gaia-triage
Triage has three workstreams — GitHub issue hygiene, skill-batch intake review, and the
cross-repo prioritization sweep. Run them independently or together depending on what the
user asks for. All three produce documented, auditable decisions rather than silent
changes.
This skill absorbed favorchurch/rock-steward's rock-triage (2026-09-01): its impact
sizing, priority matrix, four-step inspection, and prioritized-matrix output landed here
as Workstream 3 rather than shipping as a second, competing triage skill. There is one
gaia-triage. Its priority ladder is the P0-P4 ladder /gaia-meta-audit and
/gaia-meta-sweep already use — one vocabulary, not two.
Triage is not cartography. Sorting a backlog of known work is this skill; planning an
effort too big for one session as a map of decision tickets is /wayfinder. Taking one
triaged issue end-to-end to a PR is /gaia-issue-resolver.
Headquarters
Every issue lives on gaia-research/gaia-skill-tree, regardless of which repo the fix
lands in. One tracker for the whole org.
export GAIA_HQ=gaia-research/gaia-skill-tree
Tag the target repo on the issue when the work lands elsewhere — a line at the top of the
body, **Target repo:** \gaia-research/gaia-skill-heaven`, matching the /wayfinderconvention. Cross-repo references use the fullgaia-research/<repo>#<n>form; a bare#<n>` always resolves to HQ.
Classification taxonomy
Every triaged issue gets a size and a priority. Both are labels; neither is invented ad
hoc (see label hygiene below).
Impact size
Blast radius, structural complexity, and data footprint — not hours.
| Size | Means |
|---|
XS | Trivial fix, label correction, single config value, negligible blast radius |
S | Minor scoped adjustment, one predicate, isolated UI touch-up |
M | Standard scoped task — one command, one schema field, one localized workflow |
L | Cross-component change, non-trivial data backfill, multi-surface impact |
XL | Architecture-level refactor, substantial registry debt remediation |
XXL | High blast radius — core schema, authorization, or site-wide parity |
Epic | Umbrella spanning multiple distinct sub-issues across domains |
Priority — P0-P4
The same ladder /gaia-meta-audit and /gaia-meta-sweep use. Read those for the
registry-specific instances of each rung.
| Priority | Means |
|---|
P0 | Critical integrity violation or active breakage — unsupported top-rank claim, security/authorization exposure, secrets in a diff, the site dark, main red |
P1 | Structural correctness — dead evidence links, Star Bar gaps, broken data contracts, lockstep drift, a prerequisite blocking milestone delivery |
P2 | Attribution, sourcing, and feature upgrades — wrong promoted pointer, stale catalog URL, enhancement to a working surface |
P3 | Registry hygiene and nice-to-haves — over-broad mappings, duplicate clusters, backlog enhancements, exploratory design |
P4 | Documentation cleanup — placeholder bodies, generated-output drift, comment and copy fixes |
P0 interrupts. P1 schedules. P2-P4 queue.
Label hygiene
Workstream 1: GitHub Issue Triage
The goal is a clean, signal-rich backlog — not a closed one. Closing issues prematurely
hides real work; leaving stale issues open buries actionable items. For each issue, make
an explicit call: close (resolved), update (needs narrowing), or keep (valid, still open).
Step 1 — Pull the open issue list
gh issue list --repo gaia-research/gaia-skill-tree --state open --limit 100
Group by theme (CLI bug, registry data, docs drift, enhancement). This shapes the
decision pattern — CLI bugs and registry issues require code evidence; enhancement requests
require a current-relevance check against the roadmap.
Step 2 — Audit each issue against the codebase
Before commenting or closing, verify the claim. Check the evidence paths in
references/evidence-check.md for common file locations:
| Issue type | Where to look |
|---|
| CLI command request or bug | src/gaia_cli/commands/ |
| Registry data or duplicate skills | registry/gaia.json, registry/nodes/ |
| Documentation drift | Run python scripts/build_docs.py --check |
| Packaging / dependency | pyproject.toml, uv.lock |
| Test coverage | tests/ |
Step 3 — Act via gh CLI
Post a triage comment (findings without closing — use when the issue needs narrowing
or a reproduction, or when you want to flag it for a human maintainer):
gh issue comment <issue-number> --repo gaia-research/gaia-skill-tree \
--body "Triage update: <findings and recommendation>"
Close a resolved issue (only when codebase evidence confirms it is implemented):
gh issue close <issue-number> --repo gaia-research/gaia-skill-tree \
--reason completed \
--comment "Closing as implemented. Evidence: <specific file + line or command output>"
Dry-run first: the scripts/triage_batch.sh script supports --apply and
--close-resolved flags. Run without flags to preview all comments before posting:
bash .claude/skills/gaia-triage/scripts/triage_batch.sh
# Then, when ready:
bash .claude/skills/gaia-triage/scripts/triage_batch.sh --apply --close-resolved
Decision guide
| Finding | Action |
|---|
| Issue is implemented and tests pass | Close as completed with evidence pointer |
| Issue is valid but under-specified | Comment with narrowing questions; keep open |
| Issue references a stale command name | Comment updating the correct current command; keep open unless fully superseded |
| Issue is a duplicate | Close with a link to the canonical issue |
| Issue is an enhancement blocked on design | Comment with current blocker; keep open |
Workstream 2: Skill-Batch Intake Review
Draft skill proposals land in registry-for-review/skill-batches/ after a contributor
runs gaia push. This workstream gates them before they enter registry/nodes/.
Rejecting bad batches here is cheaper than demoting skills after promotion — every skill
that enters the canonical registry acquires timeline history, evidence links, and
downstream references that make later removal messy.
Step 1 — List pending batches
ls registry-for-review/skill-batches/
Each batch is a JSON file. Read its metadata to understand the contributor and scope before
evaluating individual skills.
Step 2 — Evaluate each skill proposal
For each skill in the batch, work through these checks in order:
- Schema validity — does the node satisfy the canonical JSON schema?
gaia dev validate --intake
- Nomenclature — does the skill name and description match
CONTEXT.md vocabulary?
Flag banned synonyms; the rarity field is deprecated and should not appear in new proposals.
- Evidence quality — each above-1★ claim needs graded evidence with a valid Evidence Type (arxiv|repo|github-stars) and a derived Evidence Grade;
weak/ungraded-only evidence is grounds for
needs-info. Ranking up gates on Trust Magnitude, not an evidence-class floor.
- Prerequisite existence — all
prerequisites[] IDs must resolve in the current registry.
gaia dev list --generic | grep <prerequisite-id>
- Duplication check — does a substantially equivalent skill already exist?
gaia dev list --generic | grep -i "<skill-name>"
Step 3 — Label and route
Apply one of three dispositions per skill:
| Label | Meaning | Next action |
|---|
| approve | Schema valid, evidence solid, no duplicate | Promote via gaia dev add |
| needs-info | Fixable gap (weak evidence, missing prereq, naming issue) | Comment on the batch file with specifics; do not promote yet |
| reject | Duplicate, out of scope, or unfixable schema violation | Document reason; discard batch entry |
For approved skills, promote via CLI (never hand-edit registry nodes directly):
gaia dev add "<Skill Name>" --type basic --description "<description>"
gaia dev evidence <skill-id> "<url>" --type repo --trust <value>
After promotion, regenerate Class S artifacts:
gaia dev docs
Workstream 3: Cross-repo triage sweep
Use this when the ask is "triage the backlog" at scale rather than "is this one issue
still valid". It sizes, prioritizes, links, and assigns — and it ends in a matrix a human
can act on without reopening every thread.
Four-step inspection
Run all four on every issue in scope, in order:
- State & relevance. Is it still live, or already resolved/superseded? Verify against
the codebase, merged PRs, or a read-only probe — never against the issue's own claim.
The read-only probes are the ones in
/gaia-issue-resolver Phase 2 (gaia dev audit,
scripts/verify_evidence.py, python scripts/build_docs.py --check). If resolved,
close with the evidence.
- Status update & timeline. One concise technical comment: what is true now, and
whether the work is starting soon, blocked on a prerequisite, or in backlog.
- Linkage & umbrella synthesis. Find the shared cause across issues. Where several
related sub-issues have no parent, create an umbrella and label it
epic, then nest
the children as GitHub sub-issues (gh api repos/$GAIA_HQ/issues/<parent>/sub_issues -f sub_issue_id=<child-id> — the child's id, not its number).
An umbrella that turns out to be a fog-bound effort rather than a known set of tasks is
a /wayfinder map, not an epic. The test: if the children are decisions you cannot yet
state precisely, stop triaging and chart a map.
- Assignment. Default to the original issue author unless someone is already assigned.
Assign a synthesized umbrella to the owner of its constituent issues.
Fan-out
Delegate parallel reads — pulling issue payloads, checking whether a claim still holds in
the tree, counting occurrences — to cheap subagents. Keep with yourself: deciding
priority, deciding size, judging whether two issues share a cause, and writing the
umbrella. Give each subagent the exact command and the exact thing to report back; a
subagent asked to "investigate" returns prose, one asked to "run X and report the integer
after Total:" returns data.
Output — the prioritized matrix
# Gaia Triage Summary — <date>
## P0 - Critical
- **#<n>** - <Title> | `<Size>` | repo: `<target repo>` | assignee: `@<who>` | <Active/Blocked/Ongoing>
- *Umbrella*: <link>
- *Verification*: <what was actually probed, and the result>
- *Next action*: <the immediate step>
## P1 - Structural correctness
## P2 - Attribution & upgrades
## P3 - Hygiene & backlog
## P4 - Documentation
## Umbrellas & epics
- **Epic: <Title> (#<n>)** [assignee: `@<who>`] - sub-issues: #A, #B, #C
Refer to issues by title, not by bare number — a wall of #42, #43, #44 is illegible.
Output
At the end of a triage pass, report:
- How many issues were reviewed, closed, and left open with comments
- How many batch skills were evaluated, approved, flagged needs-info, and rejected
- The prioritized matrix, if Workstream 3 ran
- Any recurring patterns worth tracking (e.g. a common prereq that's missing from the
registry) - a pattern seen a fourth time is a Class A detector in
/gaia-issue-resolver, not another issue
This summary helps the next maintainer understand what happened without re-reading every issue thread.
Optional Semantic Issue Advisory
When evaluating large backlogs or triage queues, an operator may optionally run the read-only Jev advisory runner (docs/agents/jev.md):
# Offline dry-run by default ($0 spend); add --live to query TypeSafe API
gh issue list --json number,title,body,labels > /tmp/issues.json
python scripts/jev_advisory.py --mode issues --input /tmp/issues.json
The runner provides non-binding duplicate shortlists and advisory P0–P4 priority recommendations. It never labels, comments on, or closes issues. If unauthenticated, ambiguous, or over budget, it generates a bounded handoff packet to worker-luna.