[xiaoan17]

Scope

One ceiling in the real-CLI output budget, plus the measurement that justifies it.

4b2b97b7b (#7557, "Preserve complete context in shared quota diagnostics")
replaced candidate-collection compaction in
loopx/control_plane/quota/cli_projection.py with preservation. The synthetic
model-portfolio budget for that contract was already re-qualified upstream in
98ca82923 (#7583), which raised 40,000 → 90,000 and added preflight checks that
candidates survive. This PR covers the remaining surface that still measures the
old contract.

Current gap

On main at e19d94ca0:

tests/control_plane/test_cli_output_budget.py::test_real_cli_output_stays_inside_baseline_and_growth_contracts
FAILED
AssertionError: quota_should_run/multi_agent/json emitted 25836 chars; baseline ceiling is 25500

The real quota should-run packet grew because the hot path now keeps the
capability candidate context. Nothing bounds that growth except this ceiling,
which was calibrated against the compaction contract #7557 removed.

Measurement (same fixture, base vs head)

Base is 4b2b97b7b^; head is e19d94ca0. Both arms run the existing budget
fixture.

Scenario	Base JSON	Head JSON	Delta	Ceiling
small	22,618	23,265	+647	25,000 (unchanged, 1,735 headroom)
crowded	34,732	35,836	+1,104	42,000 (unchanged, 6,164 headroom)
multi-agent	24,925	25,836	+911	25,500 → 27,000

Markdown is unchanged for crowded and multi-agent (8,148 / 6,439) and moves
7,479 → 7,481 for small, all inside their unchanged ceilings. The per-Todo
allowance (300 chars/unit) and fixed semantic allowance (9,700) are untouched
and still cover crowded-minus-small growth (12,571 measured).

Change

loopx/control_plane/testing/cli_output_budget.py — raise only the exceeded
multi-agent JSON ceiling, 25,500 → 27,000, with the base/head measurement
recorded at the spec. No other surface, growth allowance or line budget changes.

No runtime behavior changes: the constant governs a qualification fixture, not
production transport.

Validation

tests/control_plane/test_cli_output_budget.py                38 passed, 1 skipped
tests/control_plane/test_actual_default_model_behavior_portfolio.py
                                                            50 passed (upstream fix)
ruff check (changed file)                                    passed

The base/head arms were measured on the same fixture; the original 25,500
overflow is recorded here as a failure, not reclassified as a pass.

[mergify]

This pull request has merge conflicts with main and cannot be merged
until they are resolved. Please rebase or merge the base branch, @xiaoan17.

Choose the remote for the base repository, not an out-of-date fork.
For a fork clone, first inspect git remote -v; upstream must point
to https://github.com/loopx-project/loopx.git. If it is absent, add it
with git remote add upstream https://github.com/loopx-project/loopx.git.
Then run:

git fetch upstream
git rebase upstream/main
# Resolve each conflict, git add the resolved files, then git rebase --continue.
git push --force-with-lease origin HEAD

For a same-repository clone whose origin points to
https://github.com/loopx-project/loopx.git, use origin instead of
upstream for fetch/rebase. If you prefer merging the base, use
git merge <base-remote>/main and push normally.

Keep the DCO Signed-off-by trailer on every commit when you rebase.
https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[mergify]

This pull request has merge conflicts with main and cannot be merged
until they are resolved. Please rebase or merge the base branch, @xiaoan17.

Choose the remote for the base repository, not an out-of-date fork.
For a fork clone, first inspect git remote -v; upstream must point
to https://github.com/loopx-project/loopx.git. If it is absent, add it
with git remote add upstream https://github.com/loopx-project/loopx.git.
Then run:

git fetch upstream
git rebase upstream/main
# Resolve each conflict, git add the resolved files, then git rebase --continue.
git push --force-with-lease origin HEAD

For a same-repository clone whose origin points to
https://github.com/loopx-project/loopx.git, use origin instead of
upstream for fetch/rebase. If you prefer merging the base, use
git merge <base-remote>/main and push normally.

Keep the DCO Signed-off-by trailer on every commit when you rebase.
https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[loopx-agent]

Reviewer: model_agent · gpt-6.1-sol · OpenAI · runtime_reported · reasoning_effort=max

未发现当前精确 head 722f2b9 的阻断问题，结论为 APPROVE。

动机 / Motivation

运行 CLI 输出回归检查、维护 Agent 决策上下文的开发者。

同一组三 Agent fixture 的完整 JSON 已有25,836字符，原25,500上限使基线回归检查失败；本 PR 只把该上限调到27,000，保留实际返回的全部内容。

实际基线失败已复现，当前原矩阵通过；六组真实console的字段、正文和命令保持，其他预算及负向拒绝仍有效。

本次不调整生产传输硬限制、模型执行预算、冻结性能标准，也不声称真实模型效果或长程收益已经通过。 这里恢复的是有意义的回归警戒线，不把测试变绿当作生产性能或模型效果的证据。

改动思路 / Approach

这是测试回归预算的单点校准；完整上下文已有消费者，删除候选正文来满足旧阈值会损失信息，沿既有预算owner调整一个数值即可。

本 PR 只交付多 Agent JSON回归预算和配对实测注释，未改变CLI运行时、fixture、行数或增长预算。

实际调用关系是完整console输出进入measure_cli_output，再由原baseline、行数、语义键和增长门判断；没有新owner、开关或执行效果。用文字足以说明这一处校准，不额外绘图。保留25500会持续报告已实测的同一失败，统一提高所有预算或删除正文都没有本次证据支持。

具体改动 / Changes

spec_ref: docs/development/testing-and-quality.md；spec_revision: e19d94c。对应既有 Budget Failure Decisions：用相同工作负载和计量口径归因，保留有用上下文及原失败，区别回归预算与硬/冻结边界。本项为implemented；不是改写实验通过线。

完整差异为一个既有模块+5/-1，仅一个常量和实测注释。生产AST在27000替回25500后逐项相同；fixture和原断言均未改。

关键代码讲解

• CLI_OUTPUT_BUDGET_SPECS 仅将multi_agent JSON阈值从25500改到27000，Markdown仍7000；行数650、每Todo增长300及固定语义增长9700均保持。
• measure_cli_output 按实际stdout记录字符、UTF-8字节和行数并解析JSON，本身未变，不先删候选或正文。本组合fixture当前均为ASCII，所以字符数与字节数恰好相同，不能因此混用一般口径。
• assert_cli_output_baseline 继续独立拒绝超字符、超行数和缺语义键；原矩阵仍检查JSON/Markdown、规模增长及完整bootstrap路由。

同物理fixture、同运行环境、三个不可变revision的实际console测量如下；这张表只表达测量，不证明模型收益：

场景/口径	候选保留前5a7cc8195	当前基线e19d94ca0	PR head
small JSON字符	22618	23265	23265
crowded JSON字符	34732	35836	35836
multi_agent JSON字符	24925	25836	25836
multi_agent Markdown字符	6439	6439	6439

历史到当前multi_agent增量911已复现，原25500超限失败也由原测试在基线复现。当前base与head的六对完整输出仅有逐字段列明的生成时钟、rollout event id和File导入版本摘要差异；选择、候选正文、完整诊断和命令及其他字段保持。历史版本的完整原文与当前投影元数据差异保留，不冒称历史到当前全部等价。

对主干的风险 / Risks

最强风险是用提上限掩盖无价值膨胀。当前实际headroom为1164字符，只调整这一场景；未减少输入规模或输出内容，所有其他AST与预算保持。真实输出上构造27001字符、652行和删除selected_todo三种变体，均由原helper拒绝。这些是预算helper的负例，不能冒充新的生产故障注入。

PASS：预算与actual-default portfolio模块88项、标准premerge六direct及11selected、改动文件Ruff、配置17文件mypy、完整TS typecheck、diff/DCO/双语门和严格精确CQR。原矩阵在基线明确FAIL为25836>25500，在head通过。原测试另有1个native temp namespace无法容纳depth-12布局的条件skip；其fixture、条件和断言未变，不把skip或一次本地绿变成原生Windows资格。未查询或等待CI。

私有验证驱动初次用了不存在的pytest节点，未执行测试；另一次误以为所有场景都有可选capability_gate而KeyError。修正读取实际结构后才完成上述六对证据；这些驱动失败保留，不算产品测试或PR回归。

语义与 CI 对齐

沿用既有测试预算owner，保持local_only，无新状态词汇、配置或authority。既有候选完整性测试和真实console对照承担语义证据；预算门通过本身不证明文本充分或模型采用。没有松动生产硬限制、执行预算或冻结实验标准。

我的整体评价 / Verdict

APPROVE，当前单点回归预算维护目标达到。long_horizon与user_experience均为preserved：原CLI接受、选择、诊断与恢复入口保持，无新增设置、输入或确认轮次。实际CLI上下文未缩水，尚未测量模型长期效果或总成本收益。

这是测试回归预算的单点校准；完整上下文已有消费者，删除候选正文来满足旧阈值会损失信息，沿既有预算owner调整一个数值即可。 本 PR 只交付多 Agent JSON回归预算和配对实测注释，未改变CLI运行时、fixture、行数或增长预算。

有界维护已应用，不需要新框架或人为后继。已检查 quota CLI、预算consumer及现有Markdown输出；没有页面、渲染组件或交互布局变化，视觉评审不适用。 原生Windows、安装态和真实模型采用/长程效果未在本轮验证。 合并、安装及其他公共发布权仍由各自owner决定。

English verdict: APPROVE — 722f2b9. Matched real-console measurements reproduce the 25,836-character baseline failure and justify the single 27,000-character regression ceiling. Other guards remain effective; current validation passes with one preserved environment-dependent skip. Model and long-horizon qualification remain separate.

[xiaoan17]

Superseded by main — closing

The change this PR proposes is already on main, from a different session:

• 791f297d2 — test: calibrate complete multi-agent quota fixture budget —
  raised the same multi_agent.json ceiling in
  loopx/control_plane/testing/cli_output_budget.py from 25,500 to 26,500, on the
  same unchanged 18-Todo / 3-Agent fixture, with the same measurement: 25,836
  JSON characters across 96 matched surface/mode measurements and no semantic or
  route differences.
• 544aa0fee (test(heartbeat): calibrate complete thin JSON scope budget #7614) then calibrated the related heartbeat max_chars surface on
  top of it (4,500 → 4,600).

This PR proposes 27,000 for the same fixture instead. Keeping both would leave
a duplicate ceiling for the same measurement and would re-litigate a number that
main already carries with its own evidence.

The remaining CI failure on this PR was not caused by it: chat-bundle-browser (1)
fails on the shared base because
apps/presentation/dashboard/smoke/delegation-search-browser-smoke.mjs intercepts
/api/chat/app-steward and hides the fixture's later source refusal, so the
expected alert can never appear. That is fixed separately in #7599.

Closing rather than rebasing. Reopen if the 26,500 ceiling proves too tight for
the fixture; the measurement and the rationale are recorded in 791f297d2.

---

Superseded by main — closing

This PR's change is already on main: 791f297d2 raised the same
multi_agent.json ceiling from 25,500 to 26,500 on the same fixture, with the
same measurement (25,836 JSON characters, 96 matched surface/mode measurements,
no semantic or route differences), and #7614 calibrated the related heartbeat
max_chars surface on top. This PR proposes 27,000 for the same fixture instead.

The one failing check here was a base failure, not this PR's: the delegation
search smoke hides the fixture's source refusal behind a page route, so the
expected alert can never appear. Fixed separately in #7599.

Reopen if 26,500 proves too tight; the measurement lives in 791f297d2.

[xiaoan17]

Superseded by 791f297 on main (same fixture, same 25,836-char measurement, ceiling 26,500). Base CI failure fixed separately in #7599.