A row says when its last run failed

[suleimansh]

A row says when its last run failed, and three small fixes found while building "New automation" (#2032).

👤

• A run that fails is not tried again by itself. The row must say so, so I see it.

🤖 automated · Opus 5.5, effort high

Last run failed

• A row of the Automations page says "Last run failed" when the latest run of its command ended failed. Pressing it opens that run.
• It is said whatever the row's switch, and goes once a later run of the command starts.
• The scheduler reads it once per look, after its sweep, off the run records.

A prompt that opens with a dash

• A prompt typed as - fix the tests did not start: the runner's command line took it for a flag. The same held for a follow-up and for an answer.
• Now only one of the command's own flags is a flag. A text that opens with a dash and goes on is a text.
• One bare word behind a dash that is no flag (--help, a flag typed wrong) is still refused, and a flag does not take another flag for its text.
• The fix is in how agent-runner reads its command line, so every project's hook file works as it is.

Words

• A row whose shell line has no plain words said "when its check finds work". It says "when its shell line prints something".
• The page's top line said "Your choices, on this machine." though a shared row is a file of the project. It says "Each row's switch and picks are yours, on this machine."
• The form's example shell line leaves out comments that start with 🤖, so a row built from it does not answer its own agent.

Known

• While the scheduler is not running, the page shows what it last read, so "Last run failed" can be out of date then, like the rest of the row.

DECISIONS.md

One new bullet in packages/agent-scheduler/DECISIONS.md. It only writes down the pick above:

A run that fails is not started again by itself: the time a check is given moves with every start, a failed one too. The command's row says that its last run failed, until a later run of the command starts, so a person sees it and can start the work by hand. Picked over the time moving only after a run that ended well: the same failure would start an agent again each time the pace allows, and spend by itself.

Logic files not updated

By the person's choice the LOGIC.md files are not updated in this pull request (#2038). They are brought in line later, in one pass.

Checks

• Full root CI: build, typecheck, tests.
• One fresh reviewer read the diff. It found three faults, fixed here: run --help or a flag typed wrong would have started an agent; a flag could take the next flag for its text; "Last run failed" stayed a look too long after the row started again, and went missing while the records could not be pulled.