Tracker Sync

GSDL's tracker write path. Takes an already-decided payload — slice titles, a comment body, a target state — and applies it to whichever tracker the project is configured for. This skill makes no judgment calls: it does not author comments, choose slice boundaries, or pick which items to close. Callers decide; this skill executes.

Tier

SMALL. Every operation here is: resolve an ID, POST a mutation, report the result. There is no reasoning to spend a larger model on, and tracker round-trips are frequent enough across the pipeline that the cost difference is real. The orchestrator spawns a SMALL subagent for each of these.

Non-interactive. Never ask the user anything. If credentials are missing, an ID won't resolve, or a mutation fails, return a BLOCKER describing exactly what failed and what the caller should do — including the exact text a human could paste into the tracker manually.

Provider Resolution

Read .planning/gsdl.config.md (and any per-project override) per the gsdl skill's config.md. The ## Tracker section decides everything:

Use the access method recorded in the config (mcp or api). If none is recorded, try MCP first, then the credential path, and write back whichever worked. Use the same method for every call in one invocation.

Whichever method you use, the constraints below are identical — MCP is not a relaxed path.

Operation A — Create Sub-Items (called from gsdl-decompose, Step 4)

Inputs

• PARENTITEMID — the item from seed.md, or none
• PROJECT_ID — Linear team ID / Jira project key / owner/repo / Notion database ID
• SLICES — an ordered list of { title, description }, already written by the caller

Invariants for every provider

1. Apply the configured labels — ## Tracker → labels (default gsdl), so agent-generated

items stay filterable. Resolve the label; create it once if the provider requires labels to exist first, then reuse it.

1. Parent every sub-item to PARENTITEMID when one was given. With none, create standalone

items only if the caller explicitly said to — otherwise return a BLOCKER.

1. Never assign a sprint, cycle, or iteration. The parent item is the planning unit; its slices

are not. Only pass one if the caller explicitly supplied it for a slice.

1. Never invent a status. New sub-items land in the project's default/backlog state.
2. Idempotency — if the caller passes an existing identifier for a slice, update that item

instead of creating a duplicate.

Linear

MCP: saveissue (or createissue) per slice with title, description, parentId, teamId, and labelIds. Omit cycleId.

API:

# resolve a label (create it on the team once if absent)
curl -s -X POST -H "Content-Type: application/json" -H "Authorization: $LINEAR_API_KEY" \
  -d '{"query":"{ issueLabels(filter: { name: { eq: \"gsdl\" }, team: { id: { eq: \"TEAM_ID\" } } }) { nodes { id name team { id } } } }"}' \
  https://api.linear.app/graphql

# create the sub-issue
curl -s -X POST -H "Content-Type: application/json" -H "Authorization: $LINEAR_API_KEY" \
  -d '{
    "query": "mutation($input: IssueCreateInput!) { issueCreate(input: $input) { success issue { id identifier url } } }",
    "variables": { "input": {
      "teamId": "TEAM_ID", "parentId": "PARENT_ITEM_ID",
      "title": "SLICE_TITLE", "description": "SLICE_DESCRIPTION_MARKDOWN",
      "labelIds": ["LABEL_ID"]
    } }
  }' https://api.linear.app/graphql

If a configured label name resolves to a child of a label group, use the child's ID — the group itself is not assignable. If a configured label doesn't exist and can't be created at team level, return a BLOCKER rather than creating a stray workspace-level label.

Jira

API ($JIRABASEURL, $JIRAEMAIL, $JIRAAPI_TOKEN):

curl -s -X POST -u "$JIRA_EMAIL:$JIRA_API_TOKEN" -H "Content-Type: application/json" \
  -d '{"fields":{
        "project":{"key":"PROJ"},
        "parent":{"key":"PROJ-118"},
        "issuetype":{"name":"Sub-task"},
        "summary":"SLICE_TITLE",
        "description":{"type":"doc","version":1,"content":[ /* ADF */ ]},
        "labels":["gsdl"]
      }}' \
  "$JIRA_BASE_URL/rest/api/3/issue"

Jira descriptions are Atlassian Document Format, not markdown — convert the slice description (headings, bullets, checkboxes) into ADF blocks. If the project has no Sub-task issue type, fall back to a Task linked to the parent with "Relates", and say so in the return — don't silently change the hierarchy.

GitHub Issues

gh issue create --repo owner/repo \
  --title "SLICE_TITLE" --body "SLICE_DESCRIPTION_MARKDOWN

Part of #PARENT_NUMBER" \
  --label gsdl

GitHub has no universal parent field. Link both ways: reference the parent in the sub-issue body, and append the new issue to a task list in the parent's body:

gh issue view PARENT_NUMBER --repo owner/repo --json body
# append "- [ ] #NEW_NUMBER SLICE_TITLE" under a "## Slices" heading, then:
gh issue edit PARENT_NUMBER --repo owner/repo --body-file -

Use the repo's sub-issue API instead if it is available on the account — prefer native parenting when the API supports it, and report which mechanism you used.

Notion

Create one page per slice in the tracker database, with:

• the title property set to SLICE_TITLE
• the slice description as page content blocks
• a relation property pointing at the parent page (whatever the database calls it — resolve from the

database schema, don't assume a property name)

• the configured labels in the database's multi-select property, if one exists

Return the page URLs. If the database has no relation property to parent with, return a BLOCKER naming the properties it does have.

none

No API call. Confirm the caller has written the slices to tasks.md and return the slice titles with identifier: none so the caller omits tags. Do not create placeholder IDs.

Return

Each slice's identifier and url, in the order given, so the caller can tag tasks.md. On partial failure, say exactly which slices were created, with their IDs, so the caller can retry the rest without duplicating.

Operation B — Comment + Transition (called from gsdl-execute Step 6 and gsdl-close-loop Step 9)

Inputs

• ITEM_IDS — one or more item identifiers
• COMMENT — the exact markdown body to post, written by the caller
• TARGET_STATE — the state name to move to (e.g. Done, In Review), or none to comment only

If TARGET_STATE is none, skip the transition entirely — post the comment and return. none is not a state name; never match it against real states and never treat it as a resolution failure.

State names vary per team, project, and workflow — never hardcode an ID. Resolve TARGET_STATE case-insensitively against the item's real available states. If nothing matches, return a BLOCKER listing the available names; do not guess a near-match.

Linear

# available states
curl -s -X POST -H "Content-Type: application/json" -H "Authorization: $LINEAR_API_KEY" \
  -d '{"query":"{ issue(id: \"ENG-124\") { team { states { nodes { id name type } } } } }"}' \
  https://api.linear.app/graphql

# comment
curl -s -X POST -H "Content-Type: application/json" -H "Authorization: $LINEAR_API_KEY" \
  -d '{"query":"mutation($input: CommentCreateInput!) { commentCreate(input: $input) { success comment { id url } } }","variables":{"input":{"issueId":"ITEM_ID","body":"COMMENT_MARKDOWN"}}}' \
  https://api.linear.app/graphql

# transition
curl -s -X POST -H "Content-Type: application/json" -H "Authorization: $LINEAR_API_KEY" \
  -d '{"query":"mutation($id: String!, $input: IssueUpdateInput!) { issueUpdate(id: $id, input: $input) { success issue { identifier state { name } } } }","variables":{"id":"ITEM_ID","input":{"stateId":"STATE_ID"}}}' \
  https://api.linear.app/graphql

MCP equivalents: listissuestatuses, savecomment, saveissue.

Jira

# comment (body is ADF)
curl -s -X POST -u "$JIRA_EMAIL:$JIRA_API_TOKEN" -H "Content-Type: application/json" \
  -d '{"body":{"type":"doc","version":1,"content":[ /* ADF */ ]}}' \
  "$JIRA_BASE_URL/rest/api/3/issue/PROJ-119/comment"

# available transitions, then apply one
curl -s -u "$JIRA_EMAIL:$JIRA_API_TOKEN" "$JIRA_BASE_URL/rest/api/3/issue/PROJ-119/transitions"
curl -s -X POST -u "$JIRA_EMAIL:$JIRA_API_TOKEN" -H "Content-Type: application/json" \
  -d '{"transition":{"id":"TRANSITION_ID"}}' \
  "$JIRA_BASE_URL/rest/api/3/issue/PROJ-119/transitions"

Jira transitions are named by the transition, not the target status. Match TARGET_STATE against each transition's to.name.

GitHub Issues

gh issue comment 128 --repo owner/repo --body "COMMENT_MARKDOWN"
gh issue close 128 --repo owner/repo --reason completed     # for a done-like TARGET_STATE

GitHub issues have only open/closed. Map any done-like state to close, and any other state to a status label (e.g. status: in review) only if that label already exists — otherwise comment only and say in the return that the state could not be represented.

Notion

Comment → append a paragraph block to the page (or use the comments API if available). Transition → set the Status/Select property, resolved by name from the database schema.

none

Append to .planning/[project-name]/worklog.md:

## [YYYY-MM-DD] [Slice N.0 title or item reference]

**State:** [TARGET_STATE, or "comment only"]

[COMMENT body verbatim]

---

Create the file if absent. Also ensure the matching checkbox in tasks.md reflects a done-like state. The record still gets written — "no tracker" never means "no record".

Return

Per item: comment URL (or the worklog path), the resulting state name (or unchanged (comment-only)), and any failure verbatim.

Operation C — Read an Item

For callers needing current item data (state, labels, parent, project/team ID, suggested branch):

Return the fields verbatim — do not summarize or interpret.

Rules

1. Never author content — post the comment body you were given, verbatim. If it's missing,

that's a BLOCKER, not an invitation to write one.

1. Never invent a status — resolve TARGET_STATE against the item's real states and report the

options if it doesn't match. none means comment-only: no resolution, no transition, no blocker.

1. Every method enforces the same invariants — MCP, API, the local none path, and the

manual-paste blocker all carry the labels and the no-sprint rule. A path that quietly drops one is a bug in this skill.

1. Never move a parent item to a terminal done state unless the caller explicitly passed one —

GSDL's default is parent → the configured review state, sub-items → the configured done state.

1. Never assign a sprint/cycle/iteration unless the caller passed one.
2. Report partial failure honestly — if 3 of 5 sub-items were created, say exactly which 3, with

their IDs, so the caller can retry the rest without duplicating.

1. Report provider workarounds — a Jira Task-instead-of-Sub-task fallback, a GitHub state

that could only be a label, a Notion property that didn't exist: say so in the return rather than letting the caller assume a clean write.

1. Write back the working access method to the config the first time you resolve it.