/delegate

Ordinary implementation work may be handed to the Gemini CLI on gemini-3.1-flash-lite and reviewed afterwards. The loop is fixed.

For exploration, running builds or test you can also use gemini or Sonet 5 sub-agent.

Running it

1. Clean tree first. git status --short must be empty, on main. What git diff shows afterwards is then exactly what the delegate wrote.

2. Run it headless, task in a file to avoid quoting problems, and always with the workspace-trust variable exported:

   GEMINI_CLI_TRUST_WORKSPACE=true gemini -m gemini-3.1-flash-lite \
       --approval-mode yolo -p "$(cat task.txt)"
   

   yolo is required - a headless run has nobody to answer a prompt, and auto_edit stalls on the first shell command.

3. The env var is not optional. Without it the CLI first downgrades the mode ("Approval mode overridden to "default" because the current folder is not trusted") and then refuses to run at all: "Gemini CLI is not running in a trusted directory". Every headless run in an untrusted checkout dies there, before reading the task. --skip-trust is the flag-shaped equivalent; trusting the folder interactively does not carry into a headless run.

4. The workspace is the only writable root. An output path outside the repo

   • a session scratchpad under /tmp, for instance - is refused with "Path not in workspace" and the delegate quietly retargets the write to ~/.gemini/tmp/<project>/<name>, reporting success. So either name an output file inside the repo, or read the result back from that fallback directory instead of concluding the run produced nothing.

5. Re-index from the dashboard, so the review reads a graph that matches the tree rather than the one from before the edit.

What the task file must always carry

Do not commit or run any writing git command; the repository's language and encoding policy; edits must pass pre-commit run --files <paths>; an acceptance test the delegate has to check itself; and the name of any tree that is off limits.

Give the delegate the facts already established rather than letting it rediscover them - the free tier throttles by requests per minute, and every wasted turn costs a minute of backoff.

Two failure modes seen repeatedly, worth pre-empting in the prompt:

• A test that measured nothing. Where the thing being changed runs from a container image, the source is baked in by COPY, so the edit does nothing until the image is rebuilt - in this repository, a change under packages/ needs make build. A delegate that skips the rebuild is measuring the previous version.
• A substituted acceptance criterion. Asked for nodes in the graph, it answers with a file count, a log line, or a promise that they "should persist". Ask for the node list, and query it yourself regardless.

Reviewing what it wrote

Opus 5 reviews, and it reviews the actual changes - never the delegate's own summary, which is a claim rather than a result.

1. Take the real changes from git diff: the files, classes, defines, functions and parameters that actually moved.
2. For every changed entity, query the graph (search_code_nodes, get_code_graph_neighbors) for everything attached to it - callers, imports, includes, templates, data keys, whatever relations that language has - and check whether the change breaks any of them. Re-index first, or the sweep reads the tree as it was before the edit.
3. That impact sweep is exploration, so it runs on Sonnet 5. Sonnet gathers evidence; it does not rule.
4. The verdict - accept, fix on top, or revert - is Opus 5's alone, taken on the evidence gathered.
5. A migration or SQL change that breaks the database skill is reverted: a foreign key, a cascade, stored file text, or an ALTER TABLE other than a column type change the user explicitly allowed.

Then fix on top and commit with the commit skill. Local commit only; pushing is the user's call.

Recap

Say what was handed over, inside the recap the enggraph skill defines - counts, not adjectives.

• Delegations - one line each: which model ran, what it was asked for, what came back.
• Verdict - accepted, fixed on top, or reverted, and on which evidence. The delegate's own summary is a claim, so cite the diff or the graph.