Skip to content

Find code by asking what it does. A CLI for coding agents that uses Jev to discover relevant files and source context.

docs: resolve independent archive audit and retain release preparation proof

David Zhangedf7f19168e7
diff --git a/docs/architecture.md b/docs/architecture.md
index 81890e2..9a6a911 100644
--- a/docs/architecture.md
+++ b/docs/architecture.md
@@ -11,7 +11,9 @@ second coding agent or a generated-answer layer.
The accepted reference is the frozen
[unit-locators spike](../evals/implementation/swebench/hierarchy-unit-locators-spike.ts)
-with the [canonical skill](../skills/jevgrep/SKILL.md).
+with its [historical measured skill](../test/reference/accepted-skill.md).
+The [canonical production skill](../skills/jevgrep/SKILL.md) preserves those
+instructions with the executable renamed to `jg`.
The production CLI implements this strategy with the product boundaries recorded
in the implementation record. This document identifies the experimental reference,
not a substitute for the corrected package's measured quality result.
@@ -108,9 +110,8 @@ combine each task's cheapest result from different strategies.
The [acceptance audit](../specs/done/jevgrep/assets/accepted-spike-audit.json)
rechecks the frozen artifacts, official grades, and complete bills against the
-user-revised target. The [cohort result](../specs/done/jevgrep/assets/accepted-spike-audit.json)
-is the canonical numerical evidence; its original unmet-target label refers to
-the former 80% target and is retained as history.
+user-revised target and is the canonical numerical evidence for the accepted
+spike. It records the revised 70% target rather than the earlier 80% goal.
This is a small, tuned, Python-only sample with one fixed baseline per task and
model/harness. It supports accepting a useful retrieval architecture, not a
diff --git a/scripts/RELEASING.md b/scripts/RELEASING.md
index 562b21c..9f85600 100644
--- a/scripts/RELEASING.md
+++ b/scripts/RELEASING.md
@@ -30,7 +30,9 @@ asset checks. [The build](build-cli.ts) copies the authored MIT license and
[collects third-party notices](package-notices.mjs) from emitted bundle inputs.
[Retained license sources](licenses/README.md) document upstream distribution gaps.
The archive excludes test fixtures, evaluation evidence, node_modules directories, source maps,
-and repository source. Runtime parser dependencies install from exact npm pins.
+and unbundled development source. Authored Python helpers and the Node worker
+are required runtime assets and are included. Direct runtime parser dependencies
+install from exact npm pins; transitive dependencies resolve during npm installation.
On an Apple Silicon Mac, run `node scripts/test-native.mjs` after the normal
`bun install --frozen-lockfile` setup. This builds and packs the candidate, installs
diff --git a/specs/done/jevgrep/README.md b/specs/done/jevgrep/README.md
index e4c1354..e942508 100644
--- a/specs/done/jevgrep/README.md
+++ b/specs/done/jevgrep/README.md
@@ -18,13 +18,13 @@ trace comparisons and immutable aggregate; its `accepted: false` verdict remains
The earlier [passing confirmation](assets/work-clock-confirmation.md) belongs to
its own frozen package and cannot replace the corrected candidate's result.
-The user separately authorized one identical cohort repeat to double-check
-variance. **That measurement is pending at this documentation checkpoint.** It
+The user separately authorized one [identical cohort repeat](assets/variance-repeat.md)
+to double-check variance. **That measurement is pending at this documentation checkpoint.** It
is not permission to replace failures, pool outcomes, change the strategy or keep
rerunning until the gate passes. Its terminal evidence and the final independent
claim audit remain closeout work for the coordinating agent. The user authorized publication of `0.1.0` once the repeat and closeout finish.
-Versioning, release verification and publication remain pending with the
-coordinating agent. The repeated benchmark retains the frozen `0.0.0` archive;
+The [0.1.0 release candidate](assets/release-0.1.0.md) passed local validation and
+installed checks; CI publication and registry verification remain pending. The repeated benchmark retains the frozen `0.0.0` archive;
release preparation does not replace its measured artifact.
## Why this shape
@@ -89,10 +89,11 @@ The [recorded stdout](assets/stdout-example.txt) is a deterministic reference fi
not a live-provider result. There was no visual design baseline: installed-process
transcripts and source/request comparisons are the relevant review artifacts.
-The [first CPython confirmation](assets/cpython-confirmation.md),
+The [first CPython confirmation](assets/cpython-confirmation.md) retains a failed
+cohort for the bundled interpreter strategy that ships. The separate
[query trial](assets/query-framing-study.md) and
-[presentation trial](assets/presentation-study.md) retain unsuccessful directions
-and their causal limits. None earned production promotion. Lower aggregate cost
+[presentation trial](assets/presentation-study.md) retain unpromoted strategy
+changes and their causal limits. Lower aggregate cost
alone does not cure a lost solve. The tuned sample is Python-only and observed
during development; it establishes neither untouched generalization, TS/JS task
quality, universal savings nor whole-computer scalability. Windows and broader
diff --git a/specs/done/jevgrep/assets/freshness-confirmation.md b/specs/done/jevgrep/assets/freshness-confirmation.md
index 185a24f..9b72a90 100644
--- a/specs/done/jevgrep/assets/freshness-confirmation.md
+++ b/specs/done/jevgrep/assets/freshness-confirmation.md
@@ -8,6 +8,9 @@ does not compensate for the lost solves. Jev's observed subtotal is
**$1.035782622**; its full total is unknown and excluded from scored cost.
The [complete aggregate](freshness-aggregate.json) is retained byte-for-byte from
the study, SHA-256 `37f2dc0883d0e366b75e74b0f178498a393da9f8ef9cd573d5881630efc0d65a`.
+After reviewing these results, the user accepted the documented tradeoff and
+authorized finishing the product. The [implementation record](../README.md)
+records that separate acceptance decision; the measured gate remains failed.
This is a separate prospective ten-task cohort for the
[source-freshness correction](source-freshness.md), not a replacement of individual
cells from the preceding accepted study. The frozen plan, archived runner and
@@ -45,8 +48,8 @@ and test work does not isolate a causal cost effect from the freshness correctio
Root execution log: `/tmp/jg-freshness-v1-run.log`. Per-task grade and accounting
receipts, complete event streams, patches and request/response evidence remain
with the attempts. All coding, grading and accounting processes are terminal;
-no attempt was replaced and no baseline was rerun. The product quality gate
-remains open because the completed aggregate rejects the candidate.
+no attempt was replaced and no baseline was rerun. The original quantitative
+quality target remains unmet; product acceptance is a separate user decision.
Django's packet contained 51 files in 8,071 bytes without excerpts. SQLite was
listed and its critical type-only rebuild guard was read, untruncated, before the
diff --git a/specs/done/jevgrep/assets/release-0.1.0.md b/specs/done/jevgrep/assets/release-0.1.0.md
new file mode 100644
index 0000000..36e5bea
--- /dev/null
+++ b/specs/done/jevgrep/assets/release-0.1.0.md
@@ -0,0 +1,22 @@
+# 0.1.0 release evidence
+
+The user explicitly authorized publishing `0.1.0` after the variance repeat and
+closeout finish. Publication is pending; preparing an archive is not publication.
+
+The local release candidate is
+`/tmp/jg-010-release/dzhng-jevgrep-0.1.0.tgz`, SHA-256
+`a6c9061c1a7102bd192df8f1d0c2c954b470f86db3ed279818bf167660332f00`.
+It passed the release identity/content/license validator. Compared with the
+frozen benchmark archive, only package metadata and the bundled executable
+changed; the executable difference is exactly `0.0.0` to `0.1.0`. All other
+archive members are byte-identical. Proof: `/tmp/jg-010-version-parity.json`.
+
+This candidate passed both native macOS arm64 installed journeys on Node 24.14.0
+and all 24 Linux arm64 installed tests. Logs: `/tmp/jg-010-native.log` and
+`/tmp/jg-010-installed.log`. Full deterministic verification of the unchanged
+implementation is retained in [freshness verification](source-freshness.md).
+
+The tag workflow still has to run its own full verification, pack and validate
+its candidate, test those exact bytes, publish them, and verify the registry
+archive's integrity and installed behavior. The local hash above is local proof;
+the workflow's retained manifest owns the published artifact identity.
diff --git a/specs/done/jevgrep/assets/variance-repeat.md b/specs/done/jevgrep/assets/variance-repeat.md
new file mode 100644
index 0000000..6e1f029
--- /dev/null
+++ b/specs/done/jevgrep/assets/variance-repeat.md
@@ -0,0 +1,27 @@
+# Identical-runtime variance repeat
+
+Status: running. After accepting the first corrected cohort's documented
+tradeoff, the user authorized exactly one additional full cohort to examine
+variation. The original outcomes remain intact and no baseline is rerun.
+
+The repeat is `evals/runs/swebench/installed-jg-freshness-repeat-v1b/`, using the
+same frozen `0.0.0` archive as [the first cohort](freshness-confirmation.md).
+All 179 installed files, the skill, harness, registry and task inputs are
+byte-identical. The plan hash is
+`9223698cbc84abb2a247ddb29d634a86ca9ec04b0ef74c4198a0f7d7187f1ff5`.
+Ten offline runtime preflights and ten no-call validations passed before launch.
+The repeat is an additional observation, not permission to pick the better
+attempt per task or rerun until acceptance passes.
+
+The initial no-call preparation, `installed-jg-freshness-repeat-v1`, resolved a
+newer transitive `ws` dependency. Exactly two installed files differed. That
+preparation is retained with an invalid-runtime-drift disposition and made zero
+paid calls. The replacement preparation copies the original installed prefix and
+uses the unchanged archived runner. Its `preparation-method.json` and
+`/tmp/jg-freshness-repeat-v1b-freeze-proof.json` retain the method and file hashes.
+No product or evaluation policy changed to repair the setup.
+
+Live execution and sequential grading/accounting belong to sessions `67618` and
+`27936`; logs use `/tmp/jg-freshness-repeat-v1b-{run,observer}.log`. Revalidate
+their handles before any resume. Final results must be shown alongside the first
+cohort, without pooling selected cells or replacing its failed verdict.
diff --git a/specs/done/jevgrep/choices.md b/specs/done/jevgrep/choices.md
index b677fda..fec1cff 100644
--- a/specs/done/jevgrep/choices.md
+++ b/specs/done/jevgrep/choices.md
@@ -3,8 +3,8 @@
This ledger describes the decisions embodied in the implementation, not its
verification status. `jg` is the installed CLI; Jev is the relevance classifier
it calls. “Reference” means the accepted experimental retrieval implementation.
-The [spec handoff](README.md) owns unfinished work and the
-[confirmation record](assets/work-clock-confirmation.md) owns measured outcomes.
+The [implementation record](README.md) owns the closure decision and the
+[corrected confirmation](assets/freshness-confirmation.md) owns measured outcomes.
Review these medium-confidence choices first: the benchmark work clock, ordering
inside cross-file evidence, and the shared Python interpreter. They preserve the
@@ -147,8 +147,9 @@ constraint. **Confidence:** medium. Owner: [selection](../../../packages/core/sr
**When:** cache integration and performance review.
A new answer is ready to save while the cache is near its disk limit. The cache
-publishes the complete answer atomically, then scans stored entries and removes
-old entries as needed. It does not maintain a separate persistent index or run a
+publishes the complete answer atomically, then scans stored entries in filesystem
+enumeration order and removes entries beyond the retained byte budget. This is
+not oldest-first eviction. It does not maintain a separate persistent index or run a
background cleanup service. With many entries, repeating that scan for many new
answers adds overhead; an indexed eviction design would trade that work for
another stateful component to maintain.
diff --git a/specs/done/jevgrep/contracts.md b/specs/done/jevgrep/contracts.md
index 8497ad7..4ab8322 100644
--- a/specs/done/jevgrep/contracts.md
+++ b/specs/done/jevgrep/contracts.md
@@ -45,8 +45,8 @@ A downstream closed pipe ends quietly without an error stack or continued calls.
Credentials retain XDG config location and owner-only permissions; environment
key takes precedence over saved key. An explicitly empty environment key disables
-saved credentials, supporting isolated tests. Reject whitespace-containing keys
-consistently; never echo keys in either stream or provider errors. Auth saves
+saved credentials, supporting isolated tests. Trim surrounding whitespace, then
+reject internal whitespace consistently; never echo keys in either stream or provider errors. Auth saves
without a network call; doctor verifies a synthetic expected answer. Help/
version/skill/cache-clear require no provider key. Bound auth stdin input.
@@ -193,8 +193,9 @@ idempotent clear. `--no-cache` disables both reads and writes. Corrupt/unavailab
cache behaves as a miss and is summarized without turning healthy retrieval into
incomplete discovery. File layout and eviction mechanism delegated; initial size
limit 256 MiB, enforced without a daemon. Concurrent clear/writes must not crash
-or produce invalid answers; clear removes entries existing when its scan begins,
-not a global pause of other searches.
+or produce invalid answers; clear atomically detaches the current entries directory
+before cleanup. Concurrent writers may create new entries; clear is not a global
+pause of other searches.
## Output and quality