Second opinions and failures
Ask an agent on another provider to review your change, and hand a failing command straight to an agent.
NOTE
harmony review and harmony on-fail are not in 3.25.8, the current
release. They arrive with the next release.
Two commands that move work between agents without you copying text across
panes. Both deliver the same way as harmony fleet send --submit, so they
tell you if the agent could not be reached instead of assuming it got the
message.
A second opinion: harmony review
harmony review # your uncommitted changes
harmony review --base main # the commits on this branch since main
harmony review --with Tide --focus "error handling"
By default the reviewer is an agent in your workspace running a different
provider from yours, say Codex reviewing work by Claude Code. The point is
a second model's view, so Harmony never quietly picks one on the same
provider. If those are the only others running, it names them and you can
choose one with --with <name>.
The reviewer gets the diff as a file, is asked not to change anything else,
and writes its findings, most serious first, to a second file. Harmony waits
for that file and prints the findings. Both files are kept in
.harmony/reviews/ in the repository, which git ignores. Any agent that can
read and write files can answer.
| Situation | What you see |
|---|---|
| Nothing to review | Says so; new files count once added with git add |
| No second agent | Says so, and how to start or name one |
| The reviewer can't be reached | The reason, and nothing is left behind |
No answer within 20 minutes (--timeout-minutes) | That it hasn't finished, and where the findings will appear |
--no-wait asks and returns straight away, with the path to read later.
Hand a failure to an agent: harmony on-fail
harmony on-fail --to Slate -- npm test
harmony on-fail --to Slate --prompt "The build broke. Fix it." -- npm run build
The command after -- runs as usual: you see its output, and its exit code
is kept, so scripts and CI still see the failure. If it fails, Slate gets the
command, the folder, the exit code and the last 80 lines of output
(--tail to change it), and is asked to look into it. A passing run sends
nothing.
Add --watch <folder> and on-fail keeps going: it runs the command again
whenever files in that folder change, until you press Ctrl+C.
harmony on-fail --to Slate --watch src -- npm test
Changes in .git, node_modules, .harmony and build output folders are
ignored. Slate hears about a failure once, not on every save: the same
failure again isn't re-sent, a different failure is, and after a passing run
the next failure always is. A failure that couldn't be delivered is tried
again on the next run.
A name that can't be reached is flagged before the command starts, so a typo doesn't cost you a long test run. The command still runs, since the agent may start in the meantime. If the failure can't be delivered afterwards, on-fail says so and gives the reason.
Related
- Delegation: assigning tracked tasks to agents
- Verification: what counts as proof that work is done