Task Capsules

Set up and run

1. Connect the accounts you want to use under Models & Providers.
2. In Specialists & teams, configure reusable profiles with the desired account and model. An implementation profile needs an Access ceiling of Workspace edits. One profile can serve more than one role.
3. Choose Task Capsules… in the composer workflow menu, open Locus → Task Capsules…, or press ⌥⌘K in the target workspace.
4. Describe the task or customize an editable example. Choose Plan with, Implement with, and optionally Review with. Set optional allowances under Advanced · Usage limits.
5. Select Generate plan. Planning is read-only; answer any clarification in its conversation. A successful structured plan saves automatically.
6. Reopen the capsule and review instructions, constraints, checks, and design decisions. Expand steps reveals the detail. Choose Run plan when ready.
7. Review stage outcomes and available usage in run history. Review result runs a separate read-only review; Run again starts another execution of an existing plan.

If profiles are missing, Set up models opens settings while preserving the task description. Continue planning returns to a waiting planner's conversation.

Accounts remain explicit

Each stage uses the exact account assigned to its profile. An unavailable account stops the stage. ChatGPT-plan and Kimi Code membership routes never silently fall back to metered APIs. Capsule choices do not permanently replace the ordinary chat's model selection.

Saved plans and changed files

A detailed plan can contain up to 16 ordered steps. Execution preserves dependencies and keeps capsule writes sequential. The approved plan revision, content, and execution location are saved together and revalidated before dispatch. Before execution, Locus checks fingerprints for files named in those steps, including planned new files. Changed, removed, or unexpectedly created files pause the handoff.

Choose Update the plan or ask for help to inspect the current workspace and save a revision. Fingerprints cover named files, not the whole repository. Older plans without named files have no file baseline.

Limits, review, and recovery

Planning and standalone review have per-turn model-call allowances. Implementation, its automatic reviewer, and repairs share the execution allowance. Profile runtime and response limits still apply. A failed or malformed review does not count as approval.

Stage-specific API cost allowances remain estimates for their configured execution scope, not hard billing ceilings. The persistent invocation ledger separately attributes available planning, implementation, review, repair, retry, and reported tool or image usage to the task. Optional cumulative task spending limits are independent of capsule stage allowances. Unknown or unsettled charges remain visible and can block metered admission. Subscription routes record exposed calls and tokens without inventing per-token subscription prices.

Asking the planner for help is explicit. Repair and help allowances accumulate across the capsule, including later runs. Increasing a limit does not erase previous attempts.

Stopped work keeps its partial files, verified step progress, consumed allowances, repair count, and unresolved actions. Resume continues the saved attempt, Retry checks inspects existing outputs, and Run again starts another execution explicitly. If implementation is already done, recovery continues to checks or review without repeating it.

Resume compares current files with accumulated state in the actual execution checkout. Compatible verified steps can be reused; unexpected input changes invalidate affected and dependent work. An uncertain action or unsettled model call needs explicit review before recovery. Repairs rerun affected checks and the configured reviewer. The saved planner is not called again unless you request help.

Task details and acceptance

Open the capsule's task details to see the request, immutable approved plan, steps, outputs, current evidence, review findings, usage, and recovery history. Recipe settings remain available for model and allowance changes.

Declared acceptance checks must produce current tool evidence before the result is verified. Criteria requiring judgment produce Needs review and can be accepted explicitly. Historical plans with missing checks remain readable, but old completion claims are not retroactively verified.

Duo provides a quick planner/builder flow using these same capsule controls without specialist setup. In 3.0, saved-agent chats can also open capsule workflows; each planning, implementation, and review stage still uses its explicitly selected route. See Task Recovery & Restoration for selective restoration and Reusable Checks for approved project requirements.