fix stale timeout, empty tool calls and trailing-space commands

Three defects from a review of the agent loop.

A session's tool schema is built by create(), and load() then replaced the
whole object with the one on disk, so a session saved under SOLO_TIMEOUT=42
went on telling the model the 42 s limit while run() enforced whatever the
environment said now. The schema moves into mkschema() and is rebuilt after
the load, so the description and the deadline always agree.

A tool call whose arguments held no command, which a flaky router or a
truncated stream can produce, was passed to the shell as an empty string.
The shell ran it as [exit 0], the agent asked again, and a peer that always
did this pinned the loop at full speed: a mock measuring 685 requests in
five seconds of the same call. Such a call is now answered with an error
instead of a shell run, command is declared required in the schema, and
three rounds in a row that run nothing end the turn. The same mock now
stops after three requests.

dispatch() read a line without trimming it, so "/help " was taken as the
generic parameter editor and wrote a help key into the session, "/quit "
did not quit, and "/temperature 0.5   " stored trailing spaces. A line
that starts with / or ! is right-trimmed before it is routed; a task keeps
its exact text, leading space included, so the model still sees what was
typed and a leading space still keeps the line out of the history.

Eleven offline regression tests cover the three, including a mock endpoint
that sends nothing but empty tool calls. make test: 42 passed, 0 failed.

README.md

@@ -258,6 +258,11 @@ below, but it only writes a `timeout` field into the saved session, which the
 API ignores and which does not change the command timeout. Restart with the
 environment set instead.
 
+The timeout is not persisted with the session: the tool description, which
+quotes it, is rebuilt from the current environment on every start, so a session
+saved under `SOLO_TIMEOUT=42` and reopened under `SOLO_TIMEOUT=99` describes and
+enforces 99 seconds.
+
 ## Slash commands
 
 | Command            | Effect                                                          |
@@ -347,7 +352,10 @@ The loop, in order:
 
 1. `readline` reads one line. A trailing `\` continues it on the next line.
 2. `dispatch` routes the line: a slash command changes settings, `!` runs a shell
-   command locally, anything else becomes a task.
+   command locally, anything else becomes a task. The line is right-trimmed
+   first, so `/quit ` and `/help ` are still commands and a value such as
+   `/temperature 0.5   ` carries no trailing spaces; a task keeps its exact
+   text, leading space included.
 3. A task is appended to the message list as a `user` message and saved.
 4. `hc` in the forked child performs one HTTP request with libcurl, wrapping the
    transport-level result as `{"status", "success", "content"}` and writing it
@@ -383,6 +391,12 @@ The request is a plain non-streaming `POST /chat/completions` with
 
     bash(command: string)  -> combined stdout/stderr plus an exit marker
 
+`command` is required by the schema. A tool call that arrives without a usable
+one, which happens with a flaky router or a truncated stream, is answered with
+an error instead of being passed to the shell as an empty string, and three
+rounds in a row with nothing to run end the turn rather than loop on a peer that
+never sends a command.
+
 Reasoning fields are handled defensively, because providers disagree on the
 name: `reasoning` and `reasoning_content` are both read, and both are stored
 back into the transcript so that a follow-up turn can reuse them.