--- id: lab-014-resume-reconciliation title: "Reconcile a resumed task before continuing" level: L3 domain: general goal: "Reconcile the task pointer, target, branch, permissions, and side-effect state before continuing" setup: "A disposable local folder or repository with a checkpoint and two text files; no credentials, network access, production files, or irreversible commands" task: "Record the live state, compare it with the checkpoint, classify each field, and continue only when the goal, target, permission, and side-effect state agree" evidence: - "The checkpoint, live observations, commands, outputs, diff, classification table, and continuation decision" - "A clear record of matched, changed, and not_observed fields" failure_variant: "Make the task name match while the repository root or target file differs; stop before editing and identify the first unreconciled field" reflection: "Which field was easiest to assume, and which observation changed the decision to continue or stop?" status: draft last_verified: "not run" transfer_task: "Apply the reconciliation envelope to a browser or MCP session without making a remote write" transfer_domain: "browser operations, research, engineering, or content hand-off" transfer_evidence: "Save the previous request, target, approval state, observed remote-state risk, and the new checkpoint" transfer_limitations: "The disposable fixture does not prove continuity of a real account, remote resource, or resumed production task" --- # Lab 014: Reconcile a resumed task before continuing **Status:** `draft` · **Run status:** `not_run` ## Why this lab exists Public field reports describe agents returning to an older task after context compaction, capacity interruption, or resume. A new prompt can make the run look active while the task pointer, worktree, or side-effect state remains uncertain. This lab teaches reconciliation before continuation. ## Setup Use a disposable copy of a small repository or a folder containing two text files. Create a checkpoint that names the goal, target path, branch, last completed action, pending action, permission state, and evidence. Simulate an interruption by starting a second task or replacing the checkpoint with an older copy. Do not use credentials, network access, production files, or irreversible commands. ## Task 1. Record the current working directory, repository root, branch, target file, file hash or modification time, and current diff. 2. Compare those observations with the checkpoint. 3. Classify each field as `matched`, `changed`, or `not_observed`. 4. Continue only if the goal, target, permission, and side-effect state are reconciled. Otherwise create a new checkpoint and stop. ## Evidence Save the checkpoint, commands and outputs, diff, classification table, and a short decision. A successful practice record proves only that the reconciliation procedure was followed in the disposable fixture. ## Failure variant Make the visible task name match while the repository root or target file is different. The correct result is to stop before editing and identify the first unreconciled field. Do not fix the wrong checkout merely because it is writable. ## Transfer Apply the envelope to a browser or MCP session: identify the last confirmed request, target account/resource, approval state, and whether a previous call could have changed remote state. ## Acceptance checklist - [ ] I captured the actual path, repository, branch, target, and diff. - [ ] I compared the live state with a named checkpoint. - [ ] I separated changed from not observed. - [ ] I stopped when target or side-effect state was uncertain. - [ ] I did not call a resumed prompt proof of continuity. ## Reflection Write down which field was easiest to assume, which observation changed the decision, and what remains `not_observed`. ## Sources - [Field problems and prompt patterns — P2](../../docs/research/field-problems-and-prompt-patterns-p2-2026-08-11.md), FP2-01 through FP2-04 and FP2-08. - [Chapter 10: Planning and vertical slices](../chapters/10-planning-and-slicing-EN.md). - [Chapter 12: The Agent loop, state, and stop conditions](../chapters/12-agent-loop-and-stop-EN.md).

---