Mine Gemini workflows

Use session evidence to find workflows worth porting. This is workflow archaeology, not a bulk conversation export.

Non-negotiable privacy boundaries

• Require one explicitly selected Gemini project/session directory. Never crawl the entire home directory by default.
• Run the metadata-only pass first.
• Do not extract or summarize private thoughts, reasoning fields, tool arguments, secrets, or complete transcripts.
• Do not commit .context-os/migrations/ or raw Gemini recordings.
• Ask before any --include-content, --include-summaries, or --include-paths pass and explain exactly which selected sessions will be read.
• A repeated pattern is a candidate, not authorization to create or install a skill.

Procedure

1. Select evidence

Ask the user for the relevant Gemini project/session directory and optional date boundary. If they do not know the directory, help them locate candidate directories using names and modification dates only; do not read session bodies during discovery.

2. Create a metadata-only inventory

Run:

python3 scripts/mine-gemini-workflows.py \
  <selected-session-directory> \
  --output .context-os/migrations/<timestamp>/gemini-inventory.json

Add --since YYYY-MM-DD when the user supplied a date boundary. The default report includes tool names, validation status, file basenames, and session identifiers. It excludes message text, free-form workflow summaries, tool arguments, and full paths.

3. Rank candidates

Prioritize candidates that:

1. occur with positive validation in at least two sessions,
2. have successful validation evidence,
3. use a stable tool sequence,
4. solve a task the user expects to repeat.

Do not promote one-off activity or a repeated failure. Present the ranked candidates and evidence counts, then ask the user which ones to inspect.

4. Inspect only selected sessions

If metadata is insufficient, name the exact selected session IDs and ask permission to rerun with repeated --session-id <id> selectors plus only the required opt-in flag (--include-summaries, --include-paths, or --include-content). Content redaction is best-effort, not a guarantee; treat every opt-in report as sensitive. Thought/reasoning fields remain excluded. Review the output again before sharing or persisting it.

5. Draft, do not silently install

For each approved candidate:

• draft .agents/skills/<workflow>/SKILL.md as the portable core,
• add a thin .claude/commands/<workflow>.md adapter only if a Claude slash command is wanted,
• create a parity case from docs/templates/workflow-parity.json,
• record provenance using session IDs, immutable recording digests, and evidence counts, not transcript content,
• separate provider-neutral steps from Gemini-, Claude-, or Codex-specific tool adapters.

Show the draft and parity case before writing. Then run bash scripts/validate-all.sh --workspace.

6. Report limitations

State what the miner could not infer, including missing memory scratchpads, unsupported recording variants, unavailable tools, ambiguous validation, or workflows that need human judgment.

Prior art

See references/ai-data-extraction.md. The linked extractor informed the inventory-first approach, but is not a dependency and should not be run against current recordings without format and privacy review.