docs(acp): RFC for the terminal-card rendering (implemented design)

Records the verified design before implementing: keep dsh-bash agent-side
execution and render Zed's terminal tool-call card via the `_meta` convention
(terminal_info/terminal_output/terminal_exit), capability-gated on
clientCapabilities._meta.terminal_output, with the ```console text block as the
no-capability fallback. Rejects the spec's client-side terminal/create path (it
would bypass dsh-bash's sandbox/env-scrub/ownership/cwd). Studied
claude-agent-acp, codex-acp, and Zed's renderer to ground the wire contract.
Live streaming and command classification are noted as separate follow-ups.

docs/rfc/README.md

@@ -31,7 +31,6 @@ Do NOT write one for a mechanical or local choice (a variable name, a one-file r
 | [Multiplex concurrent ACP sessions over one connection](proposed/2026-06-14-acp-multi-session.md) | 2026-06-14 |
 | [Optional Code Mode — model writes TypeScript against an SDK of all tools](proposed/2026-06-15-optional-code-mode.md) | 2026-06-15 |
 | [Runtime schemas for the event vocabulary (Zod vs the merge-extensible-map pattern)](proposed/2026-06-16-typed-event-schemas.md) | 2026-06-16 |
-| [Rich ACP bash rendering — the terminal sub-protocol and command classification](proposed/2026-06-18-acp-terminal-and-tool-rendering.md) | 2026-06-18 |
 
 ## Implemented
 
@@ -56,6 +55,7 @@ Do NOT write one for a mechanical or local choice (a variable name, a one-file r
 | [Session persistence as an abstract service over `SessionEvent`](implemented/2026-06-14-session-persistence.md) | 2026-06-14 |
 | [Every session event is enclosed in a turn](implemented/2026-06-15-turn-enclosure-invariant.md) | 2026-06-15 |
 | [pnpm as the package manager instead of Yarn 4](implemented/2026-06-16-pnpm-over-yarn.md) | 2026-06-16 |
+| [Rich ACP bash rendering — the terminal card (`_meta`) and command classification](implemented/2026-06-18-acp-terminal-and-tool-rendering.md) | 2026-06-18 |
 
 ## Rejected
 

docs/rfc/implemented/2026-06-18-acp-terminal-and-tool-rendering.md

@@ -0,0 +1,40 @@
+# RFC: Rich ACP bash rendering — the terminal card via the `_meta` convention
+
+Status: implemented
+
+## Problem
+
+The ACP bridge lets each tool own its call rendering via `presentCall`/`presentResult` (see [tool-call UI presentation](../proposed/2026-06-14-acp-agent-client-protocol.md) and `packages/tools`). For `bash` we surface the model's `description` plus the command as the `tool_call` title, `kind: 'execute'`, and the completed output wrapped in a fenced ` ```console ` text block.
+
+That is a correct, capability-free baseline, but not how the reference editors render a *terminal* tool at its best. An editor like Zed has a dedicated terminal tool-call card — a header showing the working directory, the command as the label, and the command output rendered as a terminal — but it only builds that card when the `tool_call` carries terminal metadata (below). With a plain text block the output appears as static markdown and there is no cwd header. (Zed also HIDES `rawInput` for `kind: 'execute'`, which is why our command rides inside the title.)
+
+## Key finding: agent-executed terminals use a `_meta` convention, NOT `terminal/create`
+
+The ACP spec has a *client-side* terminal sub-protocol — the agent calls the client's `terminal/create` with `{ command, args, cwd, env }` and the **editor** executes the process, then the agent reads `terminal/output` / `wait_for_exit`. That model is wrong for us: our harness executes bash itself through `dsh-bash` (sandboxed env-scrub, background-task ownership, per-session cwd). Routing execution to the editor would bypass all of that and fork execution into two backends.
+
+Studying the two reference agents (2026-06-18) shows neither uses `terminal/create` for their own shell tool — **both keep agent-side execution and emit a `_meta` convention** that Zed special-cases:
+
+- **`claude-agent-acp`** (`tools.ts`, `acp-agent.ts`): gated on `clientCapabilities._meta.terminal_output`. The `tool_call` carries `content: [{ type: 'terminal', terminalId }]` and `_meta.terminal_info.{ terminal_id, cwd }`; output/exit arrive on the `tool_call_update`'s `_meta.terminal_output.{ terminal_id, data }` and `_meta.terminal_exit.{ terminal_id, exit_code, signal }`.
+- **`codex-acp`** (`CodexToolCallMapper.ts`, `TerminalOutputMode.ts`): same `terminal_info` on the call; output via `_meta.terminal_output` (full) or `_meta.terminal_output_delta` (incremental), selected from the same `_meta.terminal_output` capability.
+
+Zed's side (`crates/agent_servers/src/acp.rs`, verified): on a `ToolCall` whose `_meta.terminal_info.terminal_id` is set, it registers a **display-only** terminal (header = `terminal_info.cwd`, label = `tool_call.title`); on a `ToolCallUpdate`, `_meta.terminal_output.data` writes to that terminal and `_meta.terminal_exit.{exit_code,signal}` sets the status. It advertises the capability as `clientCapabilities._meta.terminal_output = true`. This is an off-spec `_meta` extension, but it is the de-facto contract for the Zed integration and the only way to get the terminal card while keeping execution agent-side.
+
+## Decision
+
+Keep `dsh-bash` agent-side execution; render the terminal card via the `_meta` convention, capability-gated, with the ` ```console ` text block as the guaranteed fallback.
+
+1. **Capability.** `initialize` reads `clientCapabilities._meta.terminal_output` and the bridge remembers it per connection.
+2. **Neutral presentation vocabulary.** `dsh-tools` gains a terminal-shaped presentation a tool can return — provider-neutral (`cwd`, the output `data`, an `exitCode`/`signal`), NO ACP types. `dsh-tool-bash` returns it for `bash` (cwd from the resolved workdir; output + exit parsed from the run result).
+3. **Bridge mapping.** When the client advertised the capability, the bridge maps that presentation to: on `tool_call`, `content:[{type:'terminal', terminalId}]` + `_meta.terminal_info.{terminal_id,cwd}`; on `tool_call_update`, `_meta.terminal_output.{terminal_id,data}` + `_meta.terminal_exit.{terminal_id,exit_code,signal}`. `terminalId` is derived from the harness `callId` (stable, unique per call). When the capability is absent, the bridge uses the existing ` ```console ` text content — unchanged.
+4. **No new execution path, no live streaming.** Output is attached at completion (from the agent's own `tool/result`), not streamed token-by-token. Disposal is unaffected: nothing new to tear down, since the bridge never creates a client-side terminal.
+
+## Risks / trade-offs
+
+- **Off-spec `_meta`.** The terminal card rides on a Zed-specific `_meta` extension, not the ACP terminal sub-protocol. A client that doesn't recognize it still gets the text fallback (the capability gate ensures we only emit it when the client opted in via `_meta.terminal_output`), so a non-Zed client is never worse off. If ACP later standardizes agent-executed terminals, migrate to that and drop the `_meta`.
+- **Capability honesty.** Emit terminal metadata ONLY when the client advertised `_meta.terminal_output`; the text fallback is the contract for everyone else and must never regress. Covered by a no-capability test asserting the ` ```console ` path.
+- **terminalId collisions.** Deriving it from the per-call `callId` keeps it unique within a session and stable across the call/result pair; never reuse one across calls.
+- **Provider-neutral vocabulary creep.** The terminal presentation widens the `dsh-tools` surface; keep it neutral (no ACP types leak into `dsh-tools`) and only as rich as a second UI consumer would also want.
+
+## Out of scope / non-goals
+
+The text-block baseline stays the no-capability default. Client-side `terminal/create` execution is explicitly rejected (it bypasses `dsh-bash`). Two follow-ups are deliberately NOT built here and would each warrant their own RFC when someone takes them on: **live incremental streaming** (`_meta.terminal_output_delta` as chunks arrive, which needs an incremental-output seam on `dsh-bash`), and **command classification** (parsing a `cat`/`sed` as a `read` card with a file location, a `grep` as a `search`, etc., falling back to the terminal card — display-only, must never change what executes).

docs/rfc/proposed/2026-06-18-acp-terminal-and-tool-rendering.md

@@ -1,49 +0,0 @@
-# RFC: Rich ACP bash rendering — the terminal sub-protocol and command classification
-
-Status: proposed
-
-## Problem
-
-The ACP bridge now lets each tool own its call rendering via `presentCall`/`presentResult` (see [tool-call UI presentation](2026-06-14-acp-agent-client-protocol.md) and `packages/tools`). For `bash` we surface the model's `description` plus the command as the `tool_call` title, `kind: 'execute'`, and the completed output wrapped in a fenced ` ```console ` text block.
-
-That is a correct, capability-free MVP, but it is not how the reference editors render a *terminal* tool at its best. Two gaps:
-
-1. **No live terminal card.** An editor like Zed has a dedicated terminal tool-call card — a header showing the working directory, the command, a copy button, and **streaming** output with an exit-status pill — but it only uses that card when the `tool_call`'s `content` is an ACP `terminal` block (`{ type: 'terminal', terminalId }`), not a text block. With a text block the command output appears only as static markdown *after the turn completes*; there is no live stream and no cwd header. (Zed also HIDES `rawInput` for `kind: 'execute'`, which is why our command currently has to ride inside the title.)
-
-2. **No command classification.** A bash invocation is opaque — `bash -lc "sed -n 1,40p foo.ts"` is really a file read, `rg foo` is a search. The reference adapters classify common commands and present them with a *semantic* kind/title/locations (a `read` card titled "Read file 'foo.ts'" with a follow-along file location, a `search` card), falling back to a terminal card only for an unrecognized command. This is what makes one bash call render with a search icon and "List …" while the next renders as a raw terminal.
-
-## What the reference adapters do (studied 2026-06-18)
-
-- **`codex-acp`** (`CodexToolCallMapper.ts`): classifies each command into `commandActions`. A recognized action maps to a semantic update — `read` → `{kind:'read', title:"Read file '…'", locations:[{path}]}`, `search` → `{kind:'search', title:"Search for '…' in …"}`, `listFiles` → `{kind:'read', title:"List files in '…'"}`. An `unknown` action becomes a terminal card: `{kind:'execute', title: stripShellPrefix(command), content:[{type:'terminal', terminalId}], _meta:{terminal_info:{cwd, terminal_id}}}`. The `_meta.terminal_info.cwd` is what renders the working directory as the card header.
-- **`claude-agent-acp`** (`tools.ts`): gates on `clientCapabilities._meta.terminal_output`. WITH it: a terminal content block plus `_meta.terminal_{info,output,exit}` (output + exit code). WITHOUT it: the same fenced ` ```console ` text-block fallback this bridge ships today. Title is the command; the model's `description` (when present) is shown as content.
-- **Zed** (`crates/agent_ui/.../thread_view.rs`, `crates/acp_thread/.../acp_thread.rs`): `render_terminal_tool_call` reads the terminal's `working_dir` as the header and `tool_call.label` (the title) as the command; a non-terminal text `content` block renders via `render_markdown_output`. `should_show_raw_input = !is_terminal_tool` confirms `rawInput` is suppressed for execute-kind cards.
-
-The full terminal experience is an ACP **sub-protocol**, not just a content shape: the client advertises a terminal capability, and the agent drives `terminal/create` → streams via `terminal/output` → `terminal/release`, attaching the `terminalId` to the `tool_call` content. That is a cross-seam feature (bridge ⇄ `dsh-bash` executor ⇄ client), which is why it is deferred to this RFC rather than folded into the presentation-seam PR.
-
-## Proposal
-
-Two independent, separately shippable pieces. Both build on the existing tool-owned presentation seam — neither reintroduces tool-name special-casing in the bridge.
-
-### A. Terminal content type + cwd metadata (capability-gated)
-
-1. In `initialize`, read the client's terminal capability (`clientCapabilities.terminal` / the `_meta.terminal_output` convention the references use) and remember it per connection.
-2. Extend the `dsh-tools` presentation vocabulary so a tool can ask for a terminal rendering — e.g. a `ToolResultPresentation`/`ToolCallPresentation` variant carrying `{ kind: 'terminal', cwd, terminalId? }` (provider-neutral; the bridge maps it to the ACP `terminal` content block + `_meta.terminal_info`). `dsh-tool-bash` returns it for `bash` when a cwd is known.
-3. When the client supports it, the bridge maps that to `content:[{type:'terminal', terminalId}]` + `_meta.terminal_info.{cwd,terminal_id}`; otherwise it keeps the current ` ```console ` text fallback. The fenced-text path stays the guaranteed baseline.
-4. *(Stretch)* drive live streaming through the real `terminal/*` methods so output appears as it is produced, with an exit-status pill — this needs a streaming seam on `dsh-bash` (the executor already has the process; it would push incremental output to the bridge). Scope this as a follow-up sub-step; steps 1–3 already give the cwd-header card with output attached at completion.
-
-### B. Command classification (capability-free)
-
-A small, pure classifier (in `dsh-tool-bash`, since it owns the bash schema) maps a command string to an optional semantic presentation: detect common read/search/list shapes (`cat`/`sed -n`/`head`/`tail` → `read` + a `path` location; `grep`/`rg` → `search`; `ls` → list) and return the richer `ToolCallPresentation` (`kind`, a human title, `locations`). Anything unrecognized falls through to the current execute/terminal presentation. This needs a `locations?: ToolCallLocation[]`-style field on `ToolCallPresentation` (neutral `{ path, line? }`), which the bridge maps to ACP `tool_call.locations` to drive editor "follow-along".
-
-Classification is best-effort and explicitly fallible: a misparse must degrade to the plain terminal card, never mislabel destructively (e.g. never title a `rm` as a "read"). Keep the matcher conservative and unit-test each recognized shape plus the fallthrough.
-
-## Risks / trade-offs
-
-- **Terminal sub-protocol is cross-seam and stateful.** Live streaming couples the bridge, the `dsh-bash` executor, and the client's terminal lifecycle; getting disposal/cancel right (release the terminal on turn end, abort, and disconnect) is the hard part — it must honor the same quiescence rules as the rest of the bridge. Steps A1–A3 (static cwd header + output at completion) are low-risk; A4 (live streaming) is where the lifecycle complexity lives.
-- **Capability detection must stay honest.** Advertise/emit terminal content only when the client opted in; the text fallback is the contract for everyone else, so it must never regress.
-- **Classification can mislead.** A wrong guess is worse than no guess. Bias to the terminal fallback; treat the classifier as additive polish, not a correctness path. (Security note: classification is display-only — it must never change what actually executes.)
-- **Provider-neutral vocabulary creep.** Adding `terminal`/`locations` to `ToolCallPresentation` widens the `dsh-tools` surface. Keep the additions neutral (no ACP types leak into `dsh-tools`) and only as rich as a second consumer would also want.
-
-## Out of scope / non-goals
-
-The MVP shipped in the tool-call-UI PR (description+command title, `kind:'execute'`, ` ```console ` output fallback) stays the baseline and the no-capability default. This RFC is purely additive polish on top of it.

packages/acp/README.md

@@ -46,7 +46,7 @@ How a tool call renders in the editor is owned by the TOOL, not the bridge — t
 
 The `tool/result` session event carries only `{ callId, content, isError }` — not the tool name or args — so to call a tool's `presentResult` the bridge keeps a small per-session map from `callId` to the in-flight call's `(name, args)`, populated on `tool/call` and removed as each result is presented (it holds only currently-in-flight calls, never finished ones). This is bridge-local state — NOT a change to the event schema or a core service. The map lives on the `SessionRecord`, so two concurrent sessions never cross their in-flight tool state; a `session/load` replay uses a throwaway presenter that pairs each `tool/call` with its `tool/result` as the log replays in order, so replayed tool cards render identically to live ones.
 
-A richer rendering — the ACP **terminal** content type (a live cwd-header terminal card with streaming output) and command classification (a `cat` shown as a `read`, a `grep` as a `search`) — is a capability-gated follow-up; the ` ```console ` text block here is the guaranteed baseline for clients without the terminal capability. See [the terminal-rendering RFC](../../docs/rfc/proposed/2026-06-18-acp-terminal-and-tool-rendering.md).
+A richer rendering — the ACP **terminal** content type (a live cwd-header terminal card with streaming output) and command classification (a `cat` shown as a `read`, a `grep` as a `search`) — is a capability-gated follow-up; the ` ```console ` text block here is the guaranteed baseline for clients without the terminal capability. See [the terminal-rendering RFC](../../docs/rfc/implemented/2026-06-18-acp-terminal-and-tool-rendering.md).
 
 ## Settle-exactly-once