Explore several Ways

Explore several ways runs the same bounded task through multiple Agent and model combinations. Each Attempt works from its own independent Git clone in a dedicated local Podman container. You review both the Outcome and the Route, run one repository check against every completed result, and deliberately Keep this one or finish without keeping.

Preconditions

Attempts currently require all of the following:

• a single autonomous-Agent Session (coordinators and workers are not selectable);
• a Git repository;
• the local Podman sandbox backend;
• at least two configured, usable ways;
• a task that does not depend on uploaded attachments, Skills, MCP tools, or web search.

Only one unresolved attempt set can exist for a Session. Finish or recover it before starting another.

Each Way uses its configured primary provider and model directly. A primary rate limit fails that Way; provider fallback is intentionally disabled here until retained Way identity and cost can record the effective fallback route instead of mislabeling it as the primary.

Prepare the task

1. Write a task in Conversation, then choose Explore ways.
2. Select heterogeneous Agents or model overrides. Diversity is useful only when the Ways can plausibly take different routes.
3. Open Ways to review the draft.
4. Optionally choose Plan first. One autonomous Agent proposes a shared plan; edit the resulting instruction before applying it to every way.
5. Choose Check models to preflight the selected provider/model targets.
6. Start the attempts only after the task and every target are correct.

A strong task gives every Attempt the same observable finish line:

Reduce duplicate parsing in the import path without changing its public API.
Add focused regression coverage, keep the patch within the parser module, and
leave the repository uncommitted. The result must pass: cargo test -p importer

Plan first and model preflight can fail independently of the run. Fix the specific configuration or edit the instruction; do not treat a generated plan as authoritative.

Follow the Attempts

Ways is a contextual inspector for preparation and live state. Each Attempt reports its current phase, Agent/model target, token usage when available, and failure or blocked detail. You can close Ways and continue reading Conversation; while the decision remains unresolved, the attempt state is durable and rehydrates when reopened.

Before Checks, changed paths and diffs are live-runtime evidence. After a daemon restart they remain explicitly unavailable until a successful Checks run protects the candidate identity. Completed Check evidence and its comparison then survive reload while the decision remains unresolved.

After every way reaches a terminal state, open the focused Attempts tool from More. Review:

• Outcome — the result, changed files, and verification evidence;
• Route — the actions and tool path used to reach it;
• the per-Attempt diff and any failure or blocked state;
• token and price information, including explicitly unavailable pricing.

Run Checks

Enter one non-interactive check command and choose Run checks. Axocoatl runs that command against each completed Attempt. A check is not inferred from the Agent’s prose, and a successful Agent turn is not the same as a passing check.

An Attempt is keepable only when it completed, changed at least one file, and passed Checks. If Checks are interrupted, partial results remain visible while the set is unresolved and you can retry them. You can also stop Checks and finish without keeping.

Use a bounded check command

Do not use a watcher or interactive development server as the check. The same command runs separately for every candidate and must terminate on its own.

Judge or choose directly

Judge asks a selected autonomous Agent to rank at least two passing, non-empty survivors. It is available only after Checks. Treat the recommendation and its reasoning as another review aid, not as proof.

You may instead inspect the candidates and choose Keep this one directly. Keeping:

1. applies the selected Attempt’s uncommitted delta to the Session working tree;
2. records the selected task, output, and turn attribution in the durable Session transcript;
3. removes the remaining Attempt clones and containers during cleanup.

It does not create a commit. Review the applied diff in Source Control and run the check again in the main Session when the repository or environment makes that prudent.

Keep ends the candidate record

Routes, candidate diffs, Checks, Judge ranking, and cost evidence remain available while the Ways decision is unresolved. After Keep and cleanup, durable History keeps the selected task, output, and turn attribution; it does not keep the losing Ways or the candidate comparison record. Review the current Git diff before committing.

Recover an interrupted decision

Keep is resumable. If the connection or process stops after apply begins, reopen the same attempt set and choose Finish Keep for the selected Attempt. Finishing without keeping is disabled once apply has begun because Axocoatl must finish recording and cleanup for the already-selected result.

If cleanup was interrupted before keeping, choose Finish cleanup. Failed and blocked Attempt states remain visible while the set is unresolved instead of being presented as empty successes.

Finish without keeping

Choose Finish without keeping to remove every Attempt and leave the Session working tree unchanged. During running Checks, the control becomes Stop checks & finish without keeping. Candidate review evidence is removed with the set and no selected result is added to Session History.

Attempt pricing covers the Ways where tracked usage is complete and configured prices are known. Plan first, Check models, and Judge calls are shown separately and excluded from each Way’s execution total. A successful Judge record remains with the unresolved Attempt set. Plan and model-check details are request-local; the selected planning Agent’s cumulative usage also remains visible in Settings. A failed call that returns no terminal usage is labeled as an incomplete lower bound, and remote usage without price data is unknown rather than zero.

Continue with Review Git and Last turn →