os-me Store: learning loop scores through bin/os-me-record and learns only from discussed sessions #1078

New issue

Closed

opened 2026-10-17 11:54:37 +00:00 by jared · 1 comment

jared commented

2026-10-17 11:54:37 +00:00

Owner

Problem

Step 6 of loop.md has the LLM hand-write log rows. No rule separates an accepted-verbatim hit from a discussed one. Autopilot accepts should feed the trust metric only.

Reproduce

n/a: rule change in the Store. No behavior to reproduce.

Approach

Required work:

Rewrite step 6 of ~/dev/os-me-store/learning/loop.md. Each scored prediction is written with bin/os-me-record score ID --verdict ... from the plugin. Never edit the JSONL by hand.
Add the gate to loop.md. A session is a rule-edit candidate only when a record has depth discussed or verdict miss. A record with autopilot true counts for trust only. A record with origin agent and depth accepted-verbatim counts for trust only.
Change the header of review-log.md. It receives no new rows. The header points to bin/os-me-report.
Leave token-log.md unchanged.
Update ENTRY.md wherever it names the review log.
Run the Store bin/test.

Orchestrator work: make these edits by hand, or through an agent with Store write access. Do not use the os-sdlc pipeline. Jared ruled that Store rules need no ADR.

Non-goals

Plugin code.

Blocked by #1074

Origin

Trigger: Jared asked for this ticket on 2026-10-17 while designing os-me prediction records.
Improvised this session: none
Chain: learning gate absent ← loop step 6 hand-writes rows and has no depth rule. DESIGN (~/dev/os-me-store/learning/loop.md)
Root candidate: this ticket
Where: ~/dev/os-me-store/learning/loop.md
Session: 2fc707d3-1acb-4f67-98d1-3685e5283f89
Transcript: /home/jared/.claude/projects/-home-jared-dev-cc-os/2fc707d3-1acb-4f67-98d1-3685e5283f89.jsonl

## Problem Step 6 of loop.md has the LLM hand-write log rows. No rule separates an accepted-verbatim hit from a discussed one. Autopilot accepts should feed the trust metric only. ## Reproduce n/a: rule change in the Store. No behavior to reproduce. ## Approach Required work: 1. Rewrite step 6 of ~/dev/os-me-store/learning/loop.md. Each scored prediction is written with `bin/os-me-record score ID --verdict ...` from the plugin. Never edit the JSONL by hand. 2. Add the gate to loop.md. A session is a rule-edit candidate only when a record has depth discussed or verdict miss. A record with autopilot true counts for trust only. A record with origin agent and depth accepted-verbatim counts for trust only. 3. Change the header of review-log.md. It receives no new rows. The header points to bin/os-me-report. 4. Leave token-log.md unchanged. 5. Update ENTRY.md wherever it names the review log. 6. Run the Store bin/test. Orchestrator work: make these edits by hand, or through an agent with Store write access. Do not use the os-sdlc pipeline. Jared ruled that Store rules need no ADR. ## Non-goals - Plugin code. Blocked by #1074 ## Origin - Trigger: Jared asked for this ticket on 2026-10-17 while designing os-me prediction records. - Improvised this session: none - Chain: learning gate absent ← loop step 6 hand-writes rows and has no depth rule. DESIGN (~/dev/os-me-store/learning/loop.md) - Root candidate: this ticket - Where: ~/dev/os-me-store/learning/loop.md - Session: 2fc707d3-1acb-4f67-98d1-3685e5283f89 - Transcript: /home/jared/.claude/projects/-home-jared-dev-cc-os/2fc707d3-1acb-4f67-98d1-3685e5283f89.jsonl

jared added the

project/cc-os

filed-by/user

category/prose-artifact

labels

2026-10-17 11:54:37 +00:00

jared referenced this issue

2026-10-17 12:31:36 +00:00

os-me: prediction records carry triggers, verdict, origin, and autopilot #1081

jared commented

2026-10-17 14:16:47 +00:00

Author

Owner

Resolution

Done: learning/loop.md step 6 scores each prediction with bundle exec bin/os-me-eval score-record ID --verdict hit|partial|miss (the ticket named bin/os-me-record, which does not exist) and forbids hand edits of prediction-log.jsonl; step 4 holds the rule-edit gate (depth discussed or verdict miss; autopilot true and agent+accepted-verbatim count for trust only) and no longer writes review-log rows; review-log.md header freezes the log and points to bin/os-me-report; token-log.md unchanged; ENTRY.md never names the review log, so it is unchanged.

Evidence: os-me-store commits ebe80a1 and 8e8c4bb (not pushed); Store bin/test 19 runs green. Related: Store queue fix 45b98b2 (Stop markers from #1075 no longer listed as invocations; null transcript_path no longer crashes bin/queue next); cc-os 9e57b73d lets score-record accept --verdict partial, which the review log, backfill and report already use.

Follow-ups: none

## Resolution **Done:** learning/loop.md step 6 scores each prediction with bundle exec bin/os-me-eval score-record ID --verdict hit|partial|miss (the ticket named bin/os-me-record, which does not exist) and forbids hand edits of prediction-log.jsonl; step 4 holds the rule-edit gate (depth discussed or verdict miss; autopilot true and agent+accepted-verbatim count for trust only) and no longer writes review-log rows; review-log.md header freezes the log and points to bin/os-me-report; token-log.md unchanged; ENTRY.md never names the review log, so it is unchanged. **Evidence:** os-me-store commits ebe80a1 and 8e8c4bb (not pushed); Store bin/test 19 runs green. Related: Store queue fix 45b98b2 (Stop markers from #1075 no longer listed as invocations; null transcript_path no longer crashes bin/queue next); cc-os 9e57b73d lets score-record accept --verdict partial, which the review log, backfill and report already use. **Follow-ups:** none

jared closed this issue

2026-10-17 14:16:48 +00:00