popoffvg/context-done

Context Done

Finalize an active task: summarize collected insights and distribute to each related repo's insight folder.

Usage

/context done — complete the current active task /context done <task-name> — complete a specific task

Configuration

Read insights_root from ~/.claude/memory-keeper.local.md YAML frontmatter. If the file is missing, stop and ask the user to create it with the required settings (see plugin README).

Command visibility and wiring checks

Run these checks when user reports "command not visible" or asks "for Claude?":

1. Pi skill visibility: ensure ~/.pi/agent/skills/context-done/SKILL.md exists (usually symlinked to this file).
2. Claude command visibility: ensure a Claude command file exists under harness/plugins/work-manager/claude/commands/ and that it explicitly instructs invoking context-done.
3. If Claude command is missing, do not claim the skill is available from Claude slash commands; state that only direct skill invocation works until command wiring is added.

Procedure

Step 1: Find the active task

1. Read <insights_root>/_tasks/pending.md
2. Find the task with Status: active (or match by name if provided)
3. If no active task found, tell the user and stop

Step 2: Collect task insights

1. Read <insights_root>/_tasks/<task-slug>/notes.md
2. Extract all insights with their repo tags (_(repo: <name>)_)
3. Read the task description from pending.md
4. If notes are sparse, still produce concrete repo-ready notes from available artifacts (commands run, files changed, config values) instead of skipping distribution

Step 3: Generate summary

Create a concise task summary (3-5 sentences max):

• What the task was about
• Key decisions made
• Important findings or patterns discovered

Step 4: Distribute to repo insight folders

For each unique repo in the task's Repos list:

1. Read <insights_root>/<repo>/insights.md first
2. Deduplicate: skip if an existing entry already covers the same knowledge (same topic heading, semantic overlap, or broader superset). If the new summary is a broader version of an existing entry, replace it.
3. Append to <insights_root>/<repo>/insights.md:

   ## <task-title> (task summary) — YYYY-MM-DD HH:MM
   <summary focusing on insights relevant to this specific repo>
   _(from task: <task-slug>)_
   

4. Create the repo directory if it doesn't exist

Step 5: Archive the task

1. In <insights_root>/_tasks/pending.md, change the task's status from active to done
2. Keep <insights_root>/_tasks/<task-slug>/ as-is (archive, don't delete)
3. If this run was triggered by a work-completion flow that requires cleanup, explicitly instruct the caller to remove local temporary planning notes (for example project _notes/) after insight distribution succeeds

Output

Report:

• Task name and summary
• Which repo folders received insights
• Confirmation that task is marked done

Autoresearch rules

Eval checklist:

1. Were insights distributed to every repo's insight folder that was touched during the task?
2. Does each saved insight contain a concrete fact (file path, command, config value) — not a vague summary?
3. Was the task summary written before distribution (not skipped)?
4. Did the skill read the insights_root config before attempting any file writes?

Test inputs:

• "Complete a task that touched 3 repos with debugging discoveries in each"
• "Complete a task with no significant insights worth saving"
• "Complete a task when insights_root config is missing"

Can change: summary format, distribution logic, insight deduplication, task finalization steps Cannot change: insights_root config requirement, per-repo distribution, task completion semantics Min sessions before eval: 5 Runs per experiment: 3