docs(opencode): address review feedback

Author: serghei-dev

docs/architecture.md

@@ -1427,7 +1427,7 @@ Built-in adapter summary:
 | Kind | Session model | Event surface | Resume and identity | Notable differences |
 |---|---|---|---|---|
 | `claude-code` | One subprocess per turn | Newline-delimited JSON from `claude -p --output-format stream-json --verbose` | New sessions use `--session-id`; continuation turns use `--resume`; runtime `session_id` may replace the provisional adapter-generated ID | Token usage is normalized from streamed assistant and result events. |
-| `copilot-cli` | One subprocess per turn | Newline-delimited JSON from `copilot -p --output-format json -s --autopilot --no-ask-user` | Uses `--resume <session_id>` when a session ID is known; falls back to `--continue` when a prior result omitted `sessionId` | Session identity is captured from the terminal `result` event rather than a start event. |
+| `copilot-cli` | One subprocess per turn | Newline-delimited JSON from an abbreviated `copilot -p --output-format json -s --autopilot --no-ask-user ...` invocation | Uses `--resume <session_id>` when a session ID is known; falls back to `--continue` when a prior result omitted `sessionId` | The adapter always includes `--max-autopilot-continues` and includes `--allow-all` when no tool-scoping flags are configured; session identity is captured from the terminal `result` event rather than a start event. |
 | `codex` | One persistent `codex app-server` subprocess per session | JSON-RPC 2.0 over stdio | `ResumeSessionID` maps to `thread/resume`; otherwise the adapter starts a new thread; thread ID is the session ID | Turns are started inside the persistent session with `turn/start`; tool and approval handling are part of the app-server protocol. |
 | `opencode` | One subprocess per turn | Line-delimited JSON from `opencode run --format json --dir <workspace>` | `ResumeSessionID` maps to `--session <session_id>`; the first observed `sessionID` becomes the session ID; a mismatch is `turn_ended_with_error` | The adapter maps `opencode.model`, `opencode.agent`, `opencode.variant`, `opencode.thinking`, `opencode.pure`, and `opencode.dangerously_skip_permissions` to CLI flags; parses `step_start`, `text`, `reasoning`, `tool_use`, `step_finish`, and `error`; maps plain-text permission warnings to `notification` and unknown output to `malformed`; recovers final token usage with `opencode export --sanitize <session_id>`; maps logical `error` events to `turn_failed` even when the process exits with status `0`. |
 

docs/workflow-reference.md

@@ -1414,9 +1414,9 @@ known.
 **Validation rules:**
 
 - `opencode.allowed_tools` and `opencode.denied_tools` MUST NOT overlap.
-- When either tool list is present, the adapter removes any inherited
-  `OPENCODE_PERMISSION` value before launching OpenCode and replaces it with the
-  adapter-managed JSON policy.
+- The adapter always removes any inherited `OPENCODE_PERMISSION` value before
+  launching OpenCode. If either tool list is non-empty, it replaces that value
+  with the adapter-managed JSON policy.
 
 **Custom or future adapters (illustrative example):**