artifact-resurfacing

Artifact Resurfacing

Open the drift on the table. Surface the buried artifacts — stale citations, missing referenced files, orphan plans, cross-scope memory entries pointing at paths that moved. Polish them in proposal form. Codify the lesson so the next session does not have to re-discover the same drift.

Authority boundary (read first)

This skill is propose-not-apply for any constitutional surface. It produces drift tables, edit diffs, and IRF row proposals. It does not commit edits to CLAUDE.md, MEMORY.md, governance-rules.json, registry-v2.json, or any seed.yaml without explicit conductor authorization in the same session.

• Closure for plans — defer to closeout. This skill does not assign DONE-NNN / IRF-XXX-NNN labels; it surfaces orphans for closeout's classifier.
• Memory operations — defer to consolidate-memory for memory file mutations. This skill emits proposed memory diffs.
• Cross-session continuity — defer to cross-agent-handoff. This skill writes the polish-log entry; handoff carries it forward.
• Verification of claimed transitions — defer to qa-audit.

Discovery ≠ remediation. The skill finds and proposes. The conductor applies.

When to use

• "Resurface buried artifacts in "
• "Find stale citations in this CLAUDE.md"
• "Reconcile memory drift between scopes"
• "Audit this domain for orphans"
• "Polish referenced-but-missing files"
• "Future-fix this so it doesn't drift again"
• Closeout has surfaced ≥3 follow-ups whose root cause is drift between memory and disk
• A MEMORY.md walkback flags ≥1 stale path citation
• A find for a CLAUDE.md-cited path returns zero results

Four-phase protocol

Phase 1 — Detect

Enumerate the domain's surface. For a target CLAUDE.md or MEMORY.md, extract every:

• File path (absolute or ~/-anchored)
• Function/script identifier (foo.py, bar(), --flag)
• Issue/ledger ID (IRF-XXX-NNN, DONE-NNN, GH#123)
• External reference (URL, package name, doc anchor)
• Plan filename (any ~/.claude/plans/*.md referenced by path)

Run the citation audit:

bash scripts/audit-citations.sh <path-to-CLAUDE.md>

The script emits one JSONL record per citation: {cited, kind, exists, found_at, status}. A path is present if test -f (or test -d) succeeds. It is stale if a different path on disk holds the same canonical content (e.g., the institution moved from ~/Workspace/... to ~/Code/...). It is missing if no candidate path holds the artifact.

Run the orphan-plan scan:

bash scripts/find-orphan-plans.sh [glob]

Default glob is ~/.claude/plans/*.md. Output: one JSONL record per plan, with {path, mtime, has_done_ref, has_irf_ref, has_delivered_research_marker}. Plans lacking all three markers are orphan candidates.

Surface a four-column finding table to the user before classifying:

artifact	cited-at	status	candidate-action
...	...	present | stale | missing | orphan	(proposed action, no execution yet)

Phase 2 — Classify

For each finding, assign one class:

• Present — no action.
• Stale-citation — the artifact exists on disk; the citation points at the wrong path. Polishable via citation rewrite.
• Missing-but-probably-never-written — citation appears to run ahead of an unwritten artifact (e.g., a paper title cited with no draft, no commit history, no cross-scope mention). Polishable via citation annotation ((planned — not yet written, YYYY-MM-DD)) rather than reconstruction.
• Missing-but-lost — citation points at an artifact whose creation history exists somewhere (a session transcript, an old commit, a cross-scope memory file) but the artifact itself is gone. Polishable via deep search; if search fails, propose reconstruction-from-context or citation removal.
• Orphan-plan — pass through to closeout's classifier (EXECUTED / IN-PROGRESS / ABANDONED / DELIVERED-RESEARCH). This skill does not assign closure labels.

For ambiguous classifications, surface the finding to the conductor and stop. Universal Rule #21 ("Do what is asked — never preempt") applies.

See references/buried-bodies-taxonomy.md for the full taxonomy with worked examples.

Phase 3 — Polish (propose, do not apply)

For each classified finding, generate a polish artifact. Never write directly to constitutional files. Emit diffs and proposals instead.

For stale citations, run the proposer:

python3 scripts/propose-citation-fix.py \
  --file <path-to-CLAUDE.md-or-memory-file> \
  --stale-path <wrong-path> \
  --canonical-path <correct-path>

Output: a unified diff to stdout. The conductor reviews and applies (or rejects) the diff via Edit. The script never writes the file itself.

For missing-probably-never-written, emit a one-line annotation diff:

- See ~/path/to/cited-but-missing.md for details.
+ See ~/path/to/cited-but-missing.md for details (planned — not yet written as of YYYY-MM-DD).

For missing-but-lost, emit the deep-search command set first:

find /Users/4jp -name "<artifact-name>*" 2>/dev/null | grep -v node_modules
grep -rl "<artifact-keyword>" ~/.claude/projects/*/memory/*.md 2>/dev/null
grep -rl "<artifact-keyword>" ~/Code/organvm/praxis-perpetua/prompt-corpus/ 2>/dev/null

If the search returns hits, the finding reclassifies as stale-citation and the citation-fix proposer runs. If the search returns zero hits across the deep corpus, the finding reclassifies as probably-never-written and the annotation diff applies.

For orphan plans, hand the plan path to closeout with a classification request. Do not move plans to abandoned/ from this skill — closeout owns that decision.

See references/constitutional-doc-policy.md for the full rule set on what may/may not be auto-edited.

Phase 4 — Codify (future-fix)

Polishing one drift instance is tactical. Codifying so the next drift self-corrects is strategic.

For each polished finding, append an entry to a polish-log.md in the affected repo's root (create if absent):

## YYYY-MM-DD — <one-line summary>

- **Artifact**: <cited-path-or-id>
- **Class**: stale-citation | missing-never-written | missing-lost | orphan-plan
- **Finding**: <what was wrong>
- **Action**: <what was proposed; "applied" or "deferred-to-conductor">
- **Authorization**: <session-id-and-approver, or "pending">
- **Codification**: <what's been added to prevent recurrence, or "none yet">

Then check for compound patterns:

• Same stale citation in ≥2 scopes — propose updating the workspace CLAUDE.md autogen footer (under <!-- ORGANVM:AUTO:START --> sentinels) to assert the canonical path. The next organvm refresh will re-write it, making the truth periodic rather than ad-hoc.
• Same orphan-plan class in ≥5 plans — propose adding a detector clause to closeout's SKILL.md so the class is auto-recognized in future sessions.
• Cross-scope memory drift — propose a reference_<topic>_canonical_path.md memory at the workspace scope so all sibling scopes can grep-find one truth.

Emit one IRF row proposal per unresolved finding (Universal Rule #1: N/A = vacuum). Format:

- id: IRF-<DOMAIN>-<NNN>  # next available
  title: "<artifact>: <class> drift surfaced in session <id>"
  status: open
  domain: <domain-code>
  surfaced_by: artifact-resurfacing
  resurfacing_session: <session-id-or-date>
  action_required: <one line>
  authorization_required_from: conductor

The IRF row is a proposal. Writing it into INST-INDEX-RERUM-FACIENDARUM.md requires conductor authorization (Universal Rule #21).

What this skill explicitly does NOT do

• Edit CLAUDE.md, MEMORY.md, governance-rules.json, registry-v2.json, or any seed.yaml without same-session conductor authorization
• Assign DONE-NNN or IRF-XXX-NNN closure labels (that's closeout's job)
• Move plans to ~/.claude/plans/abandoned/ (also closeout's job)
• Auto-reconstruct missing artifacts from context (proposes reconstruction; conductor authorizes)
• Run as a background daemon, cron, or LaunchAgent (Universal Rule #9 — HARD RULE)
• Walk a domain it was not asked to walk (scope visibility principle, memory rule #39)
• Batch-close orphan plans (memory rule #53)

Scripts

• scripts/audit-citations.sh — citation audit; emits {cited, kind, exists, found_at, status} JSONL
• scripts/find-orphan-plans.sh — orphan-plan scan; emits {path, mtime, has_done_ref, has_irf_ref, has_delivered_research_marker} JSONL
• scripts/propose-citation-fix.py — propose-only unified diff for stale citation rewrites

References

• references/buried-bodies-taxonomy.md — the four classes with worked examples from this skill's genesis session
• references/composition-with-closeout.md — how this skill chains with /closeout (closeout discovers; this polishes)
• references/constitutional-doc-policy.md — the propose-not-apply rule set

Examples

• examples/2026-05-17-praxis-perpetua-resurfacing.md — the genesis case: praxis-perpetua moved from ~/Workspace/meta-organvm/ to ~/Code/organvm/, leaving stale citations in pipeline CLAUDE.md + memory and three referenced-but-missing 2026-03-15 papers. The session that built this skill is the case study.

Anti-patterns

• Treating Phase 1 detection as license to Phase 3 polish without conductor sign-off
• Editing the pipeline CLAUDE.md "Academic & Institutional Context" block in the same session that discovered it was stale (discovery ≠ remediation — surface, do not silently fix)
• Inventing the canonical path when ≥2 candidates exist on disk (ask the conductor which is truth)
• Treating cross-scope memory entries as fungible — each scope has its own MEMORY.md, each is independently mutable, all need their own proposed-diff
• Running on a domain the conductor did not name — buried-body hunting is invited, not autonomous
• Calling this skill's findings "fixed" when only the propose-diff has been emitted

Universal rules this skill is built around

• Universal Rule #1 — N/A is a vacuum. Every unresolved finding becomes an IRF row proposal.
• Universal Rule #2 — Nothing local only. The skill source pushes to remote; polish-log entries commit.
• Universal Rule #6 — Fix bases, not outputs. Phase 4's "codify" leg adds the prevention to the source (autogen sentinel, closeout SKILL.md, canonical-path memory), not to the symptom.
• Universal Rule #8 — Plans are versioned, never overwritten. Orphan-plan reclassification emits -v2 revisions when polish requires content change.
• Universal Rule #9 — No LaunchAgents. Audit is on-demand only.
• Universal Rule #12 — Memory is hypothesis. Every claim re-verified against current disk state before being acted on. This skill's Phase 1 is the operationalization of #12.

Why this exists

Drift between memory claims, CLAUDE.md citations, and on-disk reality is the entropy of a multi-session, multi-scope, multi-repo workflow. Sessions name artifacts; artifacts move; sessions end; memory persists the old name; the next session re-discovers the same drift. Without a ritual, drift compounds until the system's self-description no longer matches the system.

The genesis session (2026-05-17) found four classes of drift in one domain (the institutional-authority / SGO / praxis-perpetua surface). Closeout caught the four as "follow-ups"; this skill exists to give those follow-ups a protocol instead of an ad-hoc fix-or-defer decision.

Pairs with closeout (which surfaces) and consolidate-memory (which prunes). Three skills, one entropy regime.

Related

• ~/.claude/plans/where-on-my-local-enchanted-meerkat.md — genesis plan, Phase 2 section
• ~/.claude/plans/closeout-2026-05-17-plugin-surface-reconciliation.md — genesis closeout
• ~/.claude/plans/2026-05-17-handoff-plugin-surface-reconciliation.md — genesis handoff
• references/buried-bodies-taxonomy.md, references/composition-with-closeout.md, references/constitutional-doc-policy.md