Resources
1Install
npx skillscat add brandonocasey/dotfiles/session-resume Install via the SkillsCat registry.
Use agent-task as the durable task record. Keep credentials, full transcripts,
and executable commands out of it. Stored authorization is evidence of the scope
already granted. It never grants a new push, deploy, release, or destructive action.
Reuse rules (other skills cite these): retain a resumed record's authorization only
within its exact provenance and scope. Reuse a recorded check or manual evidence only
when its code revision, inputs, and environment match.
Create a record with the objective and exact authorization provenance and scope:
agent-task create --id <id> --objective <text> --authorization <source> --scope <scope>Use --parent-pr, --target-branch, and --task-boundary when those fields
define where the resumed work belongs. The same flags update an existing record.
Before work, run agent-task show <id>. Check its repository, worktree, and HEAD
against the current checkout. Apply recorded corrections. Continue from nextStep.
Update with the revision returned by show; stale updates fail:
agent-task update <id> --revision <n> --next-step <text>
agent-task check <id> --revision <n> --name <check> --result <result> \
--code-revision <sha> --inputs <inputs> --environment <environment>Add each user correction with --correction <text> so later sessions do not
repeat the superseded approach. After a stale-revision error, run show again,
reconcile the newer record with the pending change, and retry with its revision.
Set --status complete only after all task work
is complete. Use agent-task list to find active records.
For old context, run agent-task index to update the private incremental index and see scan
coverage (Codex and Claude history, transcripts, archives, alternate profiles, plus unreadable
or malformed files). Omit --max-files and --byte-limit unless a bounded partial update is
required. After a deliberate rewrite deep inside a growing log, run agent-task index --rebuild.
Use agent-task resolve --session <id> to find an exact session. Useagent-task search --query <text> [--session <id>] to search human turns. Each
JSON-array result includes the host, session, source offset, and bounded context.
Search writes coverage JSON to standard error, including when it finds no match.
Resolve includes coverage in its JSON object. Use --file <path> to add or
refresh one known transcript.
When the user asks to transfer a task between hosts, copy only its private JSON record with ssh,
then set its mode to 0600. Run show on the destination and validate the host,
repository, worktree, and HEAD before work. Do not copy transcripts or credentials.
The skill never starts remote commands automatically.