Skip to content

release v0.1.175 - #1735

Closed
ranxianglei wants to merge 1 commit into
masterfrom
2026-09-30_release-v0.1.175
Closed

ranxianglei wants to merge 1 commit into
masterfrom
2026-09-30_release-v0.1.175

Conversation

@ranxianglei

Copy link
Copy Markdown
Owner

No-op validation release (AGENTS.md §5) — zero code changes.

Pure version bump 0.1.174 → 0.1.175; the commit touches only package.json + package-lock.json version lines.

Why this release exists

PR #1734 (fix for issue #1732 — dsh profile bundle refresh frozen on Windows) modifies the auto-update's dsh channel (src/dsh-channel.ts, imported by src/update.ts). Per AGENTS.md §5, any change to load-bearing update machinery must be validated with a no-op release first: it proves the existing upgrade path is healthy end-to-end (the currently-installed 0.1.174 auto-updates to 0.1.175 using the old, trusted updater code) before the changed code ships.

Ship order

  1. Merge this PR → CI publishes v0.1.175 to npm.
  2. Confirm a real upgrade succeeds somewhere (log line auto-updated 0.1.174 → 0.1.175).
  3. Only then merge fix(#1732): resolve dsh binary beyond bare PATH name; decode GBK subprocess output #1734 and release the fix as v0.1.176.

If step 2 shows a failed upgrade, stop — do not ship #1734; investigate the existing path first.

Changes since v0.1.174

None (no-op by design).

Pre-flight (local, branch head)

  • npm run typecheck — clean
  • npm test — 3332 tests, 0 fail (17 skipped by design)
  • npm run build — success, dist-import-annotations OK

@github-actions

Copy link
Copy Markdown
Contributor

📦 Built Package Artifact

Branch: 2026-09-30_release-v0.1.175 (39dc7dc)

Option A — Install this exact build (recommended)

npm install -g billion-context@0.0.1682-pr.1735.053c50a

This exact version is immutable and reproducible. The rolling pr tag tracks the newest test build across all PRs.

Option B — Download artifact

  1. Download the artifact from the Actions run
  2. Extract the tarball and install:
tar xzf billion-context-pr1735.tgz
npm install -g package

Install instructions are refreshed on each push.

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

Review — release v0.1.175 (no-op validation release)

① Direction check. Purpose per OP: validate the existing auto-update path end-to-end before shipping #1734 (which touches src/dsh-channel.ts, imported by src/update.ts). That is exactly the no-op-release discipline in AGENTS.md §5 — isolating the plumbing test from new code. Release mechanics all conform: branch name 2026-09-30_release-v0.1.175, isolated release v0.1.175 commit, diff = version lines only (package.json + package-lock.json, 4 lines total; acp-kernel pin untouched). ✓

② Regression audit — window v0.1.174..master (8 merged PRs). Every PR reviewed hunk-by-hunk against its pre-change behavior:

PR Existing behavior touched Conclusion
#1661/#1672 native-intercept unrouted-URL observability hook fired for every HTTP method → now POST-only; past REARM_LIMIT the guard surrendered the top slot (silent unbounded stacking → live Windows RangeError: Maximum call stack size exceeded) → guard is now permanent, the cap bounds logging only; re-entry into bili's own chain terminates the descent at the oldest live anchor (O(1) instead of walking accumulated history) Intended fix. Documented degradation: with TWO concurrently-live foreign wrappers, the older wrapper's injection is skipped for these requests
#1693 transport retry main model-request path was single-attempt → bounded retry on fail-fast pre-response network deaths only (existing BILI_REPLAY_RETRY_MAX/BASE_MS budget); HTTP responses are returned untouched — no double-delivery risk, upstream error bodies stay verbatim Intended. Worst-case latency grows by backoff × attempts
#1703 title-gen intent opencode v2 title-gen requests (carry no max_tokens) were classified as main turns and got the full compress prompt injected into the title model → now side-request passthrough via the x-bili-plugin-agent intent header, honored only when the announced plugin header is present (plain clients cannot declare side-request to dodge compression); other persona ids remain inert telemetry Intended
#1713 decompress range blank/whitespace startId/endId forced the range path → unsatisfiable retry loop; now treated as omitted (whole-block restore). Wire shape change: decompress schema loses startId/endId where CCR cannot arm (NO_RANGE variants served on the proxy lane while the session is unarmed, plugin manifest when base CCR is off; responses plugin lane always) — goldens regenerated; verified the golden deltas are exactly these description/removal lines with no unrelated drift Intended. Conservative advertise-only-what-execution-can-honor rule (#1192)
#1719 log privacy bili.log: non-loopback IPv4/IPv6 now masked [private-host] (same BILI_LOG_MASK_HOSTS=0 gate as host masking); summary fingerprints length-only in the log copy (model-facing receipt unchanged — wire fidelity intact); message ids salted per-process (cross-run/cross-machine correlation gone, within-run joins still work); secret patterns scrubbed at the logger choke point; programmatic capture hook receives raw values (documented in-process seam). Zero wire impact Intended
#1721 tag echo paired render tags wrapping non-ref prose were swallowed wholesale → now tags stripped, prose released; ref-shaped bodies and attested imitations still drop the whole span. Operates on model-prose paths only — tool-args-verbatim invariant untouched Intended
#1722 subagent sidecar opencode sessions gain additive [acp-subagent-sessions …] machine-captured child-id lines in compress receipts and rendered active-block summaries (fixes #1563 ~20% id loss through weak-model summaries). Other lanes no-op; stored in the existing extensible session.metadata field (backward compatible) Intended, additive

Cross-cutting checks: both compression modes reasoned per PR (manifest/injection split in #1713, dual carriers in #1722); no changes to src/update.ts, release/CI infra, the acp-kernel pin, or message-ref/id logic anywhere in the window. 本版本引入回归:无。

③ Code. No issues found. Two load-bearing details verified directly: fetchWithTransportRetry retries only when isFailFastUpstreamKind (nothing reached the upstream, so a replay cannot double-deliver) and never converts an HTTP verdict into a proxy-side error; the re-entry-termination argument holds because an OUR chain link reached from inside our own dispatch can only exist through a foreign wrapper that captured one of our older tops — everything below the oldest live anchor is redundant (top already rewrote/stamped or passed through).

④ Tests. All eight PRs shipped behavioral tests that reproduce their original defects (main-path-transport-retry +381, decompress-empty-range +186, subagent-sessions +217 + e2e +604, log-mask +81, tag-echo +41, native-intercept-selfheal +96, side-request-isolation +57, compress-fingerprint-refold +41) — real assertions, not tautologies. CI on head 39dc7dc9: all green, including e2e, ubuntu/windows × node 22/24, the real codex/pi/dsh/opencode-native zero-token lanes, hermetic registry e2e, build-artifact, and version-guard.

⑤ Diff. Version lines only. Clean.

Caveats (no action required)

  1. Branch is behind master by fix: preserve opencode sub-agent session ids across folds via structured-field sidecar (#1702) #1722 (cut off master just before fix: preserve opencode sub-agent session ids across folds via structured-field sidecar (#1702) #1722 landed; GitHub reports mergeable_state: behind). The merge commit folds fix: preserve opencode sub-agent session ids across folds via structured-field sidecar (#1702) #1722 in; with a version-line-only branch diff there is zero conflict risk, and release.yml re-runs the full gate on the merged tree. No rebase needed.
  2. "Changes since v0.1.174: None" is true of the release commit (no-op by design) — but the published artifact contains the eight PRs above. Major behavior changes for the release notes (each disclosed in its own thread, old→new):
    • main request path: single-attempt → bounded retry on pre-response transport failures only
    • native fetch intercept: permanent routing guard (was: surrender past REARM_LIMIT) + O(1) re-entry termination (fixes the Windows stack overflow)
    • unrouted-URL observability: all methods → POST-only
    • opencode v2 title-gen: main-turn treatment → side-request passthrough (intent header)
    • decompress: blank range fields count as omitted; schema drops startId/endId where CCR cannot arm (wire shape change)
    • logs: IP masking + secret scrubbing + length-only fingerprints + salted message ids (log surface only)
    • tag echo: paired tags with prose bodies now release the content instead of swallowing it
    • opencode: sub-agent session ids mechanically preserved across folds (additive model-visible lines)
  3. Ship-order reminder: once this merges and CI publishes, confirm a real ✔ billion-context auto-updated 0.1.174 → 0.1.175 log before merging fix(#1732): resolve dsh binary beyond bare PATH name; decode GBK subprocess output #1734 → v0.1.176. If the no-op hop fails, stop and investigate the existing path first (§5) — don't stack the updater change on a broken hop.

Verdict

✅ Mergeable. Release mechanics fully conform, no regressions found in the audited window, CI green. The merge itself is yours to execute.

中文摘要:这是发布 #1734(updater 渠道修复)之前的 no-op 验证版本——release 提交本身只含版本号变更、分支名与提交信息均合规;我对 v0.1.174 以来合入的 8 个 PR 逐一做了既有行为回归审查,未发现回归,重大行为变更已逐条列出(旧值→新值+原因)可直接用作发布说明;CI 全绿,可以合并。

@ranxianglei

Copy link
Copy Markdown
Owner Author

Closing as superseded: this PR was prepared before #1722 landed on master (base 5c9c147), while #1739 (branch 2026-09-30_release-v0.1.175-053442) carries the same version bump on current master including #1722. Merging #1739 publishes v0.1.175; keeping both same-prefixed branches around only invites confusion (#1744 附带观察).

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