite; set VITE_BACKEND_URL) · Build: npm run build
• Test: npx vitest run · Single: npx vitest run src/shell/NavSessions.test.tsx
• Typecheck: npx tsc -b · E2E: npm run e2e (Playwright)

No ESLint on the TS layers — tsc is the gate. Tests use vitest + MSW (no network).

Architecture (the parts that need multiple files to see)

frontend (React/SSE) → backend (Fastify) → pi --mode rpc → tht/harness → DWH (read-only)

• The harness owns the workflow and all persistence. The Python workflow CLI tht inside core is deterministic; harness/.pi/extensions/tht-gate.js is a Pi extension that drives an 8-phase NL→SQL workflow. The single source of workflow truth is harness/workflow.yaml; the orchestration rules the model must follow are harness/.pi/skills/tht-sessione/SKILL.md. "Current phase" is computed by folding the decision ledger (harness/tht/phase.py), not stored — read it before reasoning about phase logic.

• Persistence = phase documents, NOT chat. A session is a directory under the workspace's sessions/ path: session_manifest.yaml + per-phase artifacts (question.md, schema_linking.json, sql_final.sql, …) + review_decisions.jsonl. The contract (SKILL.md): "the persisted state is the truth — what is not recorded did not happen." There is no verbatim transcript store. A resumed Pi process rebuilds context from tht session show <id> + the on-disk artifacts.

• The backend is a thin bridge with no database. ThtRunner shells the Python workflow tht subcommands inside core; PiProcessManager runs one Pi child per session and bridges its RPC stream; SessionBridge maps Pi RPC events → client events (ui_request/text_delta/info); SseHub fans them out over SSE to the browser. App settings live in a JSON file (backend/data/settings.json), not a DB.

• Human-in-the-loop gate contract. The model proposes; a human reviewer decides at gates via widgets (reviewer_select = single pick — a chosen option carrying a decision payload auto-confirms/persists directly, an option without one only asks; reviewer_decide = multiselect, each choice IS a decision; reviewer_confirm = artifact/phase gate). The frontend renders these widget-descriptors (src/widgets/ registry) and the live transcript is rebuilt in-memory from the SSE stream (src/store/sessionStore.ts) — it is not persisted.

Project-specific gotchas

• tht's -c/--config is a PER-COMMAND option — it must follow the subcommand, never precede it (ThtRunner.buildArgv enforces this; prepending caused live 500s).
• --json output must be pristine (only valid JSON on stdout) — used as a machine contract.
• UI strings are English; document content stays the workspace language (Italian for psd) because it's the real data. Only chrome/labels are English.
• Workspaces (harness/workspaces/*.yaml) set the DB target and absolute paths.sessions/artifacts/indexes — for psd these point at a separate, uncommitted repo (tht-workspace-psd/). Secrets live ONLY in harness/.env (gitignored).
• Settings are global (backend/data/settings.json: workspace/provider/model/thinking); the New-session form is question-only.
• Resume: a resumable session re-enters at its last incomplete phase. The backend refuses resume with 409 when finalized or archived, and PiProcessManager.spawnFor must send /riprendi-sessione <id> (resume mode) vs /nuova-domanda (new) — sending the wrong prompt silently turns a resume into a new question.