diff --git a/.printing-press.json b/.printing-press.json index 47a525dc0..bced47265 100644 --- a/.printing-press.json +++ b/.printing-press.json @@ -1,7 +1,7 @@ { "api_name": "local-offload", "cli_name": "local-offload", - "version": "0.231.0", + "version": "0.232.0", "printing_press_version": "hand-built-2026-06-12", "generated_at": "2026-06-12", "description": "Local offload harness - delegate grunt work (summarize/classify/extract/triage, plus vision, OCR, transcription and media generation) to a free local Gemma-family cascade via llama.cpp; MCP + CLI; the cascade never calls cloud and defers to the caller instead. The opt-in offload_nim tool is the only remote model surface; the opt-in offload_browse lane drives the operator's own browser with typed choices from a loopback decision endpoint the operator runs.", diff --git a/CHANGELOG.md b/CHANGELOG.md index d281ccf0e..3c1398ead 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,18 @@ Releases before 0.190.0 are archived verbatim under [docs/changelog/](docs/chang ## [Unreleased] +## [0.232.0] - 2026-11-12 - the self-improvement instruments report what happened: trace-corpus guards, failure classes, selftest could-not-judge + +### Fixed - the self-improvement instruments report what happened (trace-corpus guards, failure classes, selftest could-not-judge) + +- **`trace-corpus`.** `freeze` creates its baseline exclusively and refuses a path that exists (exit 2, the way `gate freeze` refuses one), where it used to replace it silently. `harvest --since` and `--until` must be real `YYYY-MM-DD` dates with since not after until, and are refused with exit 2 otherwise: a typo such as `2026-11-1` used to keep the wrong shards, and `11/01/2026` kept all of them, both with exit 0. `help`, `--help` and `-h` print the usage with the flags of every mode and exit 0 (they exited 1 with no flag list). +- **Failure classes, written.** The remote lanes (vision, text, speech, media, compose), the composite image row and the accelerator forward now stamp the reason code from the defer class the result carries (the error class and the harness's own sentence first, the class when neither names the row). Their capacity and placement defers were filed `other`: 154 of the 157 `other` rows in a 30-day register. A deferred row whose writer gave no reason is filed with the reason `deferred: no reason recorded` instead of an empty string. +- **Failure classes, read.** Rows written before the stamp are classed from sentences the harness itself writes and that map without doubt to one existing class: `wall timeout after`, `step budget exhausted` (budget), `output failed schema` (abstention), `failed verification:`, `seat down:`, `queue deadline after` (queue_deadline, or queue_withdrawn when the node confirmed the withdrawal), `stalled: no progress` (the stall class of the phase), the lane's `no fleet node ...` placement sentences (no_eligible_node) and a dispatch refused with 503 or 429 (queue_full). A deferred row already stamped `other` is read through the same rules (`other` is the stamp for "no member named it", so it is the one stamp read again); every other stamp is read as stamped. Ambiguous sentences stay `other` and are listed in the register page. +- **`gate selftest`.** A scorer whose seat call deferred or failed (a GPU lease held, a seat down) could not judge: the verb exits 2, writes no receipt, leaves an existing `--out` file alone and names the defer reason. It used to report a mis-ranked scorer with exit 1 and a `passed: false` receipt that replaced an older `--out`. The A/B (`compaction-eval ab`) still scores such a call 0. +- **Small fixes.** The top-level `ledger` usage line lists `--failures`, `--table` and `--include-test-traffic`; the `recipe-audit` help ends its exit-code block and its usage line lists `--config`; `.gitattributes` forces LF on `internal/eval/*.json`, so the embedded gate schema, and the build that prints it, no longer depend on the builder's line-ending setting. +- **Docs.** The self-improvement, measurement-gate and failure-register pages are corrected: no part is "proposed" that is merged, the review handler's file and the recipe and reducer code are in the source map, `freeze` sorts (and refuses only an empty list, a repeated id or an empty id), the review evidence file is mode 0600 on POSIX and takes its folder's access list on Windows, the register has no owner column, and the docs lint test the pages cited no longer exists. +- **Check.** Built binary against the real delegation log and ledger: `trace-corpus freeze` onto an existing baseline is refused with the file unchanged, `harvest --since 2026-13-45` is refused with exit 2, `--help` exits 0 with the flags; `ledger --failures` puts the whole-ledger `other` share at 10.6 % (1,160 of 10,900 failures) where the installed build reads 45.4 % (4,949), and the 30-day `other` count at 55 where it reads 157; `gate selftest` against an unreachable seat exits 2 with no receipt and an untouched `--out`, where the installed build exits 1 with `passed: false`. + ## [0.231.0] - 2026-11-11 - an accelerator answers while the node's GPUs are leased, and a shared accelerator tool falls back to the next device ### Fixed - an accelerator answers while the node's GPUs are leased, and a shared accelerator tool falls back to the next device diff --git a/VERSION b/VERSION index fd611fd75..5271d25d1 100644 --- a/VERSION +++ b/VERSION @@ -1 +1 @@ -0.231.0 +0.232.0 diff --git a/changes/unreleased/fix-rsi-instruments.md b/changes/unreleased/fix-rsi-instruments.md deleted file mode 100644 index b2e4f4319..000000000 --- a/changes/unreleased/fix-rsi-instruments.md +++ /dev/null @@ -1,9 +0,0 @@ -### Fixed - the self-improvement instruments report what happened (trace-corpus guards, failure classes, selftest could-not-judge) - -- **`trace-corpus`.** `freeze` creates its baseline exclusively and refuses a path that exists (exit 2, the way `gate freeze` refuses one), where it used to replace it silently. `harvest --since` and `--until` must be real `YYYY-MM-DD` dates with since not after until, and are refused with exit 2 otherwise: a typo such as `2026-11-1` used to keep the wrong shards, and `11/01/2026` kept all of them, both with exit 0. `help`, `--help` and `-h` print the usage with the flags of every mode and exit 0 (they exited 1 with no flag list). -- **Failure classes, written.** The remote lanes (vision, text, speech, media, compose), the composite image row and the accelerator forward now stamp the reason code from the defer class the result carries (the error class and the harness's own sentence first, the class when neither names the row). Their capacity and placement defers were filed `other`: 154 of the 157 `other` rows in a 30-day register. A deferred row whose writer gave no reason is filed with the reason `deferred: no reason recorded` instead of an empty string. -- **Failure classes, read.** Rows written before the stamp are classed from sentences the harness itself writes and that map without doubt to one existing class: `wall timeout after`, `step budget exhausted` (budget), `output failed schema` (abstention), `failed verification:`, `seat down:`, `queue deadline after` (queue_deadline, or queue_withdrawn when the node confirmed the withdrawal), `stalled: no progress` (the stall class of the phase), the lane's `no fleet node ...` placement sentences (no_eligible_node) and a dispatch refused with 503 or 429 (queue_full). A deferred row already stamped `other` is read through the same rules (`other` is the stamp for "no member named it", so it is the one stamp read again); every other stamp is read as stamped. Ambiguous sentences stay `other` and are listed in the register page. -- **`gate selftest`.** A scorer whose seat call deferred or failed (a GPU lease held, a seat down) could not judge: the verb exits 2, writes no receipt, leaves an existing `--out` file alone and names the defer reason. It used to report a mis-ranked scorer with exit 1 and a `passed: false` receipt that replaced an older `--out`. The A/B (`compaction-eval ab`) still scores such a call 0. -- **Small fixes.** The top-level `ledger` usage line lists `--failures`, `--table` and `--include-test-traffic`; the `recipe-audit` help ends its exit-code block and its usage line lists `--config`; `.gitattributes` forces LF on `internal/eval/*.json`, so the embedded gate schema, and the build that prints it, no longer depend on the builder's line-ending setting. -- **Docs.** The self-improvement, measurement-gate and failure-register pages are corrected: no part is "proposed" that is merged, the review handler's file and the recipe and reducer code are in the source map, `freeze` sorts (and refuses only an empty list, a repeated id or an empty id), the review evidence file is mode 0600 on POSIX and takes its folder's access list on Windows, the register has no owner column, and the docs lint test the pages cited no longer exists. -- **Check.** Built binary against the real delegation log and ledger: `trace-corpus freeze` onto an existing baseline is refused with the file unchanged, `harvest --since 2026-13-45` is refused with exit 2, `--help` exits 0 with the flags; `ledger --failures` puts the whole-ledger `other` share at 10.6 % (1,160 of 10,900 failures) where the installed build reads 45.4 % (4,949), and the 30-day `other` count at 55 where it reads 157; `gate selftest` against an unreachable seat exits 2 with no receipt and an untouched `--out`, where the installed build exits 1 with `passed: false`. diff --git a/internal/buildinfo/buildinfo.go b/internal/buildinfo/buildinfo.go index 363f6036c..a99225fce 100644 --- a/internal/buildinfo/buildinfo.go +++ b/internal/buildinfo/buildinfo.go @@ -29,7 +29,7 @@ import ( // `version` const aliases it (keeping the VERSION-file agreement test in // main_test.go binding), and fleet/pipeline stampers read it directly so no // plumbing carries it through config. -const Version = "0.231.0" +const Version = "0.232.0" var ( buildOnce sync.Once