reorient

reorient

Restore the working contract mid-task. After a context compaction or a long run, an agent loses role identity, invents constraints that do not exist, skips required steps, and loses the task thread. reorient reloads the durable state that should have survived — the discipline contract first, the task pointer second.

Read-only and advisory. It changes nothing, produces no artifact, and emits no handoff — like next, the orientation block is the whole output; a handoff would only restate it. The third skill exempt from the agentic summary contract.

reorient is the mid-task, feature-scoped counterpart to next (project-scoped, session-start).

When to use

• Right after a context compaction (a SessionStart hook can auto-invoke it).
• Mid-task, when unsure of the role, the current checkpoint, or the next action.
• Returning to a feature after an interruption.

Not for: session-start "what should I pick up across the project" (next); loading a feature not yet started (prime-context); validating artifacts (consistency-check).

Workflow

Step 1 — Resolve and locate

Resolve the methodology root + manifest via ${CLAUDE_PLUGIN_ROOT}/scripts/dae_resolve.py (see references/resolving.md). Locate the feature (slug arg or branch name). If no feature is in scope, say so and suggest next instead.

Step 2 — Reload the discipline contract, then the task pointer

Read, read-only, in this order:

1. Role + autonomy — CHARTER.md + manifest: the autonomy level in force, and what the agent may and may not decide. Counters invented constraints.
2. Current checkpoint + exit criteria — the Checkpoint Exit Contract (Foundation Design Section 8) for the checkpoint progress.md shows in progress: its goal, its exit criteria, and which are already met. Run ${CLAUDE_PLUGIN_ROOT}/scripts/dae_progress.py <feature-dir> and show its breadcrumb — the same pipeline-position line the checkpoint skills surface at Step 0.
3. Non-negotiables — verification independence and charter-mandated mutation: steps that must not be skipped regardless of cost. The cost of a step is never a reason to skip it.
4. Current task + next action — the progress.md CURRENT header.
5. Feature contract — feature.md outcome + scope. Counters goal drift.

Also run ${CLAUDE_PLUGIN_ROOT}/scripts/dae_handoff.py <feature-dir> — report any checkpoint marked done without a complete handoff as a discipline gap.

Step 3 — Emit the orientation block

Output one tight block, nothing else:

You are <role> at autonomy <level>.
Feature NNN-slug — Checkpoint N (<goal>).
Exit criteria: <m>/<n> met — unmet: <list>.
Current task: <task> -> next action: <action>.
Must not skip: <non-negotiables>.
Constraints: <charter / autonomy limits>.

Step 4 — Stuck-loop detection

Before stopping, scan the feature's recent handoffs for a stuck loop. A loop is "stuck" when three or more consecutive handoffs from the same skill share:

• the same status (typically interrupted or complete-with-unmet-criteria), AND
• the same recommended-next, AND
• substantially the same findings_summary / failure signature (e.g. same error class, same blocker name, same infra failure type).

A stuck loop is the agent re-attempting the same operation expecting different results. nexthq saw 60+ identical ticks over 12 hours on a Functions emulator blocker — the autonomous loop never escalated to the human.

Threshold is tunable via manifest.autonomy.stuck_loop_threshold (default 3).

When detected: STOP. Do not silently re-run. Emit (just this once — break the no-handoff rule) a session-summary-style escalation handoff that:

• Names the repeated failure signature.
• Lists the N affected handoffs by filename.
• Sets human_action_needed: decision and human_action_kind: unblock-stuck-loop.
• recommended_next: "human review required — N consecutive identical failures from ".

The escalation handoff lives in .engineer/handoffs/ (it's project-scope, not feature-scope progress). The human picks the unblock — a different approach, a charter amendment, an environment fix, or a deferred re-try.

Step 5 — Stop

If Step 4 found no loop: reorient orients; it does not act. The human resumes the work. No handoff.

If Step 4 escalated: the escalation handoff is the output. The loop is broken — wait for the human.

Optional: auto-invoke on compaction

A project may add a SessionStart hook (source: compact) that nudges /engineer.reorient after every compaction. See ${CLAUDE_PLUGIN_ROOT}/examples/session-start-reorient.md. The hook is optional project config, not part of this skill.

References

• Foundation Design — the Checkpoint Exit Contract (Section 8); agentic summary exemptions (Section 5)
• Sister skill: next — the project-scoped, session-start counterpart