decision-logger

decision-logger

Surfaces decisions from IM/meetings and persists them into a Base table — so "we chốt cái này nhưng không ai nhớ ở đâu" stops happening.

Verbs (the installed lark-* skills): im +messages-search, im +threads-messages-list, vc +search, vc +notes, base +record-batch-create, base +record-search, base +table-create, docs +fetch. AI minutes artifacts → atomic lark-minutes skill (no raw +summary verb). For atomic command shapes + token flags see connectors/LARK-RECIPES.md.

Prerequisite — decisions Base table

Default schema: Title (Text), Decision (LongText), Decided by (Member), Date decided (DateTime), Context source (URL), Affected area (SingleSelect: Engineering/Sales/Product/Ops/Other), Stakeholders (Member), Status (SingleSelect: Active/Superseded/Reversed), Supersedes (Link → Decision), Review date (DateTime), Tags (MultiSelect).

Auto-create via base +table-create if missing — confirm first.

Mode A — explicit promote

1. Capture source. IM thread → im +threads-messages-list --thread <id> --jq '.data' (peek shape). Meeting → vc +notes --meeting-ids <id>. Doc → docs +fetch --api-version v2 --doc <url> --limit <n> --offset <n> (no keyword/section mode — only --limit/--offset paging).
2. Extract. Decision-keyword sentences ("chốt", "decided", "approved", "OK đi", "we'll go with", "agreed on"); decider name; affected area from group/meeting name.
3. Echo intent — show Title/Decision/Decided by/Date/Source/Area/ Stakeholders, then confirm 3 things: Status, Review date (default 90d), Supersedes? (search → pick / no).
4. On confirm: base +record-batch-create --base-token <t> --table-id <t> --json '{...}' → pre-mutate hook + audit log.
5. Cross-link (with confirm): reply in source IM thread with Base record URL; @mention stakeholders.

Mode B — scan & propose

1. Read priority-topics.md.
2. Sweep 7d: im +messages-search --query "chốt OR decided OR approve" --page-size 50 --jq '.data' (peek shape; needs search:message scope) + vc +search --start <t> --end <t> --jq '.data.items[]|{id,info:.display_info}' (data is a dict — meetings in .data.items[]).
3. Filter & cluster: drop noise (require ≥1 priority-topic match OR stakeholder count ≥3); same topic+day+group → one candidate; rank by stakeholder count.
4. Present candidates list with per-item Log/skip/merge + bulk action.
5. Per-confirm: same base +record-batch-create flow.

Hard rules

1. Never auto-create record without echo+confirm (do-not.md §6).
2. Source URL required — every decision links back to IM/meeting/doc.
3. Resolve names, never raw open_ids.
4. No @mention of stakeholders until explicit confirm.
5. Supersedes: set prior record Status=Superseded + link; never delete.
6. Review date default 90d (override in optional decision-policy.md).
7. Don't log private 1:1 decisions without explicit confirm.

Edge cases

• No table → prompt create with default schema (confirm).
• Keyword in joke context → require priority-topic match OR ≥3 stakeholders.
• Ambiguous decider → list all, user picks primary.
• Spans multiple meetings → separate records linked via Supersedes/extends.
• Reversal → search prior in Base, mark Status=Reversed, log linked new record.

Memory

Required: do-not.md. Recommended: priority-topics.md, team-roster.md, optional decision-policy.md.

Cadence

On-demand (explicit); weekly Mode B sweep Monday (via morning-brief); post-meeting offer from meeting-prep. Also called by incident-retro and sprint-retro. Base over Wiki because decisions need queryability.