| name | draft-compact |
| description | Curator-triggered context compaction. Synthesizes accumulated appends in one or more context/*/index.md files into a clean, lean current-state document. Archives the pre-compact state to log/ before overwriting. Run when any index.md reads like a change log rather than a current-state document, or exceeds ~500 words.
|
/draft:compact — Compact Context Dimensions
Fold accumulated appends into a clean, lean current-state index.md. Prevents token bloat in session injection.
When to run: When any context/*/index.md reads more like a change log than a current-state document, or when its word count is climbing past ~500 words.
What it does: Reads the current index.md + all log/ entries for a dimension → synthesizes a clean current-state document → archives the old version to log/ → overwrites. Direct write — no proposals/ step needed since this is an explicit, intentional curator action.
Step 1: Scan — measure all dimensions
Read all context/*/index.md files in $DRAFT_WORKSPACE/context/. For each, get the word count:
wc -w "$DRAFT_WORKSPACE/context/"*/index.md 2>/dev/null
Build an internal list of dimensions with their word counts and a flag for those above ~500 words.
Step 2: Select which dimension(s) to compact
Case A — called with a specific dimension (e.g. /draft:compact product):
Skip the question. Proceed directly to Step 3 for that dimension only.
Case B — called with no args (just /draft:compact):
Always ask — never auto-select. Use AskUserQuestion to surface the scan results and prompt the user to choose:
Format the question like this — list all dimensions with their word counts, flag the ones above threshold, and offer "all" as an explicit option:
Which context dimension would you like to compact?
product (847 words) ⚠️ recommended
priorities (312 words)
company (198 words)
team (95 words)
Reply with a dimension name, multiple names separated by spaces, or "all" to compact every dimension above 500 words.
Wait for the user's reply before proceeding.
If the user replies with a dimension name that doesn't exist in context/, tell them and re-prompt once.
Step 3: Compact each selected dimension
Run this full sequence for each selected dimension, one at a time:
3a. Gather source material
Read $DRAFT_WORKSPACE/context/<dim>/index.md directly in full.
For log/ files, use a scoped approach:
- List all files in
$DRAFT_WORKSPACE/context/<dim>/log/ — filenames alone are informative
(format: YYYYMMDDHHMMSS_<slug>.md). Note total count and the full filename list.
- Filter to files with a timestamp from the past 7 days (compare the 8-digit YYYYMMDD
prefix against today minus 7 days). These are the files whose full content will be read.
- Delegate reading of the filtered files to a subagent to protect the main context window:
- Prefer
draft-researcher subagent if available.
- Otherwise, use a general research/explore subagent.
- Instruct the subagent: read each selected log file and return a structured summary
(date, change type, key facts captured).
- Note the scope: total log files listed vs. files read in full (e.g. "23 total, 4 read — last 7 days").
If no log files fall within the 7-day window, synthesize from index.md alone.
Record the current word count of index.md before touching anything.
3b. Synthesize a clean current-state document
Using the current index.md content and all log/ files as inputs, produce a clean replacement document.
The compacted index.md must:
The compacted body should read like: "Here is what is true about [dimension] right now." Not "Here is everything that has happened over time."
3c. Show preview and require approval
Display the full compacted content with a before/after summary:
Compaction preview — context/<dim>/index.md
────────────────────────────────────────────
Before: [N] words
After: [N] words
[full compacted content here]
────────────────────────────────────────────
Apply? [Y/n]
If n: skip this dimension — no changes made. Move to the next selected dimension if any remain.
3d. Apply (on Y)
Execute in this exact order — do not skip any step:
1. Archive the pre-compact state before touching the file:
ARCHIVE_TS=$(date -u +%Y%m%d%H%M%S)
cp "$DRAFT_WORKSPACE/context/<dim>/index.md" \
"$DRAFT_WORKSPACE/context/<dim>/log/${ARCHIVE_TS}_pre-compact-archive.md"
2. Write the compacted version — use the Write tool to overwrite context/<dim>/index.md with the synthesized content.
3. Write a compaction log entry — create $DRAFT_WORKSPACE/context/<dim>/log/${ARCHIVE_TS}_compaction.md:
---
date: <today's date YYYY-MM-DD>
type: compaction
curator: <gh username if available via `gh api user --jq .login`, else "curator">
before_words: <N>
after_words: <N>
log_entries_folded: <count of log/ files read, excluding the archive and this entry>
---
Compaction run by curator on <date>.
Pre-compact state archived to log/<ARCHIVE_TS>_pre-compact-archive.md.
<N> log entries folded into current-state index.md.
Step 4: Confirm
Report for each compacted dimension:
✓ context/product/index.md 847 → 214 words (-75%)
Archived: context/product/log/20260522123045_pre-compact-archive.md
Log entry written.
If any dimensions were skipped (user said n): list them with "skipped — no changes made."
Then remind:
"Run /draft:publish-team to push the compacted context to your team."
Rules
- Never compact without curator selection. Always ask which dimension to compact — never auto-select or default to all.
- Never compact without curator review. Always show the full preview and require Y before writing.
- Always archive before overwriting. The pre-compact state must be recoverable via
log/. No exceptions.
- Always write a log entry. Compaction without a log entry is invisible to CHANGES.jsonl and future surfaces.
- Discard only what's superseded. Compaction is synthesis, not deletion — if older content is still accurate, keep it.
- Do not compact
tensions.md. Tensions are resolved manually by the curator, not synthesized away.
- Prefer lean over comprehensive. When in doubt, cut. The
log/ archive preserves full history.
- Read log/ directory in full, content scoped to 7 days. Always list all log/ files — filenames are descriptive and inform the synthesis. Only read full content for files timestamped within the past 7 days. Use a subagent to read them.
- This is a direct write. Compact writes directly to
index.md — it does not produce a proposals/ file. This is intentional: compact is an explicit curator action, not automated synthesis. The daemon's overwrite guard (commit-to-team-context.sh) applies only to synthesis proposals, not to this skill.
