This file preserves dated release checks. The current package version is declared in pyproject.toml and the plugin manifest; see the changelog for changes after a dated check. Historical canaries do not prove the status of a current installation. The commands near the end reproduce the repository checks on the current checkout.
The compatible signed runtime has sequence 1790433141 and SHA-256
d13fa3a736920411badd3af95cdab3d38a453efdeab2ab6ac78dc0752f050d32.
Ruff, all 260 tests, repository and plugin validation, the 25-file documentation
check, reproducible runtime checks and signature verification passed. The hook
definitions and publisher bootstrap remain unchanged.
The preview tests exercise verified nested-child relocation across UUIDv7 UTC
dates, preserved relative subdirectories and parent snapshots, unchanged direct
child/workspace output, and rejection of foreign roots, traversal, symlinks and
unproven lineage. Relocation-only PermissionError cases cover mkdir and file
creation, a single checked child-workspace fallback, the same collected snapshot
and filename, and no overwrite. Source mutations during rendering or the denied
native write are rejected before either destination is written. Independent
review reproduced these two source-recheck gaps before the final fix and verified
both regressions afterward; ordinary OSError does not trigger fallback.
A disposable Codex home installed version 0.7.3. Its manifest, hook definitions,
both runtime archives, release manifest and publisher policy matched the
checkout byte-for-byte. All 31 preview tests passed with package modules loaded
from the installed signed archive. Invoking the installed SubagentStart command
against a synthetic native catalog/transcript and then its durable preview
command produced the child's own dated path. A real OS permission denial
(EACCES, directory mode 0500) produced the checked workspace fallback; the
fixture's original permissions were restored. Both cards were mode 0600, kept
the observed child counters, and left the parent sentinel unchanged.
An offline harness executed the unmodified reader, schema and path-checking functions extracted from local Codex Desktop 26.924.22138 (build 11645), with minimal host adapters and real filesystem I/O. It accepted the installed runtime's child-root and workspace-fallback files and rejected the parent sentinel under the same child ID, with the workspace feature flag both enabled and disabled. The six direct-read checks preserved file hashes and mtimes. This establishes output-path compatibility for that installed reader; it is not native GUI painting or model-driven lifecycle verification. Existing Task runtime pins and previously emitted snapshots remain unchanged.
Aligned the reporting interface reviewed from Run Budget source range
b5bf01e739c4f11842c843333ca8793fb6573976..f3ffb509f8f07109360921d48a640f3203035438,
including the preceding footer delimiters. Usage Reports keeps its own branding,
report-only implementation, native ownership and release trust state.
Ruff, all 249 unit/integration tests, repository validation, the 25-file
documentation check and signed-release verification passed. The deterministic
runtime has sequence 1790406877 and SHA-256
e1ab602a0137fbc0bebf29ec8ce3606255303156197765a11f8c416ad5be32c0.
Tests cover eligibility before the preview command, parent/child command ownership,
own-reference wording and compact JSON round-trips for Unicode, spaces, quotes,
backslashes and control characters.
A disposable Codex home installed this checkout as version 0.7.2. Installed runtime, release manifest and hook bytes matched the checkout. Six focused tests loaded the installed signed zipapp and passed; its actual CLI read back the reporting settings. Hook definitions remained byte-identical to 0.7.1. The maintainer's active installation and native hook grants were not changed.
These are deterministic code, synthetic lifecycle and isolated installation checks. No Usage Reports model-driven conversation or Desktop painting canary was run for this change. Run Budget's six-model high-effort matrix is source-only evidence; its earlier failures and unknown results are not replaced or imported as Usage Reports validation. Existing Tasks may retain an older pinned runtime and previously injected instructions.
Checked on 2026-09-26. Ruff, all 248 tests, repository and plugin validation,
the 25-file documentation check and signed-release verification passed.
The final signed runtime has sequence 1790392814 and SHA-256
021c5b2fc5176b23290fa97ca0b97e41ba595c3bc4686922d6d9e3971ec26253.
Real SQLite regressions cover delayed request writes after revocation, conflicts before any cached request, reopen/source removal, collection interleavings, legacy schemas, bounded saturation and a broken revocation schema. Independent checks inserted an otherwise valid request as an older runtime would: the new collector still excluded the revoked child. Original Task/root/direct-parent and role checks reject drift and contradictory explicit roles at preview, Stop and completion; supported legacy baselines remain usable. Failed completion checks leave the original Stop files unchanged. Nine-language rendering tests cover child labels, selectors and escaped parent names, with unchanged root text.
A fresh disposable Codex home installed the 0.7.1 child-report candidate before
the final connection-only cleanup described below. All nine hook
definitions matched 0.7.0; only the isolated fixture received its previously
reviewed hashes. Native read-back showed nine enabled/trusted hooks. A real
codex exec spawned one child and exited successfully. Both emitted preview
references; the child's card used child-specific usage/settings labels and
showed its direct parent. Two Stop receipts and two completion revisions were
written. The child baseline's original root/direct-parent hashes matched native
lineage, and its selector matched the parent descendant row. The child observed
52,300 own-turn tokens; the parent observed 61,640 of its own and 52,300 separately
as descendant usage. Receipts contained no raw native Task UUIDs. Temporary
authentication material was removed after the run.
CI identified Python 3.13 ResourceWarnings from unclosed SQLite fixture handles and the readonly quota timing index. Both were closed explicitly without suppressing warnings or weakening CLI stderr assertions. The final complete Python 3.13 suite passed with tracemalloc and ResourceWarnings enabled, with zero warnings. A second disposable home installed the exact final runtime, with nine trusted hooks. Five tests loaded from that installed signed zipapp passed, covering revocation races and quota connection lifetime. The final runtime also accepted the canary's actual parent/child baselines and rendered the child labels, selector and parent correctly from that evidence, without another model request.
The active user installation and hook grants were not changed by this check. This proves isolated CLI hook execution and generated child-card content; Desktop painting remains unobserved. Token totals are canary observations, not plugin overhead or account billing.
Checked on 2026-09-26 with Codex CLI 0.157.1. Ruff, 229 tests, repository
validation, the 25-file documentation check, staged sensitive-data scan, and
signed-release verification passed. Runtime SHA-256:
05396ebd97e3e7a30d5525330f0f3ad2b13726f8d67b6b2c493c02019d0d6c14.
A fresh disposable CODEX_HOME installed this checkout's local marketplace and
plugin 0.7.0, with report data in a separate disposable directory. Native
hooks/list first reported nine enabled hooks needing review. Each installed
command matched the packaged hook definition; exact current hashes were trusted
only in this isolated fixture, after which all nine read back enabled and trusted.
This does not grant trust to the maintainer's disabled installation. That
installation needs its normal plugin update and /hooks review if selected.
One real codex exec turn spawned one subagent and exited successfully. The
installed hooks produced one parent and one child Stop receipt, plus completion
revisions. The child used agent_turn_stop_boundary; its 12-character hashed
selector matched the parent's child row. Its observed total was 17,329 tokens;
the parent recorded 46,378 of its own and the same 17,329 separately as
descendant usage. Neither receipt contained a raw Task UUID or agent_path.
This CLI run did not produce an inline visualization file, so it verifies native
hook execution and report correspondence, not Desktop card painting or the
original screenshot's renderer failure.
Checked on 2026-09-24 in this checkout. Ruff, all 223 unit tests, repository
validation, the 24-file documentation link check, and the staged sensitive-data
scan passed. A fresh isolated
CODEX_HOME installed this checkout's local marketplace and plugin 0.6.0.
The installed CLI reported 0.6.0, and its signed-runtime status reported an
active verified 0.6.0 runtime. No hooks were trusted in that disposable home;
an offline native trust query returned unknown. This check verifies package
installation and local CLI selection, not hook execution or Desktop rendering.
Validated on macOS, Python 3.12.8 and Codex CLI 0.154.0 on 2026-09-17.
- All 222 unit/integration tests passed, including independent OpenSSL signing, actual fixed-command execution of signed A/B runtimes, old Task pin retention, signature/payload/metadata rejection, anti-replay rollback, corrupt-cache fallback, disabled and in-flight updates, private hashed state, future entry migrations, bounded worker scheduling, and the packaged hook/CLI.
- Ruff, deterministic artifacts, signed manifest/runtime verification, plugin validation and the staged sensitive-data scan passed.
- A real native CLI canary installed
0.5.0+canary.Ain a disposable Codex home. Trust grants were seeded once in that isolated fixture only. The installedUserPromptSubmitcommand started its ordinary background worker, fetched the signed0.5.0runtime from the public GitHubmainendpoint, and activated its exact published digest. No model call was made. - A new synthetic Task's actual hook executed the published runtime. The older
Task's verified pin remained on A. Native
hooks/listreported all 8 hooks enabled and trusted, with identical hashes before and after the automatic update. The entire isolated native config file was unchanged. - A subsequent normal package reinstall from A to the published version also preserved all native hook hashes and trust states without another grant.
Published runtime SHA-256:
9aa51aad7104c062ac1704d7ca8d2d28add2ac01b5c4a28450af43a46694cdd0.
This proves real download/verification/activation, installed hook execution and native trust classification. It is not a model-driven Desktop rendering test. Migrating an existing pre-publisher installation still requires one review of the new fixed entry. Future hook/key/entry changes require their own native review. See signed update controls and boundaries.
Validated on macOS, Python 3.12.8 and Codex CLI 0.154.0 on 2026-09-17.
- All 210 tests passed, including first-prompt/installation deduplication, cache expiry and failure backoff, unknown trust, disabled hooks, scoped plugin identity, private hashed state, bounded/reaped native RPC, independent embedded commands, CLI fallback, and verified TLS with the macOS system-CA fallback.
- Ruff, repository/runtime validation and Plugin Creator/skill validation passed.
- A fresh isolated Codex home installed 0.4.0; installed runtime and hook bytes matched source. Its actual installed CLI fetched the public GitHub manifest successfully and read native pending trust. No model request was made.
- A separate native upgrade canary installed synthetic versions 9.0.0 and 9.0.1
using the real plugin CLI. Trust was seeded only in disposable isolated homes.
hooks/listobserved all 8 hooks trusted before the upgrade; afterwards the 7 changed main hooks weremodified, while the standalone reminder retained exactly the same native hash and remainedtrustedacross cache-version paths. - Invoking that installed reminder emitted both update and reauthorization notices, then no duplicate on the next prompt. After isolated re-trust, the trust warning disappeared. The cached-release/native-trust fixture took 0.137 seconds; this is an observation, not a production latency guarantee.
Installed runtime SHA-256:
d150ca31a6bde3222323818740d10f2f24ee85bafecd088542a3756034bd5d85.
The test proves native trust classification and the actual handler's output. It is not a model-driven Desktop rendering test, nor evidence that untrusted hooks can execute. First installation of the reminder still requires user review. The maintainer's active installation and hook grants were not changed.
Validated on macOS, Python 3.12.8 and Codex CLI 0.154.0 on 2026-09-16.
- All 198 unit/integration tests passed. New coverage uses synthetic catalogs and real local subprocesses for exact project scope, foreign metadata, symlinks, native/legacy counter separation, same-source resets, distinct resume launch attribution, private hashed state, hook notice deduplication, partial rendering, exit propagation, ambiguous JSON usage and nonblocking telemetry failures.
- Ruff, repository/runtime validation and Plugin Creator validation passed.
- Fresh isolated Codex homes installed the plugin from its local marketplace. The installed bootstrap and pinned runtime emitted one exec notice after a synthetic catalog insertion and none on repetition. The post-tool call took 0.0820 seconds in this small fixture; this is not a production latency bound.
- Installed CLI list/render and source identity read-back passed. An installed
launcher invoked the real Codex binary with
exec --help, forwarded its output and recorded start/exit with status 0. This smoke made no model request. - Read-only parsing of one selected real native exec session found its metadata and completion event. Parent ownership and missing usage stayed unknown. No real transcript, identifier, path or generated report is included here.
Installed runtime SHA-256:
25b65703d01bb6b08dafcfab405f86e2a15fffe00e5fd7b19de9fdea1440641a.
The hook test uses synthetic native state and actual installed hook commands; it does not prove every Desktop host delivers or displays every hook event. Native turn completion does not prove process exit. Only instrumented launches retain receipts for ephemeral runs; no background service or shared-budget enforcement was added. The maintainer's active plugin setup was not changed.
- 182 unit/integration tests passed locally. Coverage includes positive and negative offsets between native request and legacy event totals, malformed native records, true same-source resets, and a turn-counter decrease whose preceding record has fallen outside the bounded tail.
- A fresh isolated Codex home installed 0.2.1. Actual installed bootstrap and
pinned-runtime commands rendered synthetic preview delta 600, settled Stop
delta 800 in 0.0924 seconds, then reconciled delta 1,000 / Task total 2,000.
Legacy event totals were independently offset. A next-turn event of 9,990,000
was excluded, all original Stop files stayed byte-identical, and the installed
CLI returned revision 2 with
counter_source=native_request. - Read-only inspection of the selected affected native turn confirmed separate
monotonic counter streams with different historical baselines. The corrected
snapshot returned the explicit native turn counter through
task_complete. Historical settings remained partial; private records and user totals are excluded from this repository. This is parsing evidence, not a new phone UI or live hook-delivery test.
The installed runtime SHA-256 was
1158a88d63c2c170d1785beec399303e316373debd89215d1c04197e84a8669f.
Verified on macOS with Codex CLI 0.154.0 and Python 3.12.8:
- 173 unit/integration tests passed, including 14 completion-boundary and 18 background-worker cases. Ruff, packaging validation, nine-locale parity, documentation checks and the Plugin Creator validator passed.
- A fresh isolated Codex home installed 0.2.0 from the local marketplace. Its real installed bootstrap commands handled start and Stop, launched the pinned background runtime, and returned Stop in 0.0864 seconds. Synthetic turn usage increased from 200 to 500 after an explicit completion event; a following turn's counter of 9,999,000 was excluded. The installed CLI read revision 2. All original Stop JSON, HTML and Markdown bytes remained unchanged.
- Independent fault injection delayed the child catalog beyond the deadline:
the worker exited after 25.012 seconds as
expired, with zero publication calls. A dedicated deadline exception bypasses ordinary collector error handlers. Concurrent edits and symlinks at the original report are preserved because completion publication never writes original Stop files. - A read-only check of this installation's actual native completion record
returned
completion_observed=true. Its bounded tail remained partial; this check is not a complete accounting claim or a new live model conversation.
The installed runtime SHA-256 was
ad0a6916d3e33861c90f6045a921a4dc85e21dee8a52601fd6e0260674338e52.
The user confirmed the first inline card displayed synthetic version A / 100. After overwriting that same HTML path with B / 250, a new reference displayed B, but the original card still displayed A after navigating away and returning. One intermediate return also showed no card. This demonstrates readable updated source with a fresh reference, not reliable refresh of an existing phone card. The implementation therefore preserves inline snapshots and publishes a separate completion revision selected by subsequent Task-report queries.
Screenshots and raw native records stay private and are not committed. These findings apply to this tested host/session; they do not establish universal mobile behavior. The maintainer's active plugin is separate from isolated tests.
The 0.1.0 extraction was verified on macOS with Codex CLI 0.154.0 and Python 3.12.8. The automated CI matrix targets Python 3.10, 3.11, 3.12, and 3.13 on Linux. Current CI runs.
| Check | Evidence |
|---|---|
| Regression suite | 141 unit/integration tests passed locally, including foreign-Task start/Stop isolation, counter resets, partial contexts, child coverage, and nonblocking runtime failure. |
| Standalone installation | A fresh isolated Codex home accepted marketplace registration and codex plugin add codex-usage-reports@codex-usage-reports. |
| Installed hook lifecycle | The installed hook commands received a synthetic native catalog/transcript: start, preview, and Stop produced a local card and settled receipt. Task counter 1,600; turn delta 600. |
| Installed CLI read-back | The installed CLI read the receipt and returned one recorded turn, cumulative 1,600 and delta 600. |
| Real native catalog | A read-only selected-Task query against the current local Codex catalog returned usage_status=observed and a native counter. Private output stayed outside Git. |
| Python packaging | A clean virtual environment built and installed the wheel with no runtime dependencies; its console command returned report settings. |
| Public examples | Actual renderers use synthetic fixtures only. |
| Browser checks | Four locales × two widths (390, 900) × two themes = 16 cases. No horizontal overflow or page errors; each disclosure toggled with the keyboard. |
| Privacy | Staged-index and reachable-history scans run in CI, including zipapp members. |
The installed hook test invokes real installed hook commands with synthetic native data. It does not establish that every Codex host delivers every event. The real catalog query checks current native parsing, not an end-to-end model conversation. Hook trust, native schema availability, and inline rendering remain host-dependent.
The earlier documentation frame supplied its own theme variables, masking the report fragment's dependency on host foreground colors. A reproduction with an inverted foreground makes the old template's title, totals, and disclosures match the background (contrast 1:1), while muted elapsed text remains visible. This matches the reported symptom; it does not identify the iPhone host's internal implementation.
The updated fragment uses paired browser CanvasText/Canvas colors and scoped
text styles. Playwright 1.62.1 exercised Chromium and WebKit across four locales,
390/900 px widths, light/dark schemes, and missing/normal/inverted host tokens:
96 cases passed, with no horizontal overflow or page errors, readable text
(contrast at least 4.5:1), and keyboard-operable native disclosures. Both
collapsed and expanded states were checked. WebKit screenshots were inspected.
These are browser tests, not an end-to-end test on the iPhone remote app. Existing artifacts do not refresh after installation, and existing Tasks can keep their pinned runtime; start a new Task to pick up the corrected renderer.
To reproduce (optional development dependencies only):
npm install --no-save --package-lock=false playwright@1.62.1
npx playwright install chromium webkit
python3 scripts/generate_examples.py
node scripts/check_mobile_card.cjsUse BROWSER_CHANNEL=chrome for an existing Chrome installation. CI runs this
fragment matrix separately from the framed screenshot checks.
A post-update phone-created Task completed without starting an automatic report.
Native hooks/list reported all 11 upstream Run Budget hooks as enabled but
modified, with no discovery errors. This is a separate condition from card
rendering: the installed hook definitions require a new trust review before
execution. Installation checks and renderer checks alone do not prove that the
host will execute an updated hook. The installed update changed only the runtime
version and card template; changing the pinned runtime digest still changes the
hook definition hash.
Use the Codex CLI /hooks review flow after updates. Do not infer trust from
codex plugin list, and do not replace a trust review with a bypass flag or
manual trust-hash edits. See the official hook trust rules.
ruff check plugins tests scripts
python3 -m unittest discover -s tests -v
python3 scripts/build_hook_runtime.py --check
python3 scripts/validate_repo.py
python3 scripts/check_docs.py
python3 scripts/check_sensitive_data.py --index --fail-on-findings
python3 scripts/check_sensitive_data.py --history --fail-on-findingsFor manifests, also use Codex's Plugin Creator validator. Rebuild with
python3 scripts/build_hook_runtime.py after packaged source changes.
python3 scripts/generate_examples.py
node scripts/capture_examples.cjsThe screenshot step requires the optional Playwright development tool and a Chromium browser. Details.
- Local Codex catalog and rollout formats are implementation details, not a stable public reporting protocol.
- Python runtime minimum: 3.10. macOS is the locally exercised host; Linux receives CI coverage.
- Native Windows plugin hook installation is not verified in this release. POSIX file-safety primitives are used by the runtime; Windows users should not assume parity from Python installation alone.
- A compatible desktop visualization surface is required for inline cards. Saved Markdown and HTML are available independently.
- No reporting API key is needed. Optional quota reads use the user's existing Codex account context and can be unavailable.
- Existing Tasks can retain pinned hooks after a plugin update. Begin a new Task for the new installation.