| name | cairn-resync |
| description | Operator-initiated re-discovery — resolve config drift, re-cluster topics, re-curate grown areas into DEC/INV drafts. |
| when_to_use | Use when the operator asks to "resync", "re-discover", "re-curate", or
catch Cairn's ground state up to a grown/restructured tree — typically
after the config-drift baseline nudge ("your project grew"). Three
opt-in passes: deterministic config-resync, LLM topic re-cluster, and
subagent re-curation of a new area into DEC/INV drafts. Skip during an
active task, or when the operator only wants the one-tap config fix
(that's `cairn_resync` direct, no skill needed).
|
| allowed-tools | Skill(cairn:cairn-attention), Task(curator-map), Task(curator-reduce), Bash, Read |
Skill: cairn-resync
Host portability
This skill is shared by Claude Code, Cursor, and Codex. Resolve Cairn MCP
tools with the host's native discovery mechanism and use its structured
question UI, falling back to a concise A/B/C chat question. Named curator
briefs live under ../../agents/; dispatch them natively after reading the
brief, or execute inline only when the host has no subagent capability.
You are running Cairn's operator-initiated re-discovery so committed
ground state catches up to a tree that has grown or been restructured
since adoption.
Everything here is opt-in, recoverable, and never auto-applies to
committed ground: deterministic edits are review-class (the operator
commits the diff); LLM passes spend the fast tier only on genuinely-new
prose; re-curation produces _inbox/ drafts the operator drains. Confirm
scope with the operator before any pass that spends model quota (re-cluster,
re-curate).
Pass 1 — config-resync (deterministic, free)
Resolve the config-drift the sensor surfaced.
- Preview:
cairn_resync({}) (dry-run default). Summarize the proposed
.cairn/config.yaml edits — add a grown dir to componentDirs, add a
new file type to extensions, add a .gitignore entry to
off_limits, drop a dead componentDir, re-point a moved entity's
source_file.
- On the operator's OK:
cairn_resync({ apply: true }). It archives the
pre-resync config.yaml to .cairn/ground/.archive/ and is
idempotent. Tell the operator to review + commit the diff.
- Pass
area to scope either call to one subtree.
Pass 2 — topic re-cluster (LLM, opt-in)
Offer this when prose has moved/grown enough that the topic-index is
stale. Confirm first — it spends the fast tier (only on new prose; unchanged
pairs hit the cache).
- Preview:
cairn_resync({ recluster: true }) — re-walks + judges and
reports topics_before → topics_after + judge_calls_fresh, but
overwrites no map.
- On OK:
cairn_resync({ recluster: true, apply: true }) — archives the
prior topic-index.yaml + anchor-map.yaml, then rebuilds them. The
maps are gitignored, per-clone derived state, so there's nothing to
commit.
Pass 3 — re-curation of a new area (subagents, opt-in)
Re-run the curator over a grown area and surface the result as DEC/INV
drafts. Always scope to an area — a full re-curate is the adoption
pipeline's job, not this. Confirm the area + the model spend first.
Ghost mode: if the repo is registered ghost, the curator dir is
out-of-repo. The recurate:'walk' response returns the absolute
curator_dir; give the subagents absolute paths under it (mirror
cairn-adopt Step 3.5's path-resolution note).
3.1 — walk the area
Call cairn_resync({ recurate: "walk", area: "<dir>" }). It builds the
curator corpus + shards scoped to that subtree and returns shards_path,
curator_dir, records_total, and shards. If shards === 0, there's
no curatable prose in the area — tell the operator and stop (skip emit).
3.2 — dispatch curator-map (parallel rounds of 4)
Read the shard plan with node (portable — never cat/jq):
node -e "console.log(require('node:fs').readFileSync(process.argv[1],'utf8'))" <shards_path>
For each shard, compose a curator-map Task brief with shard_id, the
absolute shard_path (<curator_dir>/shards/<shard_file>), the absolute
candidates_path (<curator_dir>/candidates/<shard_id>.jsonl), module,
and project_domain. (Post-adoption the mapper's key_modules summaries
may be gone — pass an empty module_flags and a short generic
module_summary; the curation degrades gracefully.) Send up to four
briefs per assistant message so they run in parallel; await each round
before the next. Subagents write JSONL to disk and return a short
summary — read disk, not the return text.
3.3 — dispatch curator-reduce
Once every candidates/<shard_id>.jsonl exists, spawn one
curator-reduce subagent with the candidates_glob, the absolute
final_path (<curator_dir>/final.jsonl), and project_domain. It
synthesizes the de-duplicated DEC/INV set into final.jsonl.
3.4 — emit drafts
Call cairn_resync({ recurate: "emit" }). It validates each final.jsonl
entry and writes it as a reviewable draft — DEC →
.cairn/ground/decisions/_inbox/<id>.draft.md, INV →
.cairn/ground/invariants/_inbox/<id>.draft.md, each status: draft. It
rebuilds no ledger and graduates nothing. Report the dec_drafts +
inv_drafts counts.
3.5 — hand off to cairn-attention
Invoke Skill(cairn:cairn-attention) so the operator drains the fresh
drafts inline: accept graduates a DEC to the ledger / an INV to active;
reject archives a .rejected.md tombstone; edit returns the body. Nothing
the curator proposed touches committed ground until the operator accepts.
