Call the gitea-tools MCP directly instead of the parked gitea-agent #11

Merged
oleks merged 1 commits from fix/10-dead-agent-refs into main 2026-08-18 16:06:51 +03:00
6 changed files with 10 additions and 10 deletions
+1 -1
View File
@@ -1,6 +1,6 @@
{
"name": "cicd-insights",
"version": "1.4.9",
"version": "1.4.10",
"description": "CI/CD development best practices via structural insight — pipeline hygiene, blast-radius sizing, staleness, secret/env sprawl, trigger discipline, structural integrity, and retry-tag cleanup for tag-triggered CI, on a single repo or a whole fleet. Starts with pipetree (a Woodpecker CI topology scanner) and grows to other project-observation tools over time.",
"author": {
"name": "oleks",
+2 -2
View File
@@ -240,5 +240,5 @@ ______________________________________________________________________
caches by `version`), and every bump needs its `v<version>` tag.
- **Reference issues in full `owner/repo#N` form**, rendered as links to
`https://git.oleks.space/<owner>/<repo>/issues/<N>` in user-facing text.
- **File bugs via `anxious:issuer-agent`**, batched at wrap-up, not one
dispatch each.
- **File bugs via the `delivery-tools` MCP** (`prepare-filing`, then
`file-issue`), batched at wrap-up, not one dispatch each.
+2 -2
View File
@@ -167,8 +167,8 @@ git for-each-ref --sort=creatordate \
```
or, for a repo with no local clone at all, the same `<date> <tag>` shape
from `cluster:gitea-agent`'s `list_tags` (paginate; sort oldest-first).
**`cluster:gitea-agent`'s `delete_tag` is the last-resort deletion path**
from the `gitea-tools` MCP `list_tags` (paginate; sort oldest-first).
**`gitea-tools`' `delete_tag` is the last-resort deletion path**
— only when no local push-capable clone exists for that repo; it costs one
MCP call per tag, which is exactly the inefficiency `bin/retry-tag-sweep`
exists to avoid (confirmed by the user, 2026-08-01: delegating 52 tag
@@ -8,7 +8,7 @@ first — exactly what this produces for a local clone:
--format='%(creatordate:iso-strict) %(refname:short)' refs/tags
For repos with no local clone, get the same shape from gitea-tools
list_tags (name + commit + created date) via cluster:gitea-agent, sorted
list_tags (name + commit + created date) via the gitea-tools MCP, sorted
oldest-first, and feed it in the same "<date> <tag>" form.
A tag whose name ends in a bare numeric suffix (v3.3.6-19, chart-v0.1.4-2)
@@ -363,13 +363,13 @@ an assumed string is worth nothing.
This is a live gap, not a hypothetical: `cluster:ci-agent` has no
`branch_protection_read` tool, and during the founding investigation the
fact-gathering had to be re-dispatched to `cluster:gitea-agent`, which reached
it via `tea api`. Tracked as
fact-gathering had to be re-dispatched to the since-parked gitea agent, which
reached it via `tea api`. Tracked as
[oleks/claude-plugin-cluster#250](https://git.oleks.space/oleks/claude-plugin-cluster/issues/250).
Delegation targets, in order of preference:
- `cluster:gitea-agent` — owns git.oleks.space reads, has `tea api`.
- The `gitea-tools` MCP directly — `branch_protection_read` is on it.
- Any agent with `mcp__plugin_cluster_gitea-tools__branch_protection_read`.
- The user, with the exact read recipes from
[references/context-naming.md](references/context-naming.md) — a read-only
@@ -59,5 +59,5 @@ evidence-cited or marked UNVERIFIED.
Ask a session (with the skill loaded) to diagnose a stuck PR while playing an
agent that lacks branch-protection read access.
Expected: it delegates the read (naming `cluster:gitea-agent` or the documented
Expected: it delegates the read (naming the `gitea-tools` MCP or the documented
recipes) instead of guessing, citing oleks/claude-plugin-cluster#250.