From mpi-kanban
Continue active MPI work from the real current state. Use when the user says "continue this MPI plan", "MPI continue", "continue", "resume", "keep going", "pick this back up", "read a handoff and continue", "$mpi-continue", or wants implementation to proceed from an MPI plan or handoff.
How this skill is triggered — by the user, by Claude, or both
Slash command
/mpi-kanban:mpi-continueThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Continue active work intelligently. This skill replaces rigid "execute next"
Continue active work intelligently. This skill replaces rigid "execute next" behavior: it reads the active kanban entry, plan, latest handoff, and current workspace state, then proposes the next best action based on reality.
When shared coordination state exists, mpi-continue also reads
.agents/mpi-kanban/state/index.json first and follows its pointers only as
needed. The shared contract is documented in
${CLAUDE_PLUGIN_ROOT}/docs/coordination/README.md.
Lifecycle operations are documented in
${CLAUDE_PLUGIN_ROOT}/lib/coordination-ops/lifecycle.md and status values in
${CLAUDE_PLUGIN_ROOT}/lib/coordination-ops/statuses.md.
Plans are living documents. If implementation has drifted, update or annotate the plan instead of forcing the next unchecked item.
Invocation: Claude Code users may run /mpi-kanban:mpi-continue; Codex users
may run $mpi-continue or ask naturally to continue the MPI plan. References
using ${CLAUDE_PLUGIN_ROOT} mean the installed plugin root; Codex resolves
the same files relative to this plugin root.
Find the active work from the first available source:
.agents/mpi-kanban/state/handoffs/<uuid>.json or legacy
docs/handoffs/*.json.Plan file: body line.Plan file: body line.If none is visible, ask:
Which MPI plan or handoff should I continue from? Please paste the path.
Lib pointers, read only when needed:
${CLAUDE_PLUGIN_ROOT}/lib/kanban-ops/find.md - findEntry${CLAUDE_PLUGIN_ROOT}/lib/kanban-ops/mutate.md - moveEntry${CLAUDE_PLUGIN_ROOT}/lib/kanban-ops/steps.md - addSteps, markStep${CLAUDE_PLUGIN_ROOT}/lib/coordination-ops/lifecycle.md - session/task/file claim lifecycle${CLAUDE_PLUGIN_ROOT}/lib/coordination-ops/statuses.md - state vocabulary${CLAUDE_PLUGIN_ROOT}/lib/project-knowledge/indexing.md - context-budget rulesdocs/handoffs/ pointer to a
canonical .agents/ handoff, load the canonical handoff before continuing..agents/mpi-kanban/project-profile.md and
.agents/mpi-kanban/project-knowledge-index.md before the Continue Brief.
Pick the topic block matching the active plan. Load only the listed
docs/rules; do not rediscover the whole project. If the profile is
absent, fall back to the existing pre-condition behavior.lib/coordination-ops/lifecycle.md. Call ensureStateRoot() when
coordination state is relevant, then read state/index.json as the active
coordination facade.implementer session and create or attach a task
record for the active kanban entry and plan.Plan file: <planPath>.Implementation.Orient current state, Implement active work,
Verify behavior, Preserve knowledge, Close session.git status,
targeted file reads/searches). Do not run large diffs unless needed.Before proposing work, compare the plan/handoff with actual state:
## Parallel Batch now the next unit? If the next eligible unit is a
valid batch (disjoint Ownership:, per-task **Verify:**, no intra-batch
dependency, no active write claim on owned files), the default is to route to
mpi-execute-parallel, not to implement it sequentially here.If the plan is stale, edit the plan before implementation:
## Plan Drift.## Current State.## Completed or remove/rewrite remaining items.If the next eligible unit is a valid ## Parallel Batch, make the parallel path
the default. Instead of a sequential continue brief, tell the user:
Next unit is a parallel batch: "<batch title>". Default path is to run it through
$mpi-execute-parallel in Codex or /mpi-kanban:mpi-execute-parallel in Claude Code.
Say "go parallel" to route there, or "sequential" to implement it one task at a time here.
Route to mpi-execute-parallel on confirmation. mpi-continue never spawns
implementation workers itself. Fall back to a sequential continue brief only if
the user explicitly chooses sequential, or the batch fails an eligibility gate.
Before implementation, output a brief and stop:
## Continue Brief: <next action>
**Source:** <plan path and handoff path if any>
**Project mode:** <profile mode, or "no profile">
**Current state:** <1-3 bullets>
**Conventions in play:** <1-3 bullets from matched topic block, or "none">
**Plan drift:** <none or summary of plan edits made/proposed>
**Files likely touched:** <files/modules>
**Coordination:** <active claims, pending file states, or none>
**Approach:** <2-4 sentences>
**Risk:** Low | Medium | High
**Verify after:** <specific check>
Reply "go" (or "ok", "yes", "proceed") to start implementation.
Do not implement before the user approves.
After approval:
After implementation, output:
Continue step complete.
**Files changed:** <list>
**Key changes:** <summary>
**Plan updates:** <summary or none>
**Verify:**
<exact verification steps>
**Option 1 - Verified** - say "1" or "verified"
**Option 2 - Keep talking** - say "2" or "keep talking"
Stop and wait.
## Completed.## Current State.## Remaining Work accurate.complete, needs_review, needs_verification, needs_integration,
verified, or released as appropriate.Step verified. Say "continue" to keep going, "handoff" to switch sessions, or "end session" to close.
Plan complete. Suggested next step: run $mpi-end-session in Codex or /mpi-kanban:mpi-end-session in Claude Code to preserve docs/rules/memory, commit, and close the kanban entry.
Do nothing else. Stay in conversation and append once:
Context getting large? Run $mpi-handoff in Codex or /mpi-kanban:mpi-handoff in Claude Code before starting a new session.
mpi-end-session's responsibility.## Parallel Batch, default to routing it to $mpi-execute-parallel in
Codex or /mpi-kanban:mpi-execute-parallel in Claude Code; that skill is the
only worker-spawning implementation path.Guides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.
npx claudepluginhub madponyinteractive/mpi-kanban --plugin mpi-kanban