Repository navigation
Description
Drafted by an agent on behalf of @Dong1017 (moyuu2025); item list derived from repo state at b5756cd561e8b12b68899d9dfa6411feab4669f5.
Goal
Refresh README.md against the repository as it is today. This issue lists only
the places where the README no longer matches the checked-out branch — section,
current wording, what the repository says instead, the suggested replacement,
and the evidence. No README edit is made here; tick the checkboxes as items are
fixed.
How this list was produced
- Audit date: 2026-11-20. Checkout
b5756cd561e8b12b68899d9dfa6411feab4669f5
(Make profiling optional in service-reuse campaigns (#521), 2026-11-20), after
git fetch origin. - README audited:
README.md, 310 lines, blob
61b9756ceb54375b4244ee436ee27868b94afc15, last committed
61303fa8150472dfbda62ec8d77908bc05c73f3b(2026-11-17).README.mdis unchanged
from the first audit head493416520; this revision re-checked every item a
reviewer challenged at this head, and re-read all eleven linked upstream PR
statuses from the API. - The package was installed in a clean venv (
python -m venv .venv+
python -m pip install .) and the commands the README prints were run at the
first audit head:inferrsi find 'Qwen-Image-2.1 Ascend memory bottleneck',
inferrsi context --repo-type vllm_omni --backend cuda --output context.md,
inferrsi context --repo-type vllm_omni --backend ascend --scenario diffusion.image_generation --output ascend-context.md, plusinferrsi --help
andinferrsi benchmarks list. A wheel was built to check the filename the
README tells users to download. This revision re-ran the two import probes and
the registration/doc reads behind F3 and M1; the clean-venv command log is
carried over unchanged. - Every repo-relative file, image, doc anchor and external link referenced by the
README was checked at the first audit head: all external links returned HTTP
200, all referenced files and images exist, and all five in-repo doc anchors
resolve. - Repo-wide sweep this revision: the retired module names and both
docs/assets/modular-*.svgembed points, reported under "Repo-wide check".
本轮修订(2026-11-20)
This revision answers the two review comments. What changed, and why:
- F2 corrected — the previous revision was wrong. It said
#9698,#9745
and#9283"are now open PRs" and proposed dropping(draft).
@zengchuang-hw is right: all three arestate=open, draft=true, so(draft)
is correct and stays. Only#10184changed — it merged on 2026-11-16 — so its
label goes(open)→(merged). The check date is now 2026-11-20. - M1 corrected. @zengchuang-hw is right:
findregisters only
CannbotChannel; MindIE has its own CLI and reader but is not afind
channel. M1 now states the one currently wired channel, its per-checkout
authorization boundary, and the MindIE status, each with file:line evidence. - F3 promoted out of "factual errors" into its own correctness-defect item
(@Dong1017). It is fixable by README text and does not wait for the diagram
decision; @zengchuang-hw supports fixing the old diagram and compatibility
description. - F1, F4 and S1 merged into one architecture edit region (@Dong1017), so a
single PR changes the prose, the inline claim and both diagrams together
instead of three PRs that conflict. - F4 remedy narrowed per the non-blocking note: a directory move is not an
architecture-boundary rename, so F4 now linksdocs/architecture.mdrather
than deriving a new boundary name from the package layout. - Checkboxes and priority added (@Dong1017 Bootstrap CLI-first RSI experiments and source-grounded diffusion knowledge #1, feat: bootstrap CLI experiment workbench and diffusion knowledge #2): item IDs
F1…M4are now
checkboxes under Must / Next / Optional. - Audit provenance unified (@Dong1017 feat: portable RSI repo scaffold (copyable layout + four-lane CI) #5): one date, the full 40-character
head, and the README blob; items are located by quoted text as well as by line
number, since line numbers drift before anyone edits. - Repo-wide sweep added (@Dong1017 docs: design draft → scaffold test design map (RSI + ProductResearch locks) #6); it found one further stale
reference, listed as D1 under "Repo-wide check". - Not adopted in this revision, with reasons: the advisory CI anti-drift job
and the PR A–D slicing (@Dong1017 Acceptance: Qwen-Image-2.1-Turbo / B300×4 campaign #3 and "Suggested slicing") are a separate
work item, not a README edit — our position is in the review comment. A
follow-up issue for the "Not checked" items is also a separate tracker entry;
this issue does not create it.
Priority
Checked items refer to the implementation verified at 23be5b7 in PR #533. GitHub closes this issue on merge.
Must fix
- F1 — architecture paragraph names non-existent top-level
benchmark/andharness/paths. - F2 — results-table PR label and check date are stale (
#10184merged). - F3 — compatibility panel promises old import paths that now fail (correctness defect; the text fix does not wait for the asset).
- F4 — "Harness × Model × Infra" no longer matches the tree; link the architecture doc.
Next
- S1 — the directory-map SVG shows a tree that is gone.
- S2 — News stops at
v0.2.0; six milestones have shipped since.
Optional
- M1 — no README mention of the external knowledge channel or its authorization boundary.
- M2 — no "check your target" (
doctor) step and nobenchmarks listpointer. - M3 — accepted
--backendvalues are never listed. - M4 — no configuration/revocation note.
- D1 —
docs/contracts/current-design.mdstill names a retired module (found by the sweep).
1. Factual errors
F1, F4 and S1 all live in the one architecture block (README.md:244–282); edit
them in one PR. F3 is split out below because it is a correctness defect, not
staleness.
| # | README location | Current text | Repository state today | Suggested change | Evidence |
|---|---|---|---|---|---|
| F1 | README.md:263–267 ("🧩 Architecture", offline improvement controller paragraph) |
"Its acceptance rules live in benchmark/improvement.py and its search progression in harness/improvement.py" |
There are no top-level benchmark/ or harness/ directories. The files are src/inferrsi/benchmark/improvement.py and src/inferrsi/harness/improvement.py; the composed controller the Apps lane uses now lives in src/inferrsi/apps/workflows/improvement.py (#517). src/inferrsi/apps/improvement.py also exists as the Apps compatibility export. |
Point at src/inferrsi/benchmark/improvement.py and src/inferrsi/harness/improvement.py; mention apps/workflows/improvement.py if the paragraph is meant to describe the Apps lane. |
test -e benchmark/improvement.py → missing; find src -name improvement.py → src/inferrsi/{benchmark,apps,harness,apps/workflows}/improvement.py; CHANGELOG.md:7–11 |
| F2 | README.md:99–104 (results table, vLLM-Omni PR column) and README.md:120 |
#9698 (draft), #9745 (draft), #10184 (open), #9283 (draft); "Upstream PR statuses checked on 2026-11-12." |
Re-checked 2026-11-20: #9698, #9745 and #9283 are state=open, draft=true — the (draft) labels are correct and stay. #10184 merged on 2026-11-16, so (open) is wrong at README.md:101. All eleven linked PRs re-read: #8739, #8729, #9218, #9188, #9827, #8112 merged; #10243 open, not draft; the rest as stated. |
Change #10184 (open) → #10184 (merged); update the date line to 2026-11-20. Leave the three (draft) labels untouched. |
gh api repos/vllm-project/vllm-omni/pulls/<n>: #9698 open draft=true, #9745 open draft=true, #9283 open draft=true, #10184 closed merged 2026-11-16T07:02:15Z |
| F4 | README.md:246–248 ("🧩 Architecture") |
"The implemented Harness × Model × Infra boundary…" | The inferrsi.infra package now holds only artifacts.py (plus its README); the host bindings it used to own are in inferrsi.apps.host (#498 slice 3a, no forwarding module or alias). |
Do not infer a replacement name from the package layout — a directory move is not an architecture-boundary rename. Link docs/architecture.md (which already states the boundary and its limits) and either drop the third name or use the axes that doc actually names. |
git ls-tree -r --name-only b5756cd56 -- src/inferrsi/infra → README.md, __init__.py, artifacts.py; CHANGELOG.md:13–25; docs/architecture.md:407–430 |
F3 — compatibility claim (correctness defect, independent item)
| # | README location | Current text | Repository state today | Suggested change | Evidence |
|---|---|---|---|---|---|
| F3 | README.md:273 embeds docs/assets/modular-components-en.svg; its compatibility panel at docs/assets/modular-components-en.svg:46 reads "Old CLI / import paths remain available. Historical LTX helpers and quality validators stay frozen." |
Readers are told old import paths still work. | That is false for the retired root modules. import inferrsi.infra.commands → ModuleNotFoundError; import inferrsi.apps.host.commands → OK. CHANGELOG.md:13–17 records "no forwarding module or alias, and the import gate fails any import of an old path"; docs/architecture.md:418–421 states "Root Python import paths are not all retained". |
Fix by text, without waiting for the graphics decision: add a line under the diagram (or in the <details> summary) distinguishing what is retained — CLI flags, external JSON, run artifacts, reports — from retired root import paths, and link docs/architecture.md#compatibility-and-historical-evidence. Regenerating the asset is the alternative, not the blocker. |
python -c "import inferrsi.infra.commands" → ModuleNotFoundError: No module named 'inferrsi.infra.commands'; python -c "import inferrsi.apps.host.commands" → OK; docs/assets/modular-components-en.svg:46; docs/architecture.md:407 |
The same asset is embedded at docs/architecture.md:93, so a regenerated asset
changes both pages; a text fix can be confined to the README.
2. Commands that do not run
None. Every command printed in the README was executed on the first audit
checkout and succeeded:
$ inferrsi --version
0.2.6
$ inferrsi find 'Qwen-Image-2.1 Ascend memory bottleneck'
vllm_omni.ascend.diffusion.qwen_image_21_eager_t2i [record · imported · installed · not locally verified] score 150.26
open: ... -m inferrsi read package:vllm_omni.ascend.diffusion.qwen_image_21_eager_t2i --material --omit-resources
$ inferrsi context --repo-type vllm_omni --backend cuda --output context.md
{"context": ".../context.md", "records": 5}
$ inferrsi context --repo-type vllm_omni --backend ascend --scenario diffusion.image_generation --output ascend-context.md
{"context": ".../ascend-context.md", "records": 5}
$ python -m pip wheel --no-deps -w /tmp/whlcheck .
inferrsi-0.2.6-py3-none-any.whl
The wheel filename matches the inferrsi-<version>-py3-none-any.whl pattern the
README tells users to download. The --repo-type and --backend values used in
the README examples are both valid choices for inferrsi context.
3. Outdated structure
| # | README location | Current text | Repository state today | Suggested change | Evidence |
|---|---|---|---|---|---|
| S1 | README.md:280 ("Find the right directory to extend") embeds docs/assets/modular-directory-map-en.svg |
The map shows package-root cli.py, arguments.py, runspec.py, catalog.py; a model/ panel with protocols/image.py; and an infra/ panel with targets/git.py, platforms/, executors/local.py, processes.py, http.py, artifacts.py, inspection.py, environment.py, commands.py, campaign.py. |
Every one of those paths is gone from where the map puts it. The current tree has apps/cli/, contracts/runspec.py, knowledge/catalog.py, benchmark/validators/image.py, apps/host/{targets,platforms,commands,inspection,environment}, apps/workflows/, apps/mindie/, apps/cannbot/ and resources/ — none of which the map shows. docs/architecture.md:160–300 already carries an accurate text directory map. |
Belongs to the architecture edit region above. Rebuild the SVG from the docs/architecture.md directory map, or drop the image and link the section instead. |
test -e src/inferrsi/cli.py (and the other seven paths) → all missing; docs/architecture.md:160–300 |
| S2 | README.md:19–32 ("📰 News") |
Newest milestone is v0.2.0 (2026-11-14); the section says "Selected release milestones". |
Six releases exist since: v0.2.1 (11-15), v0.2.2 (11-15), v0.2.3 (11-16), v0.2.4 (11-18), v0.2.5 (11-18), v0.2.6 (11-19). The daily release bot publishes on every default-branch change, so the section drifts silently. |
Add the current milestone (or state the selection rule, e.g. "milestones only, latest vX.Y.Z") and keep one line for the automatic daily releases. | git tag --sort=-creatordate → v0.2.6 … v0.2.0; git log -1 --format=%cs v0.2.6 → 2026-11-19; docs/releases.md:1–8 |
4. Missing content
| # | README location | Current text | Repository state today | Suggested change | Evidence |
|---|---|---|---|---|---|
| M1 | README.md:205–208 ("Use InferRSI in other clients"), and the find example at README.md:227–228 |
Find is described as searching tuning knowledge only. Nothing mentions the external knowledge channels. | inferrsi find queries external knowledge through a channel registry, but on this head _channels() registers only CannbotChannel (src/inferrsi/apps/cli/__init__.py:142–160); a fresh install reports cannbot-knowledge [not_enabled] with a fix line. Enabling needs an explicitly approved local checkout and interpreter; the source has its own CANN license and is never cloned or bundled (docs/external-channels.md:1–60). MindIE is not a find channel: inferrsi mindie exists as its own CLI and apps/mindie/ has the reader, but docs/external-channels.md:216 records the channel/adapter as "not yet". |
Add one sentence under the find example: external channels are off by default; the only currently wired channel is cannbot-knowledge; it needs a per-checkout grant (inferrsi cannbot enable); MindIE has its own CLI/reader rather than being a find channel; and the grant is a local authorization, not a redistribution right (docs/external-channels.md). |
src/inferrsi/apps/cli/__init__.py:142–160; docs/external-channels.md:1–60 and :210–218; src/inferrsi/apps/cli/arguments.py:161–210 |
| M2 | README.md:210–240 ("Use the CLI") |
Jumps from install straight to inferrsi find and inferrsi context. |
The CLI ships sixteen top-level subcommands (mindie, cannbot, agents, benchmarks, doctor, trace, export, dashboard, campaign, skills, knowledge, find, collect, read, resource, context); the documented first step elsewhere is inferrsi --help then inferrsi doctor --repo . --repo-type … --backend …, and inferrsi benchmarks list exposes the packaged tasks. |
Add a two-line "check your target" step (doctor) and a pointer to inferrsi benchmarks list before the find example. |
src/inferrsi/apps/cli/arguments.py top-level add_parser calls (:161–643); docs/quickstart.md:5–17 |
| M3 | README.md:233–238 (Ascend example) |
Only --backend cuda and --backend ascend appear; no list of valid values. |
--backend accepts cpu, cuda, rocm, ascend, mlx; docs/quickstart.md documents an MLX path the README never mentions. |
State the accepted --backend values once (a short inline list is enough) so ROCm/MLX readers know they are addressable. |
src/inferrsi/knowledge/catalog.py:33 → BACKENDS = ("cpu", "cuda", "rocm", "ascend", "mlx"); docs/quickstart.md:25–45 |
| M4 | README has no configuration section | — | Per-project grants persist under ~/.config/inferrsi/cannbot-projects/ and ~/.config/inferrsi/mindie-projects/; the MindIE adapter config defaults to ~/.config/mindie-agent/codex.json. These paths appear only in docs/external-channels.md and docs/operations.md. |
Low priority: if the README gains a config line for M1, name the default location and the file that can be deleted to revoke a grant. | src/inferrsi/apps/cannbot/project.py:53; src/inferrsi/apps/mindie/project.py:40; src/inferrsi/apps/cli/arguments.py:163 |
Repo-wide check
Sweep this revision for the retired module names and both diagram embed points:
- Retired names (
inferrsi.infra.{commands,executors,protocols,targets,platforms,campaign,inspection,environment},inferrsi.model.protocols): they appear only inCHANGELOG.md(the old→new map, expected) and underevidence/(historical source pins, expected). No current README, plugin, example or otherdocs/page names a retired module — except one: - D1 (new, Optional).
docs/contracts/current-design.md:28and:35still describeinferrsi.infra.inspectionas the live facts API and the one permanent importer ofinferrsi.resources.repository_packs. task: retire theinfra/andmodel/packages into their owning modules #498 slice 3a moved that file toinferrsi.apps.host.inspection;src/inferrsi/infra/inspection.pyno longer exists.docs/module-layout.md:495already has the corrected wording, so this is a within-docs inconsistency. Either page can be the source of truth; the fix is one sentence on each line. - Diagram embed points.
docs/assets/modular-components-en.svgis embedded atREADME.md:273anddocs/architecture.md:93;docs/assets/modular-directory-map-en.svgatREADME.md:280anddocs/architecture.md:162. Any asset regeneration touches two pages and should say so in its PR.
Not checked
- Result numbers themselves (latencies, gains, image comparisons). The audit
checked links, paths and PR states, not measurements. - Whether the results table omits campaigns completed after 2026-11-12.
- Contents of external sites beyond link status (HTTP 200 only).
- Winning behaviour of the plugin install flow inside Codex/WorkSwarm/DSH
clients — the README's plugin steps were matched againstdocs/codex-plugin.md,
docs/workswarm-plugin.mdanddocs/dsh-plugin.md, not exercised in a client. - D1 is reported from the sweep only; the two
docs/contracts/current-design.md
lines were read, not the whole decision record.
Audit provenance
Audited head: b5756cd561e8b12b68899d9dfa6411feab4669f5
(Make profiling optional in service-reuse campaigns (#521), 2026-11-20).
README audited: README.md, 310 lines, blob
61b9756ceb54375b4244ee436ee27868b94afc15, last committed
61303fa8150472dfbda62ec8d77908bc05c73f3b (2026-11-17). Audit date: 2026-11-20.
Activity
Dong1017 commentedon Nov 20, 2026
Posted by Dong-bot, an automated agent acting on behalf of @Dong1017 (always-on-bot, model claude-sonnet-5-5).
Assessment. The issue asks for a README refresh against the repository as it stands, delivered as an evidence table (location, current text, actual state, suggested replacement, command/API evidence) plus an explicit "Not checked" section. The evidence discipline is what I want from a docs change: the commands the README prints were run in a clean venv, the wheel filename was built and compared, links and in-repo anchors were resolved, and the audit states the head it was taken at. I read it as internally consistent, not as verified — details below. The work is real and mostly mechanical; what it lacks is a scope cut and a rot-prevention answer.
The severity is not uniform across the four categories.
docs/assets/modular-components-en.svgtells readers "Old CLI / import paths remain available" while imports of retired paths fail. This is corroborated independently of the audit table: task: retire theinfra/andmodel/packages into their owning modules #498 states the retirement uses "no forwarding shells or module aliases" and that old paths raiseModuleNotFoundError; task: retire theinfra/andmodel/packages into their owning modules #498's own status comment (2026-11-19) records that slices 2, 3c and 3a merged andinferrsi.infranow holds onlyartifacts.py. A reader following that panel is actively misled. Treat F3 as the one item that should not wait for the graphics decision.README.md:246–282is a single architecture block whose prose, inline claim and embedded map all describe the oldinfra//model/tree. Split across three PRs and they will conflict; do them together.#10184merged 2026-11-16). S2 drifts daily because the release bot publishes on every default-branch change. Rewriting the text without changing the form reproduces this issue in a week.Gaps and risks
F1…M4or aitem → PRtable; otherwise completion is unobservable.gh apicommand next to the check date, or link the upstream PR list rather than copying four statuses. Worth one advisory CI job that fails when README's stated latest milestone is behindgit tagand when README-referenced paths/assets are missing — that catches F1, F4 and S1 mechanically, with no model calls.infra/andmodel/packages into their owning modules #498 slice 5. task: retire theinfra/andmodel/packages into their owning modules #498 records one PR for slices 3b/4/5 that deletesinfra/andmodel/outright. Land it first, or PR A's wording describes a tree that no longer exists a day later. Phrase F1/F4/S1 in terms of packages that survive (harness,benchmark,apps,knowledge).493416520(2026-11-19) andREADME.mdwas last committed61303fa81(2026-11-17) — three dates. The head is also a 9-char short SHA; task: retire theinfra/andmodel/packages into their owning modules #498/refactor(apps): group application workflows in a workflows package #517 pins full 40-char SHAs. Give one date and a full head + README blob so a later reader can tell whether a claim was already stale at audit time. Line numbers will drift before anyone edits; locate by the quoted text and re-check at edit time.docs/and in plugin listing copy that draws from the same assets. One grep for the retired names and bothdocs/assets/modular-*.svgembeds across the repo, reported as either "clean" or new items, would close the class rather than one instance.Questions that decide the shape of the fix
docs/architecture.mdanddocs/quickstart.md? This decides F4/S1 (rename the boundary and rebuild the SVG, versus drop the diagram and link the text map thatdocs/architecture.md:160-300already carries) and M2/M3.docs/assets/modular-directory-map-en.svgandmodular-components-en.svg? "Regenerate" is not actionable if they are hand-edited — in that case the link-only option is the better trade, and it retires the whole maintenance class.docs/external-channels.md/docs/quickstart.mdwith one README pointer?Suggested slicing
docs/architecture.md#compatibility-and-historical-evidence.[not_enabled]fix line, not paraphrased.Not verified here. I read this issue, #517 and #498 only. I did not run any of the audits: no checkout, no clean-venv install, no
gh apiPR-state check, no link or anchor resolution, no SVG inspection. The evidence tables are accepted as the author's recorded results, corroborated only where #498/#517 independently state the same tree (infra/holding justartifacts.py,apps/workflows/improvement.pyfrom #517, no forwarding shells). — Dong-bot (authorized delegate)zengchuang-hw commentedon Nov 20, 2026
NO PASS
支持更新 README 的旧图和兼容性描述(F3/S1),但建议先纠正审计中的两项事实,再据此改文档:
state=open, draft=true,不应去掉(draft)。#10184 确已于 11 月 16 日合入,可更新这一项和核对日期。49341652的 CLI 注册 和 外部渠道文档,Find 只注册了 CannbotChannel;MindIE 有独立 CLI/reader,不等于 Find 已接入。建议注明当前可用渠道及授权边界。非阻塞:目录搬迁不等于架构边界重命名;F4 可直接链接架构文档,避免仅按包名推导。已静态核对 main 与 PR 状态,未执行 CLI。
Dong1017 commentedon Nov 20, 2026
Posted by an agent on behalf of @Dong1017 (moyuu2025). Issue body updated at audit head
b5756cd561e8b12b68899d9dfa6411feab4669f5; re-review requested.Re-review requested. Both comments were read and each point was checked against the repository before being adopted. The body now has a "本轮修订(2026-11-20)" section mapping every change to its reviewer, a priority checklist (Must / Next / Optional) keyed
F1…M4, one architecture edit region for F1/F4/S1, F3 as a standalone correctness item, a repo-wide sweep, and unified provenance (one date, full 40-character head, README blob). Items are located by quoted text as well as line number.@zengchuang-hw — NO PASS points
#9698→open, draft=true;#9745→open, draft=true;#9283→open, draft=true. The(draft)labels stay.#10184→closed, merged2026-11-16T07:02:15Z. Its label changes(open)→(merged).The body no longer proposes dropping
(draft); it states the correction explicitly._channels()(src/inferrsi/apps/cli/__init__.py:142–160) buildsChannelRegistry([CannbotChannel()], …); no MindIE channel is registered.docs/external-channels.md:216records "MindIE as a second channel | not yet;apps/mindie/has the reader, not the adapter". M1 now sayscannbot-knowledgeis the only wired channel, states the per-checkout grant and the CANN-license boundary (docs/external-channels.md:1–60), and says MindIE has its own CLI/reader but is not afindchannel.docs/architecture.mdand states that a directory move is not an architecture-boundary rename.@Dong1017 — assessment points
Adopted:
README.md:244–282).docs/, with one exception —docs/contracts/current-design.md:28and:35still nameinferrsi.infra.inspectionas the live facts API / permanent importer, though task: retire theinfra/andmodel/packages into their owning modules #498 slice 3a moved it toinferrsi.apps.host.inspection. Reported as Optional D1. Both diagram assets are embedded inREADME.mdanddocs/architecture.md, so regenerating either touches two pages.Not adopted in this issue, with reasons:
gh api-command-next-to-date idea are worth doing inside the README edit; the CI job should be its own issue so it can be exercised against the gate it adds. I have not filed that issue — flagging it here rather than dropping it silently.@zengchuang-hw @Dong1017 please re-review the updated body.
Dong1017 commentedon Nov 20, 2026
README refresh submitted for closure
flowchart LR R["README"] --> T["Inspect a target and find knowledge"] R --> A["Locate modules and migrate imports"] R --> N["Check releases and upstream PR states"]Revision and test record
The reviewed changes cover F1–F4, S1–S2, M1–M4 and D1. The README also distinguishes task entry on the source branch from the published v0.2.6 CLI. GitHub records the closing relationship on PR #533.
On Windows with Python 3.12, CPU inspection, task discovery, packaged search and material read, both context exports and the import migration probes passed. Ruff passed. Across the targeted test run and the two cases rerun outside a Git checkout, 241 tests and 54 subtests passed. Two POSIX-only cases and an Apple Silicon probe require their respective hosts.
All 11 GitHub checks passed at this head. The Python 3.10 and 3.11 full suites each reported 5,625 passed, 31 skipped and 672 passing subtests. Full CI run.
zengchuang-hw commentedon Nov 20, 2026
PASS
回应本轮 @:上轮两项已纠正。F2 保留 #9698/#9745/#9283 的 Draft 并把 #10184 标为已合入;M1 区分已注册的 Cannbot Find 渠道与独立的 MindIE CLI/reader。F4 也不再仅因目录搬迁推断架构重命名。
非阻塞:S1 的「所有旧路径」建议改为「多数旧路径」,
infra/artifacts.py仍在原位。通过的是修订后的审计,#533 的实现还未据此算作进入 main。仅静态核对。