Lark Base Replica

Build an evidence-backed replica without claiming access to invisible platform objects. Use the run ledger as coordination evidence; live Feishu remains the source of truth.

Prerequisites

1. Read ../lark-shared/SKILL.md and ../lark-base/SKILL.md completely.
2. Read each lark-base command reference before invoking that command.
3. Read references/capability-ladder.md before selecting the extraction path.
4. Read the Browser control Skill and references/feishu-cdp-attachments.md only when visible attachments cannot be recovered through the official path.
5. Read references/acceptance-contract.md before planning or reporting, and references/release-gates.md when changing this Skill.

Non-negotiable boundaries

• Operate only on content the current signed-in user can normally view. Keep the source Base read-only.
• Never read browser cookies, local storage, profiles, password stores, tokens, or cryptographic material.
• Obtain authorization naming the exact target Base and write scope immediately before any external write. Prior read/extraction approval is not write approval.
• Do not promise 100%. Claim only the domains with zero verified mismatches; report excluded or inaccessible objects separately.
• The supported contract covers tables, semantic field definitions, supported views, supported ordinary cells, and visible attachment content. Workflows, dashboards, forms, advanced permissions, comments, sharing, formulas/lookup behavior, and original attachment binaries require separate proof.
• The CDP fallback currently accepts only decrypted PNG, WebP, and JPEG display derivatives. Recover other attachment types through the official media path or report them as unverified.
• Keep manifests, mappings, and ledgers at mode 0600. Delete signed URLs, AES material, and decrypted cache only after successful live readback and explicit cleanup confirmation.

Resumable run

Use one private run directory per migration. scripts/replica.py enforces:

NEW → PROBED → SNAPSHOTTED → PLANNED → STRUCTURE_WRITTEN
    → RECORDS_VERIFIED → ATTACHMENTS_VERIFIED → COMPLETE

1. Initialize and probe

Create the target Base only after the user authorizes that exact object. Then initialize the ledger and probe current CLI capabilities:

python3 <skill-dir>/scripts/replica.py init \
  --run-dir <private-run-dir> --source <source-base-token> --target <target-base-token>
python3 <skill-dir>/scripts/replica.py probe --run-dir <private-run-dir>

The attachment probe must explicitly find repeated --file append semantics for one attachment cell. If absent, attachment writes are blocked; do not guess or upload sequentially.

2. Capture stable manifests and plan

snapshot double-reads the source and stops if any semantic structure, record, or attachment metadata changes. It reads the target once.

python3 <skill-dir>/scripts/replica.py snapshot \
  --run-dir <private-run-dir> [--record-map <private-record-map.json>]
python3 <skill-dir>/scripts/replica.py plan --run-dir <private-run-dir>

The record map shape is documented in references/acceptance-contract.md. It normalizes source link-field table IDs and link-cell record IDs to their mapped target IDs. Never infer target IDs.

3. Write structure and ordinary data

Use the official lark-cli base +... shortcuts under the lark-base Skill. Exclude attachments and read-only/derived fields from ordinary writes. Batch at most 200 records and serialize writes within one table.

After each domain, recapture the live target with scripts/capture_manifest.py, then provide that manifest as evidence:

python3 <skill-dir>/scripts/replica.py apply \
  --run-dir <private-run-dir> --phase structure --evidence <target-after-structure.json> \
  --authorized-target <exact-target-base-token>
python3 <skill-dir>/scripts/replica.py apply \
  --run-dir <private-run-dir> --phase records --evidence <target-after-records.json> \
  --attachment-manifest <decrypted-dir/download-manifest.json> \
  --authorized-target <exact-target-base-token>

These checkpoints advance only when the relevant comparison has zero mismatches. The records checkpoint regenerates an idempotent attachment plan from live target state.

4. Recover and apply visible attachments

Use the official media path first. If it cannot recover an attachment that the user can visibly render, follow references/feishu-cdp-attachments.md. Decrypt mappings without retaining URL/key/nonce in the output manifest:

node <skill-dir>/scripts/decrypt_attachments.mjs \
  <private-cdn-map.json> <private-decrypted-dir> 8

Preview the generated attachment plan with scripts/upload_attachments.py --dry-run. Only after the user confirms the exact target and attachment write scope, run:

python3 <skill-dir>/scripts/replica.py apply \
  --run-dir <private-run-dir> --phase attachments \
  --evidence <target-after-records.json> \
  --authorized-target <exact-target-base-token> \
  --upload-root <private-run-dir/upload-ready> \
  --ephemeral-path <private-run-dir/private-cdn-map.json> \
  --ephemeral-path <private-run-dir/decrypted-dir>

The attachment manifest is merged by file_token; URL, AES, cookie, or access-token fields are rejected. Decrypted sources, registered ephemeral paths, and the upload root must be inside the run directory. The uploader verifies available SHA-256 metadata and rejects stale staging content. On the first ambiguous or retryable failure, stop. Re-read the target and retry the same attachment phase with that fresh target manifest as --evidence; the runner regenerates the plan and applies only missing files.

5. Verify and clean up

verify re-reads the live target; cached target data is never acceptance evidence.

python3 <skill-dir>/scripts/replica.py verify --run-dir <private-run-dir>
python3 <skill-dir>/scripts/replica.py cleanup --run-dir <private-run-dir> --confirmed

Cleanup is allowed only after COMPLETE and deletes only ephemeral paths explicitly registered inside the run directory. Preserve the final report and ledger.

Failure rules

• Permission error or invisible content: stop that domain; do not switch identity or retry around the denial.
• Source double-read mismatch: stop and recapture; never merge two revisions silently.
• Unknown target attachment or extra target record: stop and preserve the possible user change.
• Missing multi-file append contract: block attachment writes and request approval before any CLI update.
• Final mismatch: report partial or blocked, with counts and the exact unverified domains.