discipline:sweep-all — wave cap imports the wrong number and the shared concurrent-subagent budget is unmodeled

[kyle-sexton]

Summary

Two concurrency facts the sweep-all fan-out depends on are wrong or absent in
plugins/discipline/skills/sweep-all/SKILL.md.

F6 — the wave-cap pointer imports the wrong number

Step 1 says:

Cap concurrency in bounded waves like the -deep siblings

Following that pointer resolves to do-your-research-deep / recheck-against-upstream-deep:
"Cap concurrency to a modest wave (roughly a dozen or fewer)". Two problems:

• That number is calibrated for cheap fresh-context subagents, not for forks that each inherit
  the full transcript at the parent model. sweep-all's own Gotchas already record ~170K tokens
  per fork on a mid-length session.
• "Roughly a dozen" exceeds CLAUDE_CODE_MAX_TOOL_USE_CONCURRENCY, documented at
  https://code.claude.com/docs/en/env-vars (fetched 2026-07-26) as: "Maximum number of read-only
  tools and subagents that can execute in parallel (default: 10). Higher values increase
  parallelism but consume more resources".

The same pointer also silently skips the -deep siblings' mid-run checkpointing of the partial
ledger, which is part of what makes their fan-out survivable.

F7 — the shared concurrent-subagent budget is unmodeled

Forks are not exempt from the session's concurrency budget. Per
https://code.claude.com/docs/en/sub-agents (fetched 2026-07-26), verbatim:

By default, when 20 subagents are running in a session, spawning another with the Agent tool
fails with Concurrent subagent limit reached, and the error tells Claude not to retry.

and, on the session cap:

Every subagent Claude spawns with the Agent tool counts toward the limit: nested subagents,
forks, and background subagents […]

sweep-all models neither. The audit that produced this finding fired in a session already
holding 10+ spawned members, so the fan-out was competing for a budget it never accounted for.

Proposed fix

• State the wave cap as an explicit number chosen for forks, not a pointer to a differently
  calibrated sibling, and keep it under the documented default parallelism of 10.
• Carry over the -deep siblings' partial-ledger checkpointing rather than dropping it by
  omission.
• Model the shared budget: the sweep's forks share CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS
  (default 20) with everything else running in the session, and count toward
  CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION (default 200).

These land alongside the preflight work because they occupy the same paragraphs.

Source: plugin-quality:audit post-use audit of discipline@0.9.0 (F6, F7 — both MED).

[kyle-sexton]

This was generated by AI during triage.

Verified — both findings confirmed, all upstream quotes verbatim

F6 — the wave-cap pointer imports the wrong number. Confirmed. The sweep-all fan-out step says "Cap concurrency in bounded waves like the -deep siblings". Following that pointer, both do-your-research-deep and recheck-against-upstream-deep state "Cap concurrency to a modest wave (roughly a dozen or fewer)". https://code.claude.com/docs/en/env-vars (fetched 2026-07-26), verbatim:

CLAUDE_CODE_MAX_TOOL_USE_CONCURRENCY — Maximum number of read-only tools and subagents that can execute in parallel (default: 10). Higher values increase parallelism but consume more resources

So "roughly a dozen" does exceed the documented default, and it is calibrated for cheap fresh-context subagents rather than for forks the skill's own Gotchas price at ~170K tokens each.

The dropped checkpointing is confirmed too, and is the sharper half of F6. recheck-against-upstream-deep carries an explicit mid-run step — "Checkpoint the partial ledger mid-run, if a durable slice exists… so a crash mid-fan-out does not lose completed waves" — with a matching eval, and degrades gracefully when no durable store exists. sweep-all inherits none of it by pointing at the siblings for the cap only.

F7 — the shared concurrent-subagent budget is unmodeled. Confirmed. https://code.claude.com/docs/en/sub-agents (fetched 2026-07-26), verbatim:

By default, when 20 subagents are running in a session, spawning another with the Agent tool fails with Concurrent subagent limit reached, and the error tells Claude not to retry.

Every subagent Claude spawns with the Agent tool counts toward the limit: nested subagents, forks, and background subagents

Session cap likewise: CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION default 200, CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS default 20 — both verbatim from the env-vars page. The runbook models neither.

Why this is delegable while its siblings are not: F7 holds for any Agent-tool fan-out, fork or named, so it survives whichever way the open structural decision in #1258 goes. F6's "exceeds the documented default parallelism" half and the dropped checkpointing are likewise dispatch-independent. Nothing here writes new fork-specific machinery.

Priority: medium — matches the reporter's own MED/MED on both findings; real value, no deadline.

Role-label warning: this repository's .work-item-tracker.json has a config block with no role_labels entry, so the autonomous-eligible role resolves to the plugin default (agent-ready). The label exists live in this repo, but the resolution is a default, not a declared binding.

Agent Brief

Type: Task (documentation-of-behavior contract change to a shipped skill runbook)
Summary: Give the discipline plugin's sweep-all skill its own concurrency numbers instead of importing a differently-calibrated sibling's, carry over the mid-run checkpoint the pointer silently dropped, and model the session-shared subagent budget the fan-out competes for.

Current behavior:
The sweep-all fan-out step caps concurrency by pointing at the -deep sibling skills. That pointer resolves to a number chosen for cheap fresh-context subagents ("roughly a dozen or fewer") which also exceeds the documented default parallelism, and it silently imports none of the siblings' mid-run partial-ledger checkpointing. The runbook models no session-wide subagent budget at all, so a sweep fired in a session already holding other spawned agents competes for a limit it never accounted for.

Desired behavior:

• The runbook states its own explicit wave-cap integer, chosen for conversation-inheriting forks and justified against their per-fork cost, rather than delegating the number to a sibling. Constraint: it must sit below the documented default parallelism and be defensible against the per-fork cost the skill's own Gotchas already record. Pick the number in the PR and state the reasoning inline; do not restate the upstream default as though the skill owned it.
• The fan-out checkpoints its partial ledger mid-run when a durable slice exists and degrades gracefully when none does, matching the behavior the -deep siblings already ship rather than re-inventing a second shape.
• The runbook states that the sweep's audit subagents share the session's concurrent-subagent budget with everything else running, and count toward the per-session spawn cap. Name the governing environment variables and their documented defaults, and cite the official pages rather than asserting the numbers.

Key interfaces:

• The sweep-all skill's fan-out step — currently a pointer to sibling skills for the cap; becomes an owned number plus a checkpoint instruction.
• The -deep sibling skills' checkpoint step — the established shape to reuse, not to re-derive.
• CLAUDE_CODE_MAX_TOOL_USE_CONCURRENCY, CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS, CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION — named in the runbook with defaults attributed to the official env-vars page.

Acceptance criteria:

• The runbook no longer relies on a sibling pointer for its concurrency number; an explicit integer below the documented default parallelism is stated with its rationale.
• A mid-run partial-ledger checkpoint instruction is present, with the same graceful-degradation clause the -deep siblings use when no durable store exists.
• The runbook names the shared concurrent-subagent budget and the per-session spawn cap, with defaults cited to the official documentation rather than asserted.
• Every harness-behavior fact added is cited to a page fetched during the change, per this repository's fresh-docs mandate.
• /skill-quality:check passes on the modified skill, and its evals file still validates.

Out of scope:

• Fork-availability detection, any preflight or canary, and the sweep-all / setup degraded-mode contradiction — tracked in discipline:sweep-all — the fork stop-rule is inert (no detection mechanism) and contradicts setup's declared degraded mode #1621, blocked on the open structural decision in sweep-all-disciplines: fork subagents silently do not inherit the conversation, and two of four claimed they did #1258.
• The isolation: "worktree" containment and the "make NO writes" enforcement wording — tracked in discipline:sweep-all — "make NO writes" is unenforceable: forks receive the main conversation's exact tool pool #1622, same cluster.
• The cost gate, the situational relevance tie-break, the membership-read parse rule, and the eval gaps — tracked in discipline:sweep-all — cost gate, relevance tie-break, membership-read spec, and missing evals #1624.
• Changing the dispatch mechanism itself (forks versus named subagents) — that is sweep-all-disciplines: fork subagents silently do not inherit the conversation, and two of four claimed they did #1258's decision, not this item's.