Drew T
2.4 KiB
§411 ★★★ — THE PACK MUST CARRY THAT FUNCTION'S OWN HISTORY (P31 S71; measured 38/39 vs 124/131)
The defect, stated as a pipeline fact. Every drafting agent ends its run by writing a note — what
it tried, what it measured INERT, which lever moved the residual, where its draft sits on disk — into
subagents/workflows/wf_*/journal.jsonl. For ~30 phases nothing read them back. A function that
refused a wave went straight back into the pool, and the next wave's agent met it with no idea that
three predecessors had already burned the same four levers, or that one had left a MATCHing body on
disk. The write side felt complete because the notes were being saved.
The fix is one pack step. tools/journal_notes.py mines the journals per (binary, fn) and
appends PAST ATTEMPTS ON THIS EXACT FUNCTION; claude_wave_packs.py calls it at the end of pack
generation, so it is the default and not a step to remember. Idempotent (never appends twice) and
R48-safe: a note stamped with a DIFFERENT binary is never served, because the same name is another
function in another overlay (§238).
Measured on S71 wave 1 — 50 one-agent workflows over the 210-function real frontier, where every target had already refused at least one earlier wave:
| with notes (S71) | without (S70) | |
|---|---|---|
| MATCH at closeness 0 | 38 / 39 (97.4%) | 124 / 131 (94.7%) |
| pool | the hardest residue | strictly easier |
| cite a prior attempt as what they used | 29 / 39 (74%) | n/a |
| banked by RECOVERING an existing MATCH body | 4 / 39 (10%) | 0 |
| matched on the FIRST compile | 11 / 39 (28%) | n/a |
The two failure modes it removes are the expensive ones.
- Re-testing a measured-inert lever. §406 lists twelve, §407 fifteen, §410 four — every one of those lists was paid for by an agent, and is worthless to the next agent who never sees it.
- Re-deriving a body that already exists.
ov_SC05_003/func_80181720's pack shipped a FAILED warm-start while attempt 3's MATCH sat under.run/S70x_1/opus. The agent found it because the note named the path.
The general law, past decomp. If your agents produce a per-item report, the next agent on that item must be handed it. An archive of your own verified outcomes is training data for your own tooling — and a corpus you write but never read is indistinguishable, from the outside, from one you never wrote.