ADR 0132: A goal keeps a decision journal

Status: accepted (James, 2026-08-26). Extends ADR 0130 and ADR 0107.

Context

Goal pursuit runs across many autonomous continuations and self-wakes. The conversation's pi trees and lane logs record what Clankie said and did, but retention prunes conversation directories, and neither trail records what he decided and why while working a goal — which approach he picked, what he ruled out, what he discarded. A woken turn re-derives choices the last turn already made, and the owner auditing a finished goal has only the transcript, if it still exists.

Decision

Each operator conversation keeps an append-only decision journal beside autonomy.json, one JSONL file per conversation under ~/.clankie/captain/goal-journal/. A note_goal_decision tool appends one entry — decision, why, optional evidence pointer — host-stamped with the time, the owning goal's createdAt, and whether the turn was autonomous. get_goal returns the current goal's recent entries, so a continuation or wake starts from what was already decided instead of re-deriving it.

The journal is Clankie's own account, written by volition: the instructions ask for a line at real choices, not a protocol that demands one per turn. Entries survive goal completion and clearing — the journal is the postmortem — and recentDecisions filters to the current goal by its createdAt, so an old goal's entries never masquerade as the new goal's context. The one-writer rule holds: only the autonomy store touches the journal files.

Alternatives considered

• Episodes (remember_episode): decisions during goal work are status-shaped, and the episode ring is deliberately small personality memory; routing goal bookkeeping there would evict what the ring exists to keep.
• A host-derived trail: turn-settled.jsonl already records tool shapes, but "what he decided and why" is model-authored content no host stamp can reconstruct.
• A cross-model audit of the trail: a second model grading the journal adds a protocol and a verifier identity before there is evidence the owner's own read of the journal is not enough.
• A TSV committed into a repository: goals are not repo-scoped; many touch no repository at all.

Consequences

• A goal run can be audited and resumed from its own trail after the conversation directory is pruned.
• get_goal output grows by the recent journal tail; the full file stays on disk for trace-clankie reads.
• The TUI /goal view does not yet render the journal; the operator reads it from disk or through get_goal. That surface can follow when wanted.
• An empty journal on a finished goal is itself a signal worth asking about.