git:20260725.cd88de1 to git:20260725.9c1b406

24 added, 18 removed. Audit A to A.

---

name: minutes-graph

- description: Policy-safe relationship rankings, commitments, aliases, person profiles, and topic research. Always use Minutes' bounded native CLI surfaces; never build or read a durable graph cache.

+ description: Policy-safe relationship intelligence across normal meeting history. Use for top contacts, relationship summaries, commitments, losing-touch signals, person profiles, topic research, or questions about people across meetings. Always use Minutes' native CLI projections; never build or read a durable graph cache.

compatibility: opencode

---

# /minutes-graph

- Minutes builds relationship rankings and graph commitments from a bounded, process-private SQLite projection of stable policy-authorized Markdown. Person profiles and topic research use the separately bounded live-source search boundary. Both paths re-attest policy before returning facts. Do not fall back to a retired durable index or read meeting files directly.

+ Use Minutes' native, process-private relationship projection. It rebuilds from stable live Markdown, excludes restricted or policy-uncertain meetings, and revalidates the corpus before returning an answer. It does not retain a graph database or JSON index.

## Privacy boundary

- Never walk meeting files, parse frontmatter yourself, or read raw transcripts for this skill.

- Never run `graph_build.py` or read `~/.minutes/graph/index.json`; those are retired legacy surfaces.

- Never create a replacement graph cache, spreadsheet, JSON file, or database.

- Never pass `--include-restricted`. Restricted meetings are intentionally absent from this agent-facing skill.

- - Treat any authorization, resource-budget, correction-race, or projection error as a hard stop. Do not fall back to filesystem reads.

+ - Treat an unavailable or resource-budget error as a hard stop. Do not fall back to filesystem reads.

- ## Available commands

+ ## Choose the native command

- - `minutes people --json` — bounded relationship rankings and losing-touch signals.

- - `minutes commitments --json` — bounded graph commitments.

- - `minutes people merge <canonical> <alias...>` — confirm an identity correction in the local vocabulary; uncertain names are never merged automatically.

- - `minutes person "<name>"` — bounded person profile.

- - `minutes research "<topic>"` — bounded topic research.

+ | User intent | Command |

+ |---|---|

+ | Top contacts, relationship scores, losing-touch signals, top topics | `minutes people --json --limit 50` |

+ | One person's profile and recent history | `minutes person "<name>"` |

+ | Open or stale commitments | `minutes commitments --json` |

+ | Commitments for one person | `minutes commitments --person "<name>" --json` |

+ | Topic or question across normal meetings | `minutes research "<query>"` |

+ | Topic scoped to one attendee | `minutes research "<query>" --attendee "<name>"` |

+ Run the narrowest command that answers the question. Parse its stdout as the returned JSON where applicable; do not inspect stderr for meeting content.

## Workflow

- 1. Classify the request and use the narrowest command above.

- 2. Require exit status 0. Use only the bounded native result and never substitute filesystem reads.

- 3. For a proposed alias, show the suggestion and ask for confirmation before running `minutes people merge`; a wrong merge is worse than no merge.

- 4. Do not imply that restricted history or a relationship fact is absent when any command fails.

+ 1. Restate the scope in one short phrase when ambiguity matters, such as “normal meetings only.”

+ 2. Run one native command from the table.

+ 3. Rank or filter the returned objects in memory for this response only.

+ 4. Cite the meeting titles/dates supplied by the command. Do not open the source paths.

+ 5. If the native command cannot answer a cross-entity/co-occurrence question, say that the safe graph surface does not currently expose that relationship. Offer the nearest supported `people`, `person`, `commitments`, or `research` view; do not synthesize a durable index.

## Output

- Return only the bounded native result. Never invent rankings, commitments, or relationship signals from raw files.

+ Prefer a compact table for ranked people or commitments. Explain signals such as `losing_touch`, `meeting_count`, `top_topics`, and `open_commitments` in plain language. Do not imply that omitted restricted meetings do not exist; say the result covers the normal meeting corpus available to the agent.

## Gotchas

- - A failed person profile cannot be interpreted as “never met.” Report the source unavailable.

- - Do not imply that a later sensitivity change proves a historical fact absent; report only the current authorized projection.

- - Alias suggestions are evidence, not permission to rewrite identity. Require explicit confirmation before merging.

- - Use `minutes research "<topic>"` for bounded company, product, or topic research. If it fails, report the source unavailable; never fall back to raw corpus reads.

+ - A missing person may be absent, named differently, or present only in restricted history. Report “not available in the normal meeting graph,” not “never met.”

+ - Relationship and commitment answers are live projections, so a later sensitivity change can intentionally remove prior results.

+ - Name-merge suggestions are suggestions only. Never rewrite or merge identities without the user's explicit confirmation.

+ - Company/product deep extraction is not available through this skill because it would require reading and retaining raw corpus content outside the native policy boundary.