Skip to content

feat(spec-tree): ship the execution definitions for Codex - #657

Open
simonheimlicher wants to merge 16 commits into
mainfrom
work/change-433-codex-execution
Open

simonheimlicher wants to merge 16 commits into
mainfrom
work/change-433-codex-execution

Conversation

@simonheimlicher

@simonheimlicher simonheimlicher commented Oct 11, 2026 •

Copy link
Copy Markdown
Contributor

Summary

  • The seven execution definitions of the spec-tree plugin ship for Codex beside Claude Code, so a Codex session that holds a claimed Executable Change runs /execute-change and launches each round and Verifier by exact name.
  • /execute-change runs on Codex in place of its unavailable result.

Background

Each of the seven definitions under src/plugins/spec-tree/agents/ selected only the Claude Code target, so a Codex home held none of them and an Executable Change waited for a Claude Code session. The execution spec declared that restriction.

A Codex target carries only the restrictions Codex enforces. Each definition declares disallowedTools for Claude Code alone and states no restriction for Codex.

Changes

  • Execution spec: the four assertions of change-execution.md state that both agents run the workflow through their configured native definitions, and the ISSUES.md notes follow.
  • Definitions: change-executor, change-author, change-implementer, change-tester, change-verifier, change-skill-author, and change-subagent-author list both targets; the generated Codex definitions are in dist/codex.
  • Skill: the Codex path of execute-change replaces its unavailable surface, with fork_turns set to none on each launch.
  • Probe: the Codex protocol and the attested run of definition-invocation on the committed emitted definitions, with the retained artifacts.
  • Plugin changelog entry for the release.

Test plan

  • spx validation markdown, just eval-links, spx spec status --format json, just check-skills, just docs-check
  • just test over the target-emission, agents, native-artifact, and execution-policy nodes
  • just verify-marketplace-installation
  • Assertion audit of the execution node, skill audit of execute-change, definition audits of the seven definitions, test-evidence audits of the five implementation nodes
  • Local changeset review on the final head

Refs

  • spx/31-outcomeeng.enabler/32-changes.enabler/65-change-execution.enabler
  • spx/18-plugin-build.enabler/43-target-emission.enabler
  • spx/18-plugin-build.enabler/54-conversion.enabler/21-agents.enabler
  • spx/32-distribution.enabler/21-installation.enabler/21-repository-installation.enabler
  • outcomeeng/changes#433

Versioning

  • spec-tree: MINOR, written before merge

Validation

  • Touched-scope deterministic verification passes
  • /reload-plugins confirms the change loads in a running session

View with [code]smith Autofix with [code]smith
Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.

@claude

claude Bot commented Oct 11, 2026

Copy link
Copy Markdown

DEBT [architecture]: src/plugins/spec-tree/agents/change-author.md:9-11

Reference: spx/31-outcomeeng.enabler/32-changes.enabler/65-change-execution.enabler/change-execution.md states NEVER a round session launches a Verifier; the Executor launches every Verifier the Mutation rows name, tagged audit.
Evidence: the six producing-round and Fixer definitions, change-author, change-implementer, change-skill-author, change-subagent-author, change-tester and change-verifier, wrap their disallowedTools line in a target-equals-claude conditional, so the emitted Codex renderings such as dist/codex/spec-tree/skills/spec-tree-plugin/agents/spec-tree_change-author.toml carry no tool allowlist and no sandbox_mode at all. Every existing Codex auditor definition this generator already emits, spec-tree_spec-auditor.toml, spec-tree_pdr-auditor.toml, spec-tree_changeset-coherence-auditor.toml and others, gets sandbox_mode read-only on the codex target. In outcomeeng/distribution/agents.py the function infer_sandbox_mode only derives read-only when the source tools allowlist is a subset of the read-only and web-capable tool sets, so it has no mapping that withholds subagent-spawn while still permitting the Write and Edit tools these producing rounds need, and the CODEX_DISALLOWED_TOOLS_GUIDANCE_TEMPLATE constant already documents that disallowedTools do not enforce Codex permissions. The only thing standing between a round session and calling spawn_agent to launch its own Verifier on Codex is the prose line NEVER launch a subagent inside developer_instructions; on Claude Code the identical constraint is additionally backed by the native disallowedTools Agent denial. The spec assertion carries no per-target exception, so a round session that disregards its own prompt on Codex can do exactly what the assertion forbids, with nothing but the model own compliance preventing it.
Required: either find a Codex-native way to withhold subagent-spawn from these six definitions while keeping their write tools, or record the gap in this node ISSUES.md, which already tracks several other Codex-enforcement gaps for this same skill, for example the entry titled The Executor skill grants, verify scope and reading tag carry three open warnings.

@claude

claude Bot commented Oct 11, 2026

Copy link
Copy Markdown

No findings. Cross-checked the layers this changeset touches:

  • change-execution.md assertions (seven definitions ship for both agents; /execute-change runs on either) match the authored src/plugins/spec-tree/agents/*.md sources, which match their generated dist/claude and dist/codex renderings — verified by diffing both dist trees against the templated source and against each other.
  • The definition-invocation probe's attested Codex run (subject commit 7cc2690b1571...) is not stale relative to this PR's head (4d38caa0...): the sha256 recorded in each codex/spec-tree_change-*.result.json matches the sha256sum of the corresponding committed dist/codex/.../spec-tree_change-*.toml exactly, for all seven definitions.
  • The Codex-side omission of a structural disallowedTools equivalent (round definitions can textually, but not structurally, be kept from launching a subagent on Codex) mirrors the existing precedent in adr-auditor.md/changeset-coherence-auditor.md, is disclosed in the CHANGELOG entry and in the probe's Limitations section, and is tracked against the narrowest_allowed_tools audit rule in ISSUES.md — not a new, hidden gap.
  • src/plugins/spec-tree/CHANGELOG.md gained a 0.105.6 entry while plugin.json stays pinned at 0.105.5, consistent with this repo's bump-at-merge policy.

The seven execution definitions ship for Claude Code and Codex, and
/execute-change runs on either through its configured native definitions.
Producing rounds, Fixers, and Verifiers start in isolated native sessions,
and the invocation probe covers each definition on each agent. The node's
issue about Claude Code-only delivery narrows to the missing Codex
orchestrator start.

Refs: spx/31-outcomeeng.enabler/32-changes.enabler/65-change-execution.enabler
Verification: passing
List both targets on the seven execution definitions so the build emits
each for Codex beside Claude Code, and record the delivery in the
plugin changelog.

Verification: not-run
Refs: outcomeeng/changes#433
The skill carries out its steps in a Codex session that holds a claimed
Executable Change instead of returning unavailable. Every producing round,
Fixer, and Verifier starts in a fresh session, and the Codex build sets the
native fork_turns control to none on each launch.

Refs: spx/31-outcomeeng.enabler/32-changes.enabler/65-change-execution.enabler
Verification: passing
The definition-invocation probe gains the Codex protocol and an attested
run of the seven emitted execution definitions: discovery, loaded
configuration, child invocation, and returned result for each, with the
retained artifacts beside the protocol.

Refs: spx/31-outcomeeng.enabler/32-changes.enabler/65-change-execution.enabler
Verification: not-run
The change-author definition declares its disallowed tools for the Claude
Code target only, so the Codex definition states no restriction that Codex
cannot enforce.

Refs: spx/31-outcomeeng.enabler/32-changes.enabler/65-change-execution.enabler
The change-executor definition declares its disallowed tools for the Claude
Code target only, so the Codex definition states no restriction that Codex
cannot enforce. The plugin changelog entry records the change for the
change-executor and change-author definitions.

Refs: spx/31-outcomeeng.enabler/32-changes.enabler/65-change-execution.enabler
The change-implementer definition declares its disallowed tools for the
Claude Code target only, so the Codex definition states no restriction
that Codex cannot enforce. The plugin changelog entry records the change
for the change-executor, change-author, and change-implementer
definitions.

Refs: spx/31-outcomeeng.enabler/32-changes.enabler/65-change-execution.enabler
The change-tester definition declares its disallowed tools for the
Claude Code target only, so the Codex definition states no restriction
that Codex cannot enforce. The plugin changelog entry records the change
for the change-executor, change-author, change-implementer, and
change-tester definitions.

Refs: spx/31-outcomeeng.enabler/32-changes.enabler/65-change-execution.enabler
The change-verifier definition declares its disallowed tools for the
Claude Code target only, so the Codex definition states no restriction
that Codex cannot enforce. The plugin changelog entry records the change
for the change-executor, change-author, change-implementer,
change-tester, and change-verifier definitions.

Refs: spx/31-outcomeeng.enabler/32-changes.enabler/65-change-execution.enabler
The change-skill-author definition declares its disallowed tools for the
Claude Code target only, so the Codex definition states no restriction
that Codex cannot enforce. The plugin changelog entry records the change
for the change-executor, change-author, change-implementer,
change-tester, change-verifier, and change-skill-author definitions.

Refs: spx/31-outcomeeng.enabler/32-changes.enabler/65-change-execution.enabler
The change-subagent-author definition declares its disallowed tools for the
Claude Code target only, so the Codex definition states no restriction that
Codex cannot enforce. The plugin changelog entry now names all seven
execution definitions as declaring their withheld tools for Claude Code only.

Verification: not-run

Refs: spx/31-outcomeeng.enabler/32-changes.enabler/65-change-execution.enabler
… definitions

The attested run covers the seven execution definitions as the build emits
them for Codex with no tool denial declared, replacing the artifacts of the
earlier run on definitions that declared a deny list for Codex.

Refs: spx/31-outcomeeng.enabler/32-changes.enabler/65-change-execution.enabler
The Executor skill's open warning no longer requires the Codex render to
grant only what an unavailable path needs; the Codex render launches rounds
and Verifiers, so its grant is judged against that behavior.

Refs: spx/31-outcomeeng.enabler/32-changes.enabler/65-change-execution.enabler
…irectories

The Codex probe assembly script resolves both of its directory arguments
and reads and writes only inside them, and the file formats under the
repository formatter.

Refs: spx/31-outcomeeng.enabler/32-changes.enabler/65-change-execution.enabler
…cript

The script splits its imports onto separate lines and names its comprehension
variable line, so the repository lint step passes over it.

Refs: spx/31-outcomeeng.enabler/32-changes.enabler/65-change-execution.enabler
Record the filed test-evidence finding on the price-document test, which
reads the committed price data without driving the build. Drop the entry
about assertion helpers in the linked tests: the tests now carry their
assertions lexically.

Refs: spx/18-plugin-build.enabler/54-conversion.enabler/21-agents.enabler
@simonheimlicher
simonheimlicher force-pushed the work/change-433-codex-execution branch from 4d38caa to a479c3e Compare October 14, 2026 17:00
@sonarqubecloud

Copy link
Copy Markdown

@claude

claude Bot commented Oct 14, 2026

Copy link
Copy Markdown

No findings. Verified: the per-target Jinja conditionals added to the seven agent definitions and execute-change/SKILL.md follow the existing {!% if target == %!}/require_skill machinery in outcomeeng/distribution/build.py and match precedent in other dual-target skills (e.g. claim-change, release-change, update-instruction-block); the Codex model_reasoning_effort: high on all seven emitted .toml files matches the canonical AGENT_PROFILES registry in outcomeeng/distribution/profiles.py (CodexReasoningEffort has only HIGH), not a new per-target derivation; and the updated change-execution.md assertions, ISSUES.md settlement conditions, and the new Codex probe protocol/artifacts in probes/definition-invocation/ agree with each other (fork_turns: "none" asserted and observed on every spawn, discovery lists the seven names, each result file matches its declared blocked/not-held outcome).

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

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant