[MartinP7r]

Preview-only curation evidence

Records the first conservative Tome -> Paperclip agent assignment proposal for board review.

Proposal

• Set: tome-maintainer
• Candidate recipient: FoundingEngineer (ecadbdc0-fcf5-4d18-b62f-111999f269c0)
• Intended desired skills:
  • local/d64eaed4a6/using-tome (using-tome, company skill cfa45993-3618-49df-8d5d-4cc0cc28f7c4)
  • local/fbfe9eb2f4/rtk-command-output (rtk-command-output, company skill b01320ae-a9d0-4b4f-bc2a-11369709c574)
• Corrected-head live preview impact: add local/d64eaed4a6/using-tome; keep local/fbfe9eb2f4/rtk-command-output; remove none.

Live read-only evidence

• The preview used the real Paperclip API through company-skill catalog/detail GETs only.
• using-tome is present in the live company skill catalog as local/d64eaed4a6/using-tome.
• rtk-command-output is already a desired skill for FoundingEngineer and remains in the intended set.
• RTK provenance limitation: Paperclip detail exposes stable id/key/slug/name and attachment evidence, but source and provenance are null, so this PR does not claim a repo-owned source digest for RTK.
• Checked-in evidence: curation/paperclip-agents/MCO-163-preview.md, preview-output.txt, preview-current-state.snapshot.json, and live-skill-evidence.snapshot.json.

Safety

• No tome paperclip-agents apply, skills:sync, tome sync, model call, canonical-library mutation, runtime agent change, or Desktop/Tauri work occurred.
• Preview token state was isolated under scratch TOME_HOME.
• The preview boundary now fetches every listed company-skill detail and does not treat attachedAgentCount = 0 as authoritative emptiness.
• Regression coverage preserves a desired attachment when the list summary count is zero but the company-skill detail reports usedByAgents[].desired = true.

Corrected-head preview

• Head: c79720a5266bc5d941ac4743fc4842bfe8c37446
• Command: TOME_HOME=$PAPERCLIP_RUN_SCRATCH_DIR/mco-163-corrected-preview/tome-home cargo run -p tome -- paperclip-agents preview --catalog curation/paperclip-agents/catalog.toml --assignments curation/paperclip-agents/agents.toml --constraint paperclip-agent --paperclip-api-url "$PAPERCLIP_API_URL"
• Result: no Paperclip agent skills changed; add using-tome; keep rtk-command-output; remove none.
• Scratch-scoped confirmation token: apply-v2-2ecf1ce3fdc819538387b8ebc94aa8da.
• Token issued: 2026-10-05T11:55:10Z; expires: 2026-10-05T12:10:10Z. Because it was issued under scratch TOME_HOME, any later approved apply must first rerun preview in the intended persistent TOME_HOME and use that fresh matching token.

Verification

• cargo fmt -p tome -- --check
• cargo clippy -p tome --all-targets -- -D warnings
• cargo test -p tome paperclip_agents
• cargo test -p tome
• git -c core.fsmonitor=false diff --check HEAD~1 HEAD
• make ci reached the Rust gates and full tests, then stopped because local typos is not installed; PR CI is the source of truth for the Typos leg.

Paperclip: MCO-163

[MartinP7r]

Blocking review findings

PR #619 must be corrected before merge.

1. Never trust a zero summary count to establish an empty desired attachment set. In crates/tome/src/paperclip_agents.rs:908, attachedAgentCount > 0 skips the authoritative company-skill detail GET for listed skills whose count is zero. If that summary is stale/partial but detail includes usedByAgents[].desired = true, preview can silently omit an existing desired attachment and propose a destructive removal. Fetch every listed company-skill detail (preferred), or fail preview when a zero summary cannot prove emptiness. Add a regression where list count is zero but detail carries a desired attachment; resulting live state must preserve it.

2. Update the PR docs: record Paperclip agent curation preview #619 description to match the final evidence. It currently claims fallback-only preview, no live before/after, and absent using-tome. The current evidence documents a successful read-only live preview via company-skill attachment GETs, local/d64eaed4a6/using-tome / company skill cfa45993-3618-49df-8d5d-4cc0cc28f7c4, and local/fbfe9eb2f4/rtk-command-output preserved with no removals. Keep the explicit provenance limitation: RTK has no source/provenance field from Paperclip.

Preserve all accepted safety guarantees: GET-only preview; no apply/skills sync; isolated preview token state; rtk-command-output remains in the intended set; deterministic catalog/assignment outputs; no Desktop changes. Do not mark MCO-163 done until the corrected head has focused/full verification, green exact-head CI, independent review, and corrected PR metadata.