using-sync

Files (2)

• SKILL.md
• openai.yaml · agents/

SKILL.md (first lines)

# Using LWC Sync

Run Sync through the audited CLI only:

```text
lwc sync HOST [ABS_DIRECTORY] --mode merge|pull|push
```

Every start, resume, resolve, abort, pull, push, or merge can publish or change
durable data. Before the first execution in a bounded workflow, present a safety
notice containing the exact command, target host, absolute directory, scope, mode, affected stores, impact, risks, recovery path, and reversibility.
An explicit user instruction that already names or unambiguously accepts these facts is authorization; do not ask the human to repeat it.

One authorization covers the bounded workflow while its host, directory, scope,
mode, publication targets, and disclosed risks remain unchanged. It covers
read-only preflight, prerequisite build and installation, the initial Sync,
ordinary `--resume`, continuity or derived recovery, status checks, and an abort
before publication. Equivalent command details that do not widen the target or
risk do not require reconfirmation.

Reconfirm only when a host, directory, scope, mode, or publication target
changes; a new destructive, irreversible, privacy, or data-loss risk appears;
or a candidate resolution would discard one side. Preserve-both and idempotent
recovery within the same session remain covered. A Skill trigger alone is not
authorization.

Choose the scope explicitly before starting:

- `--scope project` synchronizes only the active project Wiki. Supply the remote project's absolute directory when it cannot be discovered from the remote login directory.
- `--scope global` synchronizes only the global Wiki; omit the directory.
- `--scope all` synchronizes project and global stores as one coordinated session. Supply the remote project's absolute directory.

Use `merge` to publish the reconciled result to both sides, `pull` to publish it only locally, and `push` to publish it only remotely. Every mode preserves unique semantic objects from both sides; direction controls the destination, not destructive authority. Scope and mode are session invariants: keep the exact original `--scope`, `HOST`, optional `ABS_DIRECTORY`, and `--mode` on every continuation.

## Continue or stop a session

- Resume with the original command plus `--resume SESSION_ID`.
- Abort without deleting audit state with the original command plus `--abort SESSION_ID`.
- Abort is available only before canonical publication. After a receipt reports
  `committed=true` or a store as published, follow its `next_action` or validated
  recovery command and resume; never restart the semantic publication.
- Treat lifecycle Hook Sync readiness as a bounded continuity cue. It never contains hosts, directories, credentials, object bodies, or conflict payloads; inspect the session through normal `lwc sync` output.

An initially missing destination is created only through Sync's staged,
validated publication path. For a single project or global scope, push from a
missing local source is an explicit `local_store_missing` no-op; pull from a
missing remote source preserves local state and does not create remote
canonical state. Pull or merge can safely create a missing local destination,
and push or merge can safely create a missing remote destination. `--scope all`
coordinates project and global units and stages every requested unit before it
publishes any of them.

Sync moves suspended sparse changesets as validated detached intent, then
replays each foreign intent as a fresh local suspended draft with a new
changeset ID; it never commits that draft into live knowledge automatically.
Queued and running Work remain machine-local. Terminal Work crosses only as a
bounded redacted origin audit, never as a raw Work directory, result, path, or
execution state. Inspect `continuity_local` / `continuity_remote` receipts. If
canonical publication succeeded but continuity did not, trust
`committed=true`, keep the same session, and follow
`next_action=resume_continuity`; resume is idempotent.

FTS is refreshed for changed semantic objects. Markdown and an already-enabled
document graph use the exact affected identifiers while the selection contains
at most 4,096 items and 256 KiB; larger selections keep bounded counts and a
digest and deliberately use `derived_selection=full`. An initialized CodeGraph
refreshes only after Git publication. A post-commit derived failure reports
`committed=true` and `next_action=resume_derived_rebuild`; resume repairs the
derived planes without replaying canonical publication.

## Resolve semantic conflicts

When Sync reports `action=conflicts`, treat the returned `conflicts` array as
the current batch. Each batch contains at most 20 conflict objects, and the
canonical resolution packet is limited to 256 KiB. Resolve ordinary content
conflicts autonomously from current source evidence, then write a packet and
resume with the original command plus `--resolve PACKET.json`:

```json
{"version":1,"decisions":[{"conflict_id":"0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef","kind":"page","logical_key":"guide","path":"title","candidate":1}]}
```

Copy `conflict_id`, `kind`, `logical_key`, and `path` from the current batch and
choose candidate `0` or `1` exactly as listed for every conflicting field. A
candidate decision has exactly these five fields:

```json
{"conflict_id":"CURRENT_64_HEX_ID","kind":"page","logical_key":"guide","path":"title","candidate":1}
```

If evidence cannot justify either candidate for an object, preserve it with
this exact four-field object decision:

```json
{"conflict_id":"CURRENT_64_HEX_ID","kind":"page","logical_key":"guide","strategy":"preserve_both"}
```

That object-level decision covers every conflicting field on the object. For
`--scope all`, wrap only the scopes represented in the current batch, for
example `{"version":1,"scopes":{"project":{"decisions":[...]}}}`.

The packet must be schema-valid and cover every conflict with either a field-level candidate or one object-level preserve-both decision. A merge must preserve both sides' non-conflicting unique objects.

- Use only decisions and object identifiers allowed by the packet schema; do not invent fields or paste object bodies into chat.
- A multi-field object legitimately repeats its `conflict_id` once per distinct
  field decision. Stale or unknown conflict IDs fail closed. Duplicate
  decisions for the same field, duplicate preserve-both decisions, mixed
  candidate/strategy shapes, unknown fields, more than 20 conflict IDs, and
  packets over 256 KiB also fail closed without publication.
- If neither candidate is justified, use normal LWC read commands to gather more evidence before choosing. If ambiguity remains, use the packet's deterministic preserve-both strategy so both valid values survive under stable variant identifiers; never guess from SQLite internals or ask a human to choose rows.
- Mark the session blocked only for a security, policy, authorization, artifact, or protocol failure that cannot be handled safely; semantic ambiguity uses deterministic preserve-both.
- Do not ask a human to inspect or resolve SQLite rows. Do not resolve conflicts by editing SQLite rows.

Resolve only one returned batch at a time. After each `--resolve`,
inspect `action`, `conflict_count`, `next_action`, and the newly returned
`conflicts`. If conflicts remain, use the exact original command with
`--resume SESSION_ID` to inspect the current status/batch when needed, build a
new packet from those current IDs, and resolve again. The existing bounded
authorization covers preserve-both and previously authorized candidate choices;

From JanYork/llm-wiki-cli, Apache-2.0.

More from JanYork/llm-wiki-cli