Repository navigation
Conversation
This was referenced Nov 2, 2026
github-actions
Bot
added
the
risk:T3
Keep Chief proposals separate from Maintainer and Owner approval. Preserve retired role context and room history without reassigning existing Tasks or inferring worker exit.
Keep exact latest Task-role native bindings and context parts, refuse unknown exit and missing handles, and use committed role catalogs when pinning execution rights. Preserve original Workflow lifecycle and legacy DAG workspace behavior.
Keep logical identity separate from artifact responsibility and SDK read-only policy. Preserve retired active claims while blocking new work until Chief selects an active target.
Preserve structured handoff identity through native delivery and restore archived context only after committed activation. Guard the actual DSH single-resource return path and keep historical implementation evidence outside the bounded active policy.
Collaborator
Author
|
Written-by: Codex Merged current #2175/#2184 into the original branch. Current composition: 130 initial targeted tests; after integration fixes,33 policy-budget/harness checks, core lint and actual DSH native return probe pass. Current full verify NOT rerun. Actual runtime reproduced logical Bot vs physical worker conflation in input budget fencing; regression/fix retained. Complete mandatory rules remain under the unchanged384k cap after completed service narratives are archived verbatim behind active memory links. No instruction source removal or silent truncation. Current head: c16a25e. Remaining composed-stack validation stays tracked under the original issue/PR. No main merge/deployment/control-plane/frontend changes. |
This was referenced Nov 3, 2026
Base automatically changed from
feat/2175-collaboration-contracts
to
delivery/ai-centric-development-service
November 3, 2026 09:14
Collaborator
Author
zy95-12
marked this pull request as ready for review
November 3, 2026 09:14
zy95-12
merged commit Nov 3, 2026
ecfd9b8
into
delivery/ai-centric-development-service
4 of 10 checks passed
zy95-12
added a commit
that referenced
this pull request
Nov 3, 2026
Written-by: Codex Latest composition at54a6bb5a8: current execution/governance/context merged;175 targeted tests and core lint pass. Current full verify NOT rerun; earlier full numbers below refer to their older named heads. Actual DSH native return with real Chief intake passed before the final stale-decision fence; final composed runtime validation remains pending. Requirements version changes now stop an old Attempt through the original pool lifecycle, and canonical claim revisions fence late transfer/split decisions before/after awaited authority. Actual SDK subprocess fixture covers version-change stop with acknowledgement/process exit. Custom Developer-capability parent resumes use child-source admission and producer Workspace mapping. Red regressions and self-review recorded. Still Draft; no main merge/deployment/frontend/control-plane change or independent Review. Refs #2179; stacked on #2225 (4101275). Head4826723ccf02b159ccdd7ca0cbbb27ebffe631d0. This is the original implementation PR for this issue; remaining acceptance stays here/dependent stages, not a replacement PR. New canonical shared Project Tasks start as proposals. Chief verifies goal, project impact and acceptance before a committed Effect can start execution. Requirements are versioned; each Attempt pins the full revision, so a scope-only change invalidates old evidence and creates a fresh rework snapshot. Requirements changes enter the existing Task inbox. A failed Attempt blocks the Task commitment instead of ending it. Member change requests ask Chief to judge revision/cancellation/reopening; public HTTP cannot submit Chief or role authority. Generic board transitions, human completion and legacy Workflow assign cannot bypass the new canonical contract. Persisted pre-contract Task delivery behavior remains covered separately. A role may split independent child proposals after confirmed source exit. The parent retains its Task, logical Session and working summary; an unsealed Attempt/workspace continues, while a sealed candidate requires one successor Attempt, releases its execution claim/quota, then waits for Chief-adjudicated child results. Failed/cancelled children do not satisfy waits. One stable outbox/Handoff request resumes the original parent role after all conditions are satisfied; children never close the parent or prove integration. Existing Effects, pools and Task construction are reused. Tests cover concurrent resume, restart replay, unconfirmed exit, failed commits and preservation of previous summaries. Actual DSH testing exposed two relevant defects: the resident native Chief used stale facts between jobs (fixed in original #2225), and the legacy Workflow gateway counted Chief intake as a development slot. The quota guard now shares released-parent semantics and excludes Chief control jobs. The runtime fixture exercises actual native Chief intake and controls only disposable host-verification gates, so queue/revocation checks do not depend on model timing or a framework delay. Validation at this head: full `TZ=UTC npm run verify` **4593 passed,0 failures,0 skips**, including4 real DSH runtime checks with deterministic inference. Selected129 Task/Chief/pool/delivery tests passed. Lifecycle coverage98.44%lines/77.97%branches; child-source module97.56%lines/76.60%branches; merge-metadata module100%lines/91.43%branches. Baseline-red evidence covers missing lifecycle, complete requirement revision and Attempt-vs-Task failure. Existing complete-role-policy budget guard caught growing PROJECT.md history: current invariants remain, exact dated reports moved to linked docs/history without raising the limit or truncating guidance. Architecture/subsystem/API notes, PROJECT/ROADMAP/LESSONS, changelog and generated patches are included. README rewrite remains last in #2189. Chief dependency judgement now supports version-bound waivers and accepted replacements with evidence; cancellation/reopening notifies Chief, and stale replacements reblock integration. The private pre-input port commits accepted child Git refs and source admissions through the existing Effects before reserving parent execution. Only matching Chief acceptance, candidate/test/Review evidence and current authority can admit sources. Public configure permission remains unchanged; moved refs, active writers, failed commits and invalidated results are refused. Parent sealing requires every required source's actual ancestry. Actual Git multi-source tests exposed first-only MERGE_HEAD inspection; all pending parents are now bounded and verified for local worktrees/resource clones. These tests use explicit deterministic child-acceptance fixtures, not a provider-built full loop. Remaining issue acceptance: full real-provider parallel collaboration, post-resume dependency-plan interruption/reconciliation, native physical Session restoration/inflight stop #2184, normalized atomic storage #2182, complete public receipts/history queries #2185, and alternate legacy creation/cutover compatibility #2186. Cancelled children remain blocked safely until resolved; this PR does not claim those remaining paths are complete. No new real-provider full Chief/child-development acceptance is claimed. No merge/deploy/restart, control-plane/UI edit or interruption of original frontend Task40/PR48. Draft; author self-review is not an independent Reviewer PASS.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Written-by: Codex
Latest composition at c16a25e: Merged current #2175/#2184 into the original branch. Current composition: 130 initial targeted tests; after integration fixes,33 policy-budget/harness checks, core lint and actual DSH native return probe pass. Current full verify NOT rerun. Actual runtime reproduced logical Bot vs physical worker conflation in input budget fencing; regression/fix retained. Complete mandatory rules remain under the unchanged384k cap after completed service narratives are archived verbatim behind active memory links. No instruction source removal or silent truncation. Prior full-check counts below belong to their explicitly named older heads, not this updated tip. Still Draft; composed Task/state/API acceptance remains in progress. No deployment, merge or independent Review.
Refs #2177; stacked on #2210. Current head: 4101275.
Bound Project/Chief/Task-role context through model/Harness budgets, full-input checks and scoped source pages with hashes/cursors. Reuse existing memory, archives, participation and Task-role Session records for relevant retrieval, versioned working summaries, bounded history views and unverified execution experience. Failed consolidation preserves the last valid summary; model experience cannot overwrite human-controlled facts or role policy. Structured native handoff input is JSON and passes current claim/resource/input-authority checks. Codex has a host-selected tool history output cap while SDK compaction defaults remain runtime-managed.
Sequential native Chief intake testing in dependent #2179 exposed a real stale-context problem: changing Bot system-prompt metadata did not guarantee an already resident Session received current job facts. This PR now delivers the current protocol and scoped facts in the actual turn input, checks its complete budget, revalidates actor/delegation before dispatch, and restores only its own prompt metadata. The same resident Session remains in use. A baseline-red/two-job regression protects this behavior; no fresh Session hides the failure.
Validation at current head: full
TZ=UTC npm run verifypassed 4560 tests, 0 failures, 0 skips, including four actual DSH runtime checks with deterministic inference. Agent suite43/43. Before the native follow-up, selected context/delivery/pool/Harness suites176/176 and the four unchanged context modules measured100% line coverage, branch73.47–91.54%; not complete branch coverage. Updated architecture/subsystem docs, PROJECT/ROADMAP/LESSONS, changelog and generated patches.Separately, actual Codex gpt-6.1-sol validation used the existing operator profile and runtime source a9ff63b: two fresh workers fixed fixture code, passed independent host tests, resumed the identical backend Session after a changed caller prompt, read their own work summary and recovered a conversation-only marker absent from repository files. Both confirmed stop/exit before release;9 checks passed. That proof names its source revision; it is not a new current-head full Chief/pool workflow, observed SDK compaction event, actual provider-window measurement or native Task-role physical restoration.
Remaining acceptance: native physical restoration #2184, storage normalization #2182, public protocol integration #2185 and full task-flow acceptance in #2179. Unknown capacity remains explicit. No production deploy/restart, control-plane/UI edits or frontend task interruption. Draft; author self-review is not an independent Reviewer PASS. README rewrite remains last in #2189.