[ e-networld.de ]

Squad-Spec-Repository-Template

Vorlage für spezifikationsgetriebene Entwicklung mit KI-Agenten: ein gemeinsames Squad, eine Regelbasis für Claude, Codex und Copilot, Profile für .NET, Go und Node.

MITClaude Code · Codex · GitHub Copilot · .NET · Go · Nodev2026.11.12 · 15.11.2026
Ablauf von Plan über Test, Code und Review bis zum Pull Request

Überblick

Das Squad-Spec-Repository-Template bündelt alles, was ein Repository für die Arbeit mit KI-Agenten braucht: ein Team aus festen Rollen, eine einheitliche Regelbasis und das Gerüst für Dokumentation und GitHub. Es besteht aus einem stackneutralen Kern und je einem Profil pro Technologie-Stack. Genutzt wird es nicht über GitHubs Schaltfläche „Use this template“, sondern über den Skill adopt-template, der die Vorlage auf ein bestehendes oder neues Repository anwendet und es später wieder auffrischt.

Das Squad arbeitet über die Skills squad-issue für Fehlerbehebungen und squad-spec für neue Funktionen. Der Ablauf: planen, hinterfragen, Sicherheitsprüfung, Tests zuerst, Umsetzung bis mindestens 80 % Abdeckung, Format- und Analyzer-Schranke, Review-Schleife, Freigabe durch den Lead und schließlich der Pull Request.

CLAUDE.md für Claude Code, AGENTS.md für Codex und .github/copilot-instructions.md für GitHub Copilot sind ab der ersten ##-Überschrift identisch. Dasselbe gilt für die Skills create-pr, review-pr, squad-issue und squad-spec, die in allen drei Umgebungen vorliegen.

Die Dateien im Ziel-Repository gehören zu drei Arten: verwaltete Dateien werden bei jeder Auffrischung überschrieben, markierte Dateien werden neu aufgebaut und behalten die Projektblöcke, Seed-Dateien werden nur einmal angelegt und danach nie wieder angefasst. So bleibt der Kern stackneutral, und projektspezifisches Wissen geht bei einer Auffrischung nicht verloren. Platzhalter der Form {{TODO: …}} lässt config-check.py so lange scheitern, bis sie ausgefüllt sind.

Funktionen

Squad aus Lead, Devil's Advocate, Security, Tester, Dev, Code Officer und Reviewer
Skills squad-issue für Fehlerbehebungen und squad-spec für neue Funktionen
Ein Regelwerk für drei Agenten, CLAUDE.md, AGENTS.md und Copilot-Anweisungen sind identisch
Qualitätsschranken vor dem Pull Request für Format, Analyzer und mindestens 80 % Testabdeckung
Profile für .NET, Go und Node mit passendem Formatter, Analyzer und Coverage-Format
Dokumentationsgerüst mit Contributing, Architektur, Entscheidungsprotokollen und Spezifikationsvorlagen
GitHub-Gerüst mit PR- und Issue-Vorlagen, CI, CodeQL und Dependabot
Wiederholbare Aktualisierung über den Skill adopt-template, projektspezifische Inhalte bleiben erhalten

Releases

v2026.11.12 · 15.11.2026
  • [Core] squad-log.py --help says --continued-from takes the step of a logged continuation too (#426)
  • [Core] A Lead decision on a failed manual test that orders no code is stamped again instead of looping through the
  • [Core] After the Lead's decision on a failed manual test, squad-next.py holds the review until the work it ordered
  • [Core] squad-next.py counts a skeleton row's Files glob, directory or bare file name as changed when a matching
  • [Core] A plan-ordered review finding applied verbatim ([NON-BLOCKING, plan, verbatim], or an untagged verbatim row
  • [Core] A manual test that fails at step 7 goes to the Lead, who adds a task group; after the group's stamp
  • [Core] A task group a Lead decision adds in steps 7–8 is logged under its own steps 4–6, never 8 Fix; `squad-log.py
  • [Core] A group's coverage pass ends with a passing full stamp after its implementation, and a deferred test whose
  • [Core] Commits that only save work end with [skip ci]; the stamped, reviewed and ready heads never do, and
  • [Core] Members run Format in the foreground under a time limit, on the changed files where the formatter takes
  • [Core] The Lead checks every planned signature against the analyzer's limits, sweeps every list of a set the change
  • [Seed] dotnet ci.yml grants checks: write to the build job only (githubactions:S8233); .squad/stack.md lists

GitHub ↗

v2026.11.11 · 14.11.2026
  • [Core] A manual: check that names the Reviewer or Security runs in its step-8 round, not as a step-7 test, and plan-check.py refuses one that names a member outside the tier's review round; a Product Manager's manual check named only in the plan's prose holds the review, and an Orchestrator 7 Manual test row records a stand-in's (#396)
  • [Core] A member's report re-sent word for word after a task notification is part of its round, not a continued launch; members send their report once, with no background command left running (#397)
  • [Core] squad-log.py names --round 1 when a round is logged late; squad-next.py asks for the verbatim row before stamping a documentation-only fix of an open review round that no Dev or Tester fix followed (#398)
  • [Core] plan-check.py skips appended challenge and security notes in its placeholder check; notes the Lead answered in Challenge are not asked for again (the template's own Challenge text answers none) (#401)
  • [Core] Security's diff verdict counts its non-blocking findings (VERDICT: APPROVED | NON-BLOCKING n, else its tags), so tiers docs and patch decide them before the PR; the patch test is written on the unfixed head before the edits (#402)
  • [Core] squad-notes.py writes a report's control characters as text escapes, so appended notes never fail plan-check.py (#400)
  • [Core] The Dev's skeleton report lists the changed files from git status and only members read back from disk, and squad-next.py sends a file a skeleton task row names but the branch has not changed back to the Dev before tests first; a plan names the test seam of a resource bound over an external system (#399)
  • [Core] lesson-label.py counts only the tools the template ships to .squad/tools/ as managed, so a repository's own tool there is labelled squad; Template-managed files names the template blocks of .editorconfig as marked (#395)
  • [Core] result.md is listed in the work folder (specs/README.md, Work folder) and posted in the working record of tiers docs and patch (#394)
  • [Profile] The seeded CI of every profile runs decision-check.py --require-history --tags-fetched in a job decision-records, so a changed released record cannot merge, and a refresh adds the job to an existing CI; a CI merged from several profiles runs it once and renames a prefixed job in needs: too (#403)
  • [Profile] dotnet: the shared global analyzer config sets MSTEST0001, a rule reported without a location that no .editorconfig section reaches, to error; dotnet-rules.py writes a pending entry for such a rule into .squad/dotnet-rules-pending.globalconfig at a higher global_level, and dotnet_config.py checks that Directory.Build.props, Dockerfiles and .dockerignore carry both global configs (#393)
  • [Profile] dotnet: .squad/dotnet-rules-pending.globalconfig keeps the line ending of its checkout, so a CRLF repository's fresh clone passes the Analyzer gate; scope-check.py lets a product PR change it with squad_settings.py, from which it is generated (#406)

GitHub ↗

v2026.11.10 · 13.11.2026
  • [Skill] adopt-template asks for the Fork scope repository settings only in the pull request that adds
  • [Core] New lesson-label.py names a squad lesson's label from the files its fix changes (squad-upstream for a
  • [Core] squad-next.py reads a non-blocking count in the verdict clause only, so "no blocking or non-blocking
  • [Core] session-usage.py counts the harness's reply to a SubagentHandback call as part of the same round, so one
  • [Core] One Lead revision logged as the plan-check revise answers a challenge and a CHANGES_REQUIRED that came back
  • [Core] squad-next.py names the fresh Lead decide launch of step 6; a member's report is one final reply (#383)
  • [Core] A Reviewer verdict on a later head that adds only documentation to the stamped code counts; a criterion
  • [Core] Tier trivial keeps a task group's order when tasks.md puts the Tester first; a reason for the spec's
  • [Core] The Lead's result of tiers docs and patch is written verbatim to result.md with `squad-notes.py
  • [Core] The Tester keeps test code, table tests included, under the analyzer's cognitive-complexity limit (#384)
  • [Profile] dotnet: the Analyzer gate reports a diagnostic without a location (MSTEST0001) for its project when a
  • [Seed] dotnet UNIT_TESTS.md names the branch conditions the Coverage gate prints; go stack.md says gocognit

GitHub ↗

v2026.11.9 · 12.11.2026
  • [Docs] Mark PiMonitor as an active consumer
  • [Core] Apply a challenge note's exact plan edit when appending it
  • [Core] Take test names only from a cell's listed spans; check a waiting PR from a worktree
  • [Skill] Split a seeded process record the project extended in place
  • [Profile] Check that the image build sees the global analyzer config
  • [Core] Work the open squad lessons and harden the seeded workflows
  • [Core] Fix what the release rehearsal of the squad-lesson changes found
  • [Docs] Release v2026.11.9

GitHub ↗

v2026.11.8 · 12.11.2026
  • [Profile] Lower pending rules the dotnet baseline does not set
  • [Profile] Switch baseline rules off in Razor code and fix rollout defects
  • [Profile] Lower pending rules that only a section of paths sets
  • [Docs] Release v2026.11.8

GitHub ↗

Alle Releases auf GitHub ↗