pythia

Oracle's MCP gives your agent a connection. pythia gives it the judgment to use it.

English · Tiếng Việt · The Complete Guide → · Website →

An Agent Skills + CLI kit for developing PL/SQL on Oracle Database with AI coding agents (Claude Code, Codex, Cursor — any of the 76 agents npx skills supports).

pythia is a harness: a book of working rules an agent studies and follows when it sits next to a developer. Not a chatbot, not an autopilot — a disciplined assistant with three properties a good one has: knowledge (it asks the live schema, never its memory of one), judgment (it measures before it proposes), and obedience to rules it can recite (gates it will quote back to you rather than quietly skip).

Why ask the database instead of reading dumps

A real mid-size system, repo export vs live database, audited in 2026:

Code that "reads fine" against the dump references types and packages the dump never heard of. Every pythia command asks the live data dictionary instead — and every truncated output says so, so an agent never mistakes a partial answer for a full one.

The operating model: Learn → Ask → Do

Give the agent a problem — "add a column and every procedure that maintains it", "why does this report double rows", "port this fix" — and the kit walks it through the same three movements a careful senior developer makes.

1 · Learn — understand before proposing

The agent studies four things, in order, with tools instead of guesses:

Nothing in this phase writes. Reading is free, so the bar is: no proposal before the blast radius is known (pythia-impact's iron law) and no line written before the neighbours have been read (pythia-write's).

2 · Ask — the questions are part of the method, not an interruption

The kit makes the agent stop at exactly the moments where a human's judgment is the missing input, and forbids it to guess past them:

• Before any write: the full preview — diff, dependents, warnings — is relayed verbatim, and the agent waits for a real yes. A compliment is not a yes. Silence is not a yes. And the wait is no longer a matter of obedience: the write cannot complete until the developer mints the approval themselves.
• When the blast radius is large: ten or more dependents, or anything cross-schema, goes to the developer before code is written, not after.
• When sources of truth disagree: a standards document says one thing, the schema does another — that gap is a question ("rule nobody follows, new-code-only, or drift?"), never a silent pick.
• When something breaks: exit 3 means written but broken. The agent reports it exactly so, with the ready rollback — reporting success here is the one sin the whole kit is built to prevent.
• When policy refuses: the refusal is relayed as information, not routed around.

3 · Do — act inside a pipeline that cannot lie

Only after Learn and Ask does anything touch the database, and then only through one door: snapshot → impact → preview → token → approve → apply → verify → report. DDL self-commits in Oracle, so the snapshot is the only real undo — it runs first and no flag disables it. A content-bound token guarantees what lands is byte-for-byte what was previewed, and a developer-minted grant guarantees a person approved it at all — through a door that is always the developer's. The pythia_approve tool (MCP, on Claude Code and Codex) shows pythia's own approval card as a select prompt, and Claude Code prompts for it in every permission mode but dontAsk — no allow rule can skip it. In chat: an AskUserQuestion whose text is the card (pythia approve --card <token>), whose Approve — written into the hook payload by Claude Code, never by the agent — mints the one-time grant. At a console: pythia approve <token>. Without that grant apply --confirm refuses. A headless agent cannot --yes its own writes, cannot mint approval, and cannot loosen policy — each of those takes a human, answering or typing.

When the developer wants the agent to just work, they delegate once: an autopilot lease — one connection, the statement groups they name, a few hours and a write budget — approved through the same doors. Until it ends, apply and deploy land covered statements without asking; every write is still snapshotted, journaled and verified. A lease never covers a protected connection, destructive statements (unless the card named them), or a file with lint errors, and it pauses itself the moment other objects break. pythia autopilot report lists every write with its undo.

The same discipline holds when the developer does the work: src and impact silently snapshot what they read, so even a change made by hand in SQL Developer has a rollback file waiting (history lists them, drift is reported when source moved with no apply behind it).

Học – Hỏi – Làm — Learn, Ask, Do. If the agent cannot show which phase it is in, it is doing none of them.

Install

npx pythia-plsql           # everything: pip install + skills picker + config scaffold + hooks

Or the same thing piecewise:

pip install pythia-plsql   # the CLI (thin driver — no Oracle Instant Client needed)
python -m pythia install   # skills + .pythia/ scaffold + Claude Code CLAUDE.md/hooks + Codex AGENTS.md/MCP approver
pythia check               # fill in connections.json first, then verify

The pip package is the whole kit: with Node.js present, pythia install runs npx skills add (77 agents, symlinked updates; --source <git-url> for internal mirrors) — without Node it copies the bundled pack itself. Prefer global skills: pythia install -g once per machine serves every project, and per-project installs then skip the skills step automatically — one copy anywhere means no skill ever shows up twice in the agent's menu. Skills alone: npx skills add thaildhe172591/pythia, or /plugin marketplace add thaildhe172591/pythia.

pip install is once per machine; pythia install is once per project — run it in each repo's root to drop the skills and a fresh .pythia/connections.json there. The CLI always reads the config of the project you are standing in (searched upward from the current directory, no global fallback), so one global CLI never mixes up databases.

Updating follows the same split:

pip install --upgrade pythia-plsql   # new CLI version (once per machine)
pythia install                       # refresh a project's skills; config never touched

(npx skills update also works for skills installed via npx.)

Running from a clone works too — python scripts/pythia.py <command>; every printed follow-up command matches how you invoked it. Windows, macOS, Linux and WSL are all CI-tested.

Setting up Claude Code and Codex

pythia install configures whichever of these you use — in one run, from the project root. Here is exactly what it wires for each, and the one or two steps each agent needs you to finish by hand.

Claude Code

1. Run python -m pythia install (or npx pythia-plsql) in the project root. It registers the write-free MCP approver in .mcp.json and merges into .claude/settings.json, keeping every key you already have there:

   • a SessionStart hook: the Learn → Ask → Do guide plus the project's live state (connections, autopilot lease, conventions, approval door);
   • a PreToolUse guard on Bash and SQLcl run-sql: pythia's read-only commands pass without a prompt that carries no decision, already-approved writes pass, and writes that skip the gate (sqlplus/SQLcl DDL) are denied;
   • PostToolUse hooks: on AskUserQuestion it mints the approval grant; on Bash it puts the verdict rules (exit 3 is broken) next to the result;
   • a deny rule so the agent cannot forge a grant, and the MCP server enabled.

   It scaffolds .pythia/connections.json (mode 600) and adds it to the project's .gitignore — commit that line — and warns if git already tracks the file: its passwords are then in the history, so untrack it and change them.

   It also writes the harness into CLAUDE.md — the file Claude Code itself loads every session — in a marker-delimited block, leaving your own text untouched. The rules hold even with --no-hooks, before a restart, or while the hooks are untrusted; when the block is there, the SessionStart hook adds only the live state instead of repeating the guide.

   The interpreter is chosen by what really starts pythia. A hook that cannot start fails open — Claude Code treats its error as non-blocking — so a bare python -m pythia goes into the shared files only when it imports pythia outside a virtual environment. Installed in a venv, with pipx, or where python is another interpreter, the hooks go to .claude/settings.local.json (this machine's file, kept out of git) with the absolute interpreter, and the MCP approver to Claude Code's local scope (claude mcp add --scope local); your teammates never get a path from your machine. Every command pythia prints uses the spelling that starts it here, the session preamble's RUN line names it for the agent (CLAUDE.md, being shared, can only say pythia), and the guard lets reads through in that spelling — this interpreter's own path, never another. pythia doctor starts every launcher it finds and fails the ones that cannot import pythia — except where this machine's own launcher already serves the same hooks: then the team's shared copy erroring here is a warning (keep it for teammates or remove it). Codex's session hook is a warning too: it carries only the session guide, and AGENTS.md the rules.

   Skills land in .claude/skills (or ~/.claude/skills with -g).

2. Restart the Claude Code session so the hooks and the MCP server load.

3. Run pythia doctor. It confirms the wiring and names every setting known to make AskUserQuestion return without showing the question — an allow rule or a skill's allowed-tools naming it, a PreToolUse hook with an empty or * matcher, dontAsk mode. That class of setting is why chat approval could fail outside bypassPermissions; the MCP door is immune to all but dontAsk.

4. Optional: copy the allow-list and autoMode from examples/claude-code-settings.example.json. Those are your permission preferences, so pythia does not set them for you (why).

Approving a change: the agent previews with pythia apply, calls the pythia_approve tool (or asks an AskUserQuestion carrying the card), you click Approve, and the agent runs apply --confirm itself.

Working autonomously: say so. The agent requests a lease (pythia autopilot request), you approve its card once, and it builds — tables, comments, packages, releases — without stopping for each write. acceptEdits mode plus a lease is the fully hands-off setup.

Codex

1. Run python -m pythia install in the project root. It:

   • writes the harness into AGENTS.md (Codex loads it every session);
   • registers a write-free MCP approver in .codex/config.toml;
   • adds a SessionStart guide hook in .codex/hooks.json.

   Skills land in .agents/skills (or ~/.agents/skills with -g) — the directory Codex reads.

2. In Codex, trust the pieces once: /mcp (the pythia server) and /hooks (the project .codex layer).

3. Restart the Codex session so the new AGENTS.md loads.

4. Check it: /skills lists the pythia skills, /mcp shows pythia: connected.

Needs Codex ≥ v0.120 for the in-chat approval (MCP elicitation); an older or headless Codex falls back to the console pythia approve <token>. Do not run Codex under danger-full-access — that sandbox auto-approves elicitations and bypasses your Approve.

Approving a change: the agent previews with pythia apply, calls the pythia_approve tool, you answer a select Approve / Reject prompt right in Codex, and on Approve the agent runs apply --confirm itself.

Commands

Every command takes --json (machine output), --conn (pick a connection), and caps output with explicit truncation markers so context windows stay intact.

The safety net covers hand edits too: src and impact snapshot the object silently into the journal, each with a runnable rollback file, so a change made later in SQL Developer still has something to go back to — pythia history <OBJECT> lists the versions. Source that moved with no apply behind it is reported as drift.

Your house style is config, not folklore: put naming patterns in .pythia/conventions.json and apply previews warn when a new object's name drifts; put the prose rules in .pythia/conventions.md and the skills make every agent read them first (pythia conventions shows both).

Security & write policy

The account is the real security layer — the policy file is an application-side fence. Give the agent its own revocable credential with proxy authentication (agent_user[schema_owner], no ANY privileges, no owner password shared):

pythia agent-user --save   # SQL for the DBA + matching credential saved as <conn>_agent
pythia check               # after the DBA ran it: proxy session, warning gone

One run does both — the password is regenerated each run, so the SQL and the saved config must come from the same run. Optional convenience: doing it by hand with examples/agent-user-setup.example.sql works just as well. pythia check warns when the session runs with more power than the task needs.

The developer approval that apply requires is minted per agent — Claude Code's AskUserQuestion hook, or Codex's pythia_approve MCP tool, or the console pythia approve <token> anywhere. How each is wired is in Setting up Claude Code and Codex above.

Per-group write policy, .pythia/policy.json (defaults shown):

The groups that cannot be snapshotted default to deny — and the refusal says that, instead of "policy forbids it". Anonymous PL/SQL blocks are refused outright. Unrecognized statements are refused, never guessed into a group.

Threat model and what a snapshot does not restore: SECURITY.md.

Reads may flow through Oracle's official SQLcl MCP server (sql -mcp, keep -R 4); it audits every interaction in DBTOOLS$MCP_LOG. Writes never do — only pythia apply has the snapshot, preview, verify and journal.

Skills

The skills teach the agent the workflow — gates, not suggestions:

• The method: using-pythia (the router) · pythia-spec (open decisions are asked, not assumed) · pythia-explore · pythia-impact (before any change) · pythia-verify (evidence before "done").
• Building on Oracle: pythia-table (tables, columns, constraints, indexes, every column commented) · pythia-package (packages, procedures, functions, triggers to industry-standard design) · pythia-write (copy the codebase's conventions) · pythia-comment · pythia-tune (measure, change one thing, measure again) · pythia-review (antipatterns).
• Landing it: pythia-apply (the gate) · pythia-deploy (release scripts, in order, one approval) · pythia-autopilot (one delegated lease instead of every approval).
• The team's own way: pythia-conventions (the LEAD's naming and lint rules, verified against real names) · pythia-skill-author (capture a workflow as a skill; pythia skill new / skill lint) · pythia-setup (connections, agent user, pythia doctor, a robust dev environment).

Compatibility

Star History

Contributing

Tests need no database — the fakes prove the safety properties (snapshot before write, deny touches nothing, stale tokens refused). See CONTRIBUTING.md.

MIT — see LICENSE.