@particle-academy/prism-acp 0.5.0 → 0.5.2

package/README.md

@@ -92,6 +92,18 @@ in the direction that spends money silently. Names are compared
 case-insensitively, because Windows environment names are case-insensitive and
 `anthropic_api_key` reaches a child exactly as the uppercase spelling does.
 
+The allow-list withholds every variable not named in it, credentials and
+non-credentials alike. If the child needs an orchestrator session or terminal
+id, a callback URL, or a feature flag, name it in `allowEnv` or it will not
+arrive. This failure is silent: the child can start and appear healthy while
+being unable to identify itself, rather than reporting a missing variable.
+An allowed variable whose parent value is `''` is forwarded as an empty string.
+That can be worse than omitting it: a referenced-but-unset value in an MCP
+config can stop every server in that file. When there is no value to give,
+omit the variable instead of passing its name with a blank value. `childEnv`
+does not filter blanks for you, because an intentionally empty value can be
+meaningful for some variable names and the package will not guess which.
+
 Nothing here mutates `process.env`. A workspace may hold an API key on purpose —
 other consumers beside this one legitimately bill per token — so the child's
 environment is constructed and the ambient one is left alone.
@@ -155,6 +167,18 @@ drivers because their measured resume behavior differs. Codex thread ids are
 the provider's own captured ids and are the same ids accepted by its resume
 method; ACP-minted ids are refused.
 
+The human half of the conversation differs too. Against a real Claude child,
+the CLI did not echo the client's prompt as `user_message_chunk`, and loading
+its session replayed zero updates. A Claude client cannot reconstruct the whole
+conversation from ACP alone; record the human turn yourself. Codex's mapping
+does emit text from `userMessage` items as `user_message_chunk` during history
+replay, so a Codex `session/load` will carry that human text. The mapping also
+does this for live items, so a live Codex turn may echo the client's own prompt
+as `user_message_chunk`; this is what our mapping allows, not a claim about
+every Codex frame. Record the human turn yourself for both providers, and treat
+a live `user_message_chunk` as the agent's record of what it received, not as
+new input to append again.
+
 ### Refusing an unknown id at load, not a turn later
 
 Pass `probeSession` and an id naming **no** conversation is refused by
@@ -298,6 +322,14 @@ A client then speaks ACP on those pipes. Several sessions run at once, each with
 its own agent process, its own in-flight turn and its own updates — tagged with
 the session they belong to, because a room of agents shares one stream.
 
+### Commit streamed messages at the turn boundary
+
+`agent_message_chunk` and `agent_thought_chunk` carry no message id. The turn
+boundary is the only delimiter between replies, so commit the accumulated
+message there. Clearing the in-flight message without committing loses the
+reply (including the whole reply in a single-message turn); not clearing it
+appends the next turn to the previous one.
+
 ## Development
 
 ```sh

package/dist/codex/driver.js

@@ -963,6 +963,8 @@ export class CodexDriver {
         this.#turnDeadlineExpiresAt = 0;
     }
     #hasOutstandingWork() {
+        // If this driver adds client-served fs/* or terminal/* calls, their in-flight
+        // ACP requests belong here too; a slow client tool is outstanding work.
         return this.#pendingApprovals.size > 0 || this.#activeItemIds.size > 0;
     }
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@particle-academy/prism-acp",
-  "version": "0.5.0",
+  "version": "0.5.2",
   "description": "Speak the Agent Client Protocol to a coding-agent CLI the user has already authenticated. No API key, no third-party adapter.",
   "license": "MIT",
   "repository": {