AGENTS.md git:20261008.f9a477dA

Immutable. This exact content never changes and is served at /api/v1/blob/e80a8fa9fb4e4af7.

# tools-registry (`treg`) - guide for every coding agent

<!-- Editors: CLAUDE.md imports this file; Codex and Cursor read it directly. -->

treg is the tool catalog for an agent: one base URL, one token, and the agent can call a curated
catalog of external endpoints plus its own team's tools without ever holding an API key. The
load-bearing mechanic is a proxy that makes the caller's **real upstream request**, injects the
credential server-side when one is required, and relays the answer verbatim. A catalog endpoint
explicitly verified as public can instead relay with no injected credential. We never model an
upstream API.

## Paired treg.to checkout

For work on the hosted treg.to service, clone the public `treg` repository and private
`treg-internal` repository as siblings with those exact directory names. When `../treg-internal`
exists, treat both repositories as one operational workspace:

- `treg` owns public product code, portable behavior and self-hosting documentation.
- `treg-internal` owns live production configuration, operational runbooks, incident evidence and
  private admin tools.
- Read both repositories before changing production behavior, but never copy credentials, live
  environment exports, customer data or raw logs between them.
- Everything committed here is public, comments and commit messages included: describe treg.to
  only by mechanism, never by its numbers, dates, hosts, instance counts or schedules. Those go to
  `../treg-internal`.
- Commit and open PRs separately. State the merge order whenever one PR links to or depends on the
  other.

Do not clone `treg-internal` inside this repository and do not make it a Git submodule. Start agents
from the repository that owns the task; the sibling path supplies the other half of treg.to context.

## Non-negotiables

Everything else in this file is guidance; these are the contract, and they win over any other passage.

1. A team's own key always wins over treg's, is never metered, and is never routed or overflowed.
2. A hold (the balance `reserve` sets aside for one call) is settled or released exactly once, on
   every path: timeout, cancellation and exceptions included.
3. A request holds zero database connections while upstream or object-storage I/O is in flight.
   Keep `reserve` and `settle` separate; read archive pointers, close the session, then fetch bytes.
4. Plain `/call/` is a faithful relay: the injected credential, the transport headers listed in
   `src/treg/infra/upstream/relay.py`, and (on treg's shared key only) the per-org and, for pinned
   agents, per-pin re-scoping of the caller's `Idempotency-Key` and treg's own `User-Agent` in place
   of the caller's are the only rewrites. A credential
   binding with `location: "json"` explicitly parses and reserializes the top-level JSON object; it
   is not byte-faithful and must never be used with an upstream that signs or hashes the raw body.
   Never add upstream-specific modeling.
   One more rewrite, on 4xx/5xx only: every spelling of an injected credential the provider echoes
   back in an error body is masked (`***`) before the caller, an idempotent replay or the evidence
   row sees it. The error body is read whole (gzip/deflate decoded) up to 8 MiB; on a credentialed
   call one that cannot be scanned is replaced whole, never relayed as a prefix.
   A live-verified free catalog endpoint may declare an anonymous fallback; its empty binding list
   omits credential injection but does not strip or rewrite caller headers.
   Routed endpoints and overflow wrap the child's answer and say so; they never alter it. With
   `X-Treg-Route-Verify`, an email find makes one more linked child call (the check), with its own
   hold and charge, and reports it in `_treg.verification`; the find's answer is not altered. Responses needing settlement or ownership evidence are buffered by the application
   up to 8 MiB; exceeding that limit fails without charging, never returns a successful prefix.
   An endpoint declaring `spooled_response` (inline media) instead reads its metered 2xx to an
   unlinked temp file under a per-body cap and per-process budget, settles from only the paths its
   row reads (usage meters, `expect` success leaf), then relays the file byte for byte; it is never
   archived or replayed.
   Authorized free final fetches needing no body evidence stream in full.
   `/table/` runs the same call through the same road and returns it as rows and columns; it never
   changes what `/call/` returns (docs/context/architecture/table.md).
5. Balances change only through money's five entries: grant, topup, reserve, settle, release.
   There is deliberately no refund or adjustment entry; an ops correction is a grant.

**Changing any invariant in this file means editing this file in the same PR.** Routed endpoints
and overflow once shipped with every other doc updated while this file still said "no router";
agents then built against a constitution that was wrong.

## Where the truth lives

- **Design docs are fragments.** `docs/context/` holds one per subsystem, each naming its
  `src/treg/*` sources in frontmatter; `docs/context/README.md` is the generated index and
  `docs/context/foundation/charter.md` the start. Read the fragment before changing an area (the
  `tools-registry-context` skill in `.agents/skills/` loads it).
- **Before pushing:** `bash .agents/skills/tools-registry-context/scripts/drift.sh` maps changed
  sources to fragments. Update them and commit the docs **in the same commit as the code**.
- **Agent-facing files are the product's front door**, not documentation: `src/treg/web/llms.txt`
  and `src/treg/web/skill.md` (installed into every agent by `install.sh`). They, `README.md` and
  this file must agree on how treg works; a behavior change asks whether all four move.
- **Three files move together** or they drift: `src/treg/web/tutorial.js` (the only interactive
  source) and its hand-kept prose mirrors `src/treg/web/tutorial.md` and `docs/TUTORIAL.md`.
- `README.md` is the overview and quickstart, `USAGE.md` the CLI reference, `CONTRIBUTING.md` the
  dev setup, `SECURITY.md` required reading before touching the proxy, runners, auth or secrets.
- If available, `../treg-internal` holds private operational configuration and tools; read its
  `README.md` before changing production settings. Public development does not depend on it.

## Architecture

### Four layers

`routers/` -> `application/` -> `domain/` -> `infra/`. Imports point inward only.

| Layer | Owns | Never |
|---|---|---|
| `routers/` | HTTP and MCP translation in, response shape out | business rules, query orchestration, money |
| `application/` | use-case sequencing, transaction boundaries, compensation, cross-domain composition | empty wrappers around one-domain CRUD |
| `domain/` | rules explainable and testable alone: `identity`, `governance`, `connections`, `tools`, `catalog`, `capacity`, `money`, `asynctasks`, `feedback`, `hub` (a maker's tool made of tools: the manifest, the reference language, the graph) | routers, application, concrete SDKs |
| `infra/` | DB engine and sessions, crypto, upstream relay and SSRF, ratestore, the shared key-value store, email, Stripe | decisions |

- Domains do not import each other, with three sanctioned edges: `governance -> identity`,
  `tools -> connections`, `capacity -> catalog` (read-only). `identity` and `money` are leaves.
  import-linter enforces the layering (`[tool.importlinter]` in `pyproject.toml`, run by CI);
  `docs/context/architecture/import-boundaries.md` explains each contract.
- `bootstrap.py` alone knows concrete implementations. `api.py` is the legacy `all`-role
  entrypoint, not where logic goes. `audit.py` is best-effort and drops rows under load, so nothing
  that must persist goes through it; `analytics` is read-only.

### Writes

- **Session discipline.** The application use case opens the session and is the only place that
  commits; domain functions never commit or roll back. A commit mid-flow silently breaks
  compensation, and no import rule can catch it. Money's public `reserve`, `settle` and `release`
  commit by design; a few other domain commits remain. Do not add another; move one out when you
  touch it.
- **Table ownership.** One writer module per table; cross-domain reads are fine. Three recorded
  exceptions: only money writes `org.balance_micro`, the daily-spend counter (`spent_today_*`) and
  the auto-top-up fields; the call runtime may persist an OAuth token refresh into `secret`; audit
  writes `callrecord`, domains only read it (`application/evidence_retention.py` also updates it,
  blanking the two evidence columns past retention).
- **Feedback handling.** This repo owns `FeedbackHandling` and `FeedbackHandlingEvent` models and
  migrations; the private admin service is their only runtime writer. Original reports remain
  owned by the feedback domain. See `docs/context/architecture/feedback.md`.
- **The call runtime is self-contained.** `src/treg/application/call/` depends on no management
  code (routes, login, OAuth consent, Stripe top-up), reads only membership, deny rules,
  credentials, catalog prices and balances, and writes only what `tests/test_call_architecture.py`
  allowlists (the ledger entries, idempotency claims, OAuth refresh, audit and telemetry, first-call
  markers, tag budgets, capacity marks, overflow spend, the member's daily-cap slot, the per-team
  archive-question marks, and durable provider-resource ownership). Extend the
  test's allowlist in the same PR as any new write, and expect the reviewer to ask why.
- **The tool hub** (`application/hub/`, behind `TREG_HUB_ENABLED`) runs a maker's recipe as ordinary
  calls: every step goes through `execute_call` under its own hold, a catalog step as the caller and
  an own-tool step as the maker; the seller's price is one more hold, settled to the maker as an
  `earned` block in one transaction (`money.settle_to_in_transaction`, the only cross-team money
  movement). A steps recipe has one fixed price (`price_usd`); a script declares `max_price_usd` and
  prices itself with `ctx.charge(usd, note)` lines, so the cap is held and the sum of the lines is
  settled, the rest refunded (`docs/hub-pricing-decisions.md`). A script runs in a separate process
  with no network; `ctx.call` is its only road out and `uses` in the manifest names every host it
  may reach. A hub tool is out of catalog search until its maker asks and treg approves (`HubListing`); once
  approved it stays under review for good, unlisted or not, and its share
  page shows a run log of outcomes only unless the maker turns `public_log` off
  (`docs/hub-listing-decisions.md`). See `docs/context/architecture/hub.md`.
- **Signup credit.** Once per new verified user, enforced by a user-level atomic claim committed
  with the grant. Team deletion never restores eligibility; legacy registration is not email proof.
- **Money.** Everything is **integer micro-USD** - never floats, never cents. The Stripe SDK lives
  only in `infra/stripe.py`, orchestration in `application/billing.py`, and `reconcile.py` is
  read-only. See `docs/context/architecture/money.md`.
- **The archive serves every tier, keyed by whose question it is.** Own-credential answers are
  recorded (bounded read, never a prefix) under an org-scoped key, or a connection-scoped key on
  an `own_account` endpoint, and reach other teams only where the endpoint itself declares
  `cache.sharing: public`; a provider's storage licence never decides that. A hit on an own key
  is free; a metered hit settles through the same hold, at `archive_hit_repeat_price_percent`
  once the team has paid for that question. See `docs/context/architecture/archive.md`.

### Security guards that look redundant on purpose

`expose_dev_code` (dev OTP only on a local sqlite database, `config.py`), the call-time SSRF check
(`infra/upstream/ssrf.py`), the fail-loud missing-Fernet-key check in `verify_db`, and the
`treg run` allow-list and rlimits (`runner.py`). Read the fragment before touching any of them.

## Development

```bash
uv run --with pytest-xdist pytest -n auto -q   # daily local default (same shape as CI)
uv run --frozen python -m pytest -q            # serial: debugging one test, or order
uv run treg --help                             # the CLI from this checkout
uv run python -m treg                          # the server
uv run lint-imports                            # the import-linter contracts (CI runs this too)
scripts/dev-local.sh up                        # live dev stack on :18790 with its own sqlite DB
```

xdist is pulled via `--with`, not the lockfile — same as CI. Every test process gets its own
database (a sqlite file per pid; under `TREG_TEST_DB_URL`, a Postgres database per xdist worker),
so parallel runs and side-by-side runs never share one.

- **Dependencies change through `uv add` or `uv lock`, never by hand.** `pyproject.toml` pins
  `required-version` so an old uv refuses to run instead of rewriting `uv.lock`; CI uses `--locked`.
- **The package is split.** The base install is the light CLI; the FastAPI/DB stack is the
  `[server]` extra, the certificate authority is `[proxy]`. Never import a heavy dependency at the
  top of a CLI-path module; the "Lightweight CLI modules" import-linter contract lists them and
  fails the build.
- **The dashboard** lives in `frontend/` (Vue components, TypeScript entry/transport, Vite).
  Build with `bash scripts/build-dashboard.sh`; generated assets in `src/treg/web/dashboard/`
  ship with Python. Run `npm --prefix frontend test` and `npm --prefix frontend run test:e2e`.
  Existing Options API use cases live in `frontend/src/state/`; preserve their session and
  navigation behavior when narrowing component state. Never put dashboard logic back into HTML.
  Manage third-party browser libraries through pinned npm packages or version-pinned CDN URLs;
  do not commit copied library builds. Keep critical app runtimes available from the npm build.
- **Schema.** Alembic owns it (`src/treg/alembic/versions/`); every schema change is a revision.
  Startup only verifies the revision and refuses to boot when behind; migrations run only via
  `python -m treg upgrade`.

## Working agreement

- Keep the suite green; add tests for new behavior. Conventional Commits (`feat(scope): ...`,
  `fix: ...`, `docs: ...`); one logical change per commit; the PR says what changed and why and
  names the fragments it updated.
- `/mcp/` and `/mcp/v2/` differ on purpose. A change to either or to shared MCP code is reviewed
  against both; do not unify them in passing.
- **A catalog data PR is a few rows and a PR body.** Cache admission, comparison declarations,
  adapters and contracts are rows in `src/treg/catalog/`; each declaration carries a one-line
  reason and nothing more. The evidence (traffic, byte sizes, change observations, bodies read)
  goes in the PR body, never into a fragment or a comment. A fragment moves only when a mechanism
  changes; `drift.sh` naming one is a prompt to check it, not an obligation to write. No dated
  per-provider sections in `docs/context/architecture/catalog.md`: a provider's quirk lives on
  its row as a `note`. No per-endpoint tests: the round-trip test over every shipped adapter and
  the validator already judge the rows, and a test that restates a list of declarations is
  deleted, not extended. Code that a data PR needs is its own PR, merged first.

## When writing user-facing copy

One concept, one word. Settled deliberately - mixed vocabulary is how the old framing creeps back.

| Thing | Word |
|---|---|
| what an agent calls | **a tool** |
| the public half | **the catalog** |
| the team's half | **your own tools** (your keys and skills) |
| the server itself | **registry**, and only for that |

**Do not** call either half a *vault*, a *marketplace*, or *the registry*. Say what the agent can
now do, not what we store. Never use a count of endpoints or providers in this file; the catalog
changes weekly and every stale number is a lie.

**Do not document what is not built.** An agent that believes a feature exists fails in a way
nobody can debug. Provider choice is the easiest thing to overstate: treg compares providers, and
chooses only in the two disclosed cases of non-negotiable 4.