discovery: explore-deep gates skill-level context: fork behind CLAUDE_CODE_FORK_SUBAGENT — wrong mechanism, unreachable fallback

[kyle-sexton]

Observation

plugins/discovery conditions its forked-skill execution on
CLAUDE_CODE_FORK_SUBAGENT=1 in four places. That environment variable does not
gate skill-level context: fork.

Per the official subagents reference (fetched 2026-07-24,
https://code.claude.com/docs/en/sub-agents) the variable gates the Agent
tool's fork subagent type and the /subtask command. The skills reference
(https://code.claude.com/docs/en/skills) documents context: fork with no
environment gate at all.

Its actual effect runs the opposite direction from what the plugin claims:

When CLAUDE_CODE_FORK_SUBAGENT is set to 1, every subagent runs in the
background and the frontmatter background field has no effect.

So unset is the normal working state; setting it removes backgrounding control.

The documented fallback is also unreachable. explore-deep/SKILL.md:7
already declares context: fork, so by the time the skill body executes it is
inside the fork and cannot detect fork-unavailability to fall back from.
evals/evals.json case 2 asserts that branch.

Separately, explore-deep/SKILL.md:24 claims "You inherit the parent's full
toolset". That is conversation-fork behavior. Per the skills reference, a
backgrounded fork runs with the narrower background-subagent tool set, and
background: false is what restores the full set. No functional breakage today
— every tool the workflow uses is inside the background set — but it misleads
anyone extending the skill.

The same variable is described correctly elsewhere in the marketplace:
plugins/re-anchor/skills/sweep-all-disciplines/SKILL.md attaches it to the
Agent-tool fork type and explicitly distinguishes the two meanings. That path
moves under the in-flight re-anchor → discipline rename; the citation is
against main as of filing.

Direction (no open decision)

• Remove the CLAUDE_CODE_FORK_SUBAGENT precondition from
  plugins/discovery/skills/explore-deep/SKILL.md:3,
  plugins/discovery/README.md:12, and
  plugins/discovery/skills/explore/SKILL.md:24. In the last, keep the preceding
  explicit-scope sentence — it is correct and load-bearing.
• Delete eval case 2 fallback-when-fork-unavailable rather than adjusting it.
  The branch cannot fire.
• Replace the toolset-inheritance claim with the background-subagent tool set.
• Version bump plus CHANGELOG entry (changelog-parity is a CI gate).

Acceptance

• No CLAUDE_CODE_FORK_SUBAGENT reference remains in plugins/discovery
• evals.json contains 5 cases; no case asserts a fork-unavailable fallback
• The toolset statement matches the documented background-subagent set
• Any retained fork claim is traceable to a doc fetched in the authoring
  session, cited inline
• plugin.json version bumped with a matching CHANGELOG entry

---

This was generated by AI during a /planning:interview session on the fact-check capability.

[kyle-sexton]

Dissolved by retirement, not fixed — recording the disposition here rather than leaving it silent

This issue was already CLOSED when discovery-subagent-dispatch (PR #1425) reached the step that would have addressed it, so this is a disposition note, not a reopen.

The line it named — explore-deep/SKILL.md:3, gating skill-level context: fork behind CLAUDE_CODE_FORK_SUBAGENT=1 — no longer exists, because the skill no longer exists. #1425 retires /discovery:explore-deep: once /discovery:explore dispatches a named agent by default, the -deep variant was a second door onto an execution path the base skill already had.

Both halves of what this issue identified were real, and both are worth keeping on the record even though the file is gone, because the same mistake is reproducible elsewhere:

• The requirement was stale. Forks have been enabled by default since v2.1.161; the variable now only forces them on or off, and the user-facing command is /subtask as of v2.1.212.
• It was attached to the wrong mechanism, which is the subtler half. CLAUDE_CODE_FORK_SUBAGENT gates the Agent tool's fork subagent type — the one that inherits the live transcript. Skill-level context: fork is a different thing entirely: a regular agent type that receives the narrow background tool set and never sees the parent conversation. Text that conflates them promises transcript inheritance a context: fork skill does not get.

Rather than let that correction die with the file, #1425 carries it forward into discovery:setup check, which now reports fork availability as a control, not a gate, and says explicitly that the older framing was stale and misattached — so the next author reads the corrected version at the point they would otherwise reproduce the old one.

Sibling issues #1268 / #1258 / #1082 track the same staleness class in other files and are untouched by this.