Skip to content

chore: track verified CookieMonster review findings #84

Closed

Description

@trmosala

Description

Reconcile CM's 18 October combined review and saved subagent reports with the current V2 split. This index records verified defects, existing work and corrections so implementation does not blindly port obsolete findings.

Reviewed CM-Split-v2@dafdcd4171466050c984d524ca0d1a09d31f100e and private v2-split@2e470e8190c319f47c86dfb83af829eb869908bf. Checked both trackers and relevant closed issues/PRs. 22 additional bugs and two decision/investigation issues were filed, alongside the five earlier issues. They are not fixes.

Verified findings and tracking

CM finding / review item Evidence Tracker
H1–H2: ordinary JSON and fenced examples become calls Reproduced CM-WPP-Bits #7
H3: XML coercion, JSON fence/escape corruption, CDATA and Unicode entities Reproduced CM-WPP-Bits #8
H4: dropped JSON calls and truncated XML batches Reproduced CM-WPP-Bits #9
H5: duplicated preamble and harmless replay differences break continuation Reproduced CM-WPP-Bits #12
H6: repeated fresh-job capability handshake Code-traced CM-WPP-Bits #13
H8: rebinding Host and blanket extension-origin allowance HTTP reproduction CM-WPP-Bits #14
H8: local API credential boundary Missing check confirmed; policy decision pending CM-WPP-Bits #15
H9: page-world controller/result forgery Trust path traced; verdict subcase reproduced CM-WPP-Bits #16
H10: shared sidecar chooses another task's session ID Threat model unresolved; not called a proven exploit opencode #83
H11: arbitrary renderer-selected open-with executable Current V2 path traced opencode #78
H12: fatal toast accepts unfinished prose Production helper reproduced CM-WPP-Bits #17
Ordinary desktop_set_panel always denied Production router reproduced opencode #81
Multiline SSE and EOF content loss Production parser reproduced CM-WPP-Bits #18
JSON-mode XHR ignored Production helper reproduced; live integration not run CM-WPP-Bits #19
Usage metrics redacted as secrets Production redactor/log writer reproduced CM-WPP-Bits #20
Logging I/O can discard a response Already reproduced and filed WPP-Bits #5
macOS activation counts hidden WPP windows Current lifecycle path traced; macOS not run opencode #80
Second main window replaces permission handlers Production helper reproduced opencode #79
Destroyed pool admits queued spawns Gated acquisition reproduced CM-WPP-Bits #21
Unused oversized browser envelope fields accepted Acceptance reproduced; DoS severity unproven opencode #83
Malformed response candidate masks valid fallback Reproduced CM-WPP-Bits #11
JSON history loses IDs; XML history changes null Reproduced CM-WPP-Bits #10
Failed inspection refresh hidden behind old data Current V2 render/control flow traced opencode #82
Detached CDP sessions retained Already reproduced and filed WPP-Bits #6
Non-ASCII titles lost Reproduced CM-WPP-Bits #22
Fenced markers stall prose streaming Gate/normalizer mismatch reproduced CM-WPP-Bits #23
Literal compaction prompt detection can drift Actual V2 initial/update mismatch reproduced CM-WPP-Bits #24
Additional review: unverified listener accepted on WPP port Already reproduced and filed WPP-Bits #4
Additional review: Linux identity collides with upstream Generated Debian archive inspected opencode #76
Additional investigation: closed-session transcripts retained 100 MiB retained in synthetic probe; eviction released it opencode #77

Corrections and items not filed as current bugs

Earlier claim Rechecked disposition
H7: injected browser permissions override user rules Legacy defect, not present in the V2 startup path. V2 no longer injects the 18 page-tool allows; plugin lifecycle defaults are prepended and permission evaluation uses the last matching rule. Legacy PR #74 remains open.
Health polling survives sidecar exit/timeout The cited V1 server.ts polling loop was replaced by the V2 service client. The claimed path is absent; do not port a fix to a nonexistent loop.
Duplicate XML IDs are never checked False for reviewed V2: a conflicting duplicate raises o1_code_conflicting_tool_call; exact transport repeats are handled separately. Preserve that during parser-port work.
Fallback selects another request's stale response Malformed-candidate masking is reproduced and tracked. Cross-request stale selection itself remains unproven.
Figma OAuth identifies as Claude Code Current V2 OAuth registers client_name=opencode. The old claim does not apply.
Packaged cwd prevents run logs being written Packaged startup changes cwd to the home directory, which is normally writable. Current log export collects the bridge's actual cwd/logs path. This is not the proved logging-I/O bug.
Opt-in console payload logging is a credential leak Page console content is intentionally gated by O1_CODE_PROXY_LOG_PAYLOADS=1. No actual credential disclosure was reproduced. Redaction is not a general prose-secret detector.
Unknown model IDs silently fall back Fallback is an explicit compatibility behavior; some newer family names fail strictly. A policy change needs a product decision, not a fabricated crash/security report.
path/old/new compatibility shims are unused Rejected: V2 edit schemas still use oldString/newString and the schema-aware compatibility path remains useful.
About 26 MB of screenshot folders are tracked The named after/before/issue49/issue58/issue62 folders contain zero tracked files in reviewed V2.
Signing/governance changes mixed into an unrelated working tree Historical workspace state, not a current runtime defect. Existing old worktrees were preserved, not modified.
Existing-release rerun fails to reset Latest Current V2 workflow deliberately publishes a beta with --prerelease --latest=false; the old release-flow claim does not describe it.
Signing secrets should be restricted to a trusted GitHub environment The workflow uses repository secrets. No unauthorized-access path was demonstrated; environment restrictions are a deployment-policy proposal, not a verified exploit.
Public/private updater contradiction and stale signing comment Repo is public; the current V2 config no longer has the cited no-certificate comment. Legacy text should not establish a current updater failure.
Unpinned chrome-devtools-mcp@latest The cited definition now lives in an unused legacy compatibility fixture, not the V2 runtime bootstrap. No active-runtime dependency issue established.
Dead AGENT_CONTRACT/HARNESS_CONTRACT and stale model/headroom comments Verified unused/stale maintenance. Optional cleanup is recorded here; preserve the live WPP prompt contract and regenerate declarations if removing exported constants.
Windows release documentation mismatch The cited legacy Windows workflow is not the reviewed V2 release path. Treat branch-specific docs separately.
Protocol incompatibility is replayable Confirmed policy: one pre-submit fresh retry is allowed. No duplicate user-prompt submission was established. The capability issue records the extra-request policy for review.

Steps to reproduce / verification record

Re-ran 12 assertion-based CM parser probes against the current private source; all 12 reproduced their expected bad outcomes. Re-ran the SSE reconstruction/thread round trip and raw HTTP guard probes. Additional production-code probes cover schema/content corruption, malformed fallback, history serialization, toast salvage, SSE framing, XHR JSON, window permissions, panel authority and queued pool destruction.

The earlier memory investigation also ran the production client with 100 synthetic 1 MiB session transcripts: about 100.3 MiB retained heap growth after GC, falling to about 1.3 MiB after explicit eviction. Its 4 client eviction tests and 30 core lifecycle/location tests passed; these do not test the missing app wiring.

The user reports that GNOME killed CookieMonster because it was using excessive memory. Which process and retention path caused that usage remains unconfirmed; no live process heap profile or OS termination log was collected.

No live WPP exploit, macOS activation or installer coexistence was established. Source tracing is marked separately from runtime reproduction. No repository source was changed.

Live memory incident follow-up

On 19 October, the installed 2.0.29-cm.1 resumed the earlier review. Its browser-inspection subagent launched the legacy browser suite; that Bun test process reached 8,514 MiB PSS, and the VM fell to 918 MiB available memory. A temporary, guarded reproduction isolates the failing dropdown assertion in browser-preview.test.ts: reporting the raw Happy DOM element drives the large allocation, while the identical boolean null predicate fails normally in 4.74 seconds at about 409 MiB RSS.

Tracked separately in #85, with legacy/V2 scope and profiling limits explicit. This is stronger incident evidence than the earlier synthetic retention probes; it does not invalidate those independently confirmed defects. The exact earlier GNOME termination was not profiled. No production fix has been applied.

Address #85 before rerunning the legacy browser suite on this box. Do not port the obsolete fixture into V2 without identifying the current contract owner.

Acceptance criteria

  • Resolve the linked functional issues against the V2 split, not just the legacy worktrees.
  • Review open WPP-Bits PRs #1, #2 and #3 for V2 conflicts and omitted acceptance cases.
  • After approved private fixes land, update the host dependency pin and run its native provider/Responses contracts.
  • Resolve the authentication and browser-boundary questions before implementing extra capability/transport architecture.
  • Record the user-reported GNOME termination for excessive memory use; establish which process and retention path caused that usage through profiling or OS evidence.

Suggested order: accidental tool execution/argument corruption and controller provenance first; then response truncation and transcript retention; then window/packaging correctness and lower-impact diagnostics. The private fix -> host pin update is an integration dependency; these issues do not otherwise require a single serial implementation.

Operating System

Ubuntu 26.04.1 LTS amd64, Bun 1.4.2; CookieMonster V2 2.0.29-cm.2 source baseline.

Activity

  1. trmosala commented on Oct 20, 2026

    @trmosala
    OwnerAuthor

    Integration update on the explicit V2 bases:

    The user explicitly approved reviewed, tested V2 merges; no releases are authorized. No issues are closed yet. #48 remains deferred and open. Memory monitoring remains active, with no watcher interruption during these checks.

  2. trmosala commented on Oct 20, 2026

    @trmosala
    OwnerAuthor

    V2 verification update, 20 October:

    • Private #26 for WPP feat(browser): recover interrupted downloads across restarts #21 is merged after independent review and 495 passing tests, type-checking and declaration generation. The host consumption candidate #88 now pins private commit 572e4494be63c6bf8cfb5ca18f2738f3dc8ec5b7, including both feat(browser): add contact and address autofill #6 and feat(browser): recover interrupted downloads across restarts #21. Its 55 host contracts and WPP production build pass. Both issues remain open until host integration is verified. The live sign-in replacement flow has not been exercised.
    • New SDK portability issue #91 has a reviewed candidate in #92. The unchanged full packed SDK check passes through Bun/Node imports, bundle guards, Wrangler dry-run and actual worker health boot. Canonical lint and all 37 type-check tasks pass. Hosted verification remains pending.
    • History paging #89 now has directly observed stale end-of-scroll callbacks resetting newer scroll writes. CM's narrow wrapper candidate passes its failing-before/passing-after regression and 27 neighbouring unit tests. Repeated production browser verification is running. No e2e assertions changed.
    • Loading fixture #90 is a verified Windows unit blocker. The SVG animation can follow browser-level reduced-motion preference independently of page media emulation. Controlled browser flags distinguish that failure from a separate local crop/raster difference. No tolerance or timeout increase has been accepted.
    • New private #27 records orphaned in-flight image-resource requests after iframe detach. A real Electron production-module probe times out after detach without loadingFinished/loadingFailed; the abort-before-detach control settles normally. Retention is source-traced, and this is not a measured heap-growth or GNOME-kill cause.
    • Live durable assistant text duplication also confirms WPP feat(browser): add bounded agent scrolling and waits #12 remains present in the installed build. It does not establish that continuation is entering fresh mode or that this caused the reported shutdown.
    • CI fix(ci): restore V2 desktop verification on hosted runners #87 now runs the full Linux unit suite successfully. The previous Windows compiler startup failure did not recur under serial execution. Later SDK, loading and component fixture failures still block integration. Two narrow CI fixture repairs pass all 11 web-search component tests and all nine What's New tests locally. Independent review and hosted confirmation are still required.

    The user explicitly permits reviewed V2 merges after checks pass. No release or installed-app replacement is authorized. All 35 scoped issues remain open. #48 remains deferred and open. The memory watcher is active, with no recent CM interruption; the installed memory problem is not declared fixed.

  3. trmosala commented on Oct 20, 2026

    @trmosala
    OwnerAuthor

    New guarded live observation on 20 October at 13:01 UTC, separate from the earlier legacy Bun assertion spike:

    • The desktop review task entered automatic compaction while a second CM task was starting.
    • The sampled CM process tree reached 7,135 MiB PSS, and VM available memory fell to 1,824 MiB. The largest two Electron processes accounted for about 1,241 MiB and 1,159 MiB. No Bun test process appears in that sample.
    • The coordinator's existing memory guard interrupted the two active CM sessions. The installed desktop and backend stayed running, and unfinished source edits were preserved.
    • Roughly five seconds later the tree was about 4,929 MiB PSS and VM available memory was 4,470 MiB.

    The raw process samples and interrupt responses are retained locally. This identifies compaction/startup overlap as a concrete profiling target; it is not a heap-retention diagnosis or proof of what caused the earlier GNOME termination.

    Work now runs one CM task at a time. The preserved desktop patch is continuing in a fresh child session with a short handoff; installed CM supports the shorter child-worker cleanup policy. The interrupted listener task has not been resumed. The 2 GiB VM floor and other memory guard limits are unchanged.

  4. trmosala commented on Oct 20, 2026

    @trmosala
    OwnerAuthor

    System-journal follow-up: the OS records confirm global memory exhaustion and kills under the desktop identity used by CM at the time. These are kernel OOM records, distinct from later memory limits on our validation jobs.

    Times below are SAST, UTC+2. Anonymous resident memory comes from the kernel's anon-rss field.

    Time Victim Anonymous RSS Recorded scope
    18 October 11:24:29 ai.opencode.des, PID 1225863 391.5 MiB GNOME Shell session scope
    19 October 19:39:51 bun, PID 104284 6.68 GiB app-ai.opencode.desktop-4147811.scope
    19 October 19:45:04 ai.opencode.des, PID 108279 916.2 MiB app-gnome-ai.opencode.desktop-106798.scope
    19 October 19:47:07 bun, PID 111184 11.25 GiB app-ai.opencode.desktop-106798.scope

    Each event has constraint=CONSTRAINT_NONE and global_oom, followed by Out of memory: Killed process. This confirms real system memory exhaustion and the large Bun victims. The kernel victim alone does not identify the process that initiated all memory pressure. CM still shared the upstream executable/launcher identity before #76; the journal did not retain an executable path for these PIDs, so these names alone cannot distinguish installations or prove which entry matches a particular reported CM shutdown.

    The journal does not retain the exact historical Bun command. The reproduced Happy DOM assertion failure in #85 independently reached about 8.3 GiB, so it is a demonstrated dangerous path in the review workflow; attribution of a particular historical PID to that command remains unproved. The transcript retention in #77 is a separate confirmed defect, not proof that transcripts caused these Bun spikes.

    The legacy DOM fix is now reviewable at f74d46a68d, on its separate bounded-dom-tests branch. Its deliberate mounted failure exits with bounded output under a memory limit. The equivalent V2 browser keeper was verified separately. The integrated V2 fixes remain in draft PR #96 pending final CI; this comment does not close issues.

    The installed CM and backend were not restarted or updated. Later CONSTRAINT_MEMCG kills belong to explicitly limited validation scopes and must not be counted as another installed-app shutdown. Raw journal evidence is retained locally as memory-historical-oom-evidence.json and memory-os-journal.jsonl in the investigation state directory.

  5. trmosala commented on Oct 20, 2026

    @trmosala
    OwnerAuthor

    Current integration status, 20 October:

    • The private backend is merged in CM-WPP-Bits PR #28, at 008fec76ed058c9a7a17a9833528405413c16ab8. Its 21 verified issues are closed. Private feat(browser): add explicit agent tab management #16 stays open pending genuine signed-in WPP capture; the installed login window still returns HTTP 401.
    • Public fixes are in draft fix(cm): integrate reviewed V2 reliability fixes #96, currently 2e63c7af917e593d8f7a3b6dd6894064f4f0bcba. Full CI has passed Linux unit, packed SDK, compiled-service, Node-build and generated-file gates. Windows unit tests hit a different five-second multi-process Git fixture budget in the server worktree test. Its assertions are unchanged in the local correction; seven targeted tests pass. The end-to-end jobs are still running.
    • The signed macOS installer passed notarization, stapling and mounted-version checks, but packaged activation timed out in the preceding run. A diagnostics-only run is collecting the failing stage and sanitized native-window state. No packaged activation pass is claimed.
    • Final component verification also found fix(app): release focus after immediate mobile drawer dismissal #97, an early mobile drawer dismissal that retains focus. A controlled animation-frame regression fails on the original code with normal motion; reduced motion passes as a control. A local Corvu patch passes that regression and the six existing overflow cases. Broader verification is ongoing. This raises the tracked backlog to 36, with 21 closed and 15 open.

    #48 remains deferred. This tracker remains open for integration, live #16 qualification and reconciliation of the new finding. The installed CM and backend have not been updated or restarted; no release has been published.

  6. trmosala commented on Oct 21, 2026

    @trmosala
    OwnerAuthor

    Merged into CM-Split-v2 as 290d7d6353c0c7a4a356af780b5687f2dcfe5cfc through PR #96. Its tree exactly matches reviewed source 68ab9205709350385471eb2d383b3cff3c18e678.

    The review findings are reconciled and their V2 implementations integrated through public PR #96 and private PR #28. Legacy private PRs #1–3 were reviewed, selectively ported and closed without merging the obsolete extraction. The host pins private source 8bb2cd4e7d0588c3ef37735453830e7206f5a455; actual host provider/Responses contracts passed 59 tests with 657 assertions before testing stopped.

    Both product decisions are implemented: WPP uses an automatically managed per-user key, and browser operations trust the authenticated backend across tasks while retaining native/page grants and a 64 KiB UTF-8 request admission limit.

    Memory investigation established two distinct problems. Closed-session transcript retention is fixed in #77. The observed review workflow also had a Bun test child reach 8,514 MiB PSS while formatting a failed mounted DOM assertion. Its bounded failure/viewport correction is recorded in #85. Kernel logs independently confirm global OOM kills, including Bun victims at 6.68 GiB and 11.25 GiB anonymous RSS under the desktop identities used at the time. Those logs did not retain the exact invocation or executable path. The exact relationship between a particular historical victim and the reproduced assertion remains undetermined. It is not claimed as proven, and further source changes cannot reconstruct missing historical process data.

    Closing the implementation/reconciliation tracker at the user's explicit request to close the issues before one final test pass. Final combined/platform validation is deferred, not reported as passed. Corrected Mac native activation and genuine signed-in WPP capture are specifically outstanding; the installed WPP window last returned HTTP 401. Any final-check failure will be recorded against its issue.

    #48 remains explicitly deferred and open. The installed CM and backend were not updated or restarted, and no release was published. The memory watcher remains active.

  7. trmosala commented on Oct 21, 2026

    @trmosala
    OwnerAuthor

    The one final verification pass ran after all 36 scoped implementation tickets were confirmed closed.

    Source: public PR #96, merge 290d7d6353c0c7a4a356af780b5687f2dcfe5cfc, private pin 8bb2cd4e7d0588c3ef37735453830e7206f5a455.

    • Public run 39533186971 completed with an overall failure due to one Windows unit fixture. Linux unit/package/SDK/compiled-service/Node/generated checks all passed. Linux production browser 449, service worker 12, session UI 135 and app components 121 all passed.
    • Windows browser/component phases passed with automatic retries: production 445 first-pass plus four flaky cases, service worker 12 passed, session UI 134 first-pass plus one flaky case, app components 120 first-pass plus one flaky RTL overflow-layout case. The new controlled early-dismissal regressions passed. These are six intermittent cases, not a clean first-attempt Windows result.
    • The only failed unit test was Session.shell > runs shells concurrently with an active model and waits for each shell's own completion, at its inner five-second process-start file gate. The remaining Windows compiled service and Node build gates passed. Windows app and CLI unit packages did not complete after Turbo stopped on the core failure, so no Windows pass is claimed for those packages.
    • PR #98, merged as ac02f1195eb5df4fef02b35d30ecb910e5a3d374, adjusts that Windows startup gate to 15 seconds and the enclosing Windows fixture to 30 seconds. Other-platform deadlines and all concurrency/completion assertions remain. It changes no production source. Canonical lint/type checks passed 37/37, but the adjustment was not retested, honoring the single-pass instruction. No further test run was dispatched.
    • Mac run 39533190446 passed signing/notarization/stapling/mounted-installer checks and real native activation with a restored visible main window and preserved hidden worker. It also passed all 630 private tests with 3,151 assertions and all 59 pinned host contracts with 657 assertions. Release publication was skipped.

    Genuine signed-in WPP capture remains unverified; the installed window returns HTTP 401 and a readiness question is pending. OS sandbox qualification of that capture is not claimed. Exact historical OOM-victim command/installation attribution remains unavailable, as recorded earlier.

    All 36 scoped implementation issues remain closed; #48 remains explicitly deferred and open. The installed CM/backend were not restarted or updated. The memory watcher remains active.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions