[wheakerd]

Scope

Refs #226, #233, #234. The three Issues ship as one installed-runtime change set, version 0.24.0, in three GitHub-signed commits on this branch: the skill rewrite, the gate behavior with the recall cases and routing contract v8, and the checked-in fresh-session observation records that the release gate requires.

• Invariants ([Runtime][Skills] Rewrite the fourteen skills for a model reader with named invariants #226): every bundled SKILL.md opens with a short direction paragraph and an ## Invariants section of three to six named rules that bind in every phase, including the references that phase loads; everything else in a skill is a default the model may adapt, saying so. The "Defaults you may adapt ... Every other rule ... binds" paragraphs are gone from the twelve skills that carried them, and agents-architect folds its Decision Latitude and Always-On Rules into six invariants. Each invariant restates a rule that already bound; no authority widens.
• Terms and sentences ([Runtime][Skills] Rewrite the fourteen skills for a model reader with named invariants #226): the gate defines route, host-native, selection surface, explicit invocation, invariant and default, handoff contract, envelope, and write set once; a skill defines the terms it introduces under ## Terms where they first appear. No sentence in bundled skill text exceeds 300 characters, measured per paragraph, list item, heading, or frontmatter field with code blocks and tables excluded; twenty-one over-long sentences were split without changing their rules.
• Invocation ([Runtime][Skills] Rewrite the fourteen skills for a model reader with named invariants #226): the gate states explicit invocation as $skill-name in Codex, /axiom:skill-name in Claude Code, or the skill named in plain words; the offline route model and a new fixture accept both host forms.
• Host facts ([Runtime][Skills] Rewrite the fourteen skills for a model reader with named invariants #226): using-axiom/references/codex-host.md and claude-code-host.md consolidate the host mechanics that were spread over agents-architect, continuous-progress, delegate-simple-task, and the permission reference, each verified against the installed host on 2026-10-30 with documentation-derived facts marked. continuous-progress keeps three references instead of five; its host-neutral supervisor and acceptance rules move into supervision-and-recovery.md and execution-and-handoff.md.
• Reconcile modes ([Runtime][Skills] Rewrite the fourteen skills for a model reader with named invariants #226): agents-architect/references/maintenance/implementation-reconciliation.md replaces the mandatory coordinator-plus-three-subagents contract with an independent review contract that one session can run as sequential passes, keeping the three independent read-only reviews, the direct recheck of decisive evidence, and the single writer.
• Readability validator ([Runtime][Skills] Rewrite the fourteen skills for a model reader with named invariants #226): axiom_validation/skill_text.py enforces the Invariants list, the sentence maximum, and the absence of a blanket binding clause in the publication aggregate, with unit tests and a documented check in the validation guide.
• Before acting ([Runtime][Gate] Restate the intended outcome, open ambiguity, and the first irreversible step before acting #233): the gate's ## Before Acting section asks the model, before an action that changes state, to say in one sentence what outcome it intends, name any material ambiguity the user left open, and mark the first irreversible or externally visible step. It adds no question, stop, or authority and stays silent for read-only requests and conversation. Two corpus cases show that the behavior selects no route and that a read-only request gets no restatement.
• Recall ([Runtime][Routing] Raise recall for verify-completion and clarify-intent and cover post-compaction route reselection #234): the verify-completion description names a completion claim carried by a compaction summary even without the word verify, and the clarify-intent description names alternatives offered across sentences or without the word either. The offline route model resolves both forms and their near-miss controls; a post-compaction route reselection case records the gate's reselection rule.
• Routing contract v8: every v7 case keeps its ID and expectation; nine cases are added, and codex-core-v8 and claude-code-core-v8 select the 40 v7 cases plus the nine new ones (49) with the v7 host profiles unchanged. The v7 corpus, benchmarks, and bindings stay frozen; candidate records of either contract validate.
• Release artifacts: manifests 0.24.0, repository policy revision 122 with the new runtime digest, routing-context record v0.24.0 (gate 7,217 bytes, reviewed growth with justification; description set 5,204 bytes), version note, Changelog entry, release status, and rendered surfaces.
• Host evidence: evals/results/v0.24.0/claude-code/linux.json and evals/results/v0.24.0/codex/linux.json record one complete passing fresh-session run per host against commit 0f3c906, 49 of 49 cases each in a single attempt: Claude Code 2.1.308 with claude-opus-5-5 at maximum effort, and Codex CLI 0.166.0 with gpt-6-astra at maximum effort. The version note's Host Evidence section also reports the two private checks: a paraphrase check in which one fresh Claude Code session per skill reproduced 78 of 78 invariants without drift, and a same-batch sample of the before-acting behavior on both hosts against the released 0.23.0 plugin as the baseline.

Affected files and surfaces

• Shared files: every plugin/skills/*/SKILL.md (fourteen); plugin/skills/using-axiom/references/codex-host.md and claude-code-host.md (new), host-permission.md; plugin/skills/agents-architect/references/inventory-audit.md, project-initialization.md, runtime-and-updates.md, maintenance/implementation-reconciliation.md; plugin/skills/continuous-progress/references/execution-and-handoff.md, supervision-and-recovery.md (changed), codex-operations.md, claude-code-operations.md (removed); plugin/skills/delegate-simple-task/references/host-evidence.md; plugin/skills/traceable-git-submit/references/baseline-and-preflight.md, checkpoint.md, consolidation.md, prepared-release-submit.md, remote-targets.md
• Codex-specific files: plugin/.codex-plugin/plugin.json (version)
• Claude Code-specific files: plugin/.claude-plugin/plugin.json (version)
• Documentation or validation files: README.md, CHANGELOG.md, docs/architecture.md, docs/examples.md, docs/trust-model.md, docs/runtime-identity.md, docs/repository-policy-revisions.md, docs/reference/skills.md, docs/maintainers/runtime-changes.md, docs/maintainers/validation.md, docs/releases/v0.24.0.md (new), docs/releases/v0.16.0.md (two links to the removed references become code spans); evals/README.md, evals/schema-v8.json, evals/routing-v8/current.jsonl, evals/benchmarks/codex-core-v8.json, evals/benchmarks/claude-code-core-v8.json, evals/context-budget/results/v0.24.0.json, evals/results/v0.24.0/claude-code/linux.json, evals/results/v0.24.0/codex/linux.json (new); evidence/runtime-identity.json, evidence/release-status.json, evidence/repository-policy-revisions-v1.json; axiom_validation/skill_text.py (new), axiom_validation/aggregate.py, axiom_validation/routing_contracts.py, axiom_validation/cases/routing.py, axiom_validation/routing_evals/__init__.py, candidates.py, constants.py, current.py, external.py, observations.py, schemas.py; tests/test_skill_text.py (new), tests/test_current_routing.py, tests/test_release_facts.py, tests/test_routing_contracts.py, tests/test_routing_evals.py, tests/test_validator_isolation.py

Routing and authorization impact

• Route-selection impact: explicit invocation is stated in each host's own syntax and both forms resolve offline; verify-completion and clarify-intent widen to the forms their descriptions now name (a summary-carried completion claim without a verification verb; alternatives across sentences or without "either"); the before-acting behavior selects no route. Every description except these two is unchanged apart from one split sentence in agents-architect. Because the gate and every skill body change, the fresh-session records cover all thirteen routes.
• Action-authorization impact: None widened. Every invariant restates a rule that already bound, and every default was already adaptable; the before-acting behavior adds no question, stop, or authority.
• New or changed stop conditions: the agents-architect reconcile modes no longer stop for a missing subagent facility; they run as sequential independent passes with one writer. No other stop condition changes.

Documentation

docs/releases/v0.24.0.md carries the reasoning, changed passages, routing contract v8, host evidence, and evidence boundary; CHANGELOG.md carries the 0.24.0 entry with its behavioral impact and required action (none). docs/architecture.md, docs/trust-model.md, docs/examples.md, docs/maintainers/runtime-changes.md, docs/maintainers/validation.md, evals/README.md, docs/reference/skills.md, docs/runtime-identity.md, and docs/repository-policy-revisions.md follow the invariants, the invocation syntax, the before-acting behavior, contract v8, and revision 122. Installation and hook commands are unchanged.

Validation

Command	Result
python3 -B scripts/check-publication.py	Passed on this exact tree: 155 required files, 167 Markdown files, 97 offline route contract fixtures, 188 black-box routing cases, 79 fixed host benchmark cases, 14 skill readability contracts, 89 canonical installed-runtime inputs, version 0.24.0
python3 -B -m unittest discover -s tests -p 'test_*.py'	OK in a disposable clone of this tree: 320 tests
git diff --check	Passed
Additional targeted checks	python3 -B scripts/measure-routing-context.py --check, python3 -B scripts/render-release-facts.py --check, and python3 -B scripts/check-runtime-identity.py --check: passed. claude plugin validate --strict for plugin and .claude-plugin/marketplace.json: passed on Claude Code 2.1.308. Release observation gate rehearsal in a disposable clone (require_route_observations from main to this tree): 13 changed routes, both host records accepted at the unchanged runtime digest. ASCII scan of the changed Markdown, JSON, and YAML files: passed; the multilingual corpus cases are non-ASCII by design.

Cross-platform parity

• Shared skill changes were reviewed against Codex and Claude Code, or this change does not affect shared skills.
• Manifest, marketplace, and hook differences remain intentional and documented.
• Optional host-native validation results include the host/tool version, or are marked unavailable.

The fresh-session records install the 0.24.0 subject on Codex CLI 0.166.0 (local marketplace install into an isolated configuration, documented hook-trust handshake) and load it on Claude Code 2.1.308 (plugin directory with the installed copy disabled per invocation); both hosts delivered the startup gate in every case. The host references label each host's mechanics in its own words and map neither onto the other. Hook commands are unchanged.

Unrelated work

• I reviewed the final diff and did not reset, hide, stage, or rewrite unrelated work.
• I did not add generated caches, disposable validation copies, local maintenance notes, or tool output.
• I called out every affected file and any routing, authorization, documentation, or parity impact above.