Repository navigation
feat(orchestrator): 1.4.0 operator regime (neutralized) - #67
Conversation
…er languages marked outdated Adds the generally valid part of a locally deployed 1.4.0 (sibling of 1.3.1): operator thinks/decides/verifies/integrates, execution goes to workers; model choice per worker; brief only with non-lookupable content; review before commit for user-judged results; active waiting on serial resources; foreign changes never discarded; readback at the destination; collecting leftovers. Personal/system-bound parts (names, incidents, tool specifics) stay local. - SKILL.md 1.3.1 -> 1.4.0 (DE), SKILL.en.md 1.1.0 -> 1.4.0 (operator section only) - es/fr/ja/ru/zh: localized "outdated" notice (as model-strategy) - registry/map regenerated Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LYHP4Js5ggpzQ78JXWRe5E
|
Welcome! Thanks for your first pull request in this repository. A maintainer will review it soon. Please make sure:
Thanks for contributing! |
…ofile pointer - "nach klarer Vorgabe" / "to a clear spec" qualifies the whole mechanics list again (as in the source wording), not only "building"; rework without a clear spec no longer reads as cheap-model work. - EN: the regime references the `burst` profile, which EN does not describe; point to the German SKILL.md for resource profiles and operating modes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LYHP4Js5ggpzQ78JXWRe5E
Second-model review (reviewer: Claude Opus 5.5; author: Claude Sonnet 5.5)1. Information fidelity: passed. I checked the operator-regime section of the local 1.4.0 line by line against this PR. All 9 points are here in neutral form: principle and exceptions, model choice per worker, brief content, review before commit, active waiting, foreign changes, readback, collecting leftovers, and why. Everything tied to a person, host, tool or incident (model names, browser and shell commands, lock file names, the serial tool and its time limit, readback examples, incident dates) is kept in the maintainer's local notes. Apart from the new section and the changelog, the body is byte-identical to 1.3.1: resource profiles, operating modes, whitelist and token automation are complete, and so is the 2. Neutrality: passed. A grep of the added lines for person, host, path, model and ticket references found nothing, and 3. Gates and conventions: passed.
4. Fixes (commit a699795):
Not blocking: the EN Verdict: approve and merge (squash). |
Neutralizes the generally valid part of the locally deployed orchestrator 1.4.0 (sibling of 1.3.1 on the 1.3.0 base, never in git) so the deploy HOLD can be lifted after merge.
Taken over from 1.4.0 (neutral wording)
New section "Operator-Regime" (DE
SKILL.md, ENSKILL.en.md), applies on user request or profileburst: operator thinks/decides/verifies/integrates, all execution to workers; explicit model per worker (cheaper for mechanics, stronger for builds/reviews); brief holds only non-lookupable content; review before commit for user-judged results (new file names per iteration); active waiting on serial resources; never discard foreign changes (rescue branch, certify/archive; security code reported, other model reviews); readback at destination; operator collects leftovers.Stayed local (not in this repo)
Person/model nicknames, memory file names, incident dates and project names, concrete tools (browser, time tool, memory CLI, lock file names, image agent). Kept in local memory
feedback_fable_nur_denker_delegieren; deployed 1.4.0 backed up locally.Other changes
Gates
registry/map --check, privacy_gate, skill_frontmatter_gate, changed_skill_gate, readme link gate green; pytest 461 passed.
language_gatereports 2 pre-existing deviations (migrate-rename, agents-bridge), unrelated and not in CI.Not deployed, no merge. Needs second-model review.
🤖 Generated with Claude Code
https://claude.ai/code/session_01LYHP4Js5ggpzQ78JXWRe5E