[loopx-agent]

Goal And Delivered Outcome

Restore the existing acceptance test for unfinished delegations across consecutive source-checkout CLI Turns. On 24027a34736b1e50c2eb431a6039009028e73d8d, all eight original cases fail before checking the delegation facts: machine-envelope consumers assume inline JSON, and native-reader consumers assume a bare loopx command. The native owner can also legitimately provide a fresh full capture on the second Turn.

The test now requests inline transport only for its machine child, executes the native producer's complete interpreter argv, and reads the current Turn's selected context through every page. The original semantic assertions remain; the test additionally carries a queued return and an unconsumed peer result alongside an already-read/adopted request whose linked Todo is done. All three remain visible on both Turns, and the collaboration receipt bytes remain unchanged.

• Basis: a reproduced public acceptance-test failure; intended base main.
• Related work: fix(collaboration): preserve complete readable returns #8139 owns readable-answer and return presentation. This independently reversible validation repair contains no duplicated product implementation.

Author Declaration

• Written by: model_agent — OpenAI GPT-6 / Codex, published as loopx-agent.
• Specification: manager-context unfinished delegation agenda, revision 24027a34736b1e50c2eb431a6039009028e73d8d.

Criterion	Disposition	Test evidence
Every managed Turn retains unfinished work after read/adoption and before return delivery or requester consumption	implemented	Two actual CLI guards per matrix case; explicit checks for all three unfinished states
Compact context is not a complete Inbox read or lifecycle transition	implemented	Required read route and incomplete-read marker retained; complete receipt snapshot unchanged
Same typed collaboration owner and existing privacy boundary	implemented	Real File/SQLite authority fixtures and existing context decoder; no production owner change

Python remains the existing test/transport consumer. There is no new decision owner, protocol vocabulary, frontend configuration or runtime language migration.

Scope And Continuation

One existing test file changes. A single subprocess adapter serves the explicit console and emitted native readers. Reader pages must advance monotonically and cover the complete UTF-8 byte range; the test no longer requires an incidental single-page result or unconditional warm delta. Parent context settings remain unchanged in both on and off cases.

This repairs source acceptance coverage. Installed desktop/WebView, actual model writing, online Lark delivery, requester adoption and full-cost qualification remain with the existing delivery work. No merge, installation, release or paid execution is included.

Validation

• Tested revision: e15a6566c49304b11e3967bbe4986a1c908f4276.
• Run state: finished. Input classes: synthetic.

Check kind	Result	Evidence / limitation
regression_parity	passed	Original baseline: 8 failures before semantic assertions. Repaired matrix: 16 cases across File/SQLite, host envelope/native files, unanswered/partial and parent context on/off; 32 actual CLI guards.
real_entrypoint / real_backend	passed	uv run --extra test python -m pytest tests/test_delegation_turn_context.py tests/test_delegation_remaining_work.py tests/test_inbox_pagination.py -q: 39 passed. Disposable local File and SQLite authority stores; no online provider or model.
static	passed	Canonical repository Ruff profile; configured mypy 17 source files; final changed-file public scan with zero errors/warnings; diff and bilingual RFC checks. The test module is outside configured mypy coverage.
static / qualification	passed	Standard premerge: five direct checks, no catalog selectors or manual holds. Goal strict quick premerge and exact-scope receipt verification pass.
manual	not_run	Full Python suite, Windows execution and installed/native model/provider qualification were not run locally. The existing POSIX-only native-reader mark is retained.

Exact-scope quality receipt: cqr_87133dea7f9ff5a10d5d, verified valid for fingerprint 87133dea7f9ff5a10d5dba83d4a55b8eb35651e7d890f02734d16c1adef08645; safe-fix allowed/applied in one bounded pass, zero blockers/warnings/advisories. Intermediate configuration failures remain failed observations and are superseded only by the later passing run.

Frontend / Visual Evidence

UI impact: none. The changed surface is the existing CLI acceptance test; no screenshot, packaged UI or installed-product improvement is claimed.

Type And Boundary

• Test update; host/control-plane acceptance integration.
• Public source contains synthetic fixture records only. No credentials, private conversation content, raw logs, internal links or local machine paths are included.
• No paired RFC changes. One DCO-signed commit. Independent review and CI remain separate from local qualification and merge authorization.