git:20260923.55727e7 to git:20260923.c296245

# 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

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. 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.

Authorized free final fetches needing no body evidence stream in full.

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` | 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.

- **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.

- **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. The Postgres CI job

(`test-postgres`) must stay serial: every worker would share one database while

`reset_db()` drops tables.

- **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.

- **Frontend rollout.** `frontend/README.md` documents account assignment and rollback.

`src/treg/web/dashboard-legacy/` is **deprecated**, retained only for temporary rollout and

rollback. Never hand-edit it or mirror new features/fixes into it; `frontend/` is the only

maintained Dashboard source. Follow the retirement checklist in `frontend/README.md` to remove

it after rollout, including anonymous entries that still use legacy at 100%.

- **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.