@kana-ai/edgar-provider-acp 0.1.2 → 0.1.4
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +3 -2
- package/package.json +3 -3
package/README.md
CHANGED
|
@@ -25,11 +25,12 @@ EDGAR is the ACP client. Each session is one child that runs the command.
|
|
|
25
25
|
- `session/request_permission` is answered by the runtime: allowed once for an engine tool, whose call the engine then decides, and rejected for anything else. The model shapes a title, such as a shell command, so only a title that is exactly an engine tool's name passes (bare, `mcp__edgar__<name>`, or `<name> (<server>)`), and never for the kinds `execute`, `edit`, `delete`, `move` and `fetch`. `fs/*`, `terminal/*` and every other request get "method not found".
|
|
26
26
|
- `resume: {handle, fork: false}` is `session/resume` when the agent supports it, else `session/load`. An agent that supports neither, a fork, and a session the agent does not have (`-32002`) give an `error` event with `session_gone`.
|
|
27
27
|
- `session.send` is `session/prompt`. `agent_message_chunk` updates make `message` events. The turn ends with the prompt's answer.
|
|
28
|
+
- A later `session/prompt` goes to the same session, so a role with `steer` gets a comment from the team as a new turn with its context kept. The engine sends it only after the turn's `result`, as a send while a turn runs is `busy`.
|
|
28
29
|
- `session.cancel` sends `session/cancel`, waits at most 5 s for the answer, then kills the child's process group. A child that sends nothing for `--idle-seconds` while a request waits is killed, and the turn ends `silence`. The clock stops while the engine runs a tool call of the agent, because the agent then waits for the answer and has nothing to send. The engine's own deadline bounds that call.
|
|
29
30
|
|
|
30
31
|
## Credentials
|
|
31
32
|
|
|
32
|
-
The child's environment is `PATH`, `HOME` and the API key under `env_key`. The engine's fallback API key, `ACP_FALLBACK_API_KEY`,
|
|
33
|
+
The child's environment is `PATH`, `HOME` and the API key under `env_key`. The engine's fallback API key, `ACP_FALLBACK_API_KEY`, and a person's own `ACP_API_KEY`, from their key file (see the credentials guide), come as an `api_key` credential too and reach the agent under the same `env_key`. The `cli` credential adds no key, and the agent uses its own sign-in, if it finds one. An `oauth` credential ends `unauthenticated` before any child starts, because `authenticate` names a method and carries no token. No credential is in an argument or a log line. A prompt the agent refuses with `-32000` ends `unauthenticated`, and its message names `ACP_API_KEY` in `[secrets]` as the fix, and on the `cli` credential the agent's own login too.
|
|
33
34
|
|
|
34
35
|
## What the runtime may reach
|
|
35
36
|
|
|
@@ -38,7 +39,7 @@ The runtime runs under Node's permission model like the other providers, with `-
|
|
|
38
39
|
## Outcomes and usage
|
|
39
40
|
|
|
40
41
|
- One mapping table in `src/outcome/failures.ts` turns stop reasons and JSON-RPC error codes into outcome kinds and keeps the raw value in `cause` (`stop:<reason>` or `rpc:<code>`). A value the table does not know is `provider_unavailable`.
|
|
41
|
-
- `usage` comes from the prompt answer's `usage`, which the protocol marks unstable: input, output, cache-read and cache-write tokens when the agent reports them. The runtime does not read `limits.max_tokens`: the engine compares each `usage` event with the limit, keeps a finished result of the turn that passed it, and ends anything else `token_budget`. An agent that reports no usage has no limit.
|
|
42
|
+
- `usage` comes from the prompt answer's `usage`, which the protocol marks unstable: input, output, cache-read and cache-write tokens when the agent reports them. The runtime does not read `limits.max_tokens`: the engine compares the total of each `usage` event, new input, cache reads, cache writes and output, with the limit, keeps a finished result of the turn that passed it, and ends anything else `token_budget`. An agent that reports no usage has no limit.
|
|
42
43
|
- `limits.max_turns`, `limits.context_tokens` and a role's `effort` are not applied: the protocol has none of them.
|
|
43
44
|
- The protocol has no rate or usage limit notification, so the runtime sends no `limit` event. A rate limit reaches the engine only as the failure kind of the stop reason or error code.
|
|
44
45
|
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@kana-ai/edgar-provider-acp",
|
|
3
|
-
"version": "0.1.
|
|
3
|
+
"version": "0.1.4",
|
|
4
4
|
"description": "The process that runs Agent Client Protocol sessions for the EDGAR engine with the agent a project names, over JSON-RPC on stdio.",
|
|
5
5
|
"license": "Apache-2.0",
|
|
6
6
|
"repository": {
|
|
@@ -19,8 +19,8 @@
|
|
|
19
19
|
"#*": "./dist/src/*.js"
|
|
20
20
|
},
|
|
21
21
|
"dependencies": {
|
|
22
|
-
"@kana-ai/edgar-contracts": "0.1.
|
|
23
|
-
"@kana-ai/edgar-sdk": "0.1.
|
|
22
|
+
"@kana-ai/edgar-contracts": "0.1.4",
|
|
23
|
+
"@kana-ai/edgar-sdk": "0.1.4",
|
|
24
24
|
"zod": "4.6.7"
|
|
25
25
|
},
|
|
26
26
|
"devDependencies": {
|