Progress guard

Execution memory and bounded recovery, not a planner, supervisor or tool watchdog. This skill grants no authority and installs no hooks, timers or background work.

Start small

Ordinary solo work stays direct; loading this skill never requires a child. Simple tasks need no ledger. For complex/stalled work, reuse the plan/todos and one existing snapshot; initialize from confirmed facts only if absent. Use storage.md for the store and optional fields. Do not delay useful work to perfect bookkeeping or create a competing plan.

If native discovery fails, read an accessible canonical SKILL.md once and record the failure, fallback path/version and owning runtime. File loading is not native discovery. Reuse that route after compaction. Retry native discovery only after a concrete runtime/catalog change. Children verify their own loading, not the parent's. Never route around a policy denial; if no permitted source is available, report the blocker.

Evidence, not activity

Progress is a verified deliverable increment, reproducible failure or observation that changes the next action. Reference its source/context; distinguish delivery from diagnostic learning and check that learning still advances the outcome. Plans, retries, pings, launches, bookkeeping and unchanged status are not progress. A passing component or child milestone is not usable end-to-end delivery.

Record only meaningful boundaries

Append one event with a complete compact snapshot at meaningful decisions, changed knowledge/failure, blocks, delivery, pause/handoff and uncertain external operation boundaries. Reuse native history for routine calls; keep logs/artifacts outside the ledger and store pointers, not transcripts or secrets.

Before long/risky operations, preserve intent, context and recovery lookup; retain the returned handle. After interruption, UNKNOWN is not failure: reconcile actual effects before retrying. Ledger writes and external effects are not atomic. A successful append receipt is sufficient; no compulsory full readback after every append. Read when recovery or uncertainty requires it, not as a ritual.

Read before acting

Read current state after resume/compaction, before a retry, on changed inputs or material user corrections, and for final acceptance/handoff. Reuse already-current state within uninterrupted work. Check actual plan, role, ownership, decisions, failed hypotheses/retry conditions and unresolved handles. New user intent/evidence outranks stale state. Reuse accepted proof until affected inputs change or inconsistency appears; check the relevant canonical index once before declaring context missing. Do not hunt credentials or reread all artifacts.

Compaction and selective recovery

Follow compaction.md when context pressure warrants it. Preserve decisions, failures, active ownership and handles; restore only current state and evidence needed for the next decision. A ledger write does not compact history or control runtime thresholds. Do not repeat compaction without benefit.

Stall triggers

Review after two failed attempts of the same hypothesis without new evidence, repeated planning/reads without a changed decision, drift from the outcome, or a long operation with no observable advancement past its expected bound. Use about 15 active no-progress minutes to review and 30 to escalate. Exclude user/offline inactivity; disclose unknown timing. These are no-progress backstops, not task expiry, phase quotas or reasons to stop a healthy advancing operation. Never invent token, turn or task budgets; explicit limits and real deadlines bind.

Recovery: one short decision, then action or stop

Identify the blocking assumption. Choose one bounded discriminating test, a different authorized approach, the missing user decision, or a blocked handoff. State expected evidence and abandon condition; no replacement-agent/research loop. If recovery adds no useful evidence, stop that branch and ask; if no answer channel exists, record needs_decision. Silence, autopilot or "continue" cannot renew an exhausted attempt. Resume only on concrete changed conditions/authorized alternative, preserving failure history and cumulative declared limits, even across sessions.

An effect-free input/command correction may use the existing ONE bounded recovery inside its mandate without a new phase approval. A missing operation ID alone does not prove nothing happened. Reconcile UNKNOWN before replay. Never override explicit retry/spend/time limits, permissions or exhausted recovery. Escalate with the precise blocker, last verified result, attempted hypotheses, evidence, recommended next action and one needed decision; elapsed time only if known.

Tools and children

Match operation, target/increment and effect to the current mandate before asking approval. Covered work proceeds; a genuinely new effect needs a decision. See approval reconciliation at that boundary, not on every step. No duplicate phase approvals; silence is not consent and implementation does not imply merge/cloud/release.

Initial waits are not timeouts. Follow native waiting/notification rules; no duplicate calls or status polling merely to stay busy. A skill cannot interrupt an in-flight MCP call: report that limit at the next opportunity. Never bypass excluded tools or stop unknown/unowned operations; a timer does not authorize cancelling a cloud mutation.

For authorized delegation, load children.md, the sole detailed assignment/receipt protocol. Persist coordinator/executor role. A coordinator routes end-to-end assignments and decisions, accepts terminal evidence and maintains the consolidated ledger; it does not implement, build/test, deeply debug, integrate/merge or deploy. Route direct-work requests to a suitable authorized executor, not a new child per step. Explicit role changes reconcile live ownership and UNKNOWN first. Preserve one-process end-to-end constraints.

Children return one completed, failed or blocked result per assignment execution; ordinary recovery and intermediate findings stay local. Genuine block resolution and explicit resume permit a new terminal revision, not repeated unchanged reports. Safety incidents, urgent cancellation and required permission/input gates remain immediate exceptions; host-generated notifications cannot be suppressed. Serialize actual shared writes/dependencies/capacity, not all service use. UNKNOWN fences its effects and interfering work, not demonstrably independent authorized tasks.

Finish

Accept against the current user-approved outcome, not activity. Distinguish implemented, locally verified, deployed and verified usable results. Append completion evidence or a precise partial/blocked handoff with unresolved handles and next decision. Coordinators assess relevant final proof, not duplicate execution. Missing/inconsistent proof needs targeted clarification, not blind acceptance. The ledger preserves claims and provenance; shape checks do not prove truth, agent compliance or runtime behavior.