coalesce-labs/research-curate
plugins/dev/skills/research-curate/SKILL.md
Walk thoughts/shared/research/ and thoughts/shared/plans/, score each doc's staleness, regenerate INDEX.md, and append LLM-surfaced contradictions to CONTRADICTIONS.md (append-only). Source docs are never modified. Classification: current (age<90d AND refs valid), needs-review (age>=90d OR broken refs), likely-stale (age>=180d AND no recent activity). Inventory is deterministic; contradiction detection runs one LLM call per cluster (CTL-467 + CTL-468 / Initiative 4 Phase 1+2).
$ install --global
skillsauthnpx skillsauth add coalesce-labs/catalyst research-curateInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
Security Scan Results
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
3
Scanners Passed
9
Scanners in report
Clean
TrivyContainer and dependency vulnerability scanner
95%
Clean
SemgrepStatic code analysis for vulnerabilities
95%
Clean
mcp-scan (Snyk)Model Context Protocol security validation
95%
Skipped
Snyk (dep)Open source security scanning
50%
Skipped
Socket.devSupply chain security analysis
50%
Skipped
VirusTotalMulti-engine malware detection
50%
Skipped
CrowdStrikeAdvanced threat intelligence
50%
Skipped
OSV-ScannerOpen Source Vulnerability database check
50%
Skipped
OWASP Dep-Check
50%
Last scanned: May 18, 2026, 2:41 AM106.4s1 file scanned
SKILL.md
- name:
- research-curate
- description:
- |
- Classification:
- current (age<90d AND refs valid), needs-review (age>=90d OR
- disable-model-invocation:
- true
- user-invocable:
- true
research-curate
Regenerates INDEX.md inventories for thoughts/shared/research/ and
thoughts/shared/plans/, then clusters docs by topic and asks an LLM to
identify contradicting claims, appending findings to CONTRADICTIONS.md.
Source markdown is never modified — only INDEX.md (overwritten) and
CONTRADICTIONS.md (appended) are written.
What it does
- Inventory — walks each directory, parses frontmatter (
date,topic,tags), counts words, extracts file:line references (file.ext:Npattern). - Score — for each doc, computes age, validates ref paths against
HEAD, and counts last-30d git log hits for the doc's top-3 tags. - Classify:
current— age < 90d AND all refs resolve at HEADneeds-review— age >= 90d OR at least one broken reflikely-stale— age >= 180d AND zero recent topic activity
- Generate — overwrites
<dir>/INDEX.mdwith three sections sorted by date descending. Output is fully deterministic. - Cluster + contradict (CTL-468) — tokenizes
tags∪topic, computes pairwise Jaccard similarity, unions docs with score ≥ 0.4 via union-find. Clusters of size 3–10 are kept (smaller dropped, larger capped at 10). For each cluster, one LLM call (defaultclaude -p) receives a token-bounded prompt (1500 chars/doc) asking for contradictions. New findings are appended to<dir>/CONTRADICTIONS.mdunder a## YYYY-MM-DD — <topic>heading. Existing entries are never modified.
Invocation
# Curate both directories in-place (inventory + contradictions)
bash plugins/dev/scripts/research-curate/run.sh thoughts/shared/research
bash plugins/dev/scripts/research-curate/run.sh thoughts/shared/plans
# Inventory only — skip the LLM-driven contradiction pass
bash plugins/dev/scripts/research-curate/run.sh --skip-contradictions \
thoughts/shared/research
# Dry run — INDEX written to /tmp; contradictions never touched
bash plugins/dev/scripts/research-curate/run.sh --dry-run thoughts/shared/research
Each invocation prints summary lines:
thoughts/shared/research: current=42 needs-review=83 likely-stale=80 → thoughts/shared/research/INDEX.md
thoughts/shared/research: contradictions clusters=5 appended=3 → thoughts/shared/research/CONTRADICTIONS.md
Projected cost for the weekly Routine: ~$0.38/run (75K input tokens × $5/M), ~$1.50/month.
Skill body
When invoked as /catalyst-dev:research-curate, run both directories
sequentially and report the summary lines back to the user.
set -euo pipefail
SCRIPT="${CLAUDE_PLUGIN_ROOT:-.}/scripts/research-curate/run.sh"
[[ -x "$SCRIPT" ]] || SCRIPT="plugins/dev/scripts/research-curate/run.sh"
bash "$SCRIPT" thoughts/shared/research
bash "$SCRIPT" thoughts/shared/plans
Source-doc immutability is enforced by inventory.sh (which only ever reads
source markdown), run.sh (which writes exclusively to INDEX.md or a
/tmp/ dry-run path), and append-contradictions.sh (which opens
CONTRADICTIONS.md in append-mode only). The test suite asserts this
property against fixture repos via before/after hash comparisons.
Out of scope (deferred)
- CMA Routine wiring (weekly cadence) — Phase 3 of Initiative 4 (CTL-469).
- LLM-generated one-line summaries — plan permits but defers for testability.
- LLM-driven clustering — fallback to deterministic Jaccard is preserved for cost predictability (Phase 3 refactor option in the plan).
Tests
bash plugins/dev/scripts/__tests__/research-curate-inventory.test.sh
bash plugins/dev/scripts/__tests__/research-curate-contradictions.test.sh
Related Skills
coalesce-labs/phase-remediate
testing
VerifiedTrustedCommunity
Phase-agent that fixes a failing verify verdict so the pipeline self-heals instead of stalling to needs-human (CTL-653). Reads `${ORCH_DIR}/workers/<ticket>/verify.json`, fixes the `findings[]` (every severity:"high" plus the regression_risk drivers) directly via Edit/Write, commits the remediation, and emits `phase.remediate.complete.<ticket>`. The scheduler's router then re-dispatches `verify` to re-check (the verify⇄remediate cycle, cap 3). Dispatched as a `claude --bg` job by `phase-agent-dispatch`, which invokes it via slash command — hence `user-invocable: true`.
13SKILL.mdUpdated May 28, 2026
coalesce-labs/phase-verify
development
VerifiedTrustedCommunity
Phase agent for the verify step of the 9-phase orchestrator pipeline (CTL-450). NEW skill — has no canonical wrapper. Runs read-only adversarial verification against the implement-phase diff: tsc, tests, lint, security scan, reward-hacking scan, code review, test coverage, silent-failure hunt. Writes ${ORCH_DIR}/workers/<TICKET>/verify.json then emits phase.verify.complete.<ticket>. Reads phase-implement.json as its prior-phase artifact. NEVER writes application code — only test files allowed. Spawned via phase-agent-dispatch via slash command — hence `user-invocable: true`.
13SKILL.mdUpdated May 18, 2026
coalesce-labs/plugins/dev/skills/phase-triage
tools
VerifiedTrustedCommunity
--- name: phase-triage description: Phase agent that triages a Linear ticket — expands acronyms, classifies (feature/bug/docs/refactor/chore), identifies dependencies, estimates scope, writes triage.json, and posts a triage analysis comment to Linear. Triage completion is signaled by that comment plus the local triage.json — there is no `triaged` label. Emits phase.triage.complete.<TICKET> on success and phase.triage.failed.<TICKET> on error. Dispatched by the phase-agent orchestrator (CTL-452)
13SKILL.mdUpdated May 18, 2026
coalesce-labs/phase-review
testing
VerifiedTrustedCommunity
Phase agent for the review step of the 9-phase orchestrator pipeline (CTL-450). Wraps the /review skill (gstack) — explicitly skips /ultrareview per user decision. Reads verify.json from the prior phase, runs /review against the diff, writes ${ORCH_DIR}/workers/<TICKET>/review.json, and creates a remediation commit for any HIGH-severity finding that has a deterministic fix. Emits phase.review.complete.<ticket>. Spawned via phase-agent-dispatch via slash command — hence `user-invocable: true`.
13SKILL.mdUpdated May 18, 2026