[Dong1017]

Governing Issue and delivery slice

Refs #248

Removes the tooling of the bounded synthetic CPU KV experiment. All retained evidence stays in place and unmodified. This is a cleanup of code that no longer has a path to run; it decides nothing about #248's acceptance, which remains open with neither completion criterion met.

Why this is removable

The slice had its own consumers only. Nothing in the general harness, the SkillEval modules, the runner or the evaluator imported it: git grep -E '\bkv_|kv\.' across src/inferrsi/harness/skill_eval*.py, evaluation.py, runner.py, improvement.py, reporting.py and tools/skill_eval_*.py returns nothing. The dependency ran one way only — from this slice onto shared parts (skill_eval_engineering.sandbox_command, skill_eval_budget, harness.runner.run_experiment, infra.artifacts, model.anthropic_agent, runspec.RunSpec). Removing it cannot reduce the general SkillEval capability.

Its only non-self consumers were the kv-task-qualification job in the ready lane — self-described there as "not a branch-protection or efficacy gate" — and four rows in docs/operations.md.

What changed

Removed	Count
tools/kv_*.py, tools/qualify_kv_*.py, tools/kv_live_authorization.json	12
src/inferrsi/harness/kv_reclamation.py, kv_edit_workbench.py	2
tests/test_kv_*.py, tests/test_qualify_kv_*.py	13
tests/fixtures/kv_reclamation/	3
docs/kv-edit-integration.md, kv-live-launcher.md, kv-pilot-admission.md	3
.github/workflows/kv-edit-pilot.yml	1

Also: the kv-task-qualification job drops out of .github/workflows/ci-ready.yml (52 lines), and the four operator rows drop out of docs/operations.md.

What is kept, and one constraint that forced it

evidence/kv-reclamation-20261030/, kv-actor-qualification-20261030/, kv-actor-qualification-20261031/ and kv-live-pilot-20261031/ are untouched — 64 files, 2.3 MiB — along with their four entries in tools/evidence_inventory.json.

docs/kv-reclamation-qualification.md stays at its path, rewritten as a retirement record rather than deleted. evidence/kv-reclamation-20261030/README.md:59 links to it, and that README's SHA-256 is pinned in its own plan/publication-integrity.json (7c6d65e4…), so the link cannot be edited out without breaking the manifest. Deleting the page would leave a dangling link inside hash-pinned evidence. The rewritten page records what was removed, where the evidence lives, and what that evidence does and does not show.

src/inferrsi/model/anthropic_edit_agent.py is also kept. Its only current consumers are the removed code and its own test, but it is a general edit-capable model adapter rather than an experiment-specific module. Removing it would widen this change beyond cleanup; flagging it here so a maintainer can decide separately.

Checks

Local, on Python 3.13 at this head:

• Test collection: 3557 → 3183, no collection errors. The 374-case drop is exactly the removed test files; no module fails to import, which is the decisive check for dangling references.
• ruff check src tests tools: passed.
• python -c "import inferrsi.harness, …skill_evaluation, …skill_eval_engineering, …runner, …evaluation": passed.
• python -m tools.validate_evidence: exit 0, 430 hashes checked, all 21 bundles including the 4 retained KV bundles verified.
• git diff --cached --check: clean. docs/operations.md shows exactly 4 removed lines, with no line-ending noise.

Not run locally: the full suite. Several of this repository's tests probe bubblewrap isolation and time out on Windows, so a full local run is not meaningful here. The ready lane runs it on Ubuntu; that is the full-suite evidence for this PR, and I am not claiming a local pass.

Limitations

This changes no behavior of any remaining feature. It is not an acceptance, rejection or rescoping of #248: both of that Issue's completion criteria are still unmet, its two 2026-10-31 attempts remain INCONCLUSIVE, and the retained evidence that records those failures is deliberately preserved here. Removing the tooling does not make the experiment repeatable or unrepeatable — the Issue's own freeze already forbids a third attempt.

Merge and any scope decision on #248 remain with the maintainers.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-06T01:47:43.388143Z	10f70e0	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[amy-why-3459]

Review at head 029878bbb000d82e6e6d38d895c0865a4adbf43b (full review of the PR, base e843b9d3). The removal itself looks clean: no remaining import of the removed modules in src/, all Ready checks are green at this head, and docs/operations.md loses exactly the four KV rows. I am not approving yet because of two points I would like resolved or confirmed:

1. Retained hash-pinned evidence still points at removed paths, and the retirement page does not say where to find them. The PR keeps docs/kv-reclamation-qualification.md specifically to avoid a dangling link from pinned evidence, but the same retained bundles also reference files this PR deletes, for example evidence/kv-live-pilot-20261031/patch/README.md:4-5 ("Both arms start from the exact unchanged tests/fixtures/kv_reclamation/candidate.py pinned by ../plan/plan.json"), and the plan/*.json / results/hosted-*/qualification.json files reference tools.qualify_kv_reclamation, tools.kv_live_job and kv_edit_workbench. Since that evidence cannot be edited, please add to docs/kv-reclamation-qualification.md (around lines 20-21 / 25-33) the last revision where the removed tools and fixtures still exist (e.g. base e843b9d31cb86e1fd12d00890c89798b0bedd1fe), so every path named in the retained evidence stays resolvable via git history.
2. Required-check confirmation for the dropped kv-task-qualification job (.github/workflows/ci-ready.yml). The job comment says it is not a branch-protection gate, but it currently reports on every open PR (e.g. it is in the green Ready set on feat: add native WorkSwarm plugin and release artifact #327). Please confirm it is not a required status check in the main ruleset; otherwise removing it would leave other PRs waiting on a check that never reports.

Non-blocking: the claims in lines 37-42 of the rewritten page (both 2026-10-31 attempts INCONCLUSIVE, the second stopping on HTTP 403, dollar figures being reserved upper bounds) are summaries of the retained evidence that I did not independently re-check; a link to the specific receipt for each would make the page self-verifying. Keeping src/inferrsi/model/anthropic_edit_agent.py for a separate maintainer decision is reasonable.

No approve or request-changes vote at this head.

[Dong1017]

Incremental read-only review at head 029878bbb000d82e6e6d38d895c0865a4adbf43b against e843b9d31cb86e1fd12d00890c89798b0bedd1fe. Scope: the required-check question in the existing review, not a fresh approval or full re-review of the removed KV tooling.

• Verified for the active ruleset: main ruleset protect-main currently declares only knowledge-model-policy under required_status_checks; kv-task-qualification is not required by that ruleset. I also read the actual .github/workflows/ci-ready.yml removal. Therefore the proposed deletion does not remove a status check required by this active ruleset.
• Remaining verification boundary: reading the separate classic branches/main/protection endpoint returned HTTP 403 (Resource not accessible by integration), so this does not prove that no classic branch-protection setting exists. A maintainer with settings access can confirm that last point if necessary.
• The existing request to preserve a discoverable historical revision for removed paths in retained evidence is not cleared by this check. No tests, model calls, or device experiments were run for this review.

This is scoped COMMENT feedback, not a merge/acceptance verdict.

[Dong1017]

Dear committer @amy-why-3459 @zengchuang-hw, can you review this pr when you have a chance?

[zengchuang-hw]

LGTM

PR review verdict — PASS

Initial independent review of 9dded7ad..029878bb (37 changed files), at head 029878bbb000d82e6e6d38d895c0865a4adbf43b. Assessed integration with current main 01c73521ced6ca760cdfa42c9b937fe985a20d64; the PR metadata still reports base e843b9d3. The merge-tree check against current main is clean. I read the earlier review and the required-check follow-up. Conclusion: no outstanding content blocker for this retirement change.

本轮已解决

• The removed modules, commands, fixtures and qualification job form the retired bounded KV slice. Searches of remaining source, tools, tests and workflows found no active consumer of the deleted modules. The general editing adapter remains, and the four operator rows are removed without unrelated changes.
• All 64 files in the four retained KV evidence directories, and the evidence inventory, have unchanged Git blobs. The retirement page preserves the pinned README's documentation target. Historical source identity is already retained in the frozen plan and actor receipt, so the deletion does not erase their source provenance.
• The dropped kv-task-qualification job is not required: the active main ruleset requires only knowledge-model-policy. A successful main branch metadata read also reports classic protection enabled=false, enforcement off and empty required-check lists. This supplies evidence beyond the earlier separate protection-endpoint 403.

批准前必做（Before I can approve）

None for this reviewed change. Issue #248 remains unfinished; this cleanup does not accept its experiment, change its frozen verdicts or authorize another attempt.

非阻塞问题或建议

Adding a direct last-available source tree link to the retirement page would make old paths easier to find. Historical analysis should still use the original revision recorded in each receipt. Links to the particular live-attempt receipts would also improve navigation. These are documentation discoverability improvements, rather than missing source provenance or a broken active operator.

Checks and limits

Independent local verification at the exact head, on Windows / CPython 3.12.1 / pytest 9.1.1:

• 25 tests passed across test_evidence_inventory.py, test_frozen_bytes.py and test_module_boundaries.py.
• 3,183 tests collected with no collection errors.
• python -m tools.validate_evidence: exit 0, 430 hashes / 21 bundles checked, no errors; pre-existing external/unavailable evidence retains its stated limits.
• Ruff 0.15.15 (src tests tools) and diff whitespace checks passed.

Current-head Ready CI passed both Python 3.10/3.11 full-check and installed-package jobs, plus PR references. The required policy result explicitly evaluates this head as successful / Not applicable. Main's intervening label/report changes do not touch the deleted KV slice or knowledge-policy scope.

I did not locally rerun the full suite or installed-wheel matrix, or execute hardware, provider or model experiments. The clean current-main merge tree was inspected, not separately run through a full test matrix. This is a PR content review, not a campaign PASS.

Direct delegation: zengchuang-hw, the account holder, authorized Codex in this chat to independently review other authors' InferMatrixLab/InferRSI PRs and submit formal Approve when content blockers are absent and applicable verification passes. Executor: Codex, with separate technical and scope/evidence cross-checks; acting GitHub account: zengchuang-hw (repository write access or higher). This account and reviewing task did not author, materially contribute to, or parent this delivery slice. Authorized action here: revision-bound APPROVE; scope: this PR's retirement diff and its verification.

[Dong1017]

@amy-why-3459 can you give the suggestion, if there is anything not fix in the pr?

[Dong1017]

Both points resolved at head 10f70e09c3874629c38a56b308fbba32ab7e93b4. Thanks — the second point had a clean answer, and chasing the non-blocking one found a real sourcing error on the page.

2. kv-task-qualification is not a required check

Confirmed against the protect-main ruleset, which is the only active ruleset on refs/heads/main. Its required_status_checks rule lists exactly one context:

knowledge-model-policy   (integration_id 15368)

kv-task-qualification is not in it, and strict_required_status_checks_policy is false. Removing the job leaves no PR waiting on a check that never reports. It reports on open PRs because it is an unconditional job in the ready lane, not because branch protection asks for it.

1. Removed paths are now addressable, and the retained references are named

Two additions to docs/kv-reclamation-qualification.md.

The history pointer, after "What was removed": every removed path still exists at e843b9d3 — the revision you named — with git show e843b9d3:<path> examples. I verified that commit is an ancestor of upstream main, so the pointer stays valid after this merges. (For the record, the branch's own parent is 9dded7ad; e843b9d3 is the current merge base. Both carry the files, and I used the one you named since it is on the mainline.)

The reference inventory, after "What was kept": I enumerated every retained file that names a deleted path rather than describing the problem in prose. It is eleven files across three bundles, found with grep -rl 'qualify_kv_reclamation\|kv_live_job\|kv_edit_workbench\|kv_reclamation/candidate' evidence/:

Bundle	Files	Reference
kv-live-pilot-20261031	patch/README.md	tests/fixtures/kv_reclamation/candidate.py
kv-live-pilot-20261031	plan/plan.json, plan/A2.json, plan/A3.json, plan/second-attempt/{plan,A2,A3}.json, plan/second-attempt-1200/plan.json	tools.qualify_kv_reclamation, tools.kv_live_job
kv-actor-qualification-20261030	results/hosted-41154288295/qualification.json, results/hosted-41182807900/qualification.json	kv_edit_workbench
kv-reclamation-20261030	patch/README.md	tests/fixtures/kv_reclamation/candidate.py

The page now states that each resolves at e843b9d3 and none resolves in the working tree, which is the expected result of the removal rather than a broken bundle.

Non-blocking: you were right, and the claims were worse than unverified

Checking the receipts for lines 37-42 found that the page attributed its outcome claims to the wrong source.

evidence/kv-live-pilot-20261031/results/README.md says:

Prospective pair: not executed … No paid model output, timing result or efficacy verdict exists in this bundle.

So the retained bundles hold frozen inputs and a not-executed marker. They contain no arm outcome at all. The INCONCLUSIVE results and the HTTP 403 come from the workflow run records (41159550621, 41194057284), assessed in issue 248 — not from the evidence the page points at. Reading the page as written, those claims looked bundle-backed, which they are not.

The section is now a per-claim table:

Claim	Receipt
No completed arm, per-arm outcome or A/A variance	results/README.md, quoted above
Dollar figures are reserved upper bounds, not observed spend	plan/second-attempt-budget.json: worst_case 162,720,000 microUSD against the PR and pilot lifetime caps
Both attempts INCONCLUSIVE, the second stopping on HTTP 403	The two run records and the #248 assessment

The page now says outright that the third row is not recorded inside the bundles, and that a workflow conclusion of success means orchestration and retention finished rather than a request being delivered or charged. One factual gap also closed: A3 never started on the second attempt, which the page did not say.

Checks at this head

Local, on the pushed tree:

Check	Result
ruff check src tests tools	All checks passed
pytest --collect-only -q	3,183 collected, 0 collection errors
python -m tools.validate_evidence	exit 0, 21 bundles, 430 hashes, 0 errors

The retirement page is not itself hash-pinned — grep finds it in no manifest, only in three retained evidence files that link to it — so editing it changes no pinned digest, and the validator result above confirms that.

No local full-suite pass is claimed; the hosted lanes are the authority for that.

@zengchuang-hw — your approval was on 029878bb and this push dismisses it under dismiss_stale_reviews_on_push. Sorry for the re-request; the change is documentation only.

[zengchuang-hw]

LGTM

PR review verdict — PASS

Incremental review 029878bbb000d82e6e6d38d895c0865a4adbf43b..10f70e09c3874629c38a56b308fbba32ab7e93b4. Reviewed head 10f70e09c3874629c38a56b308fbba32ab7e93b4 targets main (PR base metadata e843b9d31cb86e1fd12d00890c89798b0bedd1fe, current main 5c86eb57030c4aea50592b66451929c8cee705d6, actual merge base 9dded7adac0ac835f6a73493309399e328a43ab2). The retirement change has no remaining approval blocker. Previous review · Re-review request

本轮已解决

• The retirement page now points to a readable pre-retirement snapshot on main. I checked the 11 retained references named by the inventory and resolved the four historical source files. Their blobs match the actual merge base. All 64 files across the four retained bundles and the evidence inventory are unchanged in the PR. Earlier finding · Historical source tree
• The page separates the bundle’s frozen inputs and not-executed marker from the live attempt records. The budget file records a 162,720,000 microUSD reserved bound. The run records and Issue assessment identify both attempts as INCONCLUSIVE, with HTTP 403 on the second attempt and A3 unstarted. Budget record · Run and outcome assessment
• The main ruleset requires knowledge-model-policy, and classic status enforcement is off. The retired kv-task-qualification job is outside the required checks. Main rules · Branch metadata

批准前必做（Before I can approve）

None for the retirement PR. The broader task and its acceptance criteria remain open under #248. Governing Issue

非阻塞问题或建议

Call e843b9d3 a pre-retirement snapshot on main. The actual common ancestor is 9dded7ad. The inventory could also distinguish the four A2/A3 configuration files that name kv_edit_workbench.py from the plan files that name the two tools modules. The historical paths already resolve. Retirement page

Checks and limits

Reviewed the single-file delta, checked Git diff whitespace and evidence paths, and cross-checked the evidence and ancestry with an independent review agent. All five current-head Ready jobs passed, including Python 3.10 and 3.11 full-check and installed-package jobs. The policy status passed. This incremental review used hosted CI for runtime tests and the existing records for the historical attempts. It did not re-run a live experiment or verify actual billing. Ready CI · Policy

Direct delegation: zengchuang-hw, the account holder, authorized this review in the Codex chat. Codex executed the independent review using the connected GitHub account zengchuang-hw, verified with repository admin permission. The account and reviewing task are independent of the delivery slice. The delegation covers the PR content and applicable checks, responses to review requests, and a revision-bound formal Approve when no blocker remains. This submission uses that Approve authority.

[amy-why-3459]

LGTM

PR review verdict — PASS
Reviewed incremental range 029878bbb000d82e6e6d38d895c0865a4adbf43b..10f70e09c3874629c38a56b308fbba32ab7e93b4 at head 10f70e09c3874629c38a56b308fbba32ab7e93b4.

Checked:

• Diff is docs-only (docs/kv-reclamation-qualification.md, +38/−5).
• History pointer to e843b9d3 is present; that commit is an ancestor of current main; spot-checked removed paths still readable there (tools/qualify_kv_reclamation.py, tests/fixtures/kv_reclamation/candidate.py, src/inferrsi/harness/kv_edit_workbench.py) and absent at this head.
• Reference inventory names the retained files that still cite deleted paths; claim/receipt table matches evidence/kv-live-pilot-20261031/results/README.md (not-executed) and plan/second-attempt-budget.json (prior_binding.worst_case 162720000).
• Required-check claim holds: classic main protection has empty required contexts; combined status for this head reports only knowledge-model-policy (success); no kv-task-qualification check run at this head. Ready jobs (full-check / installed-package 3.10+3.11, pr-references, declarations) are green.

Prior amy points from #357 (comment) are resolved. No remaining content blocker for this retirement docs fix.