fix: feat: gardener formula — steps and recipes (#363) (#366)

Fixes #363

## Changes
Created formulas/run-gardener.toml with 7 steps: preflight (pull latest), agents-update (AGENTS.md watermark check), stale-pr-cleanup (ping/close inactive PRs), dust-bundling (group trivial issues into bundles of 3+), ci-health (detect systemic CI failures), blocked-review (triage blocked issues per #352), and commit-and-pr (single commit with all file changes). No memory, no journal. All git writes collected in the final commit-and-pr step; API calls happen during the run. Steps follow the established formula TOML format with needs dependencies.

Co-authored-by: openhands <openhands@all-hands.dev>
Reviewed-on: https://codeberg.org/johba/disinto/pulls/366
Reviewed-by: Disinto_bot <disinto_bot@noreply.codeberg.org>
This commit is contained in:
johba 2026-03-20 12:11:58 +01:00
parent 06f8e3fcba
commit 86070ceef7

302
formulas/run-gardener.toml Normal file
View file

@ -0,0 +1,302 @@
# formulas/run-gardener.toml — Gardener housekeeping formula
#
# Defines the gardener's complete run: grooming (Claude session via
# gardener-agent.sh) + CI escalation recipes (bash, gardener-poll.sh)
# + AGENTS.md maintenance + final commit-and-pr.
#
# No memory, no journal. The gardener does mechanical housekeeping
# based on current state — it doesn't need to remember past runs.
#
# Steps: preflight → grooming → blocked-review → ci-escalation-recipes
# → agents-update → commit-and-pr
name = "run-gardener"
description = "Mechanical housekeeping: grooming, blocked review, CI escalation recipes, docs update"
version = 1
[context]
files = ["AGENTS.md", "VISION.md", "README.md"]
# ─────────────────────────────────────────────────────────────────────
# Step 1: preflight
# ─────────────────────────────────────────────────────────────────────
[[steps]]
id = "preflight"
title = "Pull latest code"
description = """
Set up the working environment for this gardener run.
1. Change to the project repository:
cd "$PROJECT_REPO_ROOT"
2. Pull the latest code:
git fetch origin "$PRIMARY_BRANCH" --quiet
git checkout "$PRIMARY_BRANCH" --quiet
git pull --ff-only origin "$PRIMARY_BRANCH" --quiet
3. Record the current HEAD SHA for AGENTS.md watermarks:
HEAD_SHA=$(git rev-parse HEAD)
echo "$HEAD_SHA" > /tmp/gardener-head-sha
"""
# ─────────────────────────────────────────────────────────────────────
# Step 2: grooming — Claude-driven backlog grooming
# ─────────────────────────────────────────────────────────────────────
[[steps]]
id = "grooming"
title = "Backlog grooming — triage all open issues"
description = """
Groom the open issue backlog. This step is the core Claude-driven analysis
(currently implemented in gardener-agent.sh with bash pre-checks).
Pre-checks (bash, zero tokens detect problems before invoking Claude):
1. Fetch all open issues:
curl -sf -H "Authorization: token $CODEBERG_TOKEN" \
"$CODEBERG_API/issues?state=open&type=issues&limit=50&sort=updated&direction=desc"
2. Duplicate detection: compare issue titles pairwise. Normalize
(lowercase, strip prefixes like feat:/fix:/refactor:, collapse whitespace)
and flag pairs with >60% word overlap as possible duplicates.
3. Missing acceptance criteria: flag issues with body < 100 chars and
no checkboxes (- [ ] or - [x]).
4. Stale issues: flag issues with no update in 14+ days.
5. Blockers starving the factory (HIGHEST PRIORITY): find issues that
block backlog items but are NOT themselves labeled backlog. These
starve the dev-agent completely. Extract deps from ## Dependencies /
## Depends on / ## Blocked by sections of backlog issues and check
if each dependency is open + not backlog-labeled.
6. Tech-debt promotion: list all tech-debt labeled issues goal is to
process them all (promote to backlog or classify as dust).
For each issue, choose ONE action and write to result file:
ACTION (substantial promote, close duplicate, add acceptance criteria):
echo "ACTION: promoted #NNN to backlog — <reason>" >> "$RESULT_FILE"
echo "ACTION: closed #NNN as duplicate of #OLDER" >> "$RESULT_FILE"
DUST (trivial single-line edit, rename, comment, style, whitespace):
echo 'DUST: {"issue": NNN, "group": "<file-or-subsystem>", "title": "...", "reason": "..."}' >> "$RESULT_FILE"
Group by file or subsystem (e.g. "gardener", "lib/env.sh", "dev-poll").
Do NOT close dust issues the script auto-bundles groups of 3+ into
one backlog issue.
ESCALATE (needs human decision):
printf 'ESCALATE\n1. #NNN "title" — reason (a) option1 (b) option2\n' >> "$RESULT_FILE"
CLEAN (only if truly nothing to do):
echo 'CLEAN' >> "$RESULT_FILE"
Dust vs ore rules:
Dust: comment fix, variable rename, whitespace/formatting, single-line edit, trivial cleanup with no behavior change
Ore: multi-file changes, behavioral fixes, architectural improvements, security/correctness issues
Sibling dependency rule (CRITICAL):
Issues from the same PR review or code audit are SIBLINGS independent work items.
NEVER add bidirectional ## Dependencies between siblings (creates deadlocks).
Use ## Related for cross-references: "## Related\n- #NNN (sibling)"
Processing order:
1. Handle PRIORITY_blockers_starving_factory first promote or resolve
2. Process tech-debt issues by score (impact/effort)
3. Classify remaining items as dust or escalate
After processing, dust items are collected into gardener/dust.jsonl.
When a group accumulates 3+ distinct issues, create one bundled backlog
issue, close the source issues with cross-reference comments, and remove
bundled items from the staging file.
CRITICAL: If this step fails for any reason, log the failure and move on.
"""
needs = ["preflight"]
# ─────────────────────────────────────────────────────────────────────
# Step 3: blocked-review — triage blocked issues
# ─────────────────────────────────────────────────────────────────────
[[steps]]
id = "blocked-review"
title = "Review issues labeled blocked"
description = """
Review all issues labeled 'blocked' and decide their fate.
(See issue #352 for the blocked label convention.)
1. Look up the 'blocked' label ID (Gitea needs integer IDs for label removal):
BLOCKED_LABEL_ID=$(curl -sf -H "Authorization: token $CODEBERG_TOKEN" \
"$CODEBERG_API/labels" | jq -r '.[] | select(.name == "blocked") | .id')
If the lookup fails, skip label removal and just post comments.
2. Fetch all blocked issues:
curl -sf -H "Authorization: token $CODEBERG_TOKEN" \
"$CODEBERG_API/issues?state=open&type=issues&labels=blocked&limit=50"
3. For each blocked issue, read the full body and comments:
curl -sf -H "Authorization: token $CODEBERG_TOKEN" \
"$CODEBERG_API/issues/<number>"
curl -sf -H "Authorization: token $CODEBERG_TOKEN" \
"$CODEBERG_API/issues/<number>/comments"
4. Check dependencies extract issue numbers from ## Dependencies /
## Depends on / ## Blocked by sections. For each dependency:
curl -sf -H "Authorization: token $CODEBERG_TOKEN" \
"$CODEBERG_API/issues/<dep_number>"
Check if the dependency is now closed.
5. For each blocked issue, choose ONE action:
UNBLOCK all dependencies are now closed or the blocking condition resolved:
a. Remove the 'blocked' label (using ID from step 1):
curl -sf -X DELETE -H "Authorization: token $CODEBERG_TOKEN" \
"$CODEBERG_API/issues/<number>/labels/$BLOCKED_LABEL_ID"
b. Add context comment explaining what changed:
curl -sf -X POST -H "Authorization: token $CODEBERG_TOKEN" \
-H "Content-Type: application/json" \
"$CODEBERG_API/issues/<number>/comments" \
-d '{"body":"Unblocked: <explanation of what resolved the blocker>"}'
NEEDS HUMAN blocking condition is ambiguous, requires architectural
decision, or involves external factors:
a. Post a diagnostic comment explaining what you found and what
decision is needed
b. Leave the 'blocked' label in place
CLOSE issue is stale (blocked 30+ days with no progress on blocker),
the blocker is wontfix, or the issue is no longer relevant:
a. Post a comment explaining why:
curl -sf -X POST -H "Authorization: token $CODEBERG_TOKEN" \
-H "Content-Type: application/json" \
"$CODEBERG_API/issues/<number>/comments" \
-d '{"body":"Closing: <reason — stale blocker, no longer relevant, etc.>"}'
b. Close the issue:
curl -sf -X PATCH -H "Authorization: token $CODEBERG_TOKEN" \
-H "Content-Type: application/json" \
"$CODEBERG_API/issues/<number>" \
-d '{"state":"closed"}'
CRITICAL: If this step fails, log the failure and move on.
"""
needs = ["grooming"]
# ─────────────────────────────────────────────────────────────────────
# Step 4: ci-escalation-recipes — recipe-driven CI failure handling
# ─────────────────────────────────────────────────────────────────────
[[steps]]
id = "ci-escalation-recipes"
title = "CI escalation recipes (bash — gardener-poll.sh)"
executor = "bash"
script = "gardener/gardener-poll.sh"
description = """
NOT a Claude step executed by gardener-poll.sh before/after the Claude session.
Documented here so the formula covers the full gardener run.
gardener-poll.sh processes CI escalation entries from
supervisor/escalations-{project}.jsonl. Each entry is a dev-agent session
that exhausted its CI fix attempts and was escalated to the gardener.
The recipe engine (match_recipe function in gardener-poll.sh) matches each
escalation against gardener/recipes/*.toml by priority order, then executes
the matched recipe's playbook actions via bash functions.
Recipes (see gardener/recipes/*.toml for definitions):
- chicken-egg-ci (priority 10): non-blocking bypass + per-file fix issues
- cascade-rebase (priority 20): rebase via Gitea API, re-approve, retry merge
- flaky-test (priority 30): retrigger CI or quarantine
- shellcheck-violations (priority 40): per-file ShellCheck fix issues
- Generic fallback: one combined CI failure issue
Special cases:
- idle_timeout / idle_prompt: investigation issues (no recipe matching)
"""
needs = ["grooming"]
# ─────────────────────────────────────────────────────────────────────
# Step 5: agents-update — AGENTS.md watermark staleness check
# ─────────────────────────────────────────────────────────────────────
[[steps]]
id = "agents-update"
title = "Check AGENTS.md watermarks, update stale files"
description = """
Check all AGENTS.md files for staleness and update any that are outdated.
This keeps documentation fresh runs 2x/day so drift stays small.
1. Read the HEAD SHA from preflight:
HEAD_SHA=$(cat /tmp/gardener-head-sha)
2. Find all AGENTS.md files:
find "$PROJECT_REPO_ROOT" -name "AGENTS.md" -not -path "*/.git/*"
3. For each file, read the watermark from line 1:
<!-- last-reviewed: <sha> -->
4. Check for changes since the watermark:
git log --oneline <watermark>..HEAD -- <directory>
If zero changes, the file is current skip it.
5. For stale files:
- Read the AGENTS.md and the source files in that directory
- Update the documentation to reflect code changes since the watermark
- Set the watermark to the HEAD SHA from the preflight step
- Conventions: max ~200 lines, architecture and WHY not implementation details
6. Stage ONLY the AGENTS.md files you changed do NOT commit yet.
All git writes happen in the commit-and-pr step at the end.
7. If no AGENTS.md files need updating, skip this step entirely.
CRITICAL: If this step fails for any reason, log the failure and move on.
Do NOT let an AGENTS.md failure prevent the commit-and-pr step.
"""
needs = ["ci-escalation-recipes"]
# ─────────────────────────────────────────────────────────────────────
# Step 6: commit-and-pr — single commit with all file changes
# ─────────────────────────────────────────────────────────────────────
[[steps]]
id = "commit-and-pr"
title = "One commit with all file changes, push, create PR"
description = """
Collect all file changes from this run (AGENTS.md updates) into a single commit.
API calls (issue creation, PR comments, closures) already happened during the
run only file changes need the PR.
1. Check for staged or unstaged changes:
cd "$PROJECT_REPO_ROOT"
git status --porcelain
If there are no file changes, skip this entire step no commit, no PR.
2. If there are changes:
a. Create a branch:
BRANCH="chore/gardener-$(date -u +%Y%m%d-%H%M)"
git checkout -B "$BRANCH"
b. Stage all modified AGENTS.md files:
find . -name "AGENTS.md" -not -path "./.git/*" -exec git add {} +
c. Also stage any other files the gardener modified (if any):
git add -u
d. Commit:
git commit -m "chore: gardener housekeeping $(date -u +%Y-%m-%d)"
e. Push:
git push -u origin "$BRANCH"
f. Create a PR:
curl -sf -X POST \
-H "Authorization: token $CODEBERG_TOKEN" \
-H "Content-Type: application/json" \
"$CODEBERG_API/pulls" \
-d '{"title":"chore: gardener housekeeping",
"head":"<branch>","base":"<primary-branch>",
"body":"Automated gardener housekeeping — AGENTS.md updates.\n\nReview-agent fast-tracks doc-only PRs."}'
g. Return to primary branch:
git checkout "$PRIMARY_BRANCH"
3. If the PR creation fails (e.g. no changes after staging), log and continue.
"""
needs = ["agents-update"]