| name | workflow-resume |
| description | Resume work on an existing workflow in a new session. Use when the user says
"resume", "continue", "pick up where we left off", or names a specific
workflow ID to work on. Loads the handoff, dossier, and run-log, then begins
executing the next step.
|
workflow-resume
Purpose
Load the minimum context needed to continue working on an existing workflow in a new session. This is the most common agent entry point for returning sessions. The goal is to get oriented quickly without reading everything -- load what you need, skip what you do not.
Preconditions
- A workflow packet exists in
.agents/workflows/<state>/<workflow-id>/ where state is active, suspended, or blocked.
- The workflow ID is known (provided by the user, found in
active-workflows.md, or identified via current-focus.md).
Required Reads
Read in this exact order. Stop loading as soon as you have enough context to take the next step.
.agents/workflows/<state>/<workflow-id>/handoff.md -- read this first. It tells you the current status, the exact next step, active risks, and what else you need to read.
.agents/workflows/<state>/<workflow-id>/dossier.md -- read the Objective, Acceptance Criteria, Scope, and Plan of Record. Skip Background/Context and References unless the handoff says you need them.
.agents/state/active-workflows.md -- check for parallel workflows and write-scope conflicts before making changes.
- PM units listed in the handoff's "Required Reads Before Resuming" section, if any.
.agents/workflows/<state>/<workflow-id>/run-log.md -- read only the last 3-5 entries unless the handoff indicates you need more history.
Procedure
-
Identify the workflow. If the user names it, use that. Otherwise, read .agents/state/active-workflows.md or .agents/state/current-focus.md to find the most likely candidate.
-
Read the handoff file first. Extract: current status, next exact step, active risks/blockers, and required reads.
-
Read the dossier selectively. Load the contract-controlled sections (Objective, Acceptance Criteria, Failure Policy, Scope, Non-Goals). Load other sections only if the handoff references them.
-
Check for blockers. If the handoff lists active blockers, assess whether they are resolved. If still blocked, report to the user and stop.
-
Check for parallel work conflicts. Read active-workflows.md. If another workflow's write scope overlaps with yours, note this before proceeding.
-
Declare execution mode. Unless the user explicitly requests otherwise, default to
autonomous execution. State to the user:
"I will execute autonomously against its acceptance criteria. I will only
stop if: a stop condition from the guardrails is triggered, a blocker is encountered that
requires your input, or an acceptance criterion cannot be met."
Then proceed without waiting for confirmation. If the user has given a directive like
"execute " or "work on ", treat it as authorization for
autonomous execution.
-
Begin the next step as described in the handoff. Log the session start in run-log.md
with a new entry that includes the execution mode and checkpoint expectation:
Execution mode: autonomous. Checkpointing per guardrails § Checkpoint Protocol.
Follow the checkpoint protocol throughout the session — update handoff.md and append
to run-log.md at each trigger defined in guardrails.md.
Output
- Location: No new files are created. Updates go to:
.agents/workflows/<state>/<workflow-id>/run-log.md -- append a session-start entry.
.agents/workflows/<state>/<workflow-id>/handoff.md -- update before session ends.
- Format: Run-log entries follow the run-log template fields (timestamp, actor, planned step, action taken, result, files touched, verification run, criteria impact, blocker or risk).
Completion Criteria
- The agent has read the handoff and dossier and can state the next concrete action without guessing.
- A session-start entry has been appended to the run-log.
- The agent is ready to execute the next step from the handoff. The resume itself is complete when execution begins.
Error Handling
- If the workflow ID does not match any packet directory, list available workflows from
active-workflows.md and ask the user to clarify.
- If the handoff file is empty or missing "Next Exact Step," read the full run-log to reconstruct the last known state. Then update the handoff before proceeding.
- If the workflow state is
blocked or failed, do not silently resume. Report the state to the user and ask how to proceed.
- If the dossier's contract-controlled sections have been modified since the last session (compare with run-log), flag this to the user -- it may indicate an unauthorized spec change.