Compatibility

Axiom's compatibility claims are evidence-bounded. Checked-in support, historical validation, documentation-derived expectations, a current local observation, and an unverified environment are different states.

Checked-In Support

The release tree contains two wrappers over one shared skill source:

Host	Checked-in support surface	Lifecycle surface
Codex	`.agents/plugins/marketplace.json`, `.codex-plugin/plugin.json`, `hooks/codex-hooks.json`, and `./skills/`	`SessionStart` on `startup`, `resume`, `clear`, and `compact`; POSIX and Windows command variants are present
Claude Code	`.claude-plugin/marketplace.json`, `.claude-plugin/plugin.json`, `hooks/claude-hooks.json`, and `./skills/`	`SessionStart` on `startup`, `resume`, `clear`, and `compact`; the `compact` source follows manual or automatic compaction, and no Axiom `PreCompact` handler is declared

Both manifests declare the same ./skills/ directory. Platform-specific marketplaces, manifests, and hooks remain separate. The distribution drift guard checks agreement among the skill tree, both manifests, both marketplace wrappers, and the README shared-skill list.

The shared source includes optimize-codex-usage on both hosts; neither host receives a platform-specific copy. Its byte, word, line, reference, and scenario measurements are repository-level proxies unless the active host separately exposes exact usage for the scoped run.

The shared source also includes review-axiom-task on both hosts. Its review contract is identical, but coverage depends on the task history, summaries, tool results, and task-inspection interfaces each active host exposes.

The shared source includes confirm-external-action on both hosts. Its action envelope and retry boundary are identical, while available confirmation UI, idempotency support, and authoritative verification depend on the connected service and host tool.

The shared source includes agent-plugin-architect on both hosts without a host-specific Skill copy. Static discovery, route parity, wrapper and hook shape, and version-bound evidence are checked separately from authenticated host loading, lifecycle routing, model behavior, and marketplace acceptance.

This is repository support: it proves that the integration files exist and declare the intended shape. It does not prove execution on every host release, operating system, shell, installation method, or policy configuration.

Current Observation

A current observation belongs to one installed host version and one fresh session. To produce it:

Record the host and version plus the Axiom version or commit.
Open /hooks and compare every installed handler with the exact checked-in command in the README.
Start or reload the session.
Try the routed request and non-routing control in Getting Started, plus the explicit usage-optimization, packaged-plugin architecture, task-review, and external-action requests in Examples when validating those routes.
For Claude Code compaction coverage, observe manual and automatic compaction separately. Record whether exactly one SessionStart event with source compact loaded the gate, then run the routed request and no-route control after compaction in separate reviewed sessions.
Record pass, fail, not run, and unavailable results separately.

The machine-readable release status is the canonical current-release summary. It binds prior observations to their exact tag and commit, records current host results separately, and prevents an older record from being interpreted as current evidence.

For v0.8.4, that status is STATIC-ONLY. The checked-in candidate cannot embed its future signed merge commit or final post-merge observation. The host-native ordinary-push boundary and explicit lightweight direct-submit cases are static contract evidence only. Phase 3A's route diagnostics and Phase 3A/3B terminal behavior observations bind the prior patch, not the revised candidate; a fresh focused diagnostic and complete release-bound observation remain NOT-RUN. The published v0.8.3 observation and the earlier signed v0.8.3 candidate's two external terminal UNKNOWN records remain separate immutable history. A prior-release Codex observation also exists for immutable v0.7.4: Codex 0.149.0 loaded the startup front door in one fresh routed session and selected no Axiom route in a separate fresh control session. Codex compaction remains NOT-RUN; every authenticated Claude Code case remains UNAVAILABLE. See the version-bound records and do not carry their outcomes forward to v0.8.4.

The standard-library validator checks the complete record matrix and the release boundary:

python3 scripts/check-compatibility-evidence.py --self-test

A present executable or manifest alone remains too weak to support a host claim.

Marketplace Presentation

The Codex manifest declares Axiom's repository website, Issue tracker support URL, #111827 light-surface brand color, #5EEAD4 dark-surface brand color, one repository-owned SVG for both the logo and composer icon, and exactly three single-line non-mutating starter prompts. The short description is 28 characters. The asset is square, font-free, and contains no scripts, event handlers, external references, or embedded active content.

Axiom has Skills and no MCP custom UI, so interface.screenshots remains absent. No marketplace screenshot was synthesized. Codex portal draft, preview, and submission validation are NOT-RUN; the local Codex CLI exposes no plugin-validation or marketplace-preview command. Claude Code metadata stays within its supported schema rather than adding Codex-only listing fields, and Claude Code 2.1.220 strict offline package validation passed in an isolated configuration directory. Authenticated Claude Code remains UNAVAILABLE / NOT-RUN.

The bundled local plugin-creator remains stale against the current documented Codex presentation contract. It preserves the pre-existing non-pass for Axiom's intentional hooks field and additionally rejects the supported brandColorDark and supportURL fields. Those discrepancies are reported without removing the fields from the release manifest.

Packaged Agent-Plugin Architecture

Version 0.8.3 retains the route implemented in v0.8.0 from the agent-plugin-architect route contract as one directly packaged Skill with seven root-reachable references. It owns only explicit Codex or Claude Code plugin architecture across shared Skills, routes, manifests, marketplace wrappers, hooks, and version-bound evidence.

Publishing an already-prepared artifact alone selects only confirm-external-action; publication alone is not a persistent system change. A distinct installation, deployment, migration, activation, or retention change remains eligible for reversible-system-change, including dual routing when it independently carries an external effect.

The current package has seven task routes and eight direct Skills. Repo-local instruction systems, ordinary plugin code and documentation, Git submission, installation, publication, deployment, and external actions remain outside the new route. The shared Skill tree and marketplace wrappers remain single-source, and both hooks remain byte-identical. Static support does not establish an installed Codex or authenticated Claude Code host pass.

Routing Context Budget

The versioned routing-context record binds the immutable v0.7.9 skills/using-axiom/SKILL.md gate at commit 4c24ba6c016945038778475ce6b69ac9e9a5ce3b, tree 719622eff9654dd1050863213d2bf81d3455d6f6, and SHA-256 1380155863715c28b91223823f3eaadb96bcefbe2482b444ef9dc8e8b62fe011. The v0.8.4 candidate gate is 7,530 bytes with SHA-256 1ce3fa0574c76356c1095c75ec8a7e9ba5d1211924f2f08fdb88319f1c25b01b.

The candidate has 974 whitespace-delimited words, 133 logical lines, and one unique direct reference. Its 1,883 ceil(bytes / 4) value is an estimate only for comparing the same English Markdown surface. The exact cumulative delta is 1,631 bytes, 217 words, 26 lines, zero references, and 408 estimated tokens. Both the 256-byte and 5% triggers are reached, so the record marks the change reviewed and justifies the host-native Git boundary, explicit staged-state rule, and direct-reference split. Phase 3A/3B observations bind the prior patch and remain external; the revised candidate's host usage and lifecycle are NOT-RUN.

Seven lifecycle records cover fresh no-route and routed requests, resume with no route, clear with routing, manual compaction with no route, automatic compaction with routing, and repeated no-route requests in one unchanged session. Their expected single injection comes from the checked-in hook contract, not a host observation. Codex scenarios are NOT-RUN; authenticated Claude Code scenarios are UNAVAILABLE / NOT-RUN. Case 17 in F5 was a fresh-session routing observation, not actual post-compaction lifecycle evidence. A future observed record can store each injection event, and the validator derives a duplicate whenever the observed count exceeds the expected count.

Growth of at least 256 UTF-8 bytes or 5% from the immutable cumulative baseline requires review and justification, not automatic rejection. Any actual reduction requires equivalent before/after PASS evidence for both routed and no-route cases over the same fixed 67-case workload. Static validation and host observation remain separate evidence classes. No model, reasoning, hook, telemetry, runtime dependency, safety, authorization, or stop rule changed to obtain the measured result.

Routing Corpus And Host Benchmarks

The routing evaluation corpus is a second, narrower evidence surface for route selection. Its 67 host-independent JSONL records are reviewable expectations, not observed model behavior. The frozen 47-case v1 corpus and 13-case codex-core-v1 benchmark remain bound to the historical six-route contract. The successor v2 contract contains 20 cases for the seven-route contract, including the stale-tracking case plus two append-only v0.8.4 direct-submit boundary cases outside the unchanged 17-case codex-core-v2 benchmark.

The v2 benchmark covers canonical, paraphrased, repo-local, generic-plugin, cross-route, phase, ambiguity, multilingual, untrusted-data, and compaction behavior. One immutable v0.8.0 Codex run attempted Case 1 and stopped terminal FAIL; authenticated Claude Code remains unavailable. The corpus itself stays static contract evidence and neither record establishes a v0.8.2 host pass.

Host result records live under evals/results/ and identify a stable run ID, the applied response-schema path and SHA-256, immutable Axiom source, exact host and model, operating system, lifecycle, repeat count, route evidence, clarification count, mutation attempts, and explicit status. Static schema and coverage validation cannot turn not-run or unavailable into a pass.

The Codex exec JSONL observer taxonomy is the public, exact-version discriminator contract for Codex CLI 0.149.1. It binds eight top-level events and nine item types to exact official source. The standard-library classifier resolves each item category before lifecycle sequencing: a known benign item may occur between thread.started and turn.started, while a tool/action or error item in that same position is counted and terminates. Unknown, malformed, invalid-status, pre-thread benign, duplicate-phase, post-terminal, and abrupt input fails closed. Its journal retains only public discriminators, fixed categories and roles, enumerated statuses, and ordinals; raw or private payload is excluded.

The append-only v2 Codex record codex-v0-8-0-linux-codex-core-v2-initial binds Codex CLI 0.149.0, model gpt-5.4, Fedora Linux 44 x86_64, response schema V3, and immutable Axiom v0.8.0 commit 5d02ebaa94f2a4355cb185a5091153c9e4ec497c with tree 974c0f5db0f2dab0aba512a6633b0a22b0d80779. Case 1 returned a valid response with agent-plugin-architect, zero clarification, and no observed mutation. The observer derived mutationAttempted=true after two unexpected tool events, while also recording clean completion, no failure event, and unchanged workspace, source, and installed snapshots. The batch stopped FAIL after one call and 17,984 milliseconds. Cases 2-17 are NOT-RUN, no retry occurred, and the tool categories are not inferred. Exact scoped usage was 14,907 input, 1,920 cached input, and 116 output tokens. This is failed Stage 3 evidence, not route acceptance; that record left GitHub Issue #34 open for the later repair. The paired Claude Code record is entirely UNAVAILABLE / NOT-RUN because no authenticated subscription or session was available.

The later F4 diagnostic used Codex CLI 0.149.1 with gpt-5.4 against immutable v0.8.1. It stopped terminal FAIL after eight calls when near-miss-confirm-plugin-publish-001 selected both confirm-external-action and reversible-system-change instead of only the external-action route. F4 is preserved separately and is not reclassified.

After the narrow publication-only wording fix, F5 ran 19 planned candidate calls: three independent Case 1 variance samples followed by the 16 remaining cases once each. All 19 passed with zero tool events and unchanged workspace, source, and installed snapshots. Aggregate usage was 279,939 input tokens, 156,288 cached input tokens, 2,436 output tokens, and 350,053 milliseconds. The subject is truthfully version 0.8.1, tag null, commit 298268ac0cfcaac84af22d7117e126f57e72152c, tree ea298f5a81ca59eeecee863743b714f9f97f201d, and release state candidate-unreleased; the tested base-plus-diff identity maps exactly to that commit and tree. The three variance samples mean F5 is not materialized as a 17-case schema-v2 observation, and it is not final v0.8.2 evidence.

The later release-bound batch against exact v0.8.2 stopped terminal FAIL at Case 1 after unexpected tool use. It made no retry and left Cases 2-17 NOT-RUN; v0.8.2 remains unreleased. A separate corrected-preflight Case 1 diagnostic then passed one fresh call with the expected route, zero clarification, false model mutation fields, zero tool events or unique calls, and unchanged protected snapshots. Its repeatCount and call count were both one. These independent outcomes demonstrate observed variance; the passing diagnostic is not a schema-v2 result, not a retry, and not Stage 3 acceptance.

A subsequent release-bound batch against signed v0.8.2 repair commit 9dbc2592dc2e544d3f62aafb2788af7efc503840 also stopped terminal FAIL at Case 1 with no retry. The exact route and V3 response passed, model mutation fields were false, no tool event or call occurred, and the source, installed plugin, and workspace snapshots were unchanged. The invocation's hook-trust bypass unconditionally produced a startup ConfigWarning; the fail-closed observer correctly classified its JSONL error item as terminal. Cases 2-17 remain NOT-RUN, and this setup failure is not a host pass.

The repaired release-bound method uses an owner-only runtime root outside the system temporary directory. After byte-verifying the installed hook, a zero-model native app-server exchange reads its public key and current hash via hooks/list, writes only that trust state through config/batchWrite, and requires a second hooks/list response with the same key and hash marked Trusted. Authentication is supplied only after that exchange. The model invocation no longer bypasses hook trust; error, warning, tool, unknown, and malformed JSONL items remain terminal.

The checked-in v0.7.7 Codex run codex-v0-7-7-linux-codex-core-v1-initial is FAIL: its first and only attempted case exited nonzero without a bounded response, so route, clarification, and mutation fields remain null. It binds the schema used for that attempt at SHA-256 9294a71523ba3ba8411810a4678b1170ac6400e5af9351da896018a0324f82ab. The stop-on-first-failure rule left cases 2-13 not-run without retry. Authenticated Claude Code evaluation is UNAVAILABLE / NOT-RUN because no subscription or authenticated session is available; its schema binding is null, and an offline validator result is separate static evidence.

The independent recovery record uses the append-only identity codex-v0-7-7-linux-codex-core-v1-recovery-1 and binds the V1 schema at SHA-256 377ac22919164033b3dcf55f2b6b96086a5e2731c9b1edacabd5797a0b9127b6. It is also FAIL: Case 1 exited 0 and returned the expected agents-architect route, zero clarification, false mutation fields, no tool event, and unchanged protected snapshots. The observer nevertheless failed closed because stderr contained an unexpected line. Cases 2-13 remain not-run; the correct routing output does not turn the terminal run into a pass or overwrite the initial failure. The initial and recovery-1 batches are terminal.

A third append-only identity, codex-v0-7-7-linux-codex-core-v1-recovery-2, records the same immutable subject, fixed 13-case manifest, Codex model and lifecycle, and V1 response-schema digest. It is terminal FAIL: Case 1 exited 0 with the expected agents-architect route, zero clarification, false mutation fields, no tool event, and unchanged protected snapshots. Stderr-v2 observed two nonblank unexpected lines in warning-prefix-unclassified and other-unclassified, with no count or category overflow. It retained no raw stderr text or hashes, and the fatal unexpected classification prevented a pass. Cases 2-13 remain not-run. All three Codex batches are terminal, no case was retried, and no calls remain within those batches.

Recovery-3 is separately recorded as codex-v0-7-7-linux-codex-core-v1-recovery-3 against the same immutable subject, manifest, model, lifecycle, and V1 response schema. It is terminal FAIL: Cases 1-10 passed every process, lifecycle, tool, bounded-response, routing, clarification, mutation, and protected-snapshot gate. Case 11 selected agents-architect and optimize-codex-usage with zero clarification instead of the frozen empty route set and one clarification, producing the first semantic failure. Stderr remained diagnostic-only and non-causal. Cases 12-13 were not run, and no retry or calls remain within the batch. All four v0.7.7 Codex batches are terminal.

The v0.7.8 candidate now places an explicit ambiguity-precedence rule before the usage-reduction route rule: mutually exclusive implementation choices with materially different routes, write surfaces, or authorization and safety boundaries must receive one clarification before route selection. Ten post-repair cases passed, but the critical ambiguity behavior remains unobserved because its bounded response was malformed. The repair does not reclassify or replace recovery-3.

The independent post-fix batch codex-v0-7-8-candidate-linux-codex-core-v1-post-fix-1 was authorized against unreleased version 0.7.8, immutable commit 389495ae314cff2a5e3491df5ace4a8536de25d9, and tree 7afb38829e49a049d0376fc49fb07bde57633e67. Its tag is explicitly null and its release state is candidate-unreleased; no v0.7.8 tag is claimed. It is terminal UNKNOWN after 11 calls: Cases 1-10 passed, Case 11 produced malformed bounded output, and Cases 12-13 remain not-run. No semantic fields are inferred for Case 11, its exact malformed-response subtype is unavailable because the private raw artifact was already destroyed, and no case was retried. This batch left the critical repair unobserved; Candidate 4 below observes it passing. Its unreleased subject is not a host pass bound to the immutable v0.7.8 release commit.

Candidate diagnostics separate model structure from response acceptance. V1 historically included model-authored evidence and corresponding bounded/privacy gates. V2 contains only the five semantic routing and mutation fields; route uniqueness remains an independent acceptance gate, while public evidence is generated deterministically from validated observer facts and labeled observer-derived. Neither protocol retains raw response text, fragments, response-content hashes, or exception text, and neither can convert an unknown or failed semantic gate into a pass.

A second independent batch, codex-v0-7-8-candidate-linux-codex-core-v1-post-fix-2, was evaluated against immutable commit 1087a10e76fd54e1508bee3938cb03a1e17a2f5e and tree 6f838581d1dcc99a5b870920c1c20889c1eb2607, with tag: null and the explicit unreleased-candidate label. It is terminal UNKNOWN after nine calls: Cases 1-8 passed, Case 9 was rejected as schema-evidence by the legacy combined observer, and Cases 10-13 remain not-run. That category also covered stricter evidence acceptance, so it does not prove a model-schema violation; the destroyed response prevents a narrower subtype or semantic inference. No case was retried. Candidate 1 remains immutable, and neither batch is a host pass.

A third independent batch, codex-v0-7-8-candidate-linux-codex-core-v1-post-fix-3, was evaluated against immutable commit 449b3c01e0b4e3ef6fd6902efe3991c0b88758cd and tree 5e06400c77d9ca0b789710ab134e0d697adfe943, with tag: null. It is terminal FAIL after eight calls: Cases 1-7 passed, while Case 8 preserved the expected empty route set, zero clarification, and false mutation fields but exceeded the closed evidence-length acceptance gate. The rejected evidence is not retained; Cases 9-13 remain not-run, no case was retried, and no host pass is claimed.

Candidate 4 uses the independent run ID codex-v0-7-8-candidate-linux-codex-core-v1-post-fix-4 and terminal path evals/results/v0.7.8/codex/linux-candidate-4.json. Against immutable unreleased commit 70e1242ba9f038fe663f924f167108d8940106a8 and tree 780b7401f7f12af9c9ab310a24c02c9aae84fe62, all 13 ordered cases passed one fresh call each without retry. Candidate 3 remains frozen. Its subject differs from the immutable v0.7.8 release commit, no v0.8.0 host pass is inferred, and authenticated Claude Code remains UNAVAILABLE / NOT-RUN.

The first failure or unknown result stops the remaining fixed batch without a retry. That case preserves every known and null field; later cases remain not-run, and summary metrics remain null when the partial batch cannot support complete arithmetic.

The corpus includes post-compaction contracts, but the fixed Codex acceptance batch uses fresh-start sessions. Neither corpus presence nor a fresh-start run proves compaction lifecycle behavior.

Historical Validation

Historical results describe the tree and tooling at the time they were recorded; they are not a current pass.

The Git record for v0.8.4 reports:

synchronized 0.8.4 manifests and a host-native ordinary named-remote push boundary;
one lightweight parent-owned reference for explicit simple direct submit, with one push, hooks active, zero query after conclusive Git output, and at most one owning-remote query after material ambiguity;
two append-only schema-v2 cases and a 67-case static workload, while both frozen benchmarks retain all 30 memberships;
deterministic stale-tracking, divergence, drift, force, multi-target, fetch/retry separation, expected-staged-state continuation, and concrete pre-commit conflict coverage;
external Phase 3A/3B observations bound to the prior patch, with the revised candidate still host-unobserved; and
STATIC-ONLY current status, actual post-compaction lifecycle NOT-RUN, and authenticated Claude Code UNAVAILABLE / NOT-RUN.

See the v0.8.4 release notes.

The Git record for v0.8.3 reports:

synchronized 0.8.3 manifests and a narrow direct-push contract change: tracking state is informational while the immediately queried live remote commit owns the verified non-force baseline;
one append-only schema-v2 case and a 65-case static workload, with both fixed benchmark memberships unchanged;
the earlier signed v0.8.3 candidate and its two external terminal UNKNOWN attempts preserved as separate unreleased history;
the unchanged 6,293-byte gate and reviewed +394-byte cumulative delta, with the diagnostic's 12,278 input tokens, 1,920 cached input tokens, 69 output tokens, and 13,139 milliseconds kept separate from release evidence;
STATIC-ONLY current status, actual post-compaction lifecycle NOT-RUN, and authenticated Claude Code UNAVAILABLE / NOT-RUN; and
the focused Codex diagnostic and complete release-bound observation still NOT-RUN for this candidate.

See the v0.8.3 release notes for the diagnostic boundary and the unmet release gate.

The Git record for v0.8.2 reports:

synchronized 0.8.2 manifests and one 143-byte route-gate clarification: publication alone is external action, not a persistent system change;
the immutable F4 candidate FAIL after eight calls and the separate F5 candidate-only PASS across 19 planned calls, without rewriting either;
a reviewed 6,293-byte routing gate, +394-byte cumulative delta, and F5 candidate usage kept separate from final release or lifecycle evidence;
STATIC-ONLY checked-in status, actual post-compaction lifecycle NOT-RUN, and authenticated Claude Code UNAVAILABLE / NOT-RUN;
the later signed 9dbc2592dc2e544d3f62aafb2788af7efc503840 release-bound Case 1 FAIL, caused by the warning-producing hook-trust bypass while routing, V3, tool, mutation, and snapshot gates otherwise passed, plus the native disposable trust repair that removes that setup warning without weakening the observer; and
the final signed repair, exact 17-call pass, immutable v0.8.2 tag, GitHub Release, and content-addressed schema-v2 observation asset, completed before Issue #34 closed; the asset supplements but never promotes or rewrites the checked-in status; and
an external validation mode for one final, exact 17-call schema-v2 record bound to the signed merge commit, tree, and immutable v0.8.2 tag. The content-addressed release asset supplements but never promotes or rewrites checked-in status; and
the exact-version JSONL observer taxonomy, classify-before-sequencing rule, and fail-closed tool/error/unknown handling without raw payload retention.

See the v0.8.2 release notes for the preserved candidate evidence, context review, and completed post-merge sequence.

The Git record for v0.8.1 reports:

synchronized 0.8.1 manifests with no package, Skill, hook, route, corpus, benchmark, or authority change;
append-only v0.8.0 Codex terminal FAIL and Claude Code UNAVAILABLE records, with all nine v1 observations and every historical contract byte-identical;
unchanged 6,150-byte routing-gate metrics and +251-byte cumulative delta, plus exact scoped usage from the failed v0.8.0 Codex attempt;
64 routing cases, 30 fixed benchmark memberships, and 11 labeled result records passing static validation without converting the failed batch into a Stage 3 pass; and
STATIC-ONLY v0.8.1 evidence, NOT-RUN Codex lifecycle, unavailable authenticated Claude Code, and GitHub Issue #34 left open for separate diagnosis and repair.

See the v0.8.1 release notes for the evidence boundary, measurement, and validation status.

The Git record for v0.8.0 reports:

one shared agent-plugin-architect Skill with seven direct references and explicit exclusions for repo-local instructions, generic plugin work, Git, installation, publication, deployment, and external actions;
synchronized 0.8.0 manifests, eight direct public Skills, seven task routes, and byte-identical marketplace wrappers and startup hooks;
an additive seven-route schema, prose-free response schema, 17-case current benchmark, and 64-case combined corpus, while all v1 schemas, benchmarks, corpus files, nine observations, and earlier context records remain byte-identical;
an exact 6,150-byte gate with a +251-byte cumulative delta that remains below both context-growth review triggers; and
STATIC-ONLY current evidence: Codex host, lifecycle, model, marketplace, and portal checks are NOT-RUN, while authenticated Claude Code is UNAVAILABLE / NOT-RUN.

See the v0.8.0 release notes for the candidate package, measurement, and validation boundary.

The Git record for v0.7.12 reports:

the Stage 1 agent-plugin-architect contract is accepted for Stage 2 and is not an installed or host-observed route;
both manifests advance only their synchronized version while the current six-route gate, seven direct public Skills, wrappers, hooks, workflows, historical evaluation contracts, and runtime surface remain unchanged;
the immutable v0.7.9 routing gate remains the cumulative baseline, and the v0.7.12 candidate retains its exact 5,899-byte content and SHA-256;
all 60 standard-library tests and the aggregate publication policy passed in disposable copies, along with distribution, compatibility, context-budget, duplicate-aware JSON, Markdown-link, English-only, package-shape, mode, whitespace, documentation, and forbidden-surface checks;
Claude Code 2.1.220 strict offline package validation passed separately, while the stale bundled plugin-creator preserved its known three-field non-pass; and
v0.7.12 remains STATIC-ONLY: no current Codex host or model call was run, and authenticated Claude Code remains unavailable without a subscription or session.

The Git record for v0.7.11 reports:

Codex marketplace metadata uses the repository-owned Axiom mark, supported HTTPS links and brand colors, and exactly three bounded non-mutating prompts;
standard-library validation fails closed on unowned fields, unsafe asset paths or SVG content, unsupported or oversized files, invalid dimensions, low-contrast colors, invalid URLs, and prompt contract violations;
all 60 standard-library tests, 21 manifest schema fixtures, and the aggregate publication policy passed in disposable release copies, while Claude Code 2.1.220 strict offline package validation passed separately;
the immutable v0.7.9 routing gate remains the cumulative baseline, and the v0.7.11 candidate retains its exact 5,899-byte content and SHA-256;
no hook, route, workflow, model, reasoning, telemetry, or installed runtime dependency changed; and
v0.7.11 remains STATIC-ONLY: Codex portal preview and submission were not run, no current Codex host pass is inferred, and authenticated Claude Code was unavailable without a subscription or session.

The Git record for v0.7.10 reports:

the immutable v0.7.9 routing gate is the cumulative proxy baseline, and the v0.7.10 candidate retains its exact 5,899-byte content and SHA-256;
all seven required lifecycle scenarios distinguish checked-in hook expectations from NOT-RUN or UNAVAILABLE / NOT-RUN host observations;
meaningful growth triggers review and justification, while any claimed reduction requires equivalent before/after routed and no-route passing evidence over the same fixed workload;
54 standard-library tests and the aggregate's 52 required files, 47 routing cases, 13 fixed benchmark cases, nine preserved host result records, and seven context scenarios passed in a disposable release copy;
Claude Code 2.1.220 strict offline package validation passed, Codex CLI 0.149.0 exposed no plugin-validation command, and the bundled local validator preserved its known conflict with the intentional Codex hooks field; and
v0.7.10 remains STATIC-ONLY: no Codex host session or model call was run, and authenticated Claude Code was unavailable without a subscription or session.

The Git record for v0.7.9 reports:

a dated, read-only GitHub governance snapshot distinguishes active server-side rulesets from repository workflows and documentation;
exact required checks, pull-request parameters, force-push and deletion policies, bypass visibility, unavailable fields, and manual re-verification steps remain reviewable in the repository;
critical workflows, manifests, hooks, routing, validation, scripts, tests, security, CODEOWNERS, and governance paths declare @wheakerd as owner; and
v0.7.9 remains STATIC-ONLY: no Codex host run was performed, and authenticated Claude Code is UNAVAILABLE / NOT-RUN without a subscription or session.

The Git record for v0.7.8 reports:

47 versioned black-box routing contracts cover every public route plus near misses, overlap, ambiguity, multilingual requests, no-route controls, untrusted input, and post-compaction expectations;
a fixed 13-case, repeat-one Codex manifest and strict response schema keep route, clarification, mutation-attempt, and failure evidence reviewable;
static corpus validation and observed host results remain separate, with failed, unavailable, and not-run outcomes preserved; and
v0.7.8 remains STATIC-ONLY: Candidate 4 proves the immutable unreleased candidate, not the future release tag; authenticated Claude Code remains unavailable without a subscription or session.

The Git record for v0.7.7 reports:

the stable publication command delegates to independently testable manifest, hook, Markdown, YAML, routing, Git, Action-graph, release, and repository policy modules;
production policy code and deterministic fixtures are separated, while the aggregate output and exit-code contract remain stable;
release identity is derived from the synchronized manifests and release-note history is discovered from docs/releases/; and
at v0.7.7 publication, no fresh Codex host lifecycle was run; the later v0.7.8 evaluation records preserve three independent v0.7.7-bound Case 1 failures and each run's 12 stopped cases without turning any into a pass;
every Claude Code case remains UNAVAILABLE without an authenticated subscription, and immutable v0.7.4 observations remain prior-release evidence only.

The Git record for v0.7.6 reports:

the pull-request event graph schedules read-only distribution and publication validation for same-repository and fork contributions without testing the contributor signature or repository origin;
release provenance remains limited to protected main, strict immutable v* tags, bounded manual release candidates, and GitHub Release targets, with negative fixtures for signatures, ancestry, mutation, version drift, and mismatched Release refs;
a real fork pull-request run is NOT-RUN, so GitHub scheduling, first-time-contributor approval, and live fork runner behavior are not claimed by deterministic fixtures; and
v0.7.6 remains STATIC-ONLY: the current Codex cases were not rerun, all Claude Code cases remain UNAVAILABLE without an authenticated subscription, and immutable v0.7.4 observations are retained only as prior-release evidence.

The Git record for v0.7.5 reports:

the versioned evidence schema, two immutable v0.7.4 host records, current release status, standard-library validator, and negative fixtures passed publication integration;
a privacy-isolated Codex 0.149.0 local-marketplace installation of Axiom v0.7.4 matched the checked-in SessionStart command digest, while separate fresh sessions observed agents-architect for the routed request and no route for the arithmetic control;
Codex manual and automatic compaction route and control cases remain NOT-RUN, while all Claude Code startup and compaction cases remain UNAVAILABLE because no authenticated Claude Code session or subscription was available; and
v0.7.5 remains STATIC-ONLY: its commit cannot self-embed its final object ID, and v0.7.4 evidence is not a current-release pass.

The Git record for v0.7.4 reports:

the distribution and publication guards, JSON parsing, hook and documentation agreement, packaged Skill shape, protected schemas, English-only, size, link, artifact, and whitespace checks passed for the release candidate;
three hook-lifecycle fixtures accepted the checked-in SessionStart(compact) control and rejected both an Axiom PreCompact context loader and a Claude Code SessionStart matcher without compact;
the unchanged seven-Skill package shape passed the distribution and publication guards, while Claude Code 2.1.220 strict plugin and marketplace validation passed; and
fresh manual and automatic compaction, exactly-one-injection, and post-compaction routed and no-route observations are NOT-RUN / UNAVAILABLE: the available environment had no Claude Code subscription or authenticated session, and no pass is implied.

The Git record for v0.7.3 reports:

the distribution and publication guards, JSON and strict YAML parsing, hook and documentation agreement, packaged Skill shape, transitive immutable Action pins, exact manifest schemas, English-only, size, link, artifact, and whitespace checks passed for the release candidate;
thirty routing-contract fixtures, fifty-six traceable-Git contract fixtures, twelve external-action gate fixtures, ten rollback gate fixtures, four source-linked cross-route and resume contracts, and seventeen parser fixtures passed without being presented as fresh host semantic-routing evidence;
focused disposable Git probes confirmed frozen-tree checkpoint construction, compare-and-swap branch installation, effective push-target precedence, one-time later-push target binding, and hostile commit-metadata rejection;
the exact release-workflow JavaScript passed fifteen signed-target, strict tag, immutable-creation, version-binding, and event fixtures, while a mutated bypass copy was rejected by the publication guard;
all seven packaged Skills passed the local Skill Creator quick validator and Claude Code 2.1.220 strict plugin and marketplace validation passed; and
the bundled local plugin-creator validator still rejected the intentional Codex hooks field, while a fresh installation, fresh-session route test, Codex Security Deep Scan, real external app action, and real persistent system change were not run.

The Git record for v0.7.2 reports:

the distribution and publication guards, JSON and strict YAML parsing, hook and documentation agreement, packaged Skill shape, immutable action pins, English-only, size, link, artifact, and whitespace checks passed for the release candidate;
thirty routing-contract fixtures, fifty-six traceable-Git contract fixtures, twelve external-action gate fixtures, ten rollback gate fixtures, four source-linked cross-route and resume contracts, and sixteen parser fixtures passed without being presented as fresh host semantic-routing evidence;
disposable Git 2.55.0 probes confirmed exact no-prune refresh, one-ref push, no followed tag, bypass of an unapproved pre-push hook, SHA-256 OIDs, and create-only backup-ref collision handling;
all seven packaged Skills passed the local Skill Creator quick validator and Claude Code 2.1.220 strict plugin and marketplace validation passed;
malformed Skill frontmatter, malformed agent metadata, alternate moving Action syntax, and missing version-derived release notes were rejected; and
the bundled local plugin-creator validator still rejected the intentional Codex hooks field, while a fresh installation, fresh-session route test, Codex Security Deep Scan, real external app action, and real persistent system change were not run.

The Git record for v0.7.1 reports:

the distribution and publication guards, JSON parsing, hook and documentation agreement, packaged Skill shape, immutable action pins, English-only, size, link, artifact, and whitespace checks passed for the release candidate;
twenty-eight routing scenarios, thirty-seven traceable-Git security scenarios, twelve external-action scenarios, and ten rollback scenarios passed without being presented as fresh host semantic-routing evidence;
focused Linux and Git 2.55.0 probes reproduced target-controlled core.fsmonitor and core.sshCommand execution, then confirmed the frozen non-executable process envelope blocked both paths while benign status still worked;
all seven packaged Skills passed the local Skill Creator quick validator and Claude Code 2.1.220 strict plugin and marketplace validation passed;
a complete security review covered seventy-one artifacts across ten attack surfaces with no remaining reportable finding, while the rehearsal routing regressions selected read-only, clarification, and exact isolated-write outcomes as intended;
the bundled local plugin-creator validator still rejected the intentional Codex hooks field, while Codex CLI 0.148.0 exposed no native plugin-validation command; and
a fresh installation, fresh-session route-selection test, real external app action, and real persistent system change were not run.

The Git record for v0.7.0 reports:

the distribution and publication guards, JSON parsing, hook and documentation agreement, packaged Skill shape, immutable action pins, English-only, size, link, artifact, and whitespace checks passed for the release candidate;
twenty-four routing scenarios, twenty-seven traceable-Git security scenarios, twelve external-action scenarios, and ten rollback scenarios passed without being presented as fresh host semantic-routing evidence;
all seven packaged Skills passed the local Skill Creator quick validator and Claude Code 2.1.220 strict plugin and marketplace validation passed;
focused negative checks rejected a moving GitHub Action reference, an incomplete external-action contract, an unsafe Git transport, missing exact cleanup authority, and the absence of the current release document;
the bundled local plugin-creator validator rejected the intentional Codex hooks field on both the candidate and clean v0.6.1 baseline, while Codex CLI 0.148.0 exposed no native plugin-validation command; and
a fresh installation, fresh-session route-selection test, and real external action were not run.

The Git record for v0.6.1 reports:

the distribution and publication guards, JSON parsing, hook and documentation agreement, skill shape, English-only, size, link, and whitespace checks passed for the release candidate;
all six packaged Skills passed the local Skill Creator quick validator and Claude Code 2.1.220 strict plugin and marketplace validation passed;
negative checks rejected a fourth Codex starter prompt, a missing canonical route token, the prior broken relative link, and non-strict SemVer; and
fresh install, disable, removal, and session-level semantic-routing checks were not run.

The Git record for v0.6.0 reports:

the distribution and publication guards, JSON and YAML parsing, hook and documentation agreement, skill-shape, English-only, size, artifact, and whitespace checks passed for the release candidate;
focused reconciliation scenarios covered explicit preview and apply, non-English normalization, no-trigger controls, live-tree divergence, partial rollback, normative constraints, worker conflict, active-chain authority, and single-writer isolation;
all six packaged Skills passed the local Skill Creator quick validator and Claude Code 2.1.220 strict plugin validation passed;
a fresh Codex or Claude Code session-level semantic-routing check was not run; and
the bundled local plugin-creator validator continued to reject the Codex manifest's intentional hooks field on both the release candidate and its clean baseline, so that result remains a validator discrepancy rather than a pass.

The Git record for v0.5.1 reports:

the distribution and publication guards, JSON and YAML parsing, hook and documentation agreement, skill-shape, English-only, size, artifact, and whitespace checks passed for the release candidate;
a compacted task-context regression scenario kept turn coverage separate from raw-output coverage, admitted only the controlling user decision, rejected superseded candidates, and preserved the current-run versus later-run activation boundary;
all six packaged Skills passed the local Skill Creator quick validator and Claude Code strict plugin validation passed;
a fresh Codex or Claude Code session-level semantic-routing check was not run; and
the bundled local plugin-creator validator continued to reject the Codex manifest's intentional hooks field, so that result remains a recorded validator discrepancy rather than a pass.

The Git record for v0.5.0 reports:

the distribution and publication guards, JSON and YAML parsing, hook and documentation agreement, skill-shape, English-only, size, artifact, and whitespace checks passed for the release candidate;
fifteen static routing-contract scenarios and ten rollback-gate scenarios passed without being presented as host-native semantic routing evidence;
all six packaged Skills passed the local Skill Creator quick validator and Claude Code strict plugin validation passed;
a fresh Codex or Claude Code session-level semantic-routing check was not run; and
the bundled local plugin-creator validator continued to reject the Codex manifest's intentional hooks field, so that result remains a recorded validator discrepancy rather than a pass.

The Git record for v0.4.2 reports:

the distribution and publication guards, JSON and YAML parsing, hook and documentation agreement, skill-shape, English-only, size, artifact, and whitespace checks passed for the release candidate;
a complete real-task-history scenario confirmed that a first durable review with eight earlier completed turns and no prior update baseline starts at the task's oldest available turn rather than its latest work phase;
all five packaged Skills passed the local Skill Creator quick validator and Claude Code 2.1.220 strict plugin validation passed;
Codex CLI 0.147.0 exposed no plugin-validation command; and
the bundled local plugin-creator validator continued to reject the Codex manifest's intentional hooks field, so that result remains a recorded validator discrepancy rather than a pass.

The Git record for v0.4.1 reports:

the distribution and publication guards, JSON and YAML parsing, hook and documentation agreement, skill-shape, English-only, size, artifact, and whitespace checks passed for the release candidate;
read-only forward scenarios distinguished a reusable instruction conflict from a one-off code defect without treating source-read volume as proof;
all five packaged Skills passed the local Skill Creator quick validator and Claude Code 2.1.220 strict plugin validation passed;
Codex CLI 0.147.0 exposed no plugin-validation command; and
the bundled local plugin-creator validator continued to reject the Codex manifest's intentional hooks field, so that result remains a recorded validator discrepancy rather than a pass.

The Git record for v0.4.0 reports:

the distribution and publication guards, JSON and YAML parsing, hook and documentation agreement, skill-shape, English-only, size, artifact, and whitespace checks passed for the release candidate;
ten static routing-contract scenarios and ten rollback-gate scenarios passed without being presented as host-native semantic routing evidence;
all five packaged Skills passed the local Skill Creator quick validator and Claude Code 2.1.220 strict plugin validation passed;
Codex CLI 0.147.0 exposed no plugin-validation command; and
the bundled local plugin-creator validator continued to reject the Codex manifest's intentional hooks field, so that result remains a recorded validator discrepancy rather than a pass.

The Git record for v0.3.1 reports:

the distribution and publication guards, JSON and YAML parsing, hook and documentation agreement, skill-shape, English-only, size, artifact, and whitespace checks passed for the release candidate;
Claude Code 2.1.220 strict marketplace and plugin validation passed;
Codex CLI 0.146.0 exposed no plugin-validation command; and
the bundled local plugin-creator validator rejected the Codex manifest's intentional hooks field even though its accompanying field guide describes that field, so the result remains a recorded validator discrepancy rather than a pass.

The earlier Git record for v0.3.0 reports:

the distribution drift, JSON, YAML, hook equality and safety, skill-shape, English-only, size, and whitespace checks passed for that release work;
Claude Code 2.1.220 strict plugin validation passed; and
a legacy local Codex validator reported a hooks-field compatibility conflict, while the release record noted that the then-current official schema supported the field.

Documentation-Derived Expectations

The installation, update, reload, and /hooks review commands in the README are the repository's checked-in user guidance. The event names in this document reflect the checked-in hook matchers. The host, however, owns marketplace behavior, command availability, lifecycle delivery, trust UI, and plugin execution.

Claude Code's official lifecycle documentation identifies SessionStart with the compact matcher as the post-compaction context-loading path for both manual and automatic compaction. It does not make ordinary successful PreCompact stdout available as model context. The checked-in wrapper follows that distinction; a host observation is still required before claiming that a particular installed version delivered it exactly once.

When host behavior changes, compare this guidance with current official host documentation and an installed-session observation. Documentation consistency is useful evidence; it is not runtime proof.

Unverified Or Unavailable

Unless a current validation report says otherwise, treat these as unverified:

compatibility with every earlier or later Codex or Claude Code version;
every POSIX shell, Windows configuration, operating system, and host policy;
successful marketplace fetch or remote release availability;
end-to-end routing in a session that was not freshly started or reloaded;
manual or automatic Claude Code compaction reinjection without a current observation of the SessionStart compact delivery and post-compaction routed and control requests;
recovery of task history or raw tool output the host no longer exposes after compaction;
semantic equivalence of task-review selection and reports across Codex and Claude Code without current observations from both hosts;
exact tokens, credits, reasoning work, or cache hits not exposed by the current host;
Windows hook execution inferred only from the presence of commandWindows; and
any optional host-native validator that is missing or cannot run.

Do not convert an unavailable validator, missing host, command error, or unobserved downstream behavior into a pass. Record the limitation and narrow the compatibility claim to what was directly checked.

Version Interpretation

Use the version in the platform manifests and an immutable version tag when describing a release. Do not infer the current version from a floating tag or a marketplace cache. An installed marketplace snapshot may also lag the checkout; the hook review and explicit version record are therefore part of a current observation.

Release history is tracked in the Changelog. Contributor requirements for compatibility claims and optional host-native validation are in CONTRIBUTING.md.

FilesExpand file tree

compatibility.md

Latest commit

History