Repository navigation
feat(agentic-engineering): make a parked PR owned work, never a terminal state - #699
Conversation
…nal state Dependency-update PRs stranded for weeks once a deployment parked them: runs read the park as finished, nobody worked the blocker or re-checked the upstream release it waited on, and parked PRs filled an ecosystem's open-PR limit so no new update could open. The agent definition, the portfolio-maintenance run loop and its PR-driving procedure now say: an in-portfolio blocker is finishing work, claimed and worked ahead of new pulls; an upstream blocker is re-checked at least every 7 days with an engineering route weighed each time; a dependency PR parked for 14 days gets a recorded decision (route or why none, the upstream event awaited, a re-check date) and reaches the maintainer once past 30 days; an update ecosystem at or near its open-PR limit with parked PRs is a stall; and a mis-park is unparked and driven. The Cadence can set other bounds. The surveyor reports parked/blocker/decision facts, unknown when unreadable. A new contract test pins the clauses with mutation controls. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
🤖 Generated with Claude Code @coderabbitai review |
Action not completed
|
|
No actionable comments were generated in the recent review.
|
… agent The maintainer corrected the rule on 2026-11-13: a dependency PR parked past the bound is not handed to him. Drop the 30-day escalation. Past the bound the agent takes engineering routes itself (adapt, fork or patch, isolate or replace, contribute upstream, split the update) and keeps taking them, recording what it tried, the next route and when it looks again. A maintainer-ask is only for a genuine question of direction, never a handoff: it states the default the agent will act on, and the work continues meanwhile. No timer makes contacting the maintainer the outcome. The contract test pins the new clauses and forbids the escalation sentence. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…-PR rule From the local review round on #699: - automation-owned PRs stay outside the parked-PR rules, as the survey already leaves them; - the surveyor judges a park only from the PR's own complete label list against the park label the Portfolio map names, never a census row, and measures age from the first park; - `blocker=decision` covers a scope or design blocker, and a blocker that cannot be read or a record that names none never justifies unparking; - a lapsed route record makes the next route finishing work, never lifts the park or merge hold; - an ignore is recorded on its blocker and removed when it clears; - Dependabot security updates have their own limit, so a version-update stall does not block them; - summaries match the full stall rule, every PR row shape carries the park fields, and the README names the park label and the overridable bounds. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…viour, not survey mechanics Review round 2 on #699 returned findings of the kind round 1 already raised (park-date sourcing, label reads, ignore tracking): the surveyor was specifying park-fact mechanics that belong to each deployment's own park mechanism. Reshape once, per agent rule 3: - the surveyor and the survey procedure keep only the behaviour: a parked PR stays in the survey with its pentad and is never reported as finished; the deployment supplies park facts; - the PR-driving rule adds what the behaviour needs: a parked PR is never merged while parked; routes live in the one park record, edited in place; a blocker clears only when its fix is verifiably available to the product (closed-unfixed or not-planned still holds); an ignore is tracked on an in-portfolio issue until removed, never by writing upstream; the stall rule reads each entry's effective limit; routes continue until the PR merges or closes for a recorded reason. The contract test follows both linked references and pins the new clauses. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
🤖 Generated with Claude Code Reshape record (agent rule 3), head The local review at
The next review round counts from this head. |
…rk ends From the first review round after the recorded reshape on #699: - the pre-merge recheck now reads "not parked", parking disarms auto-merge and leaves any merge queue, and the surveyor says its MERGE-READY covers the pentad only; - a scope or design decision clears once it is recorded and its change is on the PR or merged; - a parked PR closes only when a newer update supersedes it or as a deliberate rejection, never to free a slot or clear the bound, and a close the tool records as an ignore is tracked as one; - the rules cover only PRs the Trust gate lets you drive, an unreadable age counts as past the bound, the look-again date never trails the upstream re-check, and a fork goes through the deployment's repository-creation path; - the contract test forbids broader handoff phrasings. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
main released agentic-engineering 7.7.0 (#697, the Numina MCP server), so this branch's entry moves to 7.8.0 above it. The only conflict was the changelog. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
🤖 Generated with Claude Code @coderabbitai full review |
Action performed
|
…DED (#4998) * feat(scripts): report a long-parked or unowned dependency PR as STRANDED The parked digest read every parked PR as finished rung-1 work. On 2026-11-12 all 11 open Dependabot PRs and 6 of 10 Renovate PRs were parked and the digest said actionable=0, while nine ksail gomod PRs waited weeks on one issue and filled the ecosystem's open-PR limit. `--parked-digest` now reports an exact dependabot[bot] or renovate[bot] PR whose park would otherwise hold as `STRANDED`, counted as actionable (exit 1), when: - UNCLAIMED: an open in-org issue its record names is neither parked itself nor referred to by an open non-bot PR (an assignee is not a claim); - ESCALATE: parked over 30 days (--escalate-max-age-days) on anything but an authority blocker; - NO-DECISION / DECISION-LAPSED: parked over 14 days (--stranded-max-age-days) without "awaits <event>" and one "re-check <YYYY-MM-DD>" from today to the bound in the record result. The park date is the record comment's created_at; a missing or unreadable date, author, or blocker label set is UNKNOWN (exit 2). The merge policy, work selection and run procedure now say a park settles the PR, never its blocker, and describe the new row. Aligns the deployment with devantler-tech/agent-plugins#699. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(scripts): never make contacting the maintainer the outcome of a park timer Maintainer correction 2026-11-13: "You own the code, I do not." Drop STRANDED reason=ESCALATE and --escalate-max-age-days. Past --stranded-max-age-days every parked dependency PR, whatever its blocker kind, needs a current route record in its result: "tried: <route and what it showed>", "next: <route>" and one "re-check <date>" within the bound (NO-ROUTE / ROUTE-LAPSED otherwise). An authority park is no longer exempt, because a maintainer ask never pauses the work. The merge policy says ownership stays with the agent and a maintainer-ask is only a direction question with a stated default. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(scripts): close the review round's gaps in the STRANDED verdict From the local review round on #4998: - a labelled PR whose author is not named is UNKNOWN under --input as under --org (an absent, null or empty user printed PARKED for an old dependency park); - an unusable created_at year is UNKNOWN instead of a huge age; - tests now pin the record comment's own date among several comments, a +02:00 offset that crosses midnight, and a closed blocker named after an unclaimed one; - the guide says only a PARKED row is a terminal park, that age runs from the first park, what "no open PR references it" actually checks and that agent-claim tips are not read, that a stall stops version updates (Dependabot security updates have their own limit), and that an unreadable blocker never justifies unparking. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(scripts): let only a working PR hold a blocker, and inventory the new park rules From local review round 2 on #4998 and the plugin reshape in agent-plugins#699: - a PR that is itself parked no longer holds the issue it refers to (one parked on that same issue kept the dependency PRs behind it PARKED); tested with a mutation control; - tests pin that UNCLAIMED outranks a current route record, and the token boundaries, empty values, digit count, year and case of tried:/next:/re-check; - the guide says UNCLAIMED reads upstream-kind records, a parked PR is never merged while parked, a blocker clears only when its fix is verifiably available (closed-unfixed or not-planned still holds), an ignore is tracked on an in-org issue and never written upstream, the limit is the effective one and only the full limit stops updates, `next:` must be engineering, and "no timer" is scoped to these rules; - merge-policy.sensors.tsv inventories the twelve new rule sentences (MP-101..MP-112). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * refactor(scripts): read the route record as an allow-listed segment grammar Review round 3 on #4998 found a second date hidden after a space, comma or dash slipping past the re-check regex's boundaries; the third round of free-text boundary findings. Reshape once, per agent rule 3: the record's result is split on ";" and only these segment shapes count: exactly one starting "tried:" and one starting "next:", each carrying a letter or digit, and exactly one that is exactly "re-check <YYYY-MM-DD>". Any other segment mentioning re-check is no route. Tests cover the space, comma, dash, prose, non-leading token, punctuation-only value, two next routes and a stray re-check beside a valid one, with a mutation control. Also from that round and agent-plugins#699's latest round: the guide says parking disables auto-merge and dequeues, a decision blocker clears once recorded and applied, a park closes only when superseded or the release is unwanted (never to free a slot), STRANDED rows count in actionable=, and help names the dependency-bot exclusion; MP-102 is marked prose-only and MP-113 inventories the new closing rule. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(scripts): accept natural route-record prose and say how to keep one Nits from the first review round after the reshape on #4998 (no P0/P1): - only re-check as a word of its own voids a route record, so "re-checked upstream" in prose no longer reads as a hidden date; a re-check date may end in a period; both tested; - the guide says to keep one tried:/next: pair edited in place, how a bare #N counts, that sub-issue PRs do not hold an epic, and to park and unpark only through the helper; - compose --result help and --parked-digest help state the segment grammar; the BlockerClaim comment names the parked and dependency-bot exclusions. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(scripts): allow re-check only in the route record's own segment CodeRabbit on #4998: a date inside a tried: or next: value was not seen, because those prefixes were matched before the value was inspected. The whole result now uses re-check as a word exactly once, and that use must be the exact re-check segment; occurrences are counted without consuming their neighbours, so two side by side both count. Tested with a mutation control. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Why
The maintainer asked in an interactive session on 2026-11-13: "I think some dependabot PRs got stuck as well. I would like the Agentic Engineer agents to also pick those up, fix them, and get them merge, so we do not have stranded PRs in the portfolio." Later the same night they corrected the escalation part: "It is fine you use the maintainer ask feature, but do not consider it a handoff. You own the code, I do not. I am just here to help with questions and direction when you cannot figure it out on your own. And I really want you to be able to, so I prefer no contact."
The role already treats bot dependency PRs as first-priority work and forbids calling them "self-managing". It said nothing about a PR parked through a deployment's blocked-work mechanism, so deployments read a park as a terminal state. On 2026-11-12 every open Dependabot PR in the portfolio (11) and 6 of 10 Renovate PRs were parked. Nine ksail gomod PRs waited on one blocker, the oldest since 2026-10-05, and they filled that ecosystem's open-PR limit of 10, so no new Go update could open.
What changes
A parked PR is owned work, never a resting place (
references/driving-prs.md, summarised in the agent definition's rule 4 and the run loop):parked(from the first park, never restarted by a relabel),blocker(owner/repo#n|upstream|decision|none) anddecisionfacts on every PR row shape, orunknownwhen they can't be read.Plugin version 7.7.0 → 7.8.0 (main released 7.7.0 for the Numina server in #697; merged in), with a changelog entry and refreshed desired-state digests.
Validation
parked-pr-ownership-contract.test.sh(new): 61 checks across the agent definition, the skill with its linked reference, and the surveyor. Removing any clause fails the test. So does adding any forbidden sentence (a terminal-state park, the removed "raise it once through Maintainer channels" escalation, or the overstated "security fixes included"), or dropping the skill's link to the reference. shellcheck is clean../scripts/validate-manifests.sh,check-plugin-version-bump.sh origin/main HEAD,plugin-changelog.sh check origin/main HEADandcheck-instruction-size.sh: pass.skill-owner.shreportsportfolio-maintenanceasLOCAL(authored here).1efa3bb(the surveyor keeps only the behaviour; the deployment supplies park facts); round 3 after it raised three P1s, fixed in6e3d547. The final round at the current head follows as a GitHub review.Deployment alignment: devantler-tech/monorepo#4998.
🤖 Generated with Claude Code