Release 0.8.0 — context-aware memory delivery, and the evidence that proved it

[schneiderjoseph]

Releases 0.8.0. Three commits on top of ddbdcf5 (main, v0.6.0): a7175d8 is the release, 07bbf5a and af1e029 fix a defect CI found on this branch after publication. CI is green on all four runners.

0.7.0 — skipped, not published, documented

0.7.0 prepared devia context and devia contribute but was never committed: no tag, no branch,
no stash, nothing in the reflog held a source state for it, and 0.8.0 rewrote the surfaces it
introduced (context.maxTokens → context.budget/context.mode, strict compression, the
CTX-BUDGET gate, impact-map routing, evidence hashing, and a benchmark rewritten from 6
scenarios to 352 runs) before either reached npm.

Publishing it would have meant hand-carving a source tree that never existed and was never
tested, then making its changelog claims true by deduction — for a permanent npm artifact. It is
skipped instead. Its CHANGELOG section is preserved and marked (never published) with the
reason and a pointer to 0.8.0, so an adopter looking for 0.7.0 on npm understands. No
v0.7.0 tag was created. npm goes 0.6.0 → 0.8.0.

0.8.0 — published

• npm: @schneiderjoseph/devia@0.8.0, dist-tag latest → 0.8.0
• Tag: v0.8.0 (annotated), on a7175d8 — the exact commit whose contents were published
• GitHub Release: https://github.com/schneiderjoseph/devia/releases/tag/v0.8.0

The published tarball's shasum, cbaf12cc426128862c68938d654074909b0d79a0 (449 files,
590,948 B unpacked), is identical to the tarball packed from this branch and exercised before
publication — the pre-publication verification applies to the published artifact bit for bit.

Tests performed

Command	Result
npm test	135 pass · 0 fail (52 s)
npm run validate	141 rules · 45 P0 · 56 automatable · 23 domains · 551 links in 379 files · index current
node bin/devia.mjs check --root .	15 pass · 1 warn (OPS-001, lint not referenced) · 3 skip · P0 clear · exit 0
node bin/devia.mjs validate	19 pass · exit 0

No test or threshold was modified to reach these numbers.

Benchmark verification

npm run benchmark:context — 352 runs (4 corpus shapes × 11 task types × 4 targets × 2 modes):

Critical-rule recall     100% in 352/352 runs
Routing accuracy         100% in 352/352 runs
Strict budget compliance 176/176 runs never exceeded the target
Advisory over target      68/176 runs, all because the mandatory floor exceeded it
Mean task-generic share  66.9% of selected tokens
Mean supporting filler    2.3% of selected tokens
Mean selection time       1.7 ms per run

In all 68 advisory overruns selected equals the mandatory floor to the token. Token counts are
estimates from src/lib/tokens.mjs, not a tokenizer's output. Cost per correct decision is not
measured.

Package verification

Tarball: 449 files, 191.9 kB packed. Audited for .git, .env, node_modules, tests/,
.devia/, contribution payloads, reader.html, *.pem, *.key, logs and fixtures — none
present. Top level is bin, src, rules, checklists, standard, compliance, schema,
templates, skills, scripts and the governance docs, matching files in package.json.

Installed twice into clean temporary directories — once from the local tarball before publishing,
once from the registry afterwards — and exercised both times:

devia --version      → devia 0.8.0 · standard 0.2.0
npx devia --version  → devia 0.8.0 · standard 0.2.0
devia init           → 16 memory files, devia.json, .devia/.gitignore, 5 agent adapters · exit 0
devia validate       → 18 pass · 14 warn (unfilled template placeholders) · exit 0
devia check          → BLOCKED, 2 P0 (CI-PRESENT, TST-PRESENT) · exit 1 — correct for an empty
                       directory with no tests and no CI, not a packaging defect
devia context        → advisory: selected 1642 = mandatory floor 1642, target 1200, OVER TARGET
devia context --strict --budget 600
                     → selected 598 ≤ 600, WITHIN TARGET, DEGRADED
devia contribute     → --help and list only; no remote submission was performed

Release notes

CHANGELOG.md carries the 0.8.0 entry and the annotated 0.7.0 entry. The GitHub Release notes
state explicitly that the contribution workflow is not automatic (submit writes the payload
and prints the gh command), that token counts are estimates, and that cost per correct decision
is not measured.

07bbf5a + af1e029 — a defect CI found after publication

The first CI run on this branch failed on one of four runners — ubuntu-latest, node 22 —
with not ok 58 - the benchmark passes: the saving never costs a blocking rule and
Unterminated string in JSON at position 146032. Windows node 20, Windows node 22 and ubuntu
node 20 all passed, and so did every local run.

It is not a selection defect. scripts/benchmark-context.mjs --json emits a ~180 kB report and
the test reads it through a pipe. Two things were wrong, and the first fix only found half of it:

• console.log followed by process.exit() discards whatever is still buffered. On Windows
  stdout is synchronous so the write always completed; on POSIX it is asynchronous and the exit
  cut it off. 07bbf5a replaced it with a synchronous write.
• That was still not enough. CI failed again at the same byte, 146032 of 182402 — an
  identical cut point is not a race. A single fs.writeSync on a non-blocking pipe writes what
  currently fits and returns short, and the return value was being discarded; the remaining 36 kB
  were never written. af1e029 loops until every byte is gone, treating EAGAIN as "the reader
  has not drained yet" rather than an error.

A redirect to a file never showed this either, because a file descriptor is not non-blocking.
Only a pipe — exactly what execFileSync hands the test — could.

Scope and consequence. Both commits touch scripts/benchmark-context.mjs and nothing else.
bin/ and src/ contain no process.exit() call at all — the CLI uses process.exitCode
throughout and was never exposed to this. The published 0.8.0 CLI is unaffected and the version
is not bumped.

One divergence worth stating plainly: the 0.8.0 tarball on npm carries the unfixed benchmark
script. It is a development script no adopter runs, the table form of npm run benchmark:context was never affected, and the measurements are identical before and after. If
you would rather npm and main match exactly, that wants an 0.8.1 — it is not a correctness
reason to republish.

Not verified

• devia contribute submit --yes, which is the only network path in devia — deliberately not
  exercised
• Agent compatibility beyond the five adapter files init writes; no claim of universal agent
  compatibility is made anywhere in the release notes