Commissioning-surface audit

Result

The Outcome 4 cohort was frozen at commit 20cc50f82f88f50e7b0dca71ec843d7ca28b647a. The audit found sixteen reusable commissioning or handoff cohorts. Four are revised, seven retain their exact controls, and five remain behind named evidence or design gates.

The revisions are narrow. Outcome 3 already revised the multistage writer. Outcome 4 removes repeated generic parent-lifecycle and no-nested-delegation text from the warning-fix and standalone compression-review surfaces, whose workers receive verified root Commonplace doctrine. It also corrects one Second Brain example that described task evidence as if it excluded ambient project instructions. Exact task purpose, inputs, write ownership, acceptance, return, and failure controls remain local.

No broad packet format, schema, runtime branch, or checked-text reuse mechanism is added. The audit does not reopen exact protocols merely because they contain more words.

Frozen boundary and searches

The sweep covered the 230 tracked files at the frozen commit under AGENTS.md, AGENTS.md.template, kb/instructions/, kb/types/, kb/messages/, kb/tasks/, and src/commonplace/. Canonical source files were inspected, not their .agents/skills/ and .claude/skills/ symlink projections. Run-instance messages, generated reports, source ingests, historical workshops, and vendored artifacts are evidence or task instances, not reusable packet producers. Active workshop and task contracts remain in scope.

The frozen searches were:

git grep -Il -E 'delegate|delegation|sub-agent|subagent|worker|handoff|dispatch|spawn|parallel|fresh context' 20cc50f8 -- AGENTS.md AGENTS.md.template kb/instructions kb/types kb/messages kb/tasks src/commonplace
git grep -Il -E 'prompt|brief|packet|plan|checklist|task text|command' 20cc50f8 -- AGENTS.md AGENTS.md.template kb/instructions kb/types kb/messages kb/tasks src/commonplace
git grep -Il -E 'defer|pending|follow-up|future work|revisit|resume|trigger|return condition' 20cc50f8 -- AGENTS.md AGENTS.md.template kb/instructions kb/types kb/messages kb/tasks src/commonplace
git grep -Il -E 'invoke|execute|run-review-batches|cp-skill-|review bundle|review gate' 20cc50f8 -- AGENTS.md AGENTS.md.template kb/instructions kb/types kb/messages kb/tasks src/commonplace

They returned 39, 93, 40, and 65 candidate files respectively. Every tracked instruction and type file was also enumerated, so a surface did not depend on matching one term. Screening removed criterion bodies, composition callees, documentation, validation dispatch, ordinary shell launch language, and files that merely mention agents. The only shipped code that generates a model task is src/commonplace/review/protocol/prompt.py.

Files added after the frozen commit belong to follow-up unless they invalidate the shared doctrine. No such file appeared before this audit began.

Delivery baseline used by the audit

Codex fresh workers receive root AGENTS.md as binding repository instruction. Native Claude Code 2.1.251 non-fork general-purpose workers receive the same text through CLAUDE.md → AGENTS.md. The audit treats only those direct paths as verified. Commonplace code creates and finalizes review jobs but launches no model runtime, so generated review prompts retain their self-contained controls.

A Bash-launched agent process, different working directory, custom agent definition, remote isolation, or another harness is not covered. A packet on one of those paths must carry every relied-on rule until that path is verified. Mailbox messages remain self-contained by default because their recipient path may be unknown.

Disposition records

Each record names the baseline, task intent and acceptance, inherited choices, delegated choices, retained exact controls, and any gap. Retain exact means the current wording has a task-specific or protocol reason. Defer behind evidence keeps the current bytes and names what would reopen them.

Workshop commissions — retain exact

  • Baseline: The executing session's verified Commonplace doctrine, then kb/work/COLLECTION.md.
  • Intent and acceptance: The framing states the question, who posed it, what would close it, and any evaluation or coordination premises a later session cannot reconstruct.
  • Inherited choices: Parent scheduling, integration, recovery, and no unauthorized nested delegation.
  • Delegated choices: Routes, methods, and later decisions for which the framing names discriminating evidence or bounded authority.
  • Exact controls retained: Task authority, owned mutations, accessible inputs, cross-work coordination, acceptance, and return conditions.
  • Gap and disposition: None found. Individual workshop files are task instances, not standing producers; retain the contract exact.

Mailbox requests — retain exact

  • Baseline: None by default. A message may name a verified recipient path.
  • Intent and acceptance: The request states the result and what it is for, its verification, and its return channel.
  • Inherited choices: Only rules delivered on the named verified path.
  • Delegated choices: Consequential choices the request deliberately leaves to the recipient.
  • Exact controls retained: Authority, owned outputs, inputs, coordination, and stop conditions. Posting neither launches an agent nor grants authority.
  • Gap and disposition: None found. Cross-session and cross-harness transport makes the conservative self-contained rule necessary.

Source-checkout task documents — defer behind subsystem design

  • Baseline: A future agent's repository doctrine plus the applicable kb/tasks/types/ task contract; kb/tasks/ is an operational, source-only surface without a collection contract.
  • Intent and acceptance: Active tasks carry a goal, prerequisites, context, decisions, next steps, and current state; recurring tasks carry scope, procedure, and output location.
  • Inherited choices: Repository authority and normal parent lifecycle.
  • Delegated choices: Execution route where the task's current state leaves it open.
  • Exact controls retained: Task-specific decisions, triggers, verification, and durable run-log destinations.
  • Gap and disposition: Whether this subsystem should ship or remain source-only is unsettled, and current backlog files are already complete. Do not redesign its templates until that ownership decision is made.

Multistage writing — revised in Outcome 3

  • Baseline: Verified Codex and native-Claude root doctrine, the instruction collection and type contracts, the invoked skill, and its promotion reference.
  • Intent and acceptance: Produce one grounded artifact through isolated reconstruction, staged authorship, exact-byte independent review, and guarded promotion.
  • Inherited choices: Generic parent scheduling, integration, recovery, and no unauthorized nested delegation.
  • Delegated choices: Investigative sequence, reconstruction form, disposition form, decomposition, examples, and prose inside fixed evidence and result boundaries.
  • Exact controls retained: Per-role inputs and outputs, incumbent isolation, grounding, invalidation, one repair, accepted digest, drift, mutation, rollback, and user-decision boundaries.
  • Gap and disposition: None. Four fresh evaluators found no behavioral divergence; the lean revision was adopted at commit 20cc50f8.

Source ingest and re-ingest — retain exact

  • Baseline: Verified root doctrine plus cp-skill-ingest, its standalone draft-ingest-report.md worker instruction, source contracts, and the composition-owned re-ingest.md route.
  • Intent and acceptance: Produce one current analysis from one exact snapshot while preserving snapshot identity and retained Quotes bytes.
  • Inherited choices: Generic parent lifecycle and default denial of unapproved nested delegation.
  • Delegated choices: Source analysis, value selection, and prose within the snapshot, connect report, occasion, and artifact contract.
  • Exact controls retained: Clean context, fixed paths, checksum guards, opaque Quotes, one worker replacement, byte backup and restore, anti-skill recursion, sole output, and validation. The optional occasion already carries task-specific intent without contaminating source classification.
  • Gap and disposition: None in the drafting handoff. re-ingest.md is a composition caller, not a second worker packet. The prior kept-exact disposition still applies.

Agentic-system lens workers — defer behind the active result-shape design

  • Baseline: Verified root doctrine plus analyse-agentic-system, one frozen evidence boundary, its canonical records, and the invoked epistemic method.
  • Intent and acceptance: Emit one bounded system synthesis with a mandatory runtime account and both proportionately scoped lenses.
  • Inherited choices: Generic parent lifecycle and no unauthorized nested delegation.
  • Delegated choices: Lens depth, targeted reads inside the frozen boundary, and analytical organization consistent with the canonical record set.
  • Exact controls retained: One revision, source register, ID ownership, correction/invalidation rules, no evidence upgrades, both lens outputs, and prevented-conclusion limits.
  • Gap and disposition: The skill deliberately leaves physical result layout open, so a reusable concrete lens-output packet would choose an output shape the active kb/work/analyse-agentic-system/ migration still owns. Keep the current topology exact and defer packet templating until that design settles or a failed handoff demonstrates a narrower repair.

External-literature bilateral isolation — retain exact

  • Baseline: Verified root doctrine, the multistage skill, the literature instruction, and cp-skill-ground with its authorized ingest chain.
  • Intent and acceptance: Decide the smallest faithful artifact disposition from direct claim-grained source evidence.
  • Inherited choices: Parent lifecycle and default no-nested-delegation outside the explicit grounding exception.
  • Delegated choices: Source candidacy inside the fixed search boundary, direct-evidence reconstruction, and isolated comparison judgment.
  • Exact controls retained: Conditional trigger for bilateral isolation, three named outputs, target/source blindness, exact grounding caller, admitted-source authority, mutation ownership, and user approval for an agent-nominated URL. The cp-skill-ground → ingest chain is an explicit nested exception, not authority inferred by a worker.
  • Gap and disposition: The earlier nested-authority concern is closed by the live role and composition text. Retain exact.

Warning-fix sweep — revise

  • Baseline: Verified root doctrine plus FIX-SYSTEM.md, the sweep, and the per-note fix instruction.
  • Intent and acceptance: Apply current warn findings to disjoint notes, produce an auditable fix report, validate each note, and return substantive choices instead of deciding them in a worker.
  • Inherited choices: Parent scheduling, integration, recovery, and no unauthorized nested delegation.
  • Delegated choices: Minimal wording fixes and strategy classification within each note's exact warning set.
  • Exact controls retained: Queue freshness, disjoint note/report paths, sole writes, input scope, validation and return, anti-orchestration rule, substantive-decision stop, single-use lifecycle, and parent diff review.
  • Gap and disposition: None. Remove only the repeated generic lifecycle and no-delegation wording; the root rules arrive on both supported worker paths.

revise-note tag follow-up — retain exact

  • Baseline: Verified root doctrine plus the revision instruction and the current note/tag contracts.
  • Intent and acceptance: Reconcile only the exact navigation surfaces made stale by an already-authorized tag change.
  • Inherited choices: Parent lifecycle and no unauthorized nested delegation.
  • Delegated choices: A settled disjoint follow-up only when independent judgment or parallel capacity has a named benefit.
  • Exact controls retained: Parent-held rationale, exact affected paths and deltas, authority before external edits, validation, return, and stop.
  • Gap and disposition: None. The surface already treats delegation as a task-specific delta and does not repeat generic rules.

Standalone compression review — revise

  • Baseline: Verified root doctrine plus the complete packet assembled from the captured note text, four criteria, synthesis instruction, and output contract.
  • Intent and acceptance: Produce one disposable, report-only compression assessment over authoritative captured bytes.
  • Inherited choices: Parent lifecycle and no unauthorized nested delegation.
  • Delegated choices: Criterion application, cross-finding synthesis, and analytic route.
  • Exact controls retained: Criterion order, captured-byte authority, sole output, no live-note or review-store mutation, report grammar, fresh independence, and single-use closure.
  • Gap and disposition: None. Remove only the repeated or delegate phrase; the independent reviewer still has no delegation authority through root doctrine.

Standalone critique, friction, and premise checkers — retain exact

  • Baseline: Verified root doctrine plus the selected checker instruction and caller-supplied target/output destination.
  • Intent and acceptance: Produce routed adversarial attention without editing or accepting the note. Each method has its own exact result shape.
  • Inherited choices: Parent lifecycle and no unauthorized nested delegation.
  • Delegated choices: Objection, counterexample, and inferential-joint search within the method's test.
  • Exact controls retained: Fresh independence, report-only mutation, no-verdict boundary where applicable, skeptical defaults, output shape, and single-use lifecycle.
  • Gap and disposition: None. These controls create decorrelation and are not generic packet ceremony.

Full improvement pass — defer behind its state-machine audit

  • Baseline: Verified root doctrine plus the full-pass instruction, report schema, method siblings, generated review prompts, and deterministic guards.
  • Intent and acceptance: Select a warranted contribution, apply only a bounded keep edit, and close with exact review, one recovery, or rollback and author hand-back.
  • Inherited choices: Generic parent lifecycle and no unauthorized nested delegation.
  • Delegated choices: Independent method judgment and the bounded copyedit's local flow choices.
  • Exact controls retained: Captures and hashes, phase transitions, ownership lock, job grouping, author-owned claim boundary, copyeditor packet, closing suite, recovery count, guard, and rollback.
  • Gap and disposition: The known source-capture identity question spans the wrapper, every report method, concurrent live drift, schema, and recovery. Simplification must change that state machine as a unit. Keep bytes exact until its active coherence work supplies the decision and failure tests.

Review-job dispatch and generated prompt — retain exact

  • Baseline: Verified ambient root doctrine plus the generated prompt, which is the complete job-specific contract. Parent conversation is excluded.
  • Intent and acceptance: Produce one parser-valid block for every persisted review pair, then finalize atomically under the selected model partition.
  • Inherited choices: Ambient repository governance and generic parent lifecycle.
  • Delegated choices: Criterion judgment and permitted linked-artifact reads inside the pre-resolved scope.
  • Exact controls retained: Captured note/criterion bytes, job grouping, model-partition selection, exact output path, no other writes, reading scope, consumption telemetry, sentinels, result-kind grammar, single-use contexts, all-or-nothing finalization, and freshness verification.
  • Gap and disposition: None. ADR 067 deliberately makes the generated prompt self-contained and forbids a second dispatch wrapper. Its no-delegation line remains exact because prompt portability and one-file authority are part of that accepted protocol.

Agent-memory review drafting — defer behind a focused recovery design

  • Baseline: Verified root doctrine plus the local skill's complete drafting packet, review type contract, prepared checkout, and later generated semantic review prompts.
  • Intent and acceptance: Produce a code-grounded review at a pinned commit, then pass taxonomy QA, semantic QA, and final validation.
  • Inherited choices: Generic parent lifecycle outside the explicit packet.
  • Delegated choices: Code reading, trace-learning classification, review organization, and prose inside the type contract.
  • Exact controls retained: Anti-skill recursion, sole review write, read-only checkout, pinned citations, no shell-launched agent, worker-side validation, archive lifecycle, and parent QA.
  • Gap and disposition: The incumbent is archived before worker availability and a clean replacement are known. A drafting failure can therefore leave no active review. Fixing it requires a checked candidate/promotion or exact restore protocol, including staged rename recovery; removing duplicated packet text does not solve it. Retain bytes and defer until a focused failure-path design can exercise those states or the active agentic-system corpus migration replaces this workflow. The failure path is carried as the proposal "Recoverable replacement of an incumbent review" (added 2026-08-29), so the design question sat in the proposal frontier rather than only in this deferral. ADR 083 later retired that proposal.

AutoReason revision — defer behind versioned protocol evidence

  • Baseline: Verified root doctrine plus the experimental skill's complete role prompts and blind mapping.
  • Intent and acceptance: Compare incumbent, critic-led revision, and synthesis without changing the source until user approval, while routing substantive revision outside the prose tournament.
  • Inherited choices: Repository authority and generic parent lifecycle.
  • Delegated choices: Role-local critique, revision, synthesis, audit, and blind ranking.
  • Exact controls retained: Seven-role topology, per-role files, balanced mappings, parseable decisions/rankings, hard semantic checks, Borda rule, rerun bounds, pass budget, fallback incumbent, and user apply gate.
  • Gap and disposition: No current failure justifies a protocol version change. Compression would need isolation, malformed-return, rerun, and recovery tests. Retain the versioned algorithm unchanged.

Second Brain fresh-agent example — revise

  • Baseline: The fresh worker's ambient project doctrine plus two named task artifacts.
  • Intent and acceptance: Test whether a context summary and me.md let a fresh worker explain what the operator is trying to do.
  • Inherited choices: Ordinary repository governance and parent lifecycle.
  • Delegated choices: How to synthesize the answer from those two artifacts.
  • Exact controls retained: The read-only boundary, two-file task-evidence boundary, and question used as the observable check.
  • Gap and disposition: “Loaded with only” could be read as excluding the ambient project instructions the real harness always supplies. State the actual baseline and prohibit opening other task evidence; do not add a new evaluation framework to an illustrative recipe.

Non-surfaces and composition-only hits

  • AGENTS.md, AGENTS.md.template, kb/instructions/COLLECTION.md, kb/types/instruction.md, and kb/instructions/write-instruction.md govern or generate commissioning artifacts; they are not themselves task packets.
  • cp-skill-ground, re-ingest.md, cp-skill-connect, and semantic-QA callers invoke a sibling that owns any worker packet. They do not create a second commission.
  • Criterion bodies and review gates tell an already commissioned reviewer what judgment to make. Their caller owns identity, inputs, output, lifecycle, and result transport.
  • Runtime projections are symlinks to canonical skills. Tests inspect contracts but do not commission production workers.
  • One-off mailbox messages, workshop prompts, task instances, and retained generated reports are governed instances, not reusable producers. Their standing authoring contracts are recorded above.

Complexity account

The operative changes consolidate generic rules into one verified root carrier. They add no state, branch, handoff, exception, parser field, or runtime test. The warning and compression packets keep every task-specific control. The Second Brain edit adds one accurate consumption-path qualification to an existing evaluation example.

The five deferred cohorts preserve their current bytes. Each deferral names the evidence or design decision that would reopen it; none is disguised as a completed simplification.