Refresh Research

$ARGUMENTS

Refresh Research

Orchestrate parallel research-curator agents to bulk-refresh research entries in ./research/. Detects staleness, skips fresh entries, updates only what qualifies. Safe to run repeatedly.

Arguments

`` controls scope:

• --all — Refresh every entry regardless of staleness
• --stale (default) — Refresh entries past their review date
• --category — Refresh all entries in one category (e.g., --category agent-frameworks)
• --layer — Refresh entries with matching SDLC layer metadata (0=process, 1=language, 2=stack). See [plugins/development-harness/docs/sdlc-layers/](../../../plugins/development-harness/docs/sdlc-layers/).
• --dry-run — Report what would be refreshed; do not spawn agents

Workflow

Step 1: Inventory and Staleness Detection

Glob ./research/**/*.md (exclude README.md). For each entry, parse:

1. YAML frontmatter — extract metadata.layer value (string "0", "1", or "2"; null if absent).
2. Freshness Tracking section — extract Last Verified and Next Review dates.

flowchart TD
    Start([Read entry]) --> ParseFM[Parse YAML frontmatterExtract metadata.layer]
    ParseFM --> HasFreshness{Freshness Tracking section present?}
    HasFreshness -->|No| Stale1[STALE: no tracking]
    HasFreshness -->|Yes| ComputeDays[Compute Days Old = today minus Last Verified]
    ComputeDays --> PastDue{Next Review Recommended |Yes| Stale2[STALE: past review date]
    PastDue -->|No| TooOld{Last Verified > 6 months ago?}
    TooOld -->|Yes| Stale3[STALE: too old]
    TooOld -->|No| Fresh[FRESH: N days until next review]

Build inventory table: | File | Category | Layer | Last Verified | Next Review | Days Old | Stale? |

The Days Old column holds an integer: today's date minus the Last Verified date in days. If Last Verified is absent or unparseable, render as —.

The Layer column holds the metadata.layer value or — if absent.

Step 2: Apply Scope Filter

Apply filters sequentially. Filters combine with AND logic — each filter narrows the set from the previous step.

1. Base set: Start with all inventoried entries.
2. Staleness filter (default unless --all):

• --all — keep all entries (no staleness filter)
• --stale (default) — keep only entries marked STALE in Step 1

1. Category filter (optional): --category — keep only entries whose category directory matches ``.
2. Layer filter (optional): --layer — keep only entries where metadata.layer equals the requested value. Entries without layer metadata (— in inventory) are excluded.
3. Dry-run check: --dry-run — display the filtered target list and stop without spawning agents.

If zero entries remain after all filters: report "No entries match the applied filters." and stop. When --layer was specified and zero entries match, additionally report: "No entries found for layer {N}. Entries need metadata.layer in their YAML frontmatter to be targeted by --layer."

When the --stale filter excludes entries because they are FRESH, list each excluded entry before continuing to Step 3:

Skipped (fresh):
  ./research/{category}/{name}.md — N days until next review (last: YYYY-MM-DD, vX.Y.Z)
  ./research/{category}/{name}.md — N days until next review (last: YYYY-MM-DD, vX.Y.Z)

This listing appears regardless of whether --dry-run is active. Under --all (no staleness filter), no entries are excluded by staleness, so this block does not appear.

Step 3: RT-ICA Pre-Flight

RT-ICA: Research Refresh
Goal: Refresh {N} research entries with current data from primary sources
Conditions:
1. mcp__Ref and mcp__exa available in session  (primary data gathering)
2. gh CLI authenticated                         (GitHub repo metadata)
3. Outbound network access                      (fetch fresh data)
4. ./research/ writable                         (update entry files)
5. Entry files parseable markdown               (determine what changed)
Decision: {APPROVED | BLOCKED}

If BLOCKED: report missing tools/access, suggest workarounds, stop.

Step 4: Spawn Agents in Waves

Split target entries into sequential waves of 5. Within each wave spawn agents in parallel; wait for wave completion before starting the next.

For each entry:

Agent(subagent_type: "research-curator", prompt: "--rerun ./research/{category}/{name}.md", model: "sonnet")

After each wave, collect and log results:

Wave {N} complete: {M}/{total} succeeded
  updated   -- ./research/agent-frameworks/agno.md (v0.3→v0.5, +2k stars)
  unchanged -- ./research/mcp-ecosystem/narsil-mcp.md (no changes detected)
  failed    -- ./research/developer-tools/orbstack.md -- error: [reason]

Outcome categories: Updated (content changed), Unchanged (re-verified, no changes), Failed (agent could not complete).

Step 5: Update README

After all waves complete, update ./research/README.md:

• Refresh freshness dates for updated and unchanged entries
• Add new categories if agents created them
• Regenerate category counts

Step 6: Summary Report

# Research Refresh Report

**Date**: {YYYY-MM-DD}
**Scope**: {--all | --stale | --category X | --layer N}
**Total scanned**: {N} | **Targeted**: {M} | **Skipped (fresh)**: {K} ({min}–{max} days until next review)

## Results

| Outcome | Count | Notes |
|---------|-------|-------|
| Updated | {N} | |
| Unchanged | {N} | |
| Failed | {N} | |
| Skipped (fresh) | {K} | {min}–{max} days until next review |

When K = 0, omit the Skipped (fresh) row. When K = 1, Notes column: `{N} days until next review`.
When K = 0 in the header: `**Skipped (fresh)**: 0`. When K = 1: `**Skipped (fresh)**: 1 ({N} days until next review)`.

## Updates

| Entry | Category | Change Summary |
|-------|----------|----------------|
| {name} | {category} | {version bump, stat update, etc.} |

## Failures

| Entry | Error |
|-------|-------|
| {name} | {reason} |

## Next Actions

- Due for review in 30 days: {list}
- Categories with no recent updates: {list}
- Failed entries to retry: {list}

Step 7: Post-Actions

Lint modified files before committing to prevent malformed entries reaching git history:

uv run prek run --files ./research/

Commit with a format that identifies the refresh scope for audit purposes:

git add ./research/ && git commit -m "docs(research): refresh {N} entries ({date})"
git push -u origin HEAD

Error Handling

• No entries match filter — report "All entries are fresh. Nothing to refresh." and stop
• No entries match --layer filter — report "No entries found for layer {N}. Entries need metadata.layer in their YAML frontmatter to be targeted by --layer." and stop
• Agent failures — continue remaining waves; include in summary Failures table
• Network issues mid-wave — complete current wave, report partial results, suggest retry with --stale
• README update conflict — re-read README and retry update once

Related

• /research-curator — single-entry and batch research operations; this skill wraps it with staleness detection and RT-ICA
• @research-curator agent — .claude/agents/research-curator.md — executes individual entry reruns

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Jamie-BitFlight
• Source: Jamie-BitFlight/claude_skills
• License: MIT

Install and usage instructions live in the source repository linked above.