Repository navigation
Publish independent Multi-Agent V2 arrangements and optional orchestration tools #1
Description
Outcome: maintain independently installable public Codex Multi-Agent V2 arrangements with minimal configuration, insertable instructions, optional standalone skills/hooks and a standard plugin contribution area. Public content is English with a supplementary Portuguese guide. Owner: rafaelob.
Acceptance: WHEN an arrangement is installed in a clean supported Codex environment, THE SYSTEM SHALL expose its declared roles and support parent-mediated fresh-context delegation without private dependencies, backed by demonstrated runtime evidence.
Published: PR #3 merged into main at 6568646. Version 0.2.0 provides Advanced (15 roles), Balanced (12), Lean (8), total-thread caps 5/4/3. Configuration and responsibility transfers are documented. Parent-only handoffs preserve implementation/review separation.
Evidence: catalog 3 arrangements/35 TOMLs; component preflight 2 skills/0 plugins; 37 existing behavioral tests passed locally; push and PR checks passed all eight jobs on Windows/Linux and Python 3.11/3.14. Commit-port added-line secret scan found nothing. Initial public-history Gitleaks audit is recorded in earlier comments; no new whole-history guarantee is claimed.
Remaining: complete native coverage of the revised role/model allocation, concurrency saturation and slot release are unverified. Historical native hook/writer evidence and the Windows sandbox limitation remain explicitly documented. No new paid inference or internal environment changes were performed for this revision. This issue remains open for runtime acceptance, not unfinished publication.
Next: maintainer chooses whether to authorize additional native validation; the temporary publication mission is closed after main publication and handoff. Do not change credentials, global PATH or sandbox controls as an implied next step. Authorization: maintainer explicitly requested commit/push to main, notification and mission closure in the active conversation, as recorded in issue comment 5651040807.
Activity
rafaelob commentedon Sep 12, 2026
Implementation progress
Owner: rafaelob. Next: complete the runtime checks and final review, then publish the verified branch and PR. This issue remains open; no completion claim is made.
rafaelob commentedon Sep 13, 2026
Validation checkpoint — publication remains blocked
Implemented locally on publish-v2-arrangements:
Evidence:
Remaining: successful native file artifact and independent review, complete role exposure evidence, live concurrency boundary/slot-release evidence, publishing checkpoint and hosted CI. Sanitized results are in docs/runtime-evidence.json and docs/validation.md in the local branch; these files are not yet published. No commit/push/PR or full acceptance is claimed.
Next: repair the minimal local sandbox/runner failure before repeating paid model tests. The local runtime maintainer has received the bounded reproduction; rafaelob retains ownership of this issue and the public release.
rafaelob commentedon Sep 13, 2026
The user authorized correcting the local runtime blocker (transcript:01a097d5-01e1-7311-95a7-641419acd926, message: "corrija te autorizo"). Work has resumed from the no-model sandbox echo reproduction. The sandbox will remain enabled; only evidence-backed, scoped corrections will be applied. Publication remains pending the required runtime acceptance.
rafaelob commentedon Sep 13, 2026
The minimal sandbox reproduction is now green without changing security controls. Setting only the test process TEMP and TMP to an owned empty directory changed the identical sandboxed echo from repeated 20/40-second timeouts to exit code 0; the native log measured START-to-SUCCESS at 66 ms. The failing process accumulated approximately 136,000 other (non-read/write) I/O operations over six seconds, whereas the successful process exited with 2,849 total. This identifies temporary-directory-dependent initialization, not a proven internal ACL subroutine. The console-deadlock hypothesis is superseded. The test harness is being corrected to use a fresh owned temp directory per process; native delegated artifact and review checks are being rerun before publication.
test: resolve temporary paths before permission injection (#1)
rafaelob commentedon Sep 13, 2026
Published on main through merged PR #2: commit 140fc57. Main CI passed all four Linux/Windows and Python 3.11/3.14 jobs: https://github.com/rafaelob/fleet-codex/actions/runs/34733756810 . A fresh remote-mirror audit with checksum-verified Gitleaks 8.30.1 found no secrets in all advertised refs (4 reachable commits, including 3 content-bearing commits), the 83-file main snapshot, or commit metadata. Two synthetic positive controls were correctly detected; no authentication/private-runtime paths appeared anywhere in reachable history. The merge tree equals the reviewed branch tree. This is measured pattern-based assurance, not a claim about hidden/unavailable remote objects or arbitrary undetectable secrets. The preview is published; the documented native-runtime limits remain unchanged.
rafaelob commentedon Sep 13, 2026
Follow-up scope: refresh Advanced and Lean role allocations and add Balanced Delivery, with sanitized role instructions and lower active-child caps of 4/3/2 respectively. Keep 19 distinct role contracts per arrangement; no live-home changes or new paid runtime experiments. Acceptance: WHEN one of the three arrangements is selected, THE catalog SHALL provide coherent role TOMLs, minimal V2 configuration, matching instruction snippets and documented allocation/cap tradeoffs without private dependencies. Existing runtime evidence remains historical and does not certify this revision. Local adaptation is proceeding; final source-owner validation confirmation remains pending. The initial PR #2 is already merged into main at 140fc57; the previous next-step paragraph is superseded by this follow-up.
rafaelob commentedon Sep 13, 2026
Final follow-up scope supersedes the 19-roles-per-arrangement proposal: Advanced 15 (4 Astra/4 Sol/7 Luna), Balanced 12 (2/2/8), Lean 8 (1/1/6), total-thread caps 5/4/3. Removed responsibilities have explicit owners and every child returns only to the parent; no child-to-child dependency. Three packages and 35 TOMLs pass catalog validation; 37 existing behavioral tests pass. Fresh native coverage is unmeasured and documented, not represented as passing. USER explicitly directed commit/push to main, notification and mission closure in this conversation; publication includes only sanitized catalog content.
rafaelob commentedon Sep 13, 2026
Revision 0.2.0 supersedes the prior equal-size roster proposal: Advanced 15 roles (4 Astra/4 Sol/7 Luna), Balanced 12 (2/2/8), Lean 8 (1/1/6), with total-thread caps 5/4/3. All routing is parent-mediated; removed responsibilities have explicit owners. Commit 25e0e2e is in PR #3. Catalog validation covers 3 arrangements/35 TOMLs and all 37 existing tests passed. No secrets were found by the commit-port added-line scan. Native coverage remains historical/unverified for this new matrix; this issue remains open. Next: merge PR #3 after CI and report its main SHA. The maintainer explicitly requested commit/push to main, notification and mission closure.
rafaelob commentedon Sep 13, 2026
Follow-up correction requested by the maintainer: insertable instructions must use Delegation and orchestration with concise subsections, contain no agent counts or runtime-validation link, correctly invoke the optional orchestration skill, and keep optional Pragmatic Programmer guidance in its own engineering section. Acceptance: WHEN either snippet is inserted, THE instruction section SHALL preserve parent-only delegation and complete fresh-context briefs without embedding configuration inventories or runtime-evidence navigation. Documentation-only scope; models, role cards, caps, hooks and permissions remain unchanged.
rafaelob commentedon Sep 13, 2026
Documentation correction published as 0.2.1: PR #4 merged into main at bfed8eb. All six snippets now use Delegation and orchestration, concise subsections and inline Skill-tool invocation with native fallback. No capacity section, agent counts or runtime-evidence links remain. Optional Pragmatic Programmer guidance is separate under Engineering decisions. Parent-only delegation remains because the target V2 implementation explicitly ignores V1 max_depth. All 37 tests and eight push/PR CI jobs passed; no executable runtime/model/cap change or new native validation is claimed. Publication correction complete; the existing runtime-acceptance limitation remains open.
rafaelob commentedon Sep 13, 2026
Closed at the explicit request of the maintainer in the current conversation. Catalog publication is complete: version 0.2.1 is on main at bfed8eb through merged PR #4. Closing this publication issue does not certify unperformed native runtime checks; existing limitations remain documented in docs/validation.md and docs/runtime-evidence.json.