yysworld
ADR-041·Accepted·2026-05-03

Pipeline v0.2: Minimal Context Loading and Carry-Forward Discipline

Status: Accepted Date: 2026-05-03 Affects: /pipeline-go skill, pipeline/lib/files.ts, docs/pipeline/schemas.md, package_id written into all snapshots/artifacts/comparisons going forward

Context

Each /pipeline-go run was loading more files than the prior day:

  • runs/{rootId}/snapshots/*.json (one wildcard, growing N per run)
  • runs/{rootId}/artifacts/*.json (same)
  • runs/{rootId}/inbox/queue/*.json (same)

The skill described these as "recent" but did not bound the recency. As a run grew from day 3 toward the typical monthly horizon set by ADR-003, the per-day token cost grew with it — even though the only state actually needed to write day N+1 lives in two files: the branch file (current state + identity_notes carryover) and yesterday's artifact (last day's prose, for voice continuity).

Older snapshots are an audit trail, not a context input. Older artifacts have already been distilled into the branch file's identity_notes. Older events are settled — yesterday's event-anchor lives in yesterday's artifact's world_anchor. Re-reading any of these every morning relitigates settled history and inflates token burn linearly with run length.

Decision

Bump the pipeline package to v0.2 and tighten three things:

  1. Step 2 loads a fixed, capped set of files. Always: baseline + active branch files. Per-branch, capped at one prior day: yesterday's artifact only. Optional, only if it exists for today: inbox/{targetDate}.json (no glob fallback). Everything else — snapshots/, older artifacts/, events/, comparisons/, decisions/ — is not loaded unless the user explicitly asks. Look-back hard cap: 1 day by default, never more than 2 prior days without explicit user direction.
  1. Step 10 enforces carry-forward discipline. The next day's run will only see branches/{branchId}.json + this day's artifact. So:

- state_after.identity_notes must be 2–4 lines: yesterday's tail beat, today's beat, any standing thread that outlives today. - state_after.inventory and active_burdens are the durable record of anything that persists past today. - The artifact's narrative is the only prose tomorrow's run sees. Voice cues that should propagate must appear there.

  1. `package_id` becomes `yysworld-pipeline-v0.2`. New snapshots, artifacts, and comparisons stamp v0.2; existing v0.1 files are not rewritten (immutable truth, invariant 2). The bump exists so the author can roll back or re-tighten in v0.3 if quality drops, and so the provenance trail makes the change visible.

PIPELINE_VERSION and PACKAGE_ID constants in pipeline/lib/files.ts move to 0.2 / yysworld-pipeline-v0.2. The schema reference in docs/pipeline/schemas.md and the schema block in pipeline-go.md follow.

Why Not Leave As-Is

The cost of unbounded context loading is not theoretical. By day 30 of a monthly root, a wildcard read of snapshots/ and artifacts/ is roughly 60 files of dense JSON, the largest of which contain multi-paragraph narrative prose. Most of that content cannot inform tomorrow's writing — narrative continuity is anchored by yesterday alone, and state continuity is already captured in the branch file. The pipeline was paying for re-reading content that had already been compressed into the branch state.

The carry-forward discipline is the other half of the change. If identity_notes is allowed to drift toward minimal stubs, then tomorrow's run lacks the context it needs and has to reach further back — re-introducing the cost we just removed. Pinning the discipline on the writer side (today's run) keeps the reader side (tomorrow's run) cheap.

What This Is Not

  • Not a schema break. The shape of every file is unchanged. Only package_ref.package_id changes for newly written files. Tests do not pin the old value.
  • Not a reduction in canon integrity or audit. All snapshots, comparisons, decisions, and events continue to be written and committed exactly as before. They remain available for explicit lookup, debugging, post-hoc analysis, or museum exports — they are simply not loaded into normal pipeline runs.
  • Not a change to the proposal/confirmation discipline. Steps 5–7 still pause for author input. The token savings come from what the pipeline reads, not from what the author confirms.

Rollback

If quality drops on v0.2, the change is reversible:

  • Set PACKAGE_ID = 'yysworld-pipeline-v0.1' and PIPELINE_VERSION = '0.1' in pipeline/lib/files.ts.
  • Restore Step 2's wildcard reads in .claude/commands/pipeline-go.md.
  • Roll the schema-doc reference back to v0.1.

Files written under v0.2 keep their v0.2 stamp — they are truth, not configuration. The package_id becomes the in-the-record signal of when the tightened context regime was active.