Knowledge Compact

This documentation was generated with the assistance of AI. Please report any inaccuracies.

Keep knowledge/pending and knowledge/agreements honest: one pending file per distinct piece of outstanding work, and nothing still marked open once its tracked issue says it’s done.

Purpose

iru-knowledge-compact compacts the knowledge base that iru-knowledge-process-inbox builds up meeting after meeting. Two kinds of cruft accumulate there naturally, and neither is a bug in how the base was filed: the same action item gets re-raised in a later meeting and is deliberately filed as its own pending task rather than judged in isolation, and work that has actually shipped stays marked open because nothing was watching its tracker issue. This skill sweeps both.

It groups pending tasks that describe the same piece of work — not merely the same words — and for each group keeps a single survivor: the one member that already carries a tracker issue, or the earliest task id when none do, folding whatever extra context the removed members carried (a deadline, a named owner, a narrowed scope, an extra meeting that re-raised it) into the survivor before deleting them. A group where more than one member has a tracker issue is deliberately left untouched and reported for a human decision, since collapsing it would orphan a tracked issue. It then reads the current state of every GitHub issue / Jira ticket recorded on a pending task or agreement item by iru-knowledge-to-issue / iru-knowledge-to-jira, and closes out the ones already closed as completed — delegating each to iru-knowledge-complete where a pending file still exists. An issue closed as not planned is treated as not completed and put to the user, never checked off as if it shipped. Finally it repairs drift between the two directories.

The skill only ever reduces the base. It never creates or rewords an agreement, never touches knowledge/meetings or knowledge/people (a meeting record is history), and never closes, reopens, or comments on a tracker issue — it reads issue state, it doesn’t drive it. Every deletion and edit is presented for confirmation first, because the duplicate judgement is the skill’s own reading of the text rather than something the files state outright.

Purpose group

Knowledge Base Management

Invocation

/iru-knowledge-compact

Inputs

Argument Required Description Default

(none)

No

This skill takes no arguments — it sweeps the whole of knowledge/pending and knowledge/agreements. It stops early if knowledge/ doesn’t exist yet (run iru-knowledge-setup and then iru-knowledge-process-inbox first), and reports there is nothing to compact if both directories hold nothing beyond .gitkeep.

Not applicable.

Tracker access is checked but never required: gh (authenticated, on a GitHub-hosted repository) for - GitHub issue: references, and a connected Jira MCP tool for - Jira ticket: references. If neither is reachable the run continues with duplicate detection and drift repair, skipping the completion checks rather than guessing whether anything is done.

Outputs

  • Duplicate pending files deleted, with the extra context they carried appended to the surviving knowledge/pending/pending-<timestamp>-<n>.md along with a - Supersedes: <task-id> note. The removed members' agreement items stay unchecked and gain a (superseded by task <survivor-id>) note — a duplicate is still outstanding work, just tracked elsewhere.

  • Completed items closed out: the pending file removed and its numbered agreement item checked off (via iru-knowledge-complete), or the item’s [ ] flipped to [x] directly where no pending file remained.

  • Approved drift repairs: a pending file whose agreement item is already [x] deleted.

  • Approved closed-as-not-planned items handled as the user chose — annotated (not planned — <issue url>) with the checkbox left [ ], since the work was abandoned rather than delivered.

  • A grouped report: duplicates collapsed (and those deliberately left alone, plus borderline pairs it declined to merge), items closed out and the issue that proved each one, closed-as-not-planned items, drift repairs, untracked open agreement items, and everything skipped with its reason — closing with pending/open-item counts before and after.

  • Nothing at all, if the user cancels at the confirmation step.

Execution flow

flowchart TD A[Start /iru-knowledge-compact] --> B{"knowledge/ exists\nwith content?"} B -- no --> Z1["Report: run iru-knowledge-setup +\niru-knowledge-process-inbox first, stop"] B -- "only .gitkeep" --> Z2[Report nothing to compact, stop] B -- yes --> C["Check tracker reachability\n(gh / Jira MCP) — never a hard stop"] C --> D["Build inventory: every pending file\n(task id, agreement + meeting refs,\ntracker line, extra context)\nand every numbered agreement item"] D --> E["Cross-reference the two directories"] E --> F["Group duplicates by same work;\npick survivor: the tracked one, else\nthe earliest task id"] F --> G{"More than one member\nhas a tracker issue?"} G -- yes --> H["Remove nothing from this group —\nreport for a human decision"] G -- no --> I["Work out the merge text\nthe survivor is missing"] H --> J I --> J{"Tracker\nreachable?"} J -- no --> K["Skip completion checks,\nsay so in the report"] J -- yes --> L["Look up each tracked item's state:\ncompleted / not planned /\nstill open / lookup failed"] K --> M L --> M["Add drift repairs found in the\ncross-reference"] M --> N["Present the full plan\n(AskUserQuestion)"] N -- cancel --> R N -- "apply some" --> O N -- "apply everything" --> O["Merge context into survivors first,\nthen delete duplicates, then close out\ncompleted items, then drift repairs,\nthen not-planned decisions"] O --> P{"Completed item has\na pending file?"} P -- yes --> Q["Invoke iru-knowledge-complete\nwith the task id"] P -- no --> Q2["Check off the agreement item\ndirectly"] Q --> R Q2 --> R["Report, grouped, with before/after counts"]

Dependencies

Invokes

  • iru-knowledge-complete — Step 6, once per completed item that still has a pending file, to delete that file and check off its agreement item. If it stops on an inconsistency (a missing agreement file, an item not found, an item already checked), iru-knowledge-compact records the message and carries it into its report rather than working around it by hand.

Invoked by

  • iru-knowledge-process-inbox — as its Step 11, once for the whole run after every meeting group has been filed, so the newly created pending tasks are reconciled against what earlier meetings already produced. Skipped there if the run created no agreements at all, or if this skill isn’t installed in the repository (the catalog’s skills are adopted individually).

It is also a top-level entry point, run directly by a person as periodic housekeeping.

Related agents

None — this skill does not delegate any part of its work to a built-in or custom agent.

Source

SKILL.md on GitHub — the file this page was generated from.