Investigate enabled GitHub skill missing from model-visible catalog

[shiny-code-bot]

Skill Discovery Regression Evidence

Problem

An Every Code session in the current external harness on August 16, 2026
advertised github-plan in its Available skills block but omitted the sibling
github skill. The task was to inspect and update GitHub issues and pull
requests. github-plan explicitly delegates GitHub execution to github, yet
the routing instructions said that a delegated skill must be found in the
advertised list before use.

The omission caused the agent to treat github as unavailable and fall back to
raw gh reads, even though the skill's helper-first workflow was exactly what
the task required.

This was not reproduced yet through Codex Lab's own interactive harness. The
first step is to determine whether this is a Codex Lab regression, a client
integration discrepancy, or an outer prompt-assembly problem.

Verified Environment State

• ~/.code/skills resolves to the active cbusillo/codex-skills checkout.
• github/SKILL.md exists and parses as the github skill.
• [features] skills = true is active.
• No skills.config entry disables github; the only explicit disabled local
  skill is chronicle.
• The github skill does not set allow_implicit_invocation: false.
• The external session still omitted github while including github-plan and
  other skills from the same root.
• A standalone installed-CLI skills/list diagnostic returned an empty data
  array for the same working directory, despite the external session receiving a
  populated skills list. Treat this only as a diagnostic inconsistency until the
  exact client and protocol paths are reproduced in Codex Lab.

The installed configuration therefore indicates that github should be
enabled and eligible for model-visible discovery.

Expected Behavior

• An installed, enabled, implicitly invokable skill appears in the model-visible
  Available skills list.
• A visible skill must not delegate mandatory work to another installed skill
  that the model is told is unavailable.
• If context-budget selection can omit skills, delegation closure should retain
  required delegates or the runtime should provide a reliable explicit-load
  path that overrides omission.
• Diagnostics should distinguish disabled, manual-only, invalid, budget-omitted,
  and otherwise filtered skills.

Acceptance Criteria

• Attempt reproduction through Codex Lab's interactive session and debug prompt-input paths using the same skill root and effective configuration.
• If Codex Lab does not reproduce it, identify the external harness or
  prompt-assembly boundary that produced the incomplete catalog and record
  the owning component.
• Identify why github was omitted while github-plan was included from
  the same user skill root.
• Confirm skills/list consistently reports the effective enabled state for
  user skills in both standalone and interactive clients.
• Ensure an enabled github skill is included for GitHub issue, PR, review,
  and workflow tasks.
• Prevent model-visible skills from carrying mandatory delegation to a skill
  that is absent and cannot be explicitly loaded.
• If a skill is omitted for budget reasons, surface the omitted skill name
  and reason rather than only presenting an incomplete authoritative list.
• Add a regression test with github-plan delegating to github under the
  effective skill-context budget.
• Verify that an explicit user request for the github skill can load it
  even when normal implicit-selection heuristics would not choose it.

Relationships

• Potential regression of completed issue Preserve complete model-visible skill routing under prompt budgets #564 and PR Preserve complete model-visible skill routing #567.
• Related measurement work: Measure skill-description shaping against full routing metadata #603.
• Related helper-routing investigations: cbusillo/codex-skills#364 and
  cbusillo/codex-skills#376.

Notes

This is distinct from improving the github skill's helper coverage. The skill
already exists and contains the appropriate issue, PR, comment, and
authenticated fallback helpers. The observed failure is that the enabled skill
was not made available to the model when its delegated workflow was needed. The
owning implementation may be outside Codex Lab if its own harness preserves the
catalog correctly.