Repository navigation
Description
What changed
gh 2.100.0 (2026-09-03) reaches the whole issue-relationship graph natively,
including the blocked-by / blocking half that issue-deps owns a curl script
for. But this is a gh-only capability, and the skill must say so, because
the environment where gh is absent is exactly the one the script was written
for.
Everything below was measured, not read off help text: gh version 2.100.0,
skill 0.15.0, probes run against a throwaway repo.
The client matrix, corrected
The skill's current table says sub-issues and closed_by_pull_requests are
reachable "the MCP only". In the MCP this session was given, neither is.
| Relationship | gh 2.100.0 |
This session's MCP |
|---|---|---|
| blocked-by / blocking — list | --json blockedBy,blocking |
counts only (see below) |
| blocked-by / blocking — write | --add/--remove-blocked-by, --add/--remove-blocking |
none |
| sub-issue / parent — list | --json subIssues,parent |
counts only |
| sub-issue / parent — write | --add/--remove-sub-issue, --parent, --remove-parent |
none |
| which PR closes an issue | --json closedByPullRequestsReferences |
no |
| issue type | --json issueType, --type, --remove-type |
no |
mcp__github__get_issue returns issue_dependencies_summary and
sub_issues_summary — counts with no members:
"sub_issues_summary": {"total": 0, "completed": 0, "percent_completed": 0},
"issue_dependencies_summary": {"blocked_by": 0, "total_blocked_by": 0,
"blocking": 0, "total_blocking": 0}So the MCP can tell you an issue has two blockers and cannot tell you which
two. The tool surface here is create_issue, get_issue, update_issue,
add_issue_comment, list_issues, search_issues — there is no
sub_issue_write, no issue_read, no issue_write, and update_issue takes
no parent_issue_number. The skill names all of those as the MCP path; on this
server they do not exist.
This matches reports from other sessions in the Claude cloud environment that
could not create these relationships at all.
Verified gh behaviour
Probes on a scratch repo with three issues:
--add-blocked-by 2on Scaffold the plugin: manifests, pr skill, anti-license #1 → Scaffold the plugin: manifests, pr skill, anti-license #1blockedBy=[2], and Port OpenTofu GitHub config and GHA CI from release-please-projected-releases-action #2blocking=[1].
Verified from both ends.--add-blockingworks.gh issue edit 3 --add-blocking 1→
feat: scaffold the plugin with the pr skill and an anti-license #3blocking=[1], Scaffold the plugin: manifests, pr skill, anti-license #1blockedBy=[3,2]. The skill says an edge can be stated
"from the blocked side only" becausePOST .../dependencies/blocking404s.
That remains true of REST, butghoffers both directions and inverts it
client-side. The reasoning is sound; the conclusion no longer describes the
tool a reader has.- Refusals are honest and exit non-zero. Self-edge → "Target issue cannot
be the same as the source issue (addBlockedBy)". Missing issue → "Could not
resolve to an Issue with the number of 9999". Duplicate → "Target issue has
already been taken". All exit 1, and none mutate the graph. issue_idvs#numberdoes not exist in this path.--add-blocked-by
takes a number or URL, so the script's central trap is closed by the client.
One scripting trap worth documenting
gh issue edit ... | tail swallows the status — $? is then tail's. I hit
this mid-probe and briefly believed gh exited 0 on failure. It does not.
Anything checking these writes must not pipe before reading $?.
What must NOT be dropped
-
A read against a pull request still answers empty rather than refusing.
gh issue view <pr> --json blockedByreturns{"nodes":[],"totalCount":0}—
indistinguishable from an issue with no edges.ghdoes not close this trap,
sorefuse_pull_request's reasoning survives the transport change and the
skill should keep saying that emptiness here is not evidence. -
The judgement half. Propose from evidence and confirm; never assert; a
sweep of the issue list hunting for edges is the manufacturing failure rather
than the skill working. Untouched by any of this, and the part that did the
real work in my session — it correctly stopped me inventing an edge between
two issues that merely shared a subject.
Suggested shape
Three branches, chosen on what the environment actually has:
gh>= 2.100.0 — use the flags above. Simplest, and closes the id trap.gholder, or absent butGITHUB_TOKENpresent —scripts/issue-deps.sh.
Load-bearing, not legacy: the script's header recordsghmissing on a
Claude Code web worker (2026-09-06) where the token is present.- MCP only — say plainly that the relationship graph is not writable, and
that a summary count is all a read gets. Better than a session discovering
it by failing.
Worth a version probe in the skill, since 2.100.0 is recent: mine was 2.86.0
(2026-01-21) until mid-session, and the failure mode is an unhelpful
Unknown JSON field: "issueType" that does not hint at the version at all.
The Expiry section should also widen — it says to delete the script "the day
the MCP exposes them", but the capability arrived through gh instead, and the
MCP still has not.
Context
Found while using the skill to replace prose dependency lines with structural
edges across 15 open issues in a private repo. The skill held up: the graph
turned out already complete, and nothing was invented.
Activity
[-]issue-deps: gh 2.100.0 exposes the whole graph, including blocked-by writes[/-][+]issue-deps: gh 2.100.0 writes the whole graph; the MCP cannot[/+]jmcvetta commentedon Sep 9, 2026
Taken. Work lands on claude/undertake-129-cy7rns.
Model:
claude-fable-5-1(configured and served agree).Session: https://claude.ai/code/session_01GD6mX5vZRcuFi9DiWv9Gtm
Generated by Claude Code