Chunk 1 — Distiller quality gate

import { Aside } from ‘@astrojs/starlight/components’;

Gate passed on first prompt iteration using Sonnet 4.6 via `claude -p --bare`. Two-pass A/B test on a real 828-line planning transcript: final memory 34 lines, 3/3 blind facts present, hygiene test passed (a resolved open question was correctly dropped on the incremental pass). Safe to proceed to Chunk 2.

If a future change to the prompt or model regresses output quality, **do not proceed** past this chunk. Iterate the prompt, switch model, or pivot to handoff-only.

Goal

Standalone CLI that reads a transcript + old memory, produces a new memory that passes the blind A/B bar. Zero infra, no hooks, no git.

Files

File	Purpose
distiller.mjs	Standalone CLI (not wired to any hook yet)
prompts/distill.md	The distiller system prompt — iterate on this
lib/transcript.mjs	JSONL parse + since-watermark slicer
lib/memory.mjs	Parse/serialize markdown (whole-file for this chunk)

CLI interface

node distiller.mjs \
    --transcript ~/.claude/projects/<proj>/<session>.jsonl \
    --memory .wyren/memory.md \
    --since <uuid-or-empty> \
    --out .wyren/memory.new.md

Distiller prompt (first draft)

Living in prompts/distill.md. See the full annotated prompt in the reference section.

Core rules:

1. REPLACE superseded entries — never stack contradictions.
2. REMOVE resolved workarounds — delete when transcript shows the fix.
3. NEVER append blindly — every entry must still be load-bearing.
4. NO code snippets, NO conversation quotes — conclusions only.
5. Max 30 lines per section — cull stale entries aggressively.
6. Tag each entry with [session <id>, turn <n>] for provenance.

Guiding question for every entry: “Would a new Claude opening this project in 10 minutes genuinely need this, or is this noise?”

Claude call shape

Headless Claude Code with --bare to skip global hooks/plugins/auto-memory so the distiller sees only the prompt we hand it:

echo "$PROMPT

<session-id>$SESSION_ID</session-id>

<existing-memory>
$MEMORY
</existing-memory>

<transcript-slice>
$SLICE
</transcript-slice>" | claude -p --bare --no-session-persistence \
    --tools "" --output-format text \
    --model claude-sonnet-4-6 > memory.new.md

Why --bare: without it the subprocess would inherit the team’s CLAUDE.md, SessionStart hooks (caveman, superpowers), and auto-memory — all of which would pollute the distiller’s output. With --bare, only the piped prompt reaches the model.

Fallback: Anthropic SDK with @anthropic-ai/sdk (if claude CLI unavailable).

Verification (must pass before Chunk 2)

1. Find an existing 2+ hour session JSONL under ~/.claude/projects/. Ideally one with real decisions + dead-ends.
2. Run distiller fresh (empty memory) on first half of transcript. Inspect output.
3. Run distiller again (previous output as memory) on second half. Inspect output.
4. Blind test: read only the final memory.md. Name any 3 concrete facts that should be there and check they appear.
5. Hygiene test: introduce a workaround early in transcript, resolve it later. Verify it does NOT appear in final memory.

Exit criteria

• Two independent evaluators (or self-eval with 12h gap) agree output passes blind test.
• Memory stays under 60 lines on a 2-hour transcript (tight hygiene).
• Output is deterministic enough that small transcript tweaks don’t flip unrelated entries (stability check).

Results on the gate run

Test corpus: the 828-line planning session for Wyren itself — real decisions, real rejected paths, real Windows-specific workaround discovered mid-session.

Pass	Input	Output	Time
1	Lines 0-414, empty memory	27-line memory: 9 decisions, 4 rejected paths, 2 live workarounds, 1 open question	~35s
2	Lines 415-828, pass 1 as memory	34-line memory: 12 decisions, 5 rejected paths, 5 live workarounds, 0 open questions	~80s
2b	Same as pass 2 (stability)	Same semantic content, phrasing varies	~77s

Blind test (3 concrete facts a fresh teammate would need):

• Docs site is deployed at https://ssm-08.github.io/wyren/ — ✅ captured.
• Custom rehype-mermaid-pre.mjs replaces rehype-mermaid (the latter pulled playwright) — ✅ captured.
• Windows Git Bash needs MSYS_NO_PATHCONV=1 when setting WYREN_BASE locally — ✅ captured (non-obvious gotcha worth the whole feature).

Hygiene test: pass 1 recorded “open question: .gitignore content unconfirmed” at turn 265. The second half of the transcript resolved this. Pass 2 correctly removed the open question. Self-cleaning under load. ✅

Stability (pass 2 vs 2b on identical inputs): same section structure, same decisions preserved, same rejected paths, same workarounds. Wording and turn-number estimates drift slightly (Sonnet non-determinism) but no content flips. ✅

If this chunk fails

Options in order of preference:

1. Iterate prompt. Most likely fix. Add more explicit negative examples, tighten hygiene instructions.
2. Switch model. Try Opus 4.7 if Sonnet is weak on hygiene.
3. Pivot to handoff-only. Ship without auto-distiller: /wyren-handoff becomes the only write path, humans author memory verbatim. Less magic, still useful, still demoable.

Do NOT proceed to Chunk 2 until one of these passes.