AgentsView: hold the 0.43.0 pin; record the 0.44.0 qualification attempt and its gaps

[seathatflowsinourveins]

Summary

Hold the installed AgentsView 0.43.0 pin. The build's 2026-09-27 live release
check identified v0.44.0, and its isolated official asset passes checksum
verification. The selected upstream tests pass after controlling the temporary
directory environment. Qualification still lacks four fresh native lanes, both
required independent reviews and the unchanged full-save E2E runner.

Changes

• Add the dated qualification under
  evidence/artifacts/agentsview-044-qualification-20260927/: README, decision,
  separate upstream test record, aggregate native observations and selected
  actual returned output. The repair adds explicit dated errata.
• Correct unsupported review completion, environmental failure attribution,
  historical verdict boundaries, test classes and retention statements. Add
  the retained coordinator controls and hashes for all 23 fixture files.
• Record the demonstrated errors and their verification paths in the required
  docs/harness-defaults.md anti-pattern log.
• Add seven receipt unittest checks, including shared sanitization checks that
  reject planted home-path, UUID and forbidden-key violations.
• Record future same-PR updates as a classification of every tracked file that
  names 0.43.0 (see the repair round below): manifests/stack.json, the mirrors
  that existing checks compare with it, the install references and the
  version-specific guidance. docs/stack.md is a dated record, not a target.
  Include all three worker flags, explicit-port failure handling and
  inclusive CLI usage plus ccusage for accounting.
• The pin, production recipe, host configuration, PATH and installed archive
  remain unchanged.

The coordinator harness registers the six new files in manifests/evidence.json in the branch's last commit (docs/lanes.md hot-file protocol); no generated-report changes.
This builder's repair leaves the manifest to that harness.

Evidence

• Release integrity / local integration: official Linux amd64 v0.44.0
  archive SHA-256 matches the published checksum:
  037ea7a46d52e06b20363b4aa7cd7f28e32f31d8215803d6e9a0c96bac5818e3.
• Unchanged upstream tests: coordinator execution outside the Codex sandbox,
  Go 1.27.0, isolated GOPATH/GOCACHE, GOTOOLCHAIN=local, GOTELEMETRY=off,
  TMPDIR=/var/tmp/claude-w3-agentsview-go, clean tag checkouts and the upstream
  pricing snapshot restore. The same broad MCP/parser/DB command passes 4,575
  top-level tests on v0.44.0 and 4,387 on v0.43.0; each skips 32, fails 0 and
  passes all three packages. Candidate serve-runtime selectors pass 5/5 and
  close the original #1855 omission; original parser/DB selectors pass four
  top-level tests / 13 events. JSONL and stderr are retained privately; this
  repair recomputed hashes and counts without repeating those upstream runs.
• Preserved environmental failures: the original in-sandbox broad run has
  21 failing top-level tests / 25 failing events, each returning project tmp.
  The outside-sandbox controls discriminate the environment cause. The
  confirming rerun retains only an in-session summary; its raw bytes are absent.
  No counts from overlapping attempts are summed.
• Local integration over historical copies: two copied native session logs
  from full-save E2E clones (381,416 bytes), one Claude and one Codex. Both tags
  preserve 2 inclusive sessions, 1 project and 2 bounded FTS hits. CLI and
  ccusage 20.0.24 agree: 20,965 uncached input + 58,752 cache read + 556 output
  = 80,273 tokens. Candidate MCP returns zero for this Codex exec population.
• Synthetic fixture: five populations, two children; expected automated
  counts are 0 on baseline and 2 on candidate. All 23 retained file hashes and
  explicit expectations were captured during repair, not prospectively before
  the trial. Synthetic ccusage parity still fails with zero counters, distinct
  from successful copied-log parity and from provider execution.
• Local integration, port: an owned listener occupies 17384; baseline falls
  back to 17385, candidate exits 1. The original six scratch daemon-stop calls
  return 0; the occupying socket is closed.
• Local measurement: one copied-log full sync takes 0.407672 seconds on
  baseline and 0.358587 on candidate. No performance claim.
• Review attempts: the only GPT-6 review activity is two codex exec
  launches, normal and ephemeral, each exit 1 during initialization; empty
  JSONL and absent final-message files support no review verdict. The build
  stopped stalled Claude with SIGINT/130 inside its Codex sandbox, using Opus,
  xhigh, plan permission mode, --tools Read,Agent, strict MCP configuration
  and JSON output. No returned report; usage remains unknown.
• Structural validation: python3 -m unittest tests.test_agentsview_qualification tests.test_foundation_stack_pins passes
  all 9 tests with TMPDIR outside /tmp. Before receipt fixes it failed with
  two assertions and three missing-field errors. Disarming the publication
  guard causes all three planted-input self-checks to fail (exit 1); enabling
  it passes. repair-validation.md retains output and reproduction details.
  The build's earlier FileNotFoundError run was not a discriminating control.
• Repository validation before registration: python3 scripts/validate.py
  exits 1 only for registration of the five modified evidence artifacts and
  the changed docs/harness-defaults.md hash/byte count in the coordinator-owned
  manifest. repair-final-checks.md retains all seven diagnostics for these six
  paths and the final 9-test pass. git diff --check passes. Full integrated
  validation remains with the harness; no full-suite pass is claimed here.

SOTA sources

• kenn-io/agentsview v0.44.0,
  commit 413a87f7bfbd67b2815b1119ac51abc1efbeeaba;
  SHA256SUMS
  and supported installer.
• Upstream CI,
  go.mod,
  PR #1855 and
  serve-runtime tests.
• MCP input,
  CLI usage,
  port behavior,
  Codex classification
  and commands.
• Baseline v0.43.0,
  commit 9be7745ad1906ee24e04eb05bb86c872ef0939a1.
• ccusage/ccusage Rust Codex adapter,
  v20.0.24 at ecb676cce27cb5dd0090c7804a5cecc35e8ba805.
• Go release digests, Go 1.27.0;
  Python unittest assertions
  and patch.object.
• Repository docs/acceptance-evidence-policy.md:31-69, docs/lanes.md, the
  full-save plan §2.2, and the preserved
  evidence/artifacts/sota-refresh-20260923/pins-tools/agentsview.json limits.

Review dispositions

All ten findings are accepted; none is disputed.

1. Withdraw the unsupported GPT-6 verdict and independently checked list.
   Record not_completed, only two failed launches and no retained report.
2. Preserve in-sandbox failures as environmental. Record coordinator candidate
   and same-environment baseline controls; pass the broad-test gate outside
   the sandbox and remove the upstream-fix prerequisite.
3. Reconcile the 2026-09-23 qualified receipt in dated README text and
   prior_records: its checksum/first-line version/help boundary does not
   satisfy full qualification. The 2026-09-27 hold governs any re-pin; old
   receipt, README and catalog remain untouched.
4. Add both omitted documentation paths and all mirrored install/use fields
   to the future pin and recipe guidance.
5. Correct the manifest statement to acknowledge coordinator registration;
   remove the unsupported delegated-review sentence.
6. Record the coordinator's unchanged five-test serve-runtime run, including
   the omitted #1855 test; explicitly retain remaining scope omissions.
7. Use multi_class_summary and structural_validation. Exercise planted
   violations through the publication guard with disarmed red/enabled green
   controls; missing artifact errors do not establish guard effectiveness.
8. State that the confirming rerun's raw bytes were not retained. Hash all
   retained synthetic fixture files and state automated expectations as
   repair-time additions rather than a prospective freeze.
9. Identify the build session's SIGINT/130 stop, tools, effort and sandbox
   conditions. Require the next Claude review outside the Codex sandbox.
10. Replace undefined RR shorthand with full-save E2E runner and copied-log
    descriptions, naming the private unit-relative copy location.

Repair round (2026-09-29, native-agent-stack-10)

Head reviewed: 7514c2af, by two model families, read-only. GPT-6 (gpt-6-astra, effort max, codex exec -s read-only; the model is the -m request, not an observed resolution): approve, no findings; it did not check private raw logs and fixtures, live upstream, hosted CI or the full suites. Claude evidence-reviewer (Opus, max): needs_changes, 2 minor. C1: the recorded same-PR re-pin list named five files and missed mirrors that existing checks compare with manifests/stack.json (blueprints/token-native-focus/saturation-audit.json, read by tests/test_stack_lifecycle.py) and the recipes/README.md install row. C2: the 2026-09-27 erratum reason called an unnamed, unretained review "independent".

Repair, by an Opus builder in an owned worktree and then an independent Opus verifier that re-ran the acceptance commands, four negative controls and a simulated full re-pin in scratch clones (ready_to_push true; 4 minor and 7 nit findings):

• C1. The five-file list is replaced by a classification of every file that matches the recorded search git grep -n -I -E '0\.43\.0|v0\.43\.0', 68 files at 7514c2af: 39 under dated evidence/artifacts/ and evidence/hosts/ paths, 9 re-pin files holding 18 locations, and 20 dated records or an unrelated package, each with its basis (qualification.json repository_pin_locations). docs/stack.md moves from re-pin target to dated record: its header reads "Snapshot: September 19, 2026.", its three commits are all from 2026-09-19, none of the four earlier re-pins (3058b23/Pin socraticode 1.15.0 and ccusage 20.0.26 with upstream-suite qualification receipts #446, c2ffaef/Codex 0.157.1 switched on the workstation (daemon-safe) + native quota probe and runner gate #348, ff20470, 965c9d6) changed it, and 14 of its 47 rows differ from the manifest by version (15 when the repository URL is compared). Three new tests (RepinLocationRegistryTests) run the recorded search.
• C2. The erratum reason now reads "Evidence-review repair (reviewer identity and returned report not retained)", and test_errata_do_not_name_an_unretained_reviewer fails on the old wording.
• Verifier findings fixed on this head. The classification test now skips once manifests/stack.json is re-pinned, as the location tests already did: the verifier reproduced that it failed an unrelated edit and the re-pin PR's own receipt. The scripts/native_token_ci.py match no longer includes the PINS dict's closing brace (appending a tool after agentsview failed the test). The failure message says when the counts pair but the text does not. The scripts/landscape.py span is 1373-1381, the docs/stack.md count names its basis, the recipes/README.md label names the block (line 796) and its version sentence (line 813), the README says the classification is per file, and the branch-mechanics sentence is out of the erratum.

Evidence (measured 2026-09-29 by the coordinator at feab5886, local integration checks and not upstream tests; every command exits 0):

• python3 -B -m unittest tests.test_agentsview_qualification tests.test_stack_lifecycle: 16 tests OK, 0 skipped.
• python3 -B scripts/validate.py: {"components": 69, "hashed_files": 8000, "profiles": 4, "receipts": 168, "status": "passed"}; scripts/evidence_manifest.py --check files 8000 passed; scripts/component_matrix.py --check rows 32; scripts/new_host_grand_list.py --check passed, 32 layers.
• The three repository classification tests that read whole-tree inventories (test_blind_checkout, test_osv_lockfile_coverage, test_workflow_security_coverage): 3 tests OK.
• python3 -B -m unittest tests.test_adoption_docs_consistency tests.test_ecosystem_manifest tests.test_install_claude_profile tests.test_token_e2e_preregistration: 190 tests OK, 4 skipped (environment-gated; not passes for those cases).
• Controls in a scratch clone of this head: a new tracked file that mentions 0.43.0 while the pin holds fails the registry class (2 failures, as the README now documents); the same file after a simulated re-pin of manifests/stack.json gives 3 skips and OK; appending a tool to the PINS dict after agentsview passes.

Residuals and not done

Complete four fresh native lanes with fixed model, route and effort, both
independent reviews, the unchanged full-save E2E runner and final recipe controls
before any re-pin. Run the Claude review outside the Codex sandbox. The synthetic
ccusage discrepancy remains unresolved. The selected upstream gate passes;
environmental parser failures require no upstream fix.

No full upstream ./... suite, frontend, cross-platform, complete-history,
savings, cost-parity or new provider acceptance. No host switch. Fixture hashes
are retrospective; the old confirming rerun's raw bytes cannot be recovered.

From the repair round: RepinLocationRegistryTests.hits() treats any git grep failure as "not a Git checkout" and skips, so a misconfigured checkout can pass those checks as skipped. While the pin holds, a new tracked file outside the dated evidence paths that mentions another tool's 0.43.0 fails the classification test until it is classified in qualification.json (documented in the README). qualification.json repository_tests still records the 2026-09-27 nine-test run; this repair's counts are in this body, not in the dated artifact. The re-pin PR itself is not run here.

The coordinator owns commits, re-registration of changed evidence, report
generation and validate.sh/full-suite execution against its integrated branch.
This repair performs no commit, push, rebase, reset or checkout.

Recommended lane label: lane:foundation.

Coordinator note

After an independent repair verification, the coordinator made the record commit-neutral: the registration sentence no longer names a build commit, the recipe pointer is a section anchor, and the copied logs and runner are attributed to the 2026-09-26 returned-results (RR) E2E. The branch's last commit registers this branch's own files in manifests/evidence.json (docs/lanes.md hot-file protocol); no generated report changed. Acceptance after the repair (local integration, this host): validate.sh FAILS=0; full suite 6600 tests, one known old-base failure (test_host_profile_copy_is_verbatim).

Replay onto main 364f4d9 (head 66dc69a): validate.sh FAILS=0, and the covering modules (test_agentsview_qualification, test_adoption_status, test_native_dashboard_data, test_native_token_ci, test_secret_path_guard, test_token_e2e_receipt_checks, test_sota_convergence) pass there. The replay met one content conflict: main and this branch both appended rows to the anti-pattern log in docs/harness-defaults.md. Both sides' rows are kept, main's first, and a stray blank line in main's copy, which split the table so later rows rendered as plain text, is removed.

Replayed again onto main bfdc99c (#432 merged) as 5b73121: content lines identical, 7 files re-registered, validate.sh FAILS=0, and the same covering modules ran 360 tests, OK.

Replayed once more onto main b1be50c (after #430) as feab5886 by the hot-file protocol: main's evidence manifest, the branch's files re-registered (7), generated reports regenerated by their --write commands.

🤖 Generated with Claude Code

[seathatflowsinourveins]

Review record, 2026-09-29 (native-agent-stack-10). Merge candidate: head feab5886, which is 7514c2af plus the repairs below and a merge of main at b1be50c8 by the hot-file protocol.

• Two model families reviewed 7514c2af, read-only: GPT-6 (gpt-6-astra, effort max, codex exec -s read-only; the model is the -m request, not an observed resolution) approved with no findings, and said it did not check the private raw logs, live upstream, hosted CI or the full suites. A Claude evidence-reviewer (Opus, max) returned needs_changes with two minors: the recorded same-PR re-pin list named five files and missed mirrors that existing checks compare with manifests/stack.json and the recipes/README.md install row (C1), and the 2026-09-27 erratum reason called an unnamed, unretained review "independent" (C2). Nothing was disputed.
• Repair: an Opus isolated-builder replaced the five-file list with a classification of every file matching the recorded search (68 files: 39 dated evidence paths, 9 re-pin files holding 18 locations, 20 dated records or an unrelated package, each with its basis), moved docs/stack.md from re-pin target to dated record, reworded C2, and added three tests. An independent Opus stack-verifier re-ran the acceptance commands, four negative controls and a simulated full re-pin in scratch clones and returned ready_to_push with 4 minors and 7 nits. The coordinator fixed on this head: the classification test now skips once the pin moves (it had failed an unrelated edit and the re-pin PR's own receipt), the scripts/native_token_ci.py match no longer includes the PINS dict's closing brace, and three citations and count statements. The coordinator's controls in a scratch clone: a new tracked file mentioning 0.43.0 fails the registry class while the pin holds; after a simulated re-pin the same file gives 3 skips and OK; appending a tool to the PINS dict passes. That last round was not reviewed by a second agent.
• Checks on the merge candidate (local integration checks, not upstream tests): python3 -B -m unittest tests.test_agentsview_qualification tests.test_stack_lifecycle 16 OK; scripts/validate.py, scripts/evidence_manifest.py --check, scripts/component_matrix.py --check and scripts/new_host_grand_list.py --check exit 0; the three whole-tree registry tests OK; tests.test_adoption_docs_consistency tests.test_ecosystem_manifest tests.test_install_claude_profile tests.test_token_e2e_preregistration 190 OK, 4 skipped. All 8 required hosted checks passed on this head (gh pr checks 420 --required exit 0).
• Residuals, in the body: the classification test treats a git grep failure as a skip; qualification.json repository_tests still records the 2026-09-27 nine-test run; the re-pin PR itself is not run here.