Apply this skill when code, tests, docs, or reports create, change, review, or claim safety for process-crash, power-loss, forced-termination, disk-full, short-write, flush, atomic-replace, multi-file publication, startup recovery, local journal, resumable upload or download, stale-worker fencing, or fault-injection behavior where persisted state must remain a complete old or complete new generation and incomplete work must resume, roll back, quarantine, or reconcile safely.
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Apply this skill when code, tests, docs, or reports create, change, review, or claim safety for process-crash, power-loss, forced-termination, disk-full, short-write, flush, atomic-replace, multi-file publication, startup recovery, local journal, resumable upload or download, stale-worker fencing, or fault-injection behavior where persisted state must remain a complete old or complete new generation and incomplete work must resume, roll back, quarantine, or reconcile safely.
Keep persisted work truthful and recoverable when a process, machine, network path, or storage layer
stops between any two durable steps. Prove that readers observe a complete previous generation or a
complete new generation, never a half-published mixture, and that recovery itself remains safe when
interrupted repeatedly.
Use When
Code writes or replaces durable files, local state, manifests, indexes, checkpoints, journals,
generation pointers, upload sessions, download state, or multi-file snapshots.
A process crash, forced termination, power loss, disk-full condition, partial write, flush error,
stalled I/O, restart, failover, or stale worker can interrupt persistence or publication.
Startup code discovers unfinished work, temporary files, stale RUNNING jobs, incomplete
generations, or ambiguous external effects and must resume, roll back, quarantine, or reconcile.
Upload or download work must resume without mixing chunks, versions, owners, or final objects.
A test or report claims crash safety, power-loss safety, durable completion, atomic publication,
restart safety, resumability, or recovery correctness.
Do Not Use When
The main concern is path traversal, symlink, reparse point, permission, or platform filename
safety; use cross-platform-filesystem-safety first and this skill for crash durability only.
The main concern is malicious upload content, parser isolation, download authorization, or signed
URL security; use file-upload-security-review first.
The work spans multiple business participants, callbacks, timers, compensations, or deployments;
use durable-workflow-orchestration first and this skill only for its local persistence boundary.
The main risk is duplicate logical effects, stale request replay, queue delivery, or provider
uncertainty; use idempotency-integrity-review, dual-write-consistency, or
queue-processing-integrity-review first.
The task only packages or checks an already-generated artifact; use artifact-integrity-check.
Commit-boundary ledger: intent record, data write, flush, close, rename or replace, directory-entry
persistence, pointer publication, database commit, external effect, acknowledgement, and success
state in their actual order.
Platform and storage ledger: operating system, local or network filesystem, volume boundaries,
runtime APIs, flush and replace guarantees, object-store semantics, database guarantees, and any
best-effort fallback.
Recovery-state table for each incomplete form: detectability, authority, resume, rollback,
quarantine, reconciliation, user-confirmation, retention, and cleanup decision.
Ownership ledger: operation ID, attempt ID, state version, generation, lease, fencing token,
current owner, stale-owner rejection, and conditional write result.
Transfer ledger when applicable: session ID, owner, immutable source or target version, expected
size, chunk geometry, received ranges, per-chunk digest, whole-object digest, expiry, finalize
receipt, destination publication, and cleanup policy.
Fault-injection ledger: kill points, one-way network failures, duplicate requests, hung I/O,
short writes, flush or close errors, quota or inode exhaustion, read-only transitions, repeated
crash during recovery, invariant oracle, and liveness bound.
Relevant command-intent entries for implementation, tests, docs, package, and mustflow checks.
Preconditions
The task matches the Use When conditions and does not match the Do Not Use When exclusions.
The authoritative old state, intended new state, visibility boundary, and acceptable incomplete
states are named before selecting an atomic-write or recovery technique.
Platform guarantees are verified from the actual runtime and storage target or the claim is
explicitly downgraded to best-effort.
Existing filesystem, journal, transaction, state-machine, idempotency, generation, upload, and
recovery helpers have been inspected before adding a new mechanism.
Allowed Edits
Add or tighten explicit operation states, intent journals, same-volume staging, durable flush and
replacement, generation publication, checksums, startup recovery, resumable transfer sessions,
fencing, reconciliation, cleanup, diagnostics, and deterministic fault-injection tests.
Use the database's transaction log for database-owned state. Add an application journal only for
state outside that transaction boundary or for a separately owned publication protocol.
Preserve incomplete artifacts until their recovery disposition is known. Cleanup must not erase
the only evidence needed to resume, roll back, reconcile, or diagnose.
Do not claim write, close, function return, HTTP success, rename, queue ack, or process survival
alone proves durable business completion.
Do not depend on graceful shutdown, finally, exit hooks, sleep, one successful restart, or a
process-local flag as the recovery boundary.
Procedure
Define safety and liveness separately.
Safety: after any interruption, observers see one valid old generation or one valid new
generation; no partial mixture, duplicate irreversible effect, or success without result exists.
Liveness: unfinished work reaches a declared terminal, retry, quarantine, or operator state
within a bounded policy after recovery.
Draw the durable commit sequence. Separate function return, memory mutation, buffered write,
durable data, durable namespace publication, external effect, acknowledgement, and visible
success. Mark every gap where interruption changes the surviving truth.
Model work with monotonic states and conditional transitions. Use states such as PENDING,
RUNNING, COMMITTING, SUCCEEDED, FAILED, CANCELLED, UNKNOWN, or local equivalents only
when each has a durable invariant. Allow SUCCEEDED only when required results and receipts are
durably queryable, and reject stale state, version, attempt, generation, or fence writes.
Record intent before non-transactional mutation. Persist operation ID, target, expected prior
version, intended change, expected result, and recovery policy before the side effect. Persist
completion only after the new state is durable. Do not duplicate a database transaction log with
an application journal for state already committed atomically by that database.
Publish one file through a platform-appropriate atomic-write protocol. Prefer an unpredictable
same-directory temporary file on the same volume, complete writes including short-write handling,
file-data flush, close-error handling, atomic rename or replace, and parent-directory persistence
where supported. Preserve and report weaker Windows, network-filesystem, overlay, object-store, or
runtime guarantees instead of borrowing POSIX claims.
Publish related files as one generation. Build a new immutable generation away from readers,
verify required files, sizes, digests, references, and schema compatibility, then atomically move
a small current-generation pointer. Readers must pin one generation for the whole read. Leave
incomplete generations invisible and clean them only after recovery and retention rules permit.
Coordinate database and external effects with their owning protocols. Persist outbox or effect
intent with the local transaction, reuse one idempotency identity at the provider, retain the
response receipt, and reconcile UNKNOWN outcomes before retry. Do not pretend two independent
systems share a crash-atomic commit.
Fence stale owners. A lease timeout does not kill the old worker. Issue a monotonically increasing
generation or fencing token and require the authoritative store or publication step to reject an
older token. Treat a zero-row conditional update or rejected pointer publication as lost authority,
not success.
Make startup recovery a first-class phase before normal mutation. Inventory incomplete journals,
temporary files, uncommitted generations, stale running jobs, ambiguous receipts, and transfer
sessions. Apply an explicit state table that selects resume, rollback, quarantine, reconciliation,
user confirmation, or safe cleanup. Bound recovery work and keep degraded or read-only startup
available when policy permits.
Design resumable transfers as staged publication.
Create a server-owned transfer session with owner, immutable file version, expected size,
chunk geometry, expiry, and destination identity; never expose in-progress bytes as final.
Bind each range or chunk number to length and digest. Accept identical replay, reject conflicting
replay, and acknowledge only the durability level promised by the protocol.
Resume from authoritative server state, not only a client bitmap. Pin the download source
version so resumed ranges cannot combine different files.
Make finalize idempotent: verify ranges, total size, chunk digests, whole-object digest, and
destination ownership before one final publication; replay returns the same receipt.
Keep transport checksums distinct from a whole-file content digest. Do not treat multipart
provider ETags as universal content hashes.
Design cleanup as a recoverable operation. Bind temporary objects to owner, operation,
generation, creation state, and retention. Never delete a path solely because its name looks
temporary. Cleanup must tolerate duplicate execution and a crash midway through cleanup.
Add fault-injection points at every durable boundary: before and after intent persistence, data
write, flush, close, rename, directory persistence, pointer publication, external call, local
commit, acknowledgement, finalize, and cleanup.
Exercise asymmetric and storage failures. Cover request delivered with response lost, read-only
or write-only network failure, duplicate and delayed packets, hung I/O, disk bytes exhausted,
metadata or inode exhaustion where applicable, quota failure, short writes, flush failure, close
failure, read-only filesystem transition, and destination-volume mismatch.
Crash recovery itself repeatedly. Interrupt journal replay, generation selection, transfer
finalize, outbox resend, reconciliation, and cleanup. Recovery must be idempotent and monotonic;
repeated restart must not grant, charge, publish, merge, or delete again.
Verify business invariants with an independent oracle. Check result existence, digest, counts,
money or credit totals, unique effects, queue state, state-result consistency, old-or-new
generation visibility, and stale-owner rejection. A restarted process or green health endpoint
is not a recovery proof.
Verify liveness separately. Ensure recoverable work is eventually retried or surfaced, poison or
corrupt work is quarantined, and no job, lease, temporary object, or transfer session remains
permanently stuck without an operator-visible owner and deadline.
Run the narrowest configured verification covering the changed persistence, platform, transfer,
workflow, fault-injection, docs, package, and mustflow surfaces.
Postconditions
Durable completion, visibility, acknowledgement, and business success are separate observable
boundaries with no unsupported equivalence.
Single-file and multi-file publication leave a complete old or complete new state after every
tested interruption.
Startup recovery has an explicit disposition for every detectable incomplete state and remains
safe under repeated interruption.
Resumable transfers bind one owner, session, immutable version, range map, digests, final object,
and idempotent finalize receipt.
Stale workers, ambiguous effects, unsafe cleanup, and false success are fenced, reconciled,
quarantined, or reported with a bounded owner.
Safety and liveness claims are backed by named invariants and fault-injection evidence or clearly
marked as unverified.
Verification
Use configured oneshot command intents when available:
changes_status
changes_diff_summary
lint
build
test_related
test
test_audit
docs_validate_fast
test_release
mustflow_check
Prefer deterministic crash points, fake storage faults, disposable filesystems, controlled network
proxies, and invariant oracles exposed by configured intents. Do not infer raw process killers,
virtual machines, containers, privileged disk tools, live provider calls, or destructive fault
commands outside the command contract.
Failure Handling
If the authoritative state, visibility point, or durability guarantee is unknown, stop the strong
claim and report the exact platform or storage evidence needed.
If recovery cannot distinguish not-started, committed, and unknown effects, disable automatic
replay for irreversible work and require reconciliation or operator review.
If atomic replace, durable directory publication, conditional state update, or fencing is
unavailable, narrow the design to immutable outputs plus a verified pointer or downgrade the
guarantee explicitly.
If fault injection can only prove process restart, report missing storage, network, invariant, and
repeated-recovery evidence.
If cleanup would destroy the only recovery evidence, defer cleanup and surface the retained state.
If a configured check fails, preserve the failing boundary and use failure-triage before
broadening the change.
Output Format
Persistence surface, authoritative state, platform, and durability claim
