sweep-all

Sweep all disciplines

A declared second species in this plugin: NOT a corrector. Every sibling skill re-anchors ONE discipline; this one carries no discipline of its own — it is a pure router that COMPOSES the correctors into a single batched re-anchor pass. It holds zero discipline text: the disciplines live in the correctors, the shared method lives in ${CLAUDE_PLUGIN_ROOT}/context/re-anchor-audit-correct.md, and membership + order live in each corrector's own tier metadata. This skill names no members — it globs and reads them, so the bundle cannot drift from a hand-maintained list.

Two modes

Session-start digest (cheap, default when nothing has happened yet). Derive the posture from the skill listing and each corrector's tier metadata — NO corrector bodies load, NO audit runs. Report which correctors are core (run every session), which are situational (relevance-gated), and which are never-batched. This is the conversation-start case: set posture, audit nothing.
Full batch pass (mid-session, or on explicit request). Preflight that the fan-out can inherit this conversation (mandatory — see Preflight), then fan out an audit-only subagent per in-scope corrector and apply their corrections once, on the main thread, in a fixed order (below). A failed preflight degrades to mode 1.

Resolving membership (never named inline)

Glob the sibling corrector directories and read each one's metadata.discipline-batch (core / situational / never) and metadata.discipline-batch-rank. The runbook never hardcodes member names; the tier is colocated with each corrector, so changing a shipped tier is a PR to that corrector — drift is structurally impossible.

core — in scope every session.
situational — in scope only when relevant to THIS conversation. Route from the corrector's own listing description (its trigger phrases and "at conversation start on …" clause), not from a guess. Report which situational correctors were included and which were skipped and why — a skip is always reported, never silent.
never — excluded from the batch by execution or interaction class (the -deep fan-out tiers; scrutinize-dont-coast, which needs a non-fork fresh context and stops to remediate with the user). Report that they exist and are invoked directly, not batched.

userConfig overlay (see Configuration) applies after tier resolution: exclude drops a member, promote lifts a situational corrector to always-run, demote drops a core corrector to relevance-gated. Report the net effect when an overlay changes the resolved set. Zero-config = tiers exactly as the correctors declare them.

Preflight: prove the fan-out can inherit (before step 1)

The batched pass is only meaningful if its subagents actually inherit this conversation — establish that before dispatching, never by assuming it. A subagent with no history has nothing to audit, and some share of them will invent a ledger from the system prompt rather than say so — six of eight did in the run this preflight comes from, and the two that refused are the only reason it was caught. The batched pass's step 3 then merges those ledgers and its step 4 writes their remedies to the working tree. That is the failure this preflight exists to prevent: a correctness pass whose failure mode is confident, invented corrections applied to real files.

Stage 1 — read your own tool schemas. Zero dispatch, diagnostic only. Two documented sentences pair up: fork mode "removes the run_in_background parameter from the Agent tool", while CLAUDE_CODE_DISABLE_BACKGROUND_TASKS set to 1 removes it from "Bash and subagent tools" (https://code.claude.com/docs/en/sub-agents, https://code.claude.com/docs/en/env-vars). So Agent still carrying it means fork mode is not env-var-enabled; Agent lacking it while Bash keeps it excludes the disabled-background-tasks cause and leaves fork mode as the remaining documented explanation; both lacking it says nothing about fork mode. This stage never gates — the docs tie the removal to the env-var path and say nothing about what the server-side rollout does to the parameter, so no branch is conclusive in either direction. It explains what stage 2 finds; it never replaces stage 2 and never aborts on its own.

Stage 2 — an inheritance-proof canary. The decider, and it costs one fork. Dispatch ONE fork alone, ahead of the first wave, that answers the proof question and nothing else — no corrector, no audit, no ledger. Gate the whole fan-out on it.

It is deliberately not folded into a member's real audit, which would look free and is not: a fork inherits everything the session holds when it spawns, so a member ledger returned before wave 1 would sit in every later fork's inherited context and anchor its audit — breaking the independence step 3 relies on (see "the forks stay independent by design"). A proof-only canary returns nothing that can anchor anyone. Budget the guard as one extra fork; that is what it costs to know the other ledgers are real.

Every fork — canary and members alike — answers one inheritance-proof question FIRST, before any audit content, and stops and says so plainly if it cannot. Conversations differ, so specify the question's properties, not a fixed question. All four are required:

Its answer exists only in this conversation's history — not in a file, not in a CLAUDE.md: a non-fork subagent's initial context still contains "every level of the CLAUDE.md hierarchy the main conversation loads" plus the delegation message you write (https://code.claude.com/docs/en/sub-agents). Nor derivable from this plugin.
The dispatch prompt neither contains nor paraphrases the answer — else a non-inheriting subagent answers it from the prompt alone.
It keys on ordinary inherited material — a prior user turn or tool result. Out-of-band or host-injected content is not reliably inherited (observed once in a fork-enabled session: an out-of-band advisor result was absent from a fork's inherited transcript — not documented behavior, and a proof keyed on it would have read as a false negative).
It cannot be guessed. An answer a non-inheriting subagent could hit by chance — a yes/no, a binary choice, a detail common to most sessions — clears the main thread's check without proving anything, and the blind ledgers behind it then reach the corrective write — and blind forks guess alike, since they share the question and the model, so one lucky answer is not one bad ledger. Small-domain values fail this test even when they are exact: a turn count, a file count, a finding count are all guessable. Require a long verbatim string or a specific identifier, and when in doubt mint the value (below) rather than picking one out of the history.

When the conversation offers no detail meeting all four — a thin session that opened straight into a full batch over an already-dirty tree — do not degrade. Mint one. "A fork inherits everything the main session has at the moment it spawns" (https://code.claude.com/docs/en/sub-agents), so emit a fresh high-entropy value into the transcript as an ordinary main-thread tool result BEFORE the canary spawns, and ask for it back. It is unguessable by construction, exists in no file, and a non-inheriting subagent cannot produce it — so the proof works at any conversation length, and thin history never costs the user the audit they asked for.

Verify on the main thread, and fail closed. Check the answer against what this context knows. Absent, ambiguous, or unverifiable proof counts as NOT inherited — a plausible-looking answer is not a pass, because fabrication is the exposure being defended against. Canary verified → fan the members out. Canary unproven → degrade; never re-dispatch the batch blind.

A member that returns unproven LATER, mid-fan-out, is a different case: the canary already established that inheritance works here, and earlier waves' ledgers are checkpointed and real. Discard that member's ledger, retry it once, and if it is still unproven keep every verified ledger, correct forward from those, and report the unproven members as open — do not throw away proven audits by collapsing the whole pass to the digest. Reserve the digest for a failed canary, when nothing has been proven at all.

A failed canary degrades; it does not stop empty-handed. Run the session-start posture digest (mode 1) — no fan-out, no cost — and report that the inheriting audit fan-out could not run, which signal established that, and that every corrector remains available for direct invocation, one at a time, each running the shared loop in THIS context with no fork at all. That is the same position setup reports as the full-batch prerequisite; this runbook is where it executes. The batch does not sequence the correctors itself as a substitute — that recreates the salience dilution the declared delta below exists to prevent, and yields audits weaker than the ones it declined to run.

What is gated, and what is not documented. CLAUDE_CODE_FORK_SUBAGENT set to 1 enables fork-spawning and 0 disables it "overriding any server-side rollout", and a staged rollout can enable it without the variable (https://code.claude.com/docs/en/env-vars, https://code.claude.com/docs/en/sub-agents). What the harness does when the fork type is requested while fork mode is OFF is not documented on any current page — observed once, in the failed full-batch run this preflight comes from, as subagents returning with no inherited conversation. Treat it as an observation, not a contract; the preflight does not rest on it, proving inheritance positively rather than predicting the shape of its absence.

The batched pass — a declared delta from the shared loop

Recorded here as a declared delta per the shared method doc's declared-delta rule (this runbook is not a corrector, but it orchestrates the shared loop across members and diverges from its per-corrector "correct forward now" step, so the divergence is part of this skill's contract, not silent drift). Two divergences are declared: the inserted preflight above, which sits before this list's step 1 and gates the whole pass, and the split below. The shared loop has each corrector run steps 1–3 itself, in its own context. This runbook splits them: the audit (steps 1–2) runs in per-corrector subagents; the correction (step 3) is collected and applied ONCE on the main thread, serialized in rank order. Rationale: a dozen disciplines each correcting forward independently in one context recreates the salience dilution this plugin exists to fix, and a merged pass lets order matter. The shared method's Non-negotiables are unchanged and bind every member.

Fan out, audit-only — after the preflight above has proved inheritance. For each in-scope corrector, dispatch a conversation-inheriting fork subagent — the Agent tool's subagent_type: "fork", which inherits the full conversation history the audit must read. A fresh/typed subagent receives no history, and a skill-level context: fork also discards it — neither can audit this conversation. The forks read the live working tree — do NOT isolate them (see Gotchas: isolation would hide the uncommitted work in flight, which is usually the thing under audit). Their no-writes rule is trusted, not enforced — see Gotchas, and treat a fork that wrote as untrusted output: stop rather than correct on top of it. Instruct each fork: answer the preflight's inheritance-proof question first, then load exactly this ONE corrector's SKILL.md, run shared-loop steps 1–2 only (re-anchor + self-audit), make NO writes, and return a findings ledger. Each ledger entry carries the concrete located finding AND the remedy this corrector would apply for it (the shared loop's step-3 corrective action, described not performed — the fork still writes nothing), or an honest "clean". Capturing the proposed remedy in the audit is what gives step 3 the reporter→remedy data to key on; a ledger of bare violations would leave the dedup with nothing to preserve.

Wave sizing, budget, and what counts as a failure. Prefer ONE wave: dispatch every in-scope member together, so no member's ledger is in the transcript when another member spawns. That independence is load-bearing, not decoration — step 3's dedup and step 4's ordering both assume the ledgers were formed without seeing each other, and a fork inherits everything the session holds at the moment it spawns. Splitting the fan-out is what breaks it, so do not import the -deep siblings' "roughly a dozen" wave: it is calibrated for cheap fresh-context subagents that carry no such invariant.

One wave is usually possible, and the two documented limits are not the same kind of limit. CLAUDE_CODE_MAX_TOOL_USE_CONCURRENCY ("Maximum number of read-only tools and subagents that can execute in parallel", documented default 10) caps how many run at once — it does not cap how many you dispatch. The hard one is CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS (documented default 20): past it, "spawning another with the Agent tool fails with Concurrent subagent limit reached, and the error tells Claude not to retry." Every Agent-tool subagent counts against it, forks included, shared with everything else the session is running, and against CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION (documented default 200) (https://code.claude.com/docs/en/sub-agents, https://code.claude.com/docs/en/env-vars). So even a fully-admitted set — every core plus every situational corrector — dispatches in one wave in an otherwise-quiet session.

Never drop a relevant corrector to fit a budget. Membership resolution decides what is in scope; concurrency decides only when it runs. If the session already holds enough subagents that the full set cannot be dispatched, wait for capacity and then dispatch it whole. Split only if that is not possible, and then say so in the report: every member in a later wave inherits the earlier waves' ledgers and can be anchored by them. That is a real weakening of the independence the dedup relies on, disclosed, not hidden. Only a split fan-out needs the -deep siblings' per-wave checkpointing of the collected ledgers; a single wave has no partial state to lose.

Retry, and what counts as a failure. Retry only a failed subset, once — and failure includes a ledger returned without verified inheritance proof, not only an errored dispatch: a fabricated ledger is the exposure the retry rule exists for. One dispatch error is explicitly NOT a retryable failure: Concurrent subagent limit reached, which the harness tells Claude not to retry. That is a capacity race, not a failed audit — wait for capacity and dispatch, rather than retrying into the same wall. A retry is anchored by construction, one wave or many: it spawns after other members' ledgers have landed, so it inherits them. There is no un-anchored rerun available inside this conversation — re-running the whole pass would inherit MORE, not less, since every ledger is already in the transcript any new fork copies. So do not offer a clean rerun; report each retried member's ledger as anchored and weigh it as such. A member still unproven after its one retry is reported open, and the verified ledgers are kept and corrected — see the preflight's mid-fan-out rule; only a failed canary collapses the pass to the digest.
Collect every ledger.
Dedup by root cause. Before correcting, group ledger entries across correctors that name the SAME underlying finding — distinct disciplines routinely surface one root cause as separate entries (e.g. a recall-based claim flagged independently by do-your-research, recheck-against-upstream, and mind-your-maxims). Group them into one finding carrying the union of their located evidence and, keyed BY reporting corrector, the remedy that corrector asks for — a reporter→remedy mapping, not a flat remedy list beside a separate reporter list, so step 4 knows which remedy belongs to which rank. What dedup collapses is the re-analysis and the re-reporting of one root cause — NOT the corrective work: a shared root cause can demand more than one remedy that are not interchangeable (verifying or retracting the unsupported claim satisfies the research reporters, but mind-your-maxims may still require a reader-facing uncertainty disclosure), and every reporter's remedy is still applied. This grouping is the ONLY place ledgers combine: the forks stay independent by design (no shared ledger across forks — that independence is what keeps each audit's perspective un-anchored), and grouping is by root cause, not by corrector. A finding only one corrector raised passes through unchanged.
Correct once, in rank order. Walk the in-scope members — the core and situational correctors that ran (never-tier carries no rank and was already excluded at membership resolution) — by ascending discipline-batch-rank, and correct forward each finding on the main thread now — the shared loop's step 3, batched. For a grouped finding, "once" means the root cause is analyzed and corrected in a single pass, not that its remedies collapse to one: apply every reporter's distinct remedy, each at its own reporting corrector's rank (so the research retraction and the mind-your-maxims disclosure both happen, in rank order), rather than dropping the lower-priority reporters' corrective work. The order: use-your-skills first (fix which skills govern the work) → evidence correctors (get the facts right) → structural correctors → mind-your-maxims (communication) → tighten-your-output dead last, so it never tightens text a later corrector rewrites. Correcting on the main thread does not suppress the shared method's fresh-context escalation: where a finding's suspected drift source is this context's own judgement, re-derive it in a fresh-context (non-fork) subagent blind to that reasoning, per the method doc's Non-negotiable — the batch orchestrates the correction here, it does not waive that escalation. (The fork audit inherits context on purpose: step 2 is a same-context self-audit; this carve-out is about step 4.)
Report one consolidated ledger: per corrector — corrected / clean / open; a root-cause finding merged in step 3 is listed once, attributed to every corrector that reported it; plus which situational members ran versus were skipped, and which never-members exist for direct use.

Member human-gates survive batching

The batch never converts a member's human gate into an autonomous action. The shared method's outward-artifact carve-out (no PR, issue, or published comment without explicit opt-in) holds, as do follow-our-standards' upstream-divergence routing (draft and route, do not file), pick-for-the-problem's incumbent-replacement decisions, and recheck-against-upstream's undocumented-divergence calls. The batch surfaces these for the user; it does not resolve them unasked.

Configuration

The overlay reads three personal-scalar userConfig options (configured through Claude Code's native plugin-config flow — see the plugin README — never a hand-edited member), substituted here at load:

Excluded correctors: ${user_config.batch_exclude}
Promoted to always-run: ${user_config.batch_promote}
Demoted to relevance-gated: ${user_config.batch_demote}

Each is a comma-separated list of corrector names. An unset option does not reliably substitute to empty: on a zero-config or headless install, or for a user who never ran plugin-config, the literal ${user_config.…} token survives instead. Treat BOTH an empty value AND a surviving literal placeholder as unset — no overlay from that key — and never read the literal token as a corrector name. Apply after tier resolution: batch_exclude drops a member, batch_promote lifts a situational corrector to always-run, batch_demote drops a core corrector to relevance-gated. Report the net effect whenever the overlay changes the set.

What this skill does NOT do

Not a corrector. It owns no discipline and re-anchors nothing itself.
Not use-your-skills. That corrector routes to the ONE skill fitting a task; this composes the whole posture bundle.
Not a session or SDLC orchestrator. Staged navigation and session lifecycle belong to the session-flow plugin; this only sequences the re-anchor correctors.
Does not batch the never tier — the -deep fan-out siblings and scrutinize-dont-coast are invoked directly.
Does not name its members inline — membership is each corrector's own tier metadata.

Gotchas

Fork, not fresh/typed, and not skill-level context: fork. Only the Agent-tool subagent_type: "fork" inherits the conversation the audit reads; the other two start blank.
Audit in the forks, correct on the main thread. Parallel forks that wrote would race and re-dilute salience; the value is one ordered correction pass.
The forks' no-writes rule is trusted, not enforced — say so. A named subagent's tool access can be narrowed with tools / disallowedTools; a fork's cannot. Forks "skip both filters and receive the main conversation's exact tool pool", and a fork's system prompt and tools are "Same as main session" (https://code.claude.com/docs/en/sub-agents). So every audit fork holds Write, Edit, and Bash and is only asked not to use them. Never present the audit fan-out's read-only posture as harness-enforced. If you want assurance that the fan-out honored it, capture the working tree's state before the FIRST dispatch (the canary included) and compare afterwards, and treat any difference as a fork that wrote — untrusted output, stop rather than correct on top of it. That is detection after the fact, not prevention, and a robust comparison is more than a git status diff — specifying one is tracked in #1631 rather than half-specified here.
isolation: "worktree" was considered for the audit forks and rejected. The Agent tool accepts it on a fork, and it would move a fork's file edits off the user's checkout — but a git worktree is created from a commit, so the fork would not see the uncommitted work in flight, which is usually the very thing the audit exists to inspect. It also would not bound a write addressed by an absolute path, and inherited history is full of absolute paths. It trades a real loss of audit fidelity for partial containment. Record this if it is proposed again.
The preflight is the guard, not an optimization. Skipping it does not make the sweep cheaper — it makes every ledger unfalsifiable, and the batched pass's step 4 writes those ledgers' remedies to the working tree.
Forks run at the parent model's cost. An Agent-tool fork ignores a model override and inherits the whole conversation, so each in-scope corrector's audit runs at the parent model over the full transcript; the wave cap bounds burst, not per-fork cost. Keeping the never tier out and relevance-gating the situational tier are what hold the fan-out small. Order of magnitude from a real full-batch run on a mid-length session: each fork consumed ~170K tokens (inherited transcript), so an 8-in-scope pass ran ~1.4M tokens for the audit phase alone, plus one more fork for the proof-only canary. Budget the sweep as a deliberate spend, not a reflex — on a long transcript the per-fork cost only grows.
tighten-your-output stays last — tightening before the other corrections would tighten text they then rewrite.
A situational skip is reported, not silent — the user sees what was left out and why.