---

type: adr id: adr-0088 status: accepted created: 2026-06-20 updated: 2026-06-20

ADR-0088 — Delegation provenance: a reviewable trace when an agent delegates to a subagent (neutral contract, runner-specific producers, fact-not-verdict)

Context

corpus's provenance and reviewability spine — ADR-0076 (worker provenance) and ADR-0077 (reconcile-only) — covers human/lead → worker handoffs recorded in markdown, but not agent → subagent delegation: a main agent spawning a Claude Code subagent, or handing off through the OpenAI Agents SDK. That delegation leaves no reviewable trace today — who was delegated to, why, with what inputs, what context was filtered, what tools it held, whether it could edit, what evidence it returned. Two gaps: corpus run records only the launch envelope ({task_id, adapter, worktree, branch, source, exit}; the changed_files[]/commands[] stream is the deferred D1 item), and in-session subagents — spawned by the parent through the runner's own Agent tool — never touch the CLI at all.

corpus-works #43 (the corpus-agents exploration) identified a reviewable delegation trace as the genuinely-new, agent-neutral contribution a corpus-agents member would make. RFC-delegation-provenance argued the approach; this ADR freezes the contract.

Decision

1. A delegation-provenance trace is a neutral, frozen contract. Per agent → subagent handoff, a trace records exactly these fields, extending ADR-0076's Provenance line and the reserved run-record form (docs/reference/future-cli.md):

   • worker — the subagent / role identity (the adapter name, or the subagent definition's name).
   • reason — why it was delegated to (the task or instruction it received).
   • inputs — what it was given (the prompt / task; references it was pointed at).
   • filtered — what history / context was withheld (fresh-context, inherited, or a description).
   • tools — the tools it was granted.
   • could_edit — whether it could write / edit source (true / false).
   • evidence — what it returned (the result digest or an artifact reference).
   • started / finished / exit — timing and outcome, where the producer knows them.

2. A trace is a record, never a verdict (ADR-0077 Decision 8). No field carries Pass/Fail/Unverified/Blocked. A missing or thin trace is a human-attention fact, not a corruption — the human still owns the review result. Level: convention (ADR-0063) this cycle — no checker mints it.

3. The contract is neutral; producers are runner-specific. Two producers emit conformant data for the surface each can see:

   • Producer 1 — corpus run --agent launches. corpus-cli extends its run-record toward the D1 form: changed_files[] (the worktree diff after the agent exits) plus a provenance block carrying the fields it knows (worker/adapter, reason/task, worktree, isolation, exit). Additive — old run-records stay valid; no field becomes required. commands[] (the agent's internal commands) the interactive launcher cannot observe; it stays deferred.
   • Producer 2 — in-session subagents. A Claude Code SubagentStart / SubagentStop hook recipe appends a trace line to .corpus/work/delegations.ndjson — because in-session subagents bypass corpus run entirely. Runner-specific (Claude-Code-first); shipped as an opt-in recipe in the kit's advanced/, lower-staleness than agent frontmatter (a shell hook keyed to a stable event name).

4. The trace audits delegation; it never runs it. The producers ship records, never an executor. The only sanctioned launcher remains corpus run --agent, which never becomes the agent (ADR-0077). A launched run that bypassed both producers leaves no trace — so the absence of orchestration is observable. This is the structural defense against a hidden multi-agent runtime.

5. Convention-first, not a check (yet). No C0xx mint and no contract-version bump this cycle (the trace is a record, not a spec/review structure corpus check reconciles). A check ("a launched run recorded a provenance trace") is deferred to a separate decision, minted only if review keeps catching missing traces — the ADR-0063 anti-enforcement-theater rule and the ADR-0086/0087 measure-first precedent.

Alternatives considered

Alternative	Why weaker
Do nothing (status quo)	Agent → subagent delegation stays unreviewable; the provenance/reviewability spine degrades as multi-agent work grows — the exact gap #43 found.
A multi-agent runtime / orchestrator that owns delegation	Violates ADR-0077 (no model loop, no center) and breaks the constellation (members depend on a hub). The trace audits delegation; it never runs it.
Provenance only in corpus run's record (no hook producer)	Misses in-session subagents entirely — the majority of Claude-Code delegation never touches the CLI. The two-producer split is what makes the contract complete.
Mint a check (C0xx) requiring a trace now	Premature enforcement theater (ADR-0063); start as a convention/record and measure first (the ADR-0086/0087 precedent).

Consequences

Accepted: the delegation-provenance contract is canon — a neutral, verdict-free trace with a frozen schema and two runner-specific producers. Honors ADR-0077 D8 (a fact, never a verdict), ADR-0063 (convention this cycle; toolable/enforced only on a later measured decision), and extends ADR-0076 plus the future-cli.md run-record form. The producers — the corpus-cli run-record D1 extension and the Claude Code hook recipe — land per CHANGE-delegation-provenance; the experimental corpus-reviewer probe is the first subject to measure. Nothing is enforced — a missing trace never blocks anything — until a future gate decides otherwise.

Propagation

docs/reference/future-cli.md (the run-record provenance block references this contract) · docs/adrs/README.md (the index row) · corpus-cli (corpus run run-record extension — producer 1, CHANGE-delegation-provenance W2) · ../corpus-starter-kit/advanced/ (the Claude Code hook recipe — producer 2, W3). No checks.yaml change (convention-first). The corpus-works workspace cuts and reviews the producer tasks.