Merge slice/020-tool-protocol-registry

Slice 020: tool protocol and registry. PROOF.md at slices/020-tool-protocol-registry/PROOF.md; gate and postgres jobs green on the closing tree.

Signed-off-by: Ayla Croft <aylacroft@proton.me>

Author: HackTuah

ROADMAP.md

@@ -37,7 +37,7 @@ standards register names the rows that ask for them.
 | 011 | LLM provider layer (req_llm behind `Trinity.LLM` behaviour) | 1 Core loop | M | 010 | approved |
 | 012 | Session process + agent loop (gen_statem, DynamicSupervisor, rehydration) | 1 Core loop | L | 010, 011 | approved |
 | 013 | LiveView chat UI with streaming | 1 Core loop | M | 012 | approved |
-| 020 | Tool protocol + registry | 2 Tools | M | 012 | planned |
+| 020 | Tool protocol + registry | 2 Tools | M | 012 | done |
 | 021 | Permission gate + approval UI (M2 fingerprint-bound, M7) | 2 Tools | M | 020, 013 | planned |
 | 022 | Core tools: filesystem, web fetch/search, shell (MuonTrap) | 2 Tools | L | 021 | planned |
 | 023 | Context compaction + session lineage | 2 Tools | M | 012 | planned |

VERSIONS.md

@@ -135,6 +135,7 @@ never pin a version hex marks as retired or vulnerable.
 | `telegex` | **not pinned** | 🔍 not a single package | Telegram. ⚠️ The latest release on hex is a release candidate, roughly two years old, and this file's own rule forbids pinning an `-rc`. Alternative: ex_gram. Slice 071 decides with the measurement. |
 | `phoenix_streamdown` | not used (measured at Slice 013, 2026-09-20) | 🔍 not a single package | Streaming markdown renderer for LiveView. Measured at Slice 013: 1.0.0-beta.4 (2026-05-03) was still the latest release with no stable behind it, and this file's own rule forbids a pre-release. It is 801 lines over `mdex`, whose own `streaming: true` option completes fragments; the rest is a rendering rule Slice 013 keeps anyway. The row stays so the decision is visible where a reader would look for the package. |
 | `mdex` | ~> 0.13 | ✅ in `mix.lock` | Markdown renderer for the chat (`TrinityWeb.Markdown`), streaming fragments completed by its `streaming: true` option, raw HTML omitted (`unsafe: false`) and the default sanitizer on top. Added at Slice 013 after the measurement in its NOTES.md: `earmark` 1.4.49 is retired on hex with an open XSS advisory (EEF-CVE-2026-48591), which the gate refuses. ⚠️ A Rust NIF (`mdex_native`): the first in the bundle. Measured at Slice 013 (NOTES finding 13): neither precompiled artifact loads in Burrito's musl ERTS on Linux, so the linux package builds it from source for musl with Zig as the linker (`rustler` below, `scripts/zig-cc-musl`, the three settings in config/config.exs); macOS and Windows load the precompiled artifact. The `--smoke` path prints whether it rendered, and the `package` workflow reads that line on every target. |
+| `jsv` | ~> 0.23 | ✅ in `mix.lock` | JSON Schema (2020-12) validation of tool arguments in `Trinity.Tools.Schema`, with `cast: false` so a malformed call is refused and never repaired (docs/07). Was transitive through req_llm; direct since Slice 020 because a module of ours calls it (ADR-0009: Trinity's own tool behaviour, jsv for its schemas). |
 | `rustler` | ~> 0.38 | ✅ in `mix.lock` | Build time only (`runtime: false`): what `rustler_precompiled` needs to compile `mdex_native` from source when `MDEX_NATIVE_BUILD=1`, which the linux package sets (owner decision 2026-09-20, Slice 013 NOTES finding 14). Nothing in the tree calls it. |
 
 ### Dev and quality

config/test.exs

@@ -7,6 +7,18 @@ import Config
 # The MIX_TEST_PARTITION environment variable can be used
 # to provide built-in test partitioning in CI environment.
 # Run `mix help test` for more information.
+# Slice 020: the test tools, and the toolset the core ones belong to. AC1's one config line
+# is the `modules:` entry; a tool module in test/support plus this line is the whole change.
+config :trinity, :tools,
+  modules: [
+    Trinity.TestTools.Echo,
+    Trinity.TestTools.Sleep,
+    Trinity.TestTools.Crash,
+    Trinity.TestTools.Big
+  ],
+  toolsets: %{core: ["echo", "sleep", "crash", "big"]},
+  timeout_ms: 2_000
+
 # Slice 012: sessions hibernate and stop quickly in tests so AC8 is observable in seconds.
 config :trinity, :sessions, idle_hibernate_ms: 200, idle_stop_ms: 60_000
 

coverage.tsv

@@ -5,3 +5,4 @@ slice_id	percent	sha	date
 011	51.57	ec5334a	2026-09-20
 012	60.82	019d9cf	2026-09-20
 013	64.41	080c543	2026-09-20
+020	67.18	8a5b7ae	2026-09-20

docs/01-architecture.md

@@ -23,7 +23,9 @@ Trinity.Application
 ├── Trinity.Sessions.Supervisor (DynamicSupervisor) # one Trinity.Sessions.Session per conversation. Slice 012, as built
 │     └── Trinity.Sessions.Session (gen_statem)   # states: idle → thinking → tool_wait → approval_wait → compacting → error
 │           └── Task.Supervisor (started by the Session, linked, unnamed) # the model call and the tool calls of one turn
-├── Trinity.Tools.Supervisor                      # tool runtime (ports, browsers). Slice 020/022
+├── Trinity.Tools.Supervisor                      # Slice 020, as built: Trinity.Tools.TaskSupervisor (every tool
+│     │                                         # call of a turn runs under it) and Trinity.Tools.Registry
+│     │                                         # (GenServer over ETS). 022 adds the stateful runtimes beside them
 ├── Trinity.Permissions.Gate                      # approval requests + allowlist cache. Slice 021
 ├── Trinity.Receipts.Supervisor                   # Slice 024
 │     └── Trinity.Receipts.ChainWriter (one per chain_scope, :unique in Trinity.Registry; ADR-0013)
@@ -113,6 +115,17 @@ UI/Gateway ──user_message──▶ Session(gen_statem)
 Every state transition is persisted before it is broadcast. A crash between persist and broadcast is safe
 (rehydrate re-broadcasts the last state).
 
+**Tool calls (Slice 020):** at `{:done, :tool_calls}` the Session hands the turn's calls to
+`Trinity.Sessions.ToolRunner.run_all/2`, the seam whose implementation in force is `Trinity.Tools.Runner`
+(config, so a test can put the stub back). The runner runs every call at once under
+`Trinity.Tools.TaskSupervisor`, each with its tool's timeout: lookup, `jsv` validation of the arguments
+(refused, never repaired), `Trinity.Permissions.decide/3` once, `execute/2`, the result cap. A crash, a timeout
+and an unknown name are error results the model reads; the Session writes one `tool` row per answer with the
+tool's definition digest. Each turn's request carries the declared surface (`Trinity.Tools.to_llm_tools/0`) and
+the assistant row records it (`provider_meta.tool_surface`); `Trinity.Tools.surface_diff/1` over a history names
+the calls a turn made outside it. Sessions depends on Tools; Tools depends on Permissions and never on Sessions
+(the runner implements the seam's functions without naming the behaviour, which would close a cycle).
+
 **Effect path (Slice 024):** `Session → Permissions.decide → Effects.execute → Authority → tool.execute/2 (local) or a proposal (external adapter) → Receipts.append`. `Effects` is the only caller of `execute/2` for effectful tools; a census test enforces it. Reads emit query receipts.
 
 **The page (Slice 013):** `TrinityWeb.SessionLive.Show` subscribes to `session:<id>` on mount, calls

docs/03-conventions.md

@@ -27,6 +27,21 @@
   previous slice fails until a NOTES.md justification names the reason. The rule as originally written stored no
   baseline, so nothing could check it, which is the pattern CLAUDE.md §8 forbids.
 
+## Tools (slice 020)
+
+- A tool is a module implementing `Trinity.Tools.Tool` (`name/0`, `description/0`, `schema/0` as a JSON Schema
+  map with string keys, `risk/0`, `effect/0`, `execute/2`, optional `timeout/0` and `format_result/1`) plus one
+  line in `config :trinity, :tools` (`modules:`; `toolsets:` groups names). No core module changes; the
+  registry test asserts `git grep TestTools lib/` finds nothing.
+- Tool modules are stateless. A runtime with state (a shell, a browser) is a child of `Trinity.Tools.Supervisor`
+  the tool looks up.
+- `effect/0` is `:none` (a read), `:artifact` (a local write) or `:catalog` (an external effect). A `:catalog`
+  tool is listed in `Trinity.Effects.Catalog`'s module attribute or it does not start; nothing registered at
+  runtime may claim it. A runtime tool's name is namespaced (`mcp:<server>:<tool>`, `skill:<name>`); core
+  names are reserved; the tier map in `Trinity.Permissions` is code and lists core names only.
+- A tool returns `{:ok, %Trinity.Tools.Result{}}` or `{:error, reason}`; the runner caps the content at
+  `result_cap_bytes` (64 KB) with a marker. Arguments arrive validated; a tool never repairs them either.
+
 ## UI (decided at slice 013)
 
 - Tokens live in `assets/css/app.css` and nothing else names a colour, a radius or a font: two daisyUI themes

docs/05-data-model.md

@@ -36,7 +36,7 @@ with adapter-specific `execute/1` guarded by `repo().__adapter__()`.
 | role | string | "system" \| "user" \| "assistant" \| "tool" |
 | content | text | plain text or JSON for tool payloads |
 | parts | map | structured content parts (text, tool_call, tool_result, image ref). Every part carries `origin`, `source_ref`, `digest` and `taint ∈ {trusted, untrusted, blocked}`; summaries inherit the maximum taint of their inputs (Slice 022) |
-| tool_call_id | string, nullable | |
+| tool_call_id | string, nullable | the assistant row's call id a `tool` row answers (Slice 012); its `parts` carry `tool`, `ok`, `tool_result` (`content`, `truncated`, `meta`, `artifacts`, or `error`) and `tool_definition_digest`, and the assistant row's `provider_meta.tool_surface` maps every declared tool name to its digest (Slice 020) |
 | usage | map, nullable | prompt/completion tokens, cost |
 | provider_meta | map | model, finish reason, latency |
 Append-only. Editing is a new message with `parts.supersedes`. One edit is allowed and named (slice 012): an

lib/trinity.ex

@@ -7,11 +7,25 @@ defmodule Trinity do
   # application locks the data directory before the Repo starts), Repo and UUID (the contexts
   # use them), and the Sessions sub-boundary: a context TrinityWeb may call (docs/01). Slice
   # 013 exports the schemas the chat renders, `Sessions.Message` and `Sessions.SessionRow`,
-  # which the Sessions boundary exports itself; Store stays inside.
+  # which the Sessions boundary exports itself; Store stays inside. Slice 020 exports the
+  # Tools and Permissions sub-boundaries and `Effects.Catalog`, a plain module of this
+  # boundary the tool registry reads (Effects becomes its own boundary at 024).
   use Boundary,
     deps: [],
     exports:
-      [Paths, Repo, UUID, Config, Sessions, Sessions.Message, Sessions.SessionRow, LLM] ++
+      [
+        Paths,
+        Repo,
+        UUID,
+        Config,
+        Sessions,
+        Sessions.Message,
+        Sessions.SessionRow,
+        LLM,
+        Tools,
+        Permissions,
+        Effects.Catalog
+      ] ++
         if(Mix.env() == :test, do: [DataCase, NetworkGuard, Factory], else: [])
 
   @moduledoc """

lib/trinity/application.ex

@@ -37,6 +37,9 @@ defmodule Trinity.Application do
           {Task.Supervisor, name: Trinity.LLM.TaskSupervisor},
           # Slice 012: one session process per conversation, found by id.
           {Registry, keys: :unique, name: Trinity.Registry},
+          # Slice 020: the tool registry and the task supervisor tool calls run under, before
+          # the sessions that call them.
+          Trinity.Tools.Supervisor,
           Trinity.Sessions.Supervisor,
           # Start to serve requests, typically the last entry
           TrinityWeb.Endpoint

lib/trinity/effects/catalog.ex

@@ -0,0 +1,24 @@
+# SPDX-FileCopyrightText: Sudo Apt Holdings LLC
+# SPDX-License-Identifier: Apache-2.0
+defmodule Trinity.Effects.Catalog do
+  @moduledoc """
+  The effect catalog, resolved at compile time (docs/07, M4). Slice 020 opens it empty.
+
+  Every tool whose `effect/0` is `:catalog` (an external effect: send, spend, a provider
+  mutation) is listed here by name with its risk tier, in a module attribute and nowhere
+  else. `Trinity.Tools.Registry` admits a core `:catalog` tool only if its name is in this
+  list and refuses a runtime registration claiming `:catalog` outright; the census test
+  (slice 020 AC8) walks the tree and asserts no other path admits one. Slice 024 makes the
+  membrane read it.
+  """
+
+  @catalog []
+
+  @doc "Every catalog tool as `{name, tier}`."
+  @spec all() :: [{String.t(), atom()}]
+  def all, do: @catalog
+
+  @doc "The names alone."
+  @spec names() :: [String.t()]
+  def names, do: Enum.map(@catalog, &elem(&1, 0))
+end