Repository navigation
release v0.1.168 - #1594
release v0.1.168#1594
Conversation
📦 Built Package ArtifactBranch: Option A — Install this exact build (recommended)npm install -g billion-context@0.0.1535-pr.1594.716527bThis exact version is immutable and reproducible. The rolling Option B — Download artifact
tar xzf billion-context-pr1594.tgz
npm install -g packageInstall instructions are refreshed on each push. |
[bot] 🏷 Release v0.1.168 audit complete — verified, no regressions found, ready for human merge. This is an agent-created release PR ( ① PR integrity ✅
② CI (all green on head 7f3963b) ✅CI · CI Windows codex overlay (#686) · E2E (real codex) · e2e-fake-codex · e2e-local-registry · e2e-native-pi/opencode/dsh · PR Build Artifact — all ③ Local pre-flight on the release head ✅Checked out ②b Version-range audit —
|
| Change | Scope | Verdict |
|---|---|---|
| #1542 — strip leaked chain checkpoints from incoming history at any position | request pipeline (src/server.ts ×4 prepare paths, new stripEmbeddedChainCarriers) |
audited in depth below |
| #1579/#1586 — omp native identity stamp fixes (landed after the cut) | agent-side only (src/agent/native-intercept.ts, src/agent/pi.ts) |
audited below |
The issue body's "Changes since v0.1.167" lists only #1542 — stale. The auto-generated GitHub Release notes (git log since last tag) will correctly include both, so the published artifact and release notes stay consistent; only this thread's list is incomplete. No action required unless you want the body updated.
#1542 deep-dive (the pipeline change)
- Root cause is correct:
extractChainCarriersonly recognizes carriers in the trailing user-run slot; a carrier that leaked into a client transcript and sank into mid-history is invisible to it, rides every resend, and stacks monotonically. The fix strips well-formed whole-part carriers from ANY position on the normal-processing path. - Recognition strictness matches the frozen ID contract (feat: chain checkpoint step 2 — shadow recognition (parser, JCS digest, carrier contract) #1397): whole-PART exact match via
parseChainCheckpoint(all 5 attrs, strict regexes, ≤512 chars). Tag-shaped text embedded in prose is NOT stripped — the [bug] 插件透传会剥掉工具调用参数里的 render tag(arguments 流挂了 tag-echo 过滤器) #1039/design: Chain Checkpoint — first-processor-wins idempotent interop (Step 2/3) #1395 boundary is preserved (test asserts it). - Roles bounded: user+assistant (google: user+model).
tool/system/function_call/compaction_triggerare never touched — tool-args-verbatim discipline intact. - Chain-verdict flow untouched:
evaluateChainruns beforeprepare*(server.ts:1453); valid/recent-mismatch/stale-matched setchainSkipand the entire processing block is guarded by!chainSkip(server.ts:1704) → verbatim-forwarded chained-bili bodies never reach the strip. stale-unmatched/invalid/none fall through to normal processing, which now also strips mid-history carriers — the intended extension. - Wire invariants hold: kernel
coreToAnthropic/coreToGooglefuse same-side role runs, so removing a mid-history carrier cannot produce a non-alternating Anthropic/Google body; OpenAI/Responses have no alternation requirement. One theoretical residual (not blocking): if a leading google user-carrier entry were removed and no system injection follows, first-content could bemodel— practically unreachable since a stamp only ever exists after a real turn, and flagged here for completeness. - Both compression modes covered: the e2e test drives a real
startServer+ fake upstream in plugin-mode AND proxy-mode and asserts exactly ONE fresh carrier reaches the LLM (trailing slot), old digests absent, strip logged. - State safety: stripping happens before kernel projection, so removed carriers never enter ref assignment (no re-issue window — kernel cursor =
highestUsedIndex+1); prefix-affinity matching stays consistent across resends because the strip is deterministic per version. - Behavior change (disclosed): incoming bodies carrying a well-formed whole-part
<bili-chain>tag in user/assistant content at any position are now modified before kernel projection (previously passed through verbatim). Logged per request when >0. This is the fix itself.
#1579/#1586 (agent-side, landed after the cut)
- Restores
prompt_cache_keyidentity stamping for native omp sessions, where [Bug] bili omp 扩展 stampPromptCacheKey 无条件注入顶层 prompt_cache_key —— 严格 schema 上游(opencode zen anthropic 端点)全量 400 #1403's launcher-shaped routing gate had silently killed the stamp for every native session. Conservative double-gate: native-intercept flag AND live proxy claim (detectProxyBase) AND model-API URL shape — [Bug] bili omp 扩展 stampPromptCacheKey 无条件注入顶层 prompt_cache_key —— 严格 schema 上游(opencode zen anthropic 端点)全量 400 #1403's "never stamp where the proxy can't see it" guarantee is preserved for everything the interceptor wouldn't rewrite anyway. session_startno longer awaits native-origin resolution across its captured ctx (stale-ctx guard from review); a round-1 miss self-heals on round 2; launcher mode unaffected (synchronous base resolution).- Green on its own PR CI and on post-merge master; +115 lines of dedicated tests.
- Behavior change (disclosed): native omp sessions now emit pck stamps where they were previously suppressed — the intended restoration.
Conclusion
本版本引入回归:无。 Both shipped changes are intentional, conservatively gated, test-pinned, and CI-green across the full gate set (incl. real-codex e2e). No config surface, persistence format, or acp-kernel pin changes. Per §7.4 the #1542 pipeline change is human-merge territory regardless — please merge yourself when ready: #1594 — CI publishes automatically on merge.
中文摘要:审核了 release v0.1.168(仅版本号 diff,CI 全绿,本地 typecheck/测试/构建复验通过);版本区间审计发现分支切出后 master 又合入了 #1586,故本次发布实际包含 #1542(链式检查点泄漏剥离,已逐路径核实无回归)与 #1579(omp 原生身份戳恢复,双重门控保守安全)两项修复,均有意变更且已披露,可以合并。
Auto-fallback from the one-click release workflow: direct push to master was blocked by branch protection. Merging this PR publishes via the standard release.yml flow.
Changes since v0.1.167