| name | handoff |
| description | Create or execute a context-complete cross-domain plan as an orchestrator. Use when another agent must continue without prior context, or when a multi-domain plan needs coordinated execution while this skill retains decision ownership. For coding-session persistence, use essential:handover. |
Handoff
Create or execute a context-complete cross-domain plan. This skill owns
portable planning and coordinated execution; essential:handover instead
persists the current coding session in continuation files.
Boundaries
- Use for: writing a plan another agent can execute without any of this
session's context, and orchestrating a multi-domain plan's execution while
retaining decision ownership.
- Do not use for: persisting a coding session for later continuation
(
essential:handover), or doing the planned work inline — execution is always
delegated.
Inputs
- Required: the work to plan, or an existing plan to execute.
- Optional: the
Workflow tool for multi-phase execution; coding:*
skills when available — confirm availability before routing to one,
otherwise name the equivalent action or files without invoking it.
- Required evidence for a persisted plan: current repository root, active
branch, base and HEAD commit SHAs, worktree status, relevant tool versions,
exact command lines already run with outcomes, the
capability_id of each
actor whose run produced recorded evidence (Essential's truth.md), and
absolute paths to every source artifact the next executor must use.
Before creating or materially rewriting a project artifact, read the absolute
engineering-work.md path injected by Essential. If unavailable, stop artifact
writes and report the missing contract. Resolve the active work directory from
that contract.
Workflow
- Resolve material uncertainty. Separate user-stated intent, observed
evidence, inferences, accepted assumptions, and unresolved questions. Ask
only questions whose answers change scope, architecture, acceptance
criteria, sequencing, or another material decision; give a recommended
answer and reason. Remaining uncertainty must be low-impact and reversible,
explicitly deferred with an owner and decision deadline, or marked blocking.
- Write the plan as a zero-context handoff. Open with a
self-contained Goal block the user can copy and paste verbatim to
initiate the work — it states the intended outcome fully on its own, without
relying on your context. Then write the plan so an execution agent never has
to rediscover basics. The persisted plan must include these named sections:
- Baseline — exact repository root, active branch, base SHA, HEAD SHA,
dirty-status summary, environment constraints, and every test, coverage,
lint, build, or inspection command already run, with timestamps or run
order, exit status, and concise output/result.
- Immutable Inputs — exact absolute paths and repo-relative paths,
filenames, artifact IDs, URLs, issue/PR IDs, data snapshots, and commit
SHAs that define the work. Include file checksums when a referenced file
lives outside the repository or may change independently.
- Current Decisions and Assumptions — user-stated intent, observed
evidence, accepted assumptions, unresolved questions, owners, deadlines,
and the precise evidence that should trigger a plan pivot.
- Execution Environment and Tooling — concrete shell commands for
setup, dependency installation, code generation, tests, coverage, linters,
formatters, previews, and verification. Commands must be copy-pasteable,
include working directories and required environment variables, and name
the expected pass/fail signal.
- Step-by-Step Implementation Plan — ordered phases with exact files to
read or edit, the reason each file matters, acceptance criteria per phase,
and rollback/stop conditions.
- Dynamic Workflow Script — when
Workflow is appropriate, embed either
a complete plain-JavaScript Workflow script or an exact durable script path
plus its SHA-256 checksum and invocation arguments. The script must be
concrete enough for the next agent to run as-is: no placeholders, no
hidden context, deterministic inputs, explicit agent types, and validation
against plugins/essential/references/workflow-tool.md. If Workflow is
unavailable or inappropriate, include an equivalent sequential command plan
instead.
Avoid pronouns like "this", "that", "above", or "the current task" unless
the noun is named in the same sentence. Prefer absolute paths plus
repo-relative paths for handoff-critical files. When the plan must persist,
write it to a lowercase proposals/<plan-slug>.md child with status open;
the PM reconciles the lazy proposals.md overview. Update the child to
accepted only after user approval.
- Execute as orchestrator (when execution is requested). Run a
multi-phase plan as a
Workflow — one phase per stage, fanning out to
subagents where a phase allows — instead of doing the work inline. Act as
the orchestrator and decision maker only: route each phase to the right
agent with complete context, synthesize the results, and make the calls.
Delegate all execution — reading, writing, running, testing — to
subagents; never do the work yourself. To pause a coding session rather
than hand a plan off, use essential:handover.
- Run the verification below; when a check fails, fix the cause and re-run
that check. Repeat until every check passes or a concrete blocker remains,
then report the blocker instead of looping.
Verification
- The Goal block stands alone: pasted into a fresh session, it fully states
the intended outcome.
- The plan has a Baseline section with exact paths, branch, base/HEAD SHAs,
dirty status, commands, outcomes, and relevant output.
- A reader without this session's context could execute the plan — no step
depends on unstated knowledge or rediscovery of paths, SHAs, filenames,
commands, tools, or source artifacts.
- The plan embeds a complete dynamic Workflow script, or names the durable
script path, checksum, and args; if Workflow is not used, it embeds the
concrete sequential execution commands.
- Every residual unknown is accepted and reversible, explicitly deferred with
an owner and deadline, or blocking; the plan names evidence that requires a
pivot.
- When executed: every phase was delegated with complete context, and each
phase's results were checked against the plan's success criteria before the
next phase started.
Completion
Report the plan's location and its Goal block, the decisions made while
planning, and — when executed — each phase's outcome and any open questions
or deviations from the plan. A blocked execution names the phase, what was
attempted, and what decision or input is needed to continue. Return explicit
final paths generated or materially rewritten as generated_files; the PM
size-checks only eligible work Markdown inside the target .state/.