scoville-handoff

Create one compact, factual continuation prompt for an explicitly requested transfer to another agent or session, including Scoville Handoff or "handoff to a new session". Do not activate for ordinary

SKILL.md

GitHub

Scoville Handoff

Give the receiver accurately labeled context to continue safely from one copy-ready snapshot. An explicit transfer applies even when work is empty, completed or not started. Without a transfer request, perform the requested task without a handoff. If future reuse is requested but transfer intent is unclear, ask one question before reading sources or producing a handoff. When asked to explain or assess a hypothetical handoff, answer that question. Produce the continuation artifact only for a requested transfer of an actual task.

Read within the transfer scope

Any already required and authorized pre-transfer update belongs to its active owner, such as Plan for Step status, and must finish before Handoff starts. The transfer request grants no additional editing authority.

Use task facts already established in the conversation. Additional reads are limited to sources explicitly named or already established as task sources in the conversation, optional read-only version-control inspection, and the continuation template. Do not search for more sources. Throughout this workflow, do not advance the task: no edits, builds, tests, probes, dummy or other task commands, unrelated reads, stat or list operations, or external effects. Missing facts do not authorize additional actions. Only for preparing and delivering this handoff, interpreter discovery, temporary artifact writes, size checks and complete-file publication are permitted. This exception grants no task edits, tests or other task commands.

Retain each source's exact path and read state:

Read state Action
Complete Use the established facts.
Partial Read permitted ordered ranges or cursors within their limits. Never attempt an oversized read or recover text after truncation.
Failed range, plausibly transient Retry that range once; explicit user read limits prevail.
No progress, repeated failure or read limit Stop recovery. Preserve usable facts, recovery, gaps and unread ranges beside the source. A partial source is not wholly unavailable.
  1. Finish the permitted recovery sequence before composing, including stopping when the table requires it; do not defer these attempts to the receiver.
  2. Compose the snapshot from usable facts, preserving each source's read state, recovery, gaps and unread ranges. Never label a partial source complete.
  3. Conflicting source revisions and material unread ranges block receiver verification and actions that depend on them, not snapshot creation.

Preserve continuation facts

Read the shared writing rules when composing the continuation prompt or assessing a disputed factual or technical conclusion. They grant no permission to read more task sources. Select the relevant continuation facts below.

Capture the goal, deliverable, acceptance, scope and authority, canonical owners, user-owned changes, accepted decisions, active work and running handles, observed evidence and its limits, blockers, hazards and next safe action. Preserve exact IDs, paths, commits, URLs, commands, errors, quoted decisions, assumptions, unknowns and time-sensitive details where needed for continuation. Transfer known material facts, not just pointers or instructions to reread them. The handoff request itself is not a task decision. Keep preferences and proposals distinct from requirements and accepted decisions.

Use unknown for missing information and none known when no instances are known, such as no known blockers. Neither proves absence. Set Status: not_started only when the conversation or an explicitly named or already established task source establishes it; otherwise a missing status stays unknown.

For file-based work, use the task working directory established by the conversation or an explicitly named or already established canonical project source. A task repository already verified during the work is sufficient; the user need not name it again. If no task location is established, include Working directory: unknown. Runtime CWD alone does not establish the task location. Include temporary workspace or host state only when established as task state.

Output secret-bearing facts only in redacted form: NAME=[redacted]. Keep each established variable name. Redact every secret occurrence in quotes, warnings and replacement instructions too; never show the original value when explaining its removal.

For a tight output limit, remove repetition and unrelated history first, then shorten explanations. Preserve authority, ownership, hazards, evidence limits and the safe first step. An explicit lossless request preserves every in-scope non-secret fact. If required content still cannot fit, publish the complete unchanged fenced artifact through the shared complete-file route. Return its absolute path and an instruction to read the entire file before continuing.

Compose and check the prompt

Keep the template's four H2 sections: Receiver Instructions, Objective, State and Resume Steps, with all three Resume Steps. Keep Receiver Instructions intact except for translation and secret redaction. Use the requested language, otherwise the conversation language; translate headings, labels and instructions consistently. Preserve technical identifiers, literal markers such as unknown and exact quotations except secret values. Use one outer Markdown fence with at least four backticks and more backticks than any run inside the prompt; match its opening and closing length, including when saving the artifact to a file. Include every Objective field. Under State, always preserve Status and the required continuation facts, including relevant unknowns. Other template labels are optional; omit empty or inapplicable categories. Name each source once beside its facts and identify conversation facts as such. Repeat a fact only when a hazard or first step needs it.

Choose Resume Step 1 in this priority order:

Condition First step
Goal unknown, or Acceptance unknown and next action depends on it Ask before execution or recovering a running command.
Known blocker Resolve the first blocker.
In-flight work Recover its state.
Otherwise Take the next safe action.

Make the remaining steps concrete and end with an observable completion criterion. Active or incompletely accepted work names its decisive next check. For completed work with current evidence, the three steps cover only checking current state, reconciling contradictions and confirming no material mismatch. Do not invent more work or repeat a current check to fill a step.

  1. Compare the full prompt with captured facts and sources: required facts, exact identifiers, attribution, Objective fields, Receiver Instructions and first safe step. For lossless requests, check every in-scope non-secret fact.
  2. Correct omissions and contradictions; remove placeholders, secrets, invented facts and tool details used only to prepare the snapshot.
  3. Return exactly the fenced artifact without surrounding text, or the specified path and complete reading instruction for file delivery.

Handoff owns the snapshot; preserve active sibling state in it.

This Skill works independently. Other Scoville Skills are optional. Use an available, active sibling only for its applicable concern; do not install, simulate or require an absent sibling. Honor explicit user exclusions.

Family owners, in suite order:

  • scoville-code: engineering scope, implementation, risk, and validation.
  • scoville-plan: durable Plans, Work Items, Decisions, and lifecycle state.
  • scoville-ui: framework UI implementation and acceptance, including supported WordPress admin surfaces.
  • scoville-handoff: active-work transfer.
  • scoville-project-context-cleanup: requested project-rule and index wording, information quality and structure.

Reuse an already verified interpreter meeting this Skill's Python 3.11+ requirement. Otherwise check py -3 on Windows or python3 elsewhere; try python if needed. Choose it locally, without asking the user. Verify its version before the first helper operation. Use that executable wherever examples say python or <verified-python>, including Python commands after --run --. Report a missing runtime only when no suitable installed interpreter is found.

Before processing machine reports or logs, use the shared writing rules. If their complete text is not loaded, read it in a separate call before reading the data. Reuse retained rules under their existing reload conditions. Before helper calls, follow shell command rules.

Runtime helpers

Use the bundled helpers for their operations. Read their invocation instructions, not their source, unless diagnosing a failure.

Condition Required route
Suitable Python 3.11+ is available Use the bundled helper; do not load manual references.
No suitable Python 3.11+ Read only the matching optional reference below.
Missing script, missing dependency or helper error Stop the affected operation; this never enables the manual route.
Helper Optional no-Python reference
scripts/check_text_size.py check_text_size