glacier-sync
Optional skill. Syncs work tracking between the current repo and Glacier board. Auto-invoke after branch creation (step 1 of implement workflow), PR open, PR merge, or when the user mentions board syn
SKILL.mdskills/glacier-sync/SKILL.md
GitHubGlacier Sync (explicit transitions)
Keeps the Glacier board in sync with repo activity. Callers invoke it at known workflow points with a named transition. No hooks.
Fully optional. If env vars are missing, skip silently. Never block the parent workflow.
Why explicit, not hooks
The previous version relied on a FileChanged hook on .git/HEAD. It was unreliable and cannot work with parallel runs:
FileChangedmatchers take literal filenames, andifonly applies to tool events, so the shell-test condition was ignored.- Skill hooks register only after the skill is invoked, so the first branch creation in a session was missed.
- In a git worktree
.gitis a file and HEAD lives under the main repo's.git/worktrees/<name>/, so the hook never fires for worktree-isolated agents.
Explicit calls at fixed workflow steps are deterministic and work the same in the main checkout, in worktrees, and in parallel batches.
Activation conditions
All must be true, otherwise silent skip:
GLACIER_ENABLED=trueGLACIER_WORKSPACE_IDis setGLACIER_PROJECT_IDis set
Configuration
In .env.local (gitignored — list it in .worktreeinclude so worktrees get a copy):
GLACIER_ENABLED=true
GLACIER_WORKSPACE_ID=<uuid from Project Settings>
GLACIER_PROJECT_ID=<uuid from Project Settings>
MCP server URL is hardcoded: https://www.getglacier.ai/api/mcp
Column IDs resolve at runtime via Glacier:list_columns — no stored IDs, board can be restructured freely.
Invocation contract
Callers pass:
| Input | Required | Values |
|---|---|---|
transition |
yes | in-progress, in-review, done |
card_id |
preferred | Glacier card UUID (skip matching when provided) |
issue |
fallback | GitHub issue number, used to find the card |
narrate |
no | true when called from /implement-v (demo formatting) |
Who calls what
| Transition | Caller | When |
|---|---|---|
in-progress |
implement agent / /implement-v |
Right after the feature branch is created (step 1) |
in-progress |
/triage |
For every card in a batch, before agents are dispatched |
in-review |
implement agent / /implement-v / /triage |
After gh pr create succeeds |
done |
/glacier (PR sync) |
After merge — merges happen outside Claude Code |
Single writer rule for batches: when /triage dispatches parallel agents, it passes glacier: orchestrator to each agent, and the agents make no Glacier writes. The orchestrator performs all transitions. This avoids concurrent writes and WIP-limit races.
Transition rules
- Resolve columns (cached per session): match names case-insensitively — "Backlog", "Ready", "In Progress", "In Review", "Done".
- Resolve the card:
card_idif given, else match viaissue(see Card matching). - Never regress a card. Allowed moves:
in-progress: from Backlog or Readyin-review: from Backlog, Ready, or In Progressdone: from In Review (or In Progress if the PR is already merged)- Card already in the target column or later → no-op, print nothing.
- Check the WIP limit of the target column. If at limit, warn in one line and still move only if the caller said
force_wip: true; otherwise skip and report. - Move with
Glacier:update_card.
Card matching strategy
In order of reliability:
- Explicit
card_idfrom the caller - GitHub issue link —
Glacier:list_cards+Glacier:get_card_github_status, match the issue URL/number - Title fuzzy match — manual
/glacierruns only, ask for confirmation
No confident match → silent skip. Never create cards from a transition call.
Output formatting
Default
Single compact line, only on actual moves:
Glacier: "Stripe webhook retry logic" → In Progress
narrate: true (called from /implement-v)
↳ 🧊 Glacier: "Stripe webhook retry logic" → In Progress ✓
↳ 🧊 Glacier: move failed (continuing) — <one-line reason>
/implement-v prints the (pending) line itself before calling this skill. Print nothing for no-ops, unmatched cards, or disabled Glacier — silence is correct.
Manual operations (via /glacier)
- Status: board overview (cards per column, WIP limits, blockers)
- PR merged → Done: scan recent merges, move matching cards
- TODO scanning: scan
// TODO(glacier):comments in branch diff, create cards - Issue linking: link a GitHub issue to an existing card
MCP call pattern
EVERY Glacier MCP call must include workspace_id from GLACIER_WORKSPACE_ID (OAuth tokens are user-scoped, not workspace-scoped).
Example: Glacier:list_columns(project_id: $GLACIER_PROJECT_ID, workspace_id: $GLACIER_WORKSPACE_ID)
MCP tools used
Glacier:list_columns— column IDs and WIP statusGlacier:list_cards— find cards by projectGlacier:get_card— card detailsGlacier:get_card_github_status— verify GitHub issue/PR linksGlacier:update_card— move between columnsGlacier:create_card— from TODOs (manual only)Glacier:link_card_to_github— link issues (manual only)
Rules
- Always pass
workspace_idto every MCP call. - Never move a card without a confident match.
- Never regress a card.
- Respect WIP limits.
- One line per move. No stack traces.
- NEVER block the parent workflow. If anything fails, warn in one line and continue.