feat(image): add NDJSON feedback log (MK-5 Layer 3, feedback half) #27

Merged

nrupard merged 3 commits from feat/image-feedback-log-MK-5 into main 2026-05-21 17:34:33 +02:00

Conversation 0 Commits 3 Files changed 6 +488 -34

nrupard commented 2026-05-21 17:17:23 +02:00

Member

Copy link

What

Feedback-log half of MK-5 Layer 3. The external tuner is deferred to its own follow-up issue/PR per the umbrella plan (it only earns its keep once dozens of rated runs exist).

• monkey image auto --record runs.jsonl appends one NDJSON line per run: ts_unix_ms, input blake3, class, recipe sha, steps, output blake3, duration_ms.
• monkey image rate <input> --score 1..5 [--notes "..."] --record runs.jsonl hashes the input and appends a verdict keyed by the same blake3, so a rating joins its run.

Design choices (confirmed before building)

• Scope: feedback log only; tuner is a separate follow-up.
• Rate command lives under image (monkey image rate), consistent with classify/auto.
• --record requires an explicit path. No default, no XDG_DATA writes, no hidden state.

Notes

• The log is append-only NDJSON: grep-able, git-able, and a partial write loses at most the trailing line.
• Recording is unified across the filtered and the byte-copy passthrough paths, so an unknown passthrough records an output hash equal to its input hash.
• recipe_sha is the blake3 of the recipe source text, so the exact recipe version (including a --recipe-dir override) is captured. Recipe::load now returns that source; recipe::Step is pub + Serialize.
• New deps: blake3, serde_json.

Testing

just pre-commit green (fmt, clippy -D warnings, build, 72 tests). New unit tests cover hash stability, score-range rejection, and one-NDJSON-line-per-record with notes omitted when absent. Manually verified in the builder image: auto --record then rate produce two joined records; an unknown passthrough records output == input; --score 9 errors with exit 1.

Closes MK-5 Layer 3 (feedback log). Remaining: the tuner + recipe promote (separate follow-up).

🤖 Generated with Claude Code

## What Feedback-log half of **MK-5 Layer 3**. The external tuner is deferred to its own follow-up issue/PR per the umbrella plan (it only earns its keep once dozens of rated runs exist). - `monkey image auto --record runs.jsonl` appends one NDJSON line per run: `ts_unix_ms`, input blake3, class, recipe sha, steps, output blake3, `duration_ms`. - `monkey image rate <input> --score 1..5 [--notes "..."] --record runs.jsonl` hashes the input and appends a verdict keyed by the same blake3, so a rating joins its run. ## Design choices (confirmed before building) - **Scope:** feedback log only; tuner is a separate follow-up. - **Rate command** lives under `image` (`monkey image rate`), consistent with `classify`/`auto`. - **`--record`** requires an explicit path. No default, no `XDG_DATA` writes, no hidden state. ## Notes - The log is append-only NDJSON: grep-able, git-able, and a partial write loses at most the trailing line. - Recording is unified across the filtered and the byte-copy passthrough paths, so an `unknown` passthrough records an output hash equal to its input hash. - `recipe_sha` is the blake3 of the recipe source text, so the exact recipe version (including a `--recipe-dir` override) is captured. `Recipe::load` now returns that source; `recipe::Step` is `pub` + `Serialize`. - New deps: `blake3`, `serde_json`. ## Testing `just pre-commit` green (fmt, clippy `-D warnings`, build, 72 tests). New unit tests cover hash stability, score-range rejection, and one-NDJSON-line-per-record with `notes` omitted when absent. Manually verified in the builder image: `auto --record` then `rate` produce two joined records; an `unknown` passthrough records `output == input`; `--score 9` errors with exit 1. Closes MK-5 Layer 3 (feedback log). Remaining: the tuner + `recipe promote` (separate follow-up). 🤖 Generated with [Claude Code](https://claude.com/claude-code)

nrupard added 1 commit 2026-05-21 17:17:23 +02:00

feat(image): add NDJSON feedback log (MK-5 Layer 3, feedback half) ... 7d468d5da2

Add the feedback-log half of Layer 3. The external tuner is deferred to its own follow-up, since per the design it only earns its keep once dozens of rated runs exist.

`monkey image auto --record runs.jsonl` appends one NDJSON line per run: timestamp, input blake3, resolved class, recipe sha (blake3 of the recipe source), the applied steps, output blake3, and duration. `--record` takes an explicit path; there is no default location and no hidden state. Recording is unified across both the filtered and the byte-copy passthrough paths, so an `unknown` passthrough logs an output hash equal to its input hash.

`monkey image rate <input> --score 1..5 [--notes ...] --record runs.jsonl` hashes the input and appends a verdict keyed by that same blake3, so ratings join their runs. Out-of-range scores are rejected.

The log is append-only newline-delimited JSON: grep-able, git-able, and a partial write loses at most the trailing line rather than corrupting the file. New deps: `blake3` (hashing) and `serde_json` (NDJSON). `recipe::Step` is now public and `Serialize`, and `Recipe::load` returns the recipe source text so the exact recipe version can be hashed.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

nrupard added 1 commit 2026-05-21 17:27:25 +02:00

fix(image): hash canonical recipe steps for recipe_sha; honest durability doc ... ec91012f35

Address review of the feedback log:

`recipe_sha` now hashes the recipe's parsed, serialized steps (`Recipe::sha`) instead of the raw source text. Two recipes that differ only in comments, whitespace, or key order now produce the same sha. This matters because MK-10's tuner groups runs by class + recipe_sha; hashing source text would have let a cosmetic edit fragment a class's rated samples across two shas. `Recipe::load` reverts to returning just the recipe (the source-text plumbing is gone).

Correct the durability claim. The log doc and README previously implied a partial write "loses at most the trailing line". A single `O_APPEND` `write_all` is not crash-atomic: a process killed mid-write can leave a partial line that the next append concatenates onto. The docs now say durability is best-effort and that readers must skip a malformed trailing line rather than assume every line parses.

Adds a test that two source-formatting variants of the same steps hash equal.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

nrupard added 1 commit 2026-05-21 17:31:07 +02:00

fix(image): set run kind internally, warn on orphan rating, stream hash, self-cleaning test tmpfiles ... 6faff5b0c1

Clean up the remaining review items on the feedback log:

`RunRecord` fields are now private and built via `RunRecord::new`, which stamps `kind = "run"` internally. A caller can no longer set the wrong discriminator, mirroring how `RatingRecord` is constructed.

`monkey image rate` now warns (on stderr, without failing) when the log holds no `run` record matching the input hash, so a rating for an unrecorded or mis-pathed input is flagged rather than silently orphaned. The lookup skips malformed lines, matching the log's documented skip-bad-trailing-line contract.

`hash_file` streams the file through `blake3::Hasher` via `io::copy` instead of reading it fully into memory, so a large input (or a future video record) is not buffered.

Test temp files now use a self-deleting `TmpFile` drop guard, so a panicking assertion no longer leaves stale files in the tempdir. Adds tests for streamed hashing and the run-lookup helper.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

nrupard merged commit 25fdbfa8f1 into main 2026-05-21 17:34:33 +02:00

nrupard deleted branch feat/image-feedback-log-MK-5 2026-05-21 17:34:33 +02:00

nrupard referenced this pull request from a commit 2026-05-21 17:34:35 +02:00

Merge pull request 'feat(image): add NDJSON feedback log (MK-5 Layer 3, feedback half)' (#27) from feat/image-feedback-log-MK-5 into main

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

pandoras-box/monkey!27

No description provided.