Voir en action
Audit a false native-evidence handoff with positive and negative controls
Historical author audit: a real Joplin archive, available/missing native-byte controls, simulated native Evidence and exact CLI/MCP checks expose a false identity link. The full report retains partial coverage, unresolved timing variability and validation failures; it is not a site test or speedup claim.
Exemple de résultat
# Native handoff byte-identity verification
Tracks #3926 under #4027. Baseline: `9f8659e0536e4e50a438f5b9d5661f3c9624f6ac`.
Final integration base: `62bf09809ef7bb224028916424f8f07652a20f4f`.
The analyst already knows an ASAR-relative native path and supplies prior native
Evidence. The question is which Evidence can be linked to that component's exact
bytes. Discovery and native-provider execution are outside this task.
## Observed behavior and correction
The native-node producer distinguishes a content digest from a canonical path
whose bytes are unavailable. Its unavailable observation legitimately cites the
containing archive. The handoff consumer previously treated that supporting
artifact digest as the native component's digest.
Public CLI/stdio MCP regression: the missing-companion case failed and the
available-byte control passed. Real Joplin reproduced the same result: missing
keytar bytes produced an `evidence-linked` handoff with the ASAR digest and the
simulated container observation. CLI/MCP agreed on that incorrect answer.
The consumer now uses only the node owner's content-digest identity. The fallback
observation scan is removed. Unavailable native nodes retain their canonical
path, availability, container-backed provenance and partial coverage in the
inline graph; they yield no false actionable handoff or native Evidence link.
An empty handoff list describes eligibility, not absence of native components.
The per-query node index is allocated only after an eligible target exists; no
persistent cache, schema, tool or target authority is added.
Final real checks passed both cases with exact CLI/MCP trace parity. The available
control preserved the complete parent/trace Evidence IDs, native facts, links,
coverage and all token projections. The missing control preserved its parent,
native node, source provenance and coverage, while removing the wrong handoff
and container Evidence link.
## Real input and scope
[Official Joplin v3.7.21 Linux AppImage](https://github.com/laurent22/joplin/releases/download/v3.7.21/Joplin-3.7.21.AppImage),
published 2026-09-25, 201,462,481 bytes, SHA-256
`5d202dcaca31e14206826aa3f04125e7c218dbc3dc3d98479d811efbd6b856c1`.
Read-only 7z 26.02 inspection found 6,324 ASAR files, 46 unpacked entries and
three `.node` components. Extract only the ASAR and those three binaries:
```sh
input_root="$(mktemp -d -t rea-native-handoff-XXXXXX)"
curl --fail --location --output "$input_root/Joplin-3.7.21.AppImage" \
https://github.com/laurent22/joplin/releases/download/v3.7.21/Joplin-3.7.21.AppImage
sha256sum "$input_root/Joplin-3.7.21.AppImage"
7z x -mmt=1 "-o$input_root/extracted" "$input_root/Joplin-3.7.21.AppImage" \
resources/app.asar \
resources/app.asar.unpacked/node_modules/keytar/build/Release/keytar.node \
resources/app.asar.unpacked/node_modules/sqlite3/lib/binding/napi-v6-linux-glibc-x64/node_sqlite3.node \
resources/app.asar.unpacked/node_modules/onnxruntime-node/bin/napi-v6/linux/x64/onnxruntime_binding.node
python3 - "$input_root" <<'PY'
import os, sys
from pathlib import Path
root = Path(sys.argv[1])
source = root / 'extracted/resources'
for mode in ('available', 'missing'):
case = root / 'cases' / mode
case.mkdir(parents=True)
os.link(source / 'app.asar', case / 'app.asar')
if mode == 'available':
for native in (source / 'app.asar.unpacked').rglob('*.node'):
target = case / native.relative_to(source)
target.parent.mkdir(parents=True, exist_ok=True)
os.link(native, target)
PY
```
Verify the asset digest above before extraction. The ASAR is 233,575,014 bytes,
SHA-256 `767df409b18effe60421fbefcca873de237c7e813eac744ecdeebbc5e825a33a`.
Keytar is 76,960 bytes, SHA-256
`0700016bcd761cbe02d275f2046c7cc42f39f730a50d66e3d3307aa9d413373a`.
All extracted component hashes are in the JSON record. 7z reported one
trailing-data warning for the AppImage's embedded SquashFS and exited 0.
Both cases select only `package.json` for static source analysis while retaining
the inventory. The available case stages all three native companions; the other
43 unpacked entries are not staged. The missing case stages none. Whole-application
coverage is therefore partial in both cases, independently of the selected
native component's known or unknown byte identity. Neither the application nor
a native analysis provider is executed. Native Evidence is explicitly simulated
to test exact subject linkage. Live Electron, real Ghidra/Hopper/IDA behavior and
Windows/macOS verification are **NOT_ENOUGH_INFO**.
## Reproduce the public workflow
Use pinned Node 24.18.0/npm 11.16.0, `npm ci`, then `npm run build:cached`.
The verifier returns compact records and deletes its private exported parent and
CLI input after owned cleanup. No downloadable application or large Evidence is
archived.
```sh
for mode in available missing; do
node scripts/verify/javascript/native-handoffs.mjs \
--artifact "$input_root/cases/$mode/app.asar" \
--addon-path node_modules/keytar/build/Release/keytar.node \
--native-sha256 0700016bcd761cbe02d275f2046c7cc42f39f730a50d66e3d3307aa9d413373a \
--source-path package.json --rounds 3 --cli-parity --require-safe
done
```
Add `--tokenizer-python /absolute/path/to/python-with-tiktoken-0.12.0` for the
reported BPE projections. The npm alias `verify:javascript:native-handoffs`
builds first; recorded costs use the direct Node invocation above after building.
For a pre-fix observation, run the saved verifier on the baseline build without
`--require-safe`; `safe_identity: false` is an incorrect answer, not a passing
analyst task. Paths affect serialized requests and some Evidence identities.
The exact resource runs additionally used the preserved
[8 GiB observer](https://github.com/morluto/rea/blob/b3bdabab694af1f9801eff7caf75001b92314a58/experiments/analysis-efficiency/resource-run.py)
and adjacent `resource_payload.py`; both hashes are pinned in the JSON.
## Completed workflow costs and limits
| Case | Revision | Wall s | CPU s | Peak cgroup bytes | Identity oracle |
| ---------------------- | -------- | ------: | ---------: | ----------------: | -------------------------- |
| Native bytes available | Before | 60.369 | 59.806670 | 2,662,629,376 | pass |
| Native bytes available | After | 258.885 | 106.221003 | 2,678,456,320 | pass, exact baseline facts |
| Native bytes missing | Before | 57.930 | 57.479109 | 2,719,457,280 | wrong identity/linkage |
| Native bytes missing | After | 59.738 | 59.254776 | 2,571,235,328 | pass, unknown retained |
Whole-verifier resources include input hashing, analysis, three retained trace
queries, export, CLI parity and tokenization. One CPU, 8 GiB hard cap, 6 GiB early
stop, no swap; cleanup verified and no OOM. Memory includes charged file cache,
not summed RSS. Counters are at payload exit while the observer remains alive.
The available control's final wall and CPU times are substantially higher than
baseline. Earlier post-fix controls on the original base took 61.657/74.956 s wall
and 61.135962/60.352428 s CPU for available/missing cases; their complete receipts
remain separately recorded. The final missing case took 59.738 s wall. This
variability is unresolved and cannot be attributed to the native consumer change.
These runs establish neither speedup nor performance equivalence. Input hashing warms the ASAR
cache, build/OS caches and shared-host scheduling are uncontrolled. Fresh MCP
sessions do not establish OS-cold conditions. First post-analysis and repeated
warm-query samples are preserved separately; query CPU/memory are unmeasured.
One defined task is one analysis summary plus one retained trace, with the path
and prior Evidence already supplied. Static `cl100k_base`/tiktoken 0.12.0 counts
include both requests, replies and name/description/inputSchema contracts:
| Available-byte task, unchanged before/after | BPE tokens |
| ----------------------------------------------- | ---------: |
| Selected catalog + structured replies | 12,340 |
| Selected catalog + complete tool replies | 18,906 |
| Full catalog projection + structured replies | 202,292 |
| Full catalog projection + complete tool replies | 208,858 |
These are alternative reply/catalog projections. The selected catalog alone is
5,570 BPE; the full projection is 195,522. Output schemas, annotations, host
rendering/replay, reasoning, final answers and model billing are unmeasured.
Repeated queries/export/CLI/tokenization are validation overhead outside this
one-task token count, but inside resource receipts. Missing-case structured-task
counts change 11,372 → 11,103 by removing unsupported claims; the incorrect
baseline earns no efficiency credit.
## Regression and verification record
The initial four-file focused selection passed 24 tests. The two public cases
were rerun after adding named summary-result parsing and correcting an optional
architecture factory input. Omitting that field instead of passing null produced
byte-identical normalized simulated Evidence in all four real-case inputs.
The initial CLI-selector harness error and first typecheck failure are retained
as harness/validation attempts, not product reproductions.
After syncing the reviewed upstream changes, the updated application workflow
CLI acceptance file passed 34 tests, bringing the focused selection to 58 distinct
cases across five files. Both Joplin controls were then repeated on the final
integration base. Canonical byte identity is also used by the related version
matching/content comparison consumers; no equivalent observation-digest fallback
was found there.
Final `npm run check && npm run docs:check` passed on the integration base,
including five documentation consistency tests: 128.732 s wall, 115.518920 s CPU,
4,301,684,736 B peak cgroup memory, clean owned exit and no OOM.
Final public and static/docs outcomes are recorded in the
[compact measurements and source pins](javascript-native-handoff-identity-2026-10-15.json).Licence
Complete historical author report preserved verbatim. It illustrates the same-repository audit method; the source does not establish that this specific Skill invocation produced it. No application, provider, verifier or download command in the report was executed here. Simulated native Evidence and real static-input checks remain distinct.
MIT License
Copyright (c) 2026 morluto
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
Mode d’emploi
- Copiez la commande et examinez le paquet avant de l’installer dans votre environnement agent.
- Consultez le document original pour les outils, identifiants et environnements requis.
- Décrivez les entrées et le livrable souhaité ; comparez la sortie aux sources.
npx skills add https://github.com/morluto/rea --skill rea-ci-tests-auditAvant de commencer
- L’installation n’inclut ni abonnement au modèle, ni quota API, ni accès aux fichiers.
- Étoiles et installations reflètent l’activité, pas un test indépendant de qualité ou de sécurité.
Rendu depuis le dépôt source en conservant titres, exemples, code, tableaux, liens et images.
REA CI and tests audit
Establish which checks still catch real regressions and which costs can be removed without losing that protection. The outcome may be a verified fix, an evidence-backed recommendation, or an explicit decision to stop.
Follow the checkout's AGENTS.md, testing guide, and contribution rules. Use their current commands and pinned toolchain, rather than commands remembered from an earlier audit. Preserve the requested scope: investigation alone does not authorize edits; this skill does not grant permission to publish issues, push, or merge. Existing user authorization still applies. Keep audit notes and raw results outside the tracked tree, normally in a task-specific /tmp directory.
Establish the executed baseline
Fetch when the user asks for latest main. Pin the baseline SHA and inspected workflow revision; use an isolated worktree when changing or comparing code. Do not modify another task's checkout. Record dirty changes and the exact command, toolchain, host, selected tests, instrumentation, and prerequisites. An old run is evidence for its executed revision, even if main advances.
Trace workflow triggers, changed-path selection, job dependencies, reusable actions, package scripts, test discovery, artifact producers, and report gates through to the commands that actually execute. Include scripts, generated data, runtime registration, and subprocess boundaries missing from import graphs. Check positive and negative selection paths: required files must be discovered once per intended test configuration; deliberate platform/configuration repeats are separate coverage. A missing explicit test or missing required report must not turn green. Optional lanes may skip; preserve their stated prerequisites.
Use existing timing and failure artifacts to select suspicious families. Run relevant checks to establish a baseline; reserve a full local suite for broad CI diagnosis or uncertainty that focused checks cannot resolve. Do not rebuild or rerun an unchanged complete suite after every documentation edit.
Prove the failure or protection
For each candidate, state the expected observable behavior and its owner, then construct the smallest reproduction against production behavior. Follow the producer and affected consumers before deciding where to fix it.
- For drift, compare fixtures with current admission requirements, protocol
replies, named contracts, tool versions, and actual public entrypoints. A plausible-looking mock can fail before the behavior its test claims to cover. Search the whole affected fixture family, including other version strings. Preserve intentionally malformed fixtures and test their intended rejection.
- For a suspicious green test, perturb the relevant behavior in an isolated
copy or inject a controlled failure. Verify that the existing assertion or gate detects it. A test may merely repeat its mock, import abandoned code, or hide a subprocess/checker failure. Restore mutations after the experiment.
- For shared setup or pruning, map distinct assertions to their retained
scenarios. Preserve malformed inputs, cancellation, cleanup failure, permissions, capacity, identity, and meaningful concurrency interactions. A faster run that skips a case or weakens its oracle is not an improvement.
- For provider coverage, separate fake-producer, boundary, and actual-engine
evidence. Inspect named skips and prerequisites; fixtures do not establish real Ghidra, LLDB, browser, or other provider behavior.
Classify findings as reproduced defects, supported costs, hypotheses, or unavailable evidence. Record the baseline command, observed output/exit status, expected result, affected scope, and competing explanation. A failure on main may be pre-existing; that does not by itself make it irrelevant to a CI audit. Attribute PR failures using the executed revision and a clean baseline, not just their resemblance to a previous failure. Retry only to test a specific transient explanation; retain the original failure and stop when repetition adds no evidence. A passing retry does not establish a root cause or a fix.
Assess efficiency without losing the gate
Read measurement and CI checks when investigating slowness, caching, runner waits, shared artifacts, or log noise. Compare equivalent completed work before claiming a speedup. Distinguish setup/build time, test execution, coverage instrumentation, dependency waits, and runner scheduling; state which costs remain unmeasured.
Prefer removing proven duplicate preparation, reusing immutable setup, and reducing repeated subprocess work before shrinking correctness coverage. Preserve independent sessions where lifecycle or concurrency is the oracle. Keep live tests fresh; a cached previous pass cannot validate current host, installed provider contents, permissions, or external state.
Fix, verify, and hand off within scope
If implementation is authorized, make the smallest structural change at the owning boundary and add a regression that fails on the baseline for the right reason. Retest affected consumers and meaningful negative paths; verify cache invalidation or discovery when those mechanisms change. Read the installed Turbo documentation before changing its configuration or commands.
Run the repository's applicable checks and inspect complete current-head CI results before an authorized merge. Preserve required coverage and discovery gates. Do not suppress a failure, add blanket retries, raise timeouts, or delete distinct assertions solely to make CI green. Changes to these controls need their own demonstrated reason and retained failure detection. A slow-host failure at a literal test, child, or wait deadline belongs to the test deadline policy: remove the literal rather than raise it, and keep promptness oracles explicit.
Report each confirmed issue with exact revision, reproduction, consequence, proposed correction, and validation limits. Report performance with workload and measurements, not just a percentage. Use English issue/PR records when requested; follow the user's assignment, commit identity, review, and merge preferences without assuming them for other users. Open ready-for-review PRs only when that lifecycle is authorized. Resolve necessary review findings and check final main against the tested tree after merging.
Stop when the authorized work is verified, remaining benefits are unproven, or further changes would trade correctness for speculative speed. List genuine uncertainties without turning them into additional optimization rounds.