Submit Isaac feedback asynchronously in Smart Router sessions

docs/claude-mod.md

@@ -1,7 +1,8 @@
 # Claude mod entry point
 
-UG's smart-routed Claude launcher loads `hooks/register.ts` through the existing
-launch-scoped `ug-smart-router` plugin. Claude Code 2.1.290 or newer is required.
+UG's Claude launcher loads `hooks/register.ts` through the existing launch-scoped
+`ug-smart-router` plugin in sessions launched with smart routing. Claude Code 2.1.290 or
+newer is required.
 
 The existing `_write_routed_claude_plugin()` copies the packaged TypeScript entry point
 from `src/ucode/agents/claude_mods/` and writes `hooks/hooks.json` into that plugin.
@@ -63,13 +64,48 @@ rollout setting or the user's explicit `1`/`0` or `true`/`false` override. As wi
 boolean environment flags, `1`, `true`, `yes`, and `on` enable it (case-insensitive,
 with surrounding whitespace ignored); other values disable it. It controls Isaac feedback
 availability, so UG's routing presets and on/off toggles leave it unchanged.
-The native form keeps routing on until Enter accepts a valid reason, then disables
-routing and uploads the exact text in the background with `isaac feedback`. The
-transcript shows the feedback ID the mod passes with `--feedback-id`; a toast reports
-the result. Cancel or Escape leaves routing on and sends nothing. A missing
-`isaac` CLI or an unavailable form leaves routing on. A feedback submission failure after a valid
+The native form keeps routing on until Enter accepts a valid reason. It closes the
+form and shows the sending message and feedback ID before starting Isaac. The existing
+native background callback runs `isaac feedback --with-evidence` to capture and upload
+the report while the user continues working. Routing is turned off independently of
+that process. Delivery produces both a persistent message and a success toast;
+a failed upload offers `/isaac-feedback retry <id>` when a local snapshot was saved.
+Before submitting, Cancel or Escape leaves routing on and sends nothing. A missing
+`isaac` executable at launch or an unavailable form leaves routing on. A feedback submission failure after a valid
 reason leaves routing off and reports the failure in the transcript. The form makes no model requests.
 
+In these sessions, with the same flag enabled, `/isaac-feedback` uses the same native
+form, capture, and background uploader as `/smart-router off`. It remains available
+after routing is turned off so users can report on an earlier routed task. It asks
+for a category when none was supplied, and for feedback text when the invocation is empty.
+`/isaac-feedback --category smart-routing <text>` submits directly. Smart-routing
+reports pin the current session's most recent task by its native user-message ID.
+The `session.append` hook remembers that ID; a session with no observed prompt uses
+Isaac's existing target inspection before opening the form. Empty sessions use an
+unmatched `no-prior-task` selector so subsequent work cannot become their target.
+The native flow does not generate an agent note. Other reports retain their text-only
+transport and do not receive an evidence feedback ID. `/isaac-feedback retry <id>` uploads the saved
+snapshot in the background without capturing it again.
+
+The submission and its sending/completion messages make no model requests and
+need no updated Isaac skill instructions. Explicit inspection and local-preview
+requests still use the existing skill. Earlier targets, agent notes, and additional
+files remain available through the existing Isaac CLI options.
+
+Claude sessions launched without smart routing use the existing feedback skill.
+
+Each submission passes Claude's native session ID and the current routing controls
+to Isaac. Valid session-file overrides are copied into that child process's environment,
+so turning routing off or back on cannot change the report's recorded settings.
+These settings describe submission time; subagent execution still requires routing-log
+evidence. An invalid controls file remains unverified.
+
+These uploads belong to the running Claude session. Keep it open until completion;
+if interrupted, retry the saved evidence ID on the same machine. If capture did not
+finish, submit the feedback again. No Isaac background worker or new CLI options are
+required. Evidence is collected in the background: the selected task stays fixed,
+while transcript and file snapshots reflect capture time and available file history.
+
 `isaac-feedback.ts` owns this Databricks-specific policy, form, and submission.
 `register.ts` binds the routing callbacks before Isaac's command guard, then
 handles toggles the guard passes through. Claude's mod compiler requires engine