Skip to content

fix(learning): a session log named by its file alone is found - #42

Merged
DEVfancybear merged 3 commits into
masterfrom
fix/log-signal-name
Nov 21, 2026
Merged

DEVfancybear merged 3 commits into
masterfrom
fix/log-signal-name

Conversation

@DEVfancybear

Copy link
Copy Markdown
Owner

Why

Live dream.sh on the released 0.9.16 ha.exe (sha256 dda5154c…) failed section 5 (6 checks): the meditation is shown each log as <log name=".harness/sessions/...md">, the model named both logs by their file alone (2026-11-06-15_30-parser-edge-cases.md), and the host refused every signal with "not a session log the meditation read". The code is unchanged since 0.9.15, where the model happened to give the full path.

Change

named_log in lessons.rs: the log of that exact name (with \ read as /), else the one log whose file is that name; a file name two logs share names neither. Docs (LEARNING_LOOP, MEMORY_STORE, en + vi) say so.

Tests

  • New a_log_named_by_its_file_alone_is_found_when_one_log_has_it (bare name, \, shared name, unknown, empty, and through log_signals); the_meditation_finds_signals_in_the_session_logs still passes.
  • Local gate and live dream.sh on the candidate are running; results will follow here.

Based on fix/g13-clippy (#41) so the Rust jobs are not red from the g13 clippy error on master.

🤖 Generated with Claude Code

The meditation is shown each log as `<log name=".harness/sessions/...md">`.
On 0.9.16 the model named both logs by their file alone and the host refused
every signal from them ("not a session log the meditation read"), so
dream.sh section 5 failed. The host now takes the log of that exact name
(with `\` read as `/`), and else the one log whose file is that name; a file
name two logs share names neither.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@DEVfancybear

Copy link
Copy Markdown
Owner Author

Evidence on 465c6b4

Gate (fmt, clippy -D warnings, candidate, workspace tests): clean; 2690 passed, 0 failed.

Live, openai-codex/gpt-6-luna, thinking max:

  • candidate ha.exe sha256 74fc1428958660496bd96149c40011a032c2b7e10103e6534822a0dc7c432b6f: dream.sh RESULT: 22 passed, 0 failed (section 5: both logs' signals kept, les_justtest promoted, 3 signals from 3 conversations).
  • control, released 0.9.16 ha.exe sha256 dda5154c7bc848da91312f4b6efafd90cbf877a05b4dd1ffe93b8bc9b021dda1: dream.sh RESULT: 16 passed, 6 failed (all section 5: both log signals refused, "not a session log the meditation read"); memory-safety.sh 22/0, memory.sh 44/0.

Which form of the name the model gives is its choice per run; the diagnostics record the log's own name, so this live run does not show which form it used. The bare-name path itself is covered by the new unit test.

🤖 Generated with Claude Code

…he log

`learning log_signal` now names the log by its own name and keeps what the
meditation wrote in `named` when that was otherwise, so a live run shows
whether a log named by its file alone was found. dream.sh prints how many
logs were named so and checks that no seeded log was refused for being named
by its file (0 kept / 2 refused on the 0.9.16 evidence, 0 / 0 on the fix).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@DEVfancybear

Copy link
Copy Markdown
Owner Author

Live proof of the bare-name path, on 534a08a

The run on 465c6b4 did not exercise it: the model named both logs by their path. 534a08a records how the meditation named a log (named on learning log_signal) and dream.sh checks that no seeded log is refused for being named by its file alone.

Gate on 534a08a: fmt, clippy, candidate, workspace tests clean (2690 passed, 0 failed).

Live dream.sh, candidate ha.exe sha256 693509043e4b98d3478cf327cded522a8e9a8b55008dde426345f33231a51ddd, openai-codex/gpt-6-luna thinking max, run until the model named a log by its file:

  • runs 1-4: RESULT 23 passed, 0 failed each; model named both logs by their path (0 named by file).
  • run 5: RESULT 23 passed, 0 failed; both logs named by their file alone and kept:
    • log: .harness/sessions/2026-11-06-15_30-parser-edge-cases.md, named: 2026-11-06-15_30-parser-edge-cases.md
    • log: .harness/sessions/2026-11-05-10_00-fixture-test-failures.md, named: 2026-11-05-10_00-fixture-test-failures.md
    • les_justtest promoted, 3 signals from 3 conversations.

Control: the new check on the released 0.9.16 run's evidence counts 2 seeded logs refused for their name (it fails there); 0 on the fix.

Evidence: F:/ha-live/fix0917b/run1..run5.

🤖 Generated with Claude Code

P0-P6 now take 19-28 minutes on windows-latest and P7 was canceled at its
30-minute limit on master (a23d6a8) and twice on this branch.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@DEVfancybear
DEVfancybear merged commit 0508322 into master Nov 21, 2026
35 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant