sweep-all-disciplines: fork subagents silently do not inherit the conversation, and two of four claimed they did

[kyle-sexton]

What happened

Running /re-anchor:sweep-all-disciplines in a session with substantial history, the batched fan-out dispatched its per-corrector audit agents with subagent_type: "fork". None of the four forks inherited the conversation. All four reported, in substance, that they had no history to audit.

The skill's own runbook says to stop rather than audit blind, so the batched form is not merely degraded here — it cannot run at all in the configuration it selects by default.

The harder half

Two of the four forks asserted they could see the conversation before checking, then self-corrected. That is worse than the inheritance gap itself: a fork that confidently reports "I have reviewed the session history" and only afterwards discovers it has none will, in any run where it is not asked to re-verify, return a fabricated audit. The batched sweep's output is a merge of those returns, so one unchecked fork silently poisons the whole pass.

Why this is ours

re-anchor is an org-owned plugin. The dispatch mechanism, the blind-audit stop rule, and the batched merge are all in our control.

Suggested shape of a fix

Not prescribing the fix, but the pieces that look load-bearing:

• The sweep needs a positive inheritance check before it trusts any audit return — something the agent can only answer from real conversation content, not a self-report that it has context.
• A fork that cannot see history must fail closed (return an explicit "no context, did not audit") rather than proceed. The runbook's stop rule needs a mechanism, not just a sentence.
• Worth deciding whether the batched form should dispatch forks at all, given the observed behavior, or whether named subagents plus explicit context hoisting is the only shape that works.

Reproduction

Observed in session 29e54148-1236-458c-85cb-fbf4c11314f8 on a long-running session. Not reduced to a minimal case — filing on the observation rather than sitting on it. A minimal repro is the first triage step.

Context

Surfaced while auditing the Fable field guide against the fable-5 playbook; the sweep was abandoned as a detour and the work proceeded with two cross-vendor reviews instead. Recorded in docs/topics/fable-field-guide-audit/ (see the handoff's "Tried and ruled out").

[kyle-sexton]

This was generated by AI during triage.

Verified — with one correction to the report

The mechanism is confirmed against the repo source at this commit, but the skill has been renamed since this was filed. re-anchor became discipline (commit 7cad5cae), and sweep-all-disciplines is now plugins/discipline/skills/sweep-all/. The substance is unaffected; the paths in the body are stale.

Confirmed in plugins/discipline/skills/sweep-all/SKILL.md:

• :73-74 — the fan-out dispatches a "conversation-inheriting fork subagent — the Agent tool's subagent_type: "fork"".
• :78-81 — the skill already knows forks can be unavailable and states the stop rule: "so if forks are unavailable, report that the inheriting audit fan-out cannot run and stop, rather than auditing blind."
• :181-182 — "Fork, not fresh/typed, and not skill-level context: fork."

The gap the item identifies is real and precisely stated: the stop rule exists only as prose instructing the fork to self-report. There is no positive inheritance check anywhere in the file — no mechanism that can distinguish a fork that genuinely inherited the conversation from one that asserts it did. That is exactly the failure mode observed, where two of four forks claimed context before checking and self-corrected only when pressed. In a run without that prompt, the fabricated audit merges into the batched output silently.

Related and worth reading together: commit 6d6d101a dropped the CLAUDE_CODE_FORK_SUBAGENT gate from forked explore in the discovery plugin. sweep-all:78 still reasons about that env gate, so the two plugins now hold different assumptions about fork availability.

Classification

• Work class: C4 structural — the fix changes the skill's dispatch mechanism and its trust contract, and the item explicitly leaves the shape open.
• Priority: high. Recorded reasoning, since the reporter self-labeled no severity: this is a silent-fabrication defect, not a degradation. A tool that returns a confident audit it did not perform is worse than one that fails, because the batched merge launders it into the result. It also blocks the discipline-sweep step that the loop lanes depend on, which is the "blocks other work" criterion.

Human-gated — genuinely open design space

Routed human-gated rather than briefed for autonomous work. The item itself declines to prescribe a fix ("Not prescribing the fix, but the pieces that look load-bearing"), and the load-bearing question it raises is a real fork in the design:

whether the batched form should dispatch forks at all, given the observed behavior, or whether named subagents plus explicit context hoisting is the only shape that works.

That is not a reversible implementation detail — it decides whether conversation-inheriting audit is a supported capability of this skill at all, and the answer partly depends on harness behavior this repo does not own. Not defaultable.

First triage step named in the body — a minimal repro — has not been done. No reduced case exists yet; the evidence is one observed session (29e54148-1236-458c-85cb-fbf4c11314f8). Whoever takes this should establish whether the non-inheritance is environment-specific before designing around it, because the fix differs entirely depending on whether forks never inherit here or failed under specific conditions.