Skip to content

feat: output-side compression (verbosity steering + effort routing) - #1096

Closed
ranxianglei wants to merge 5 commits into
masterfrom
2026-09-21_output-side-compression
Closed

ranxianglei wants to merge 5 commits into
masterfrom
2026-09-21_output-side-compression

Conversation

@ranxianglei

Copy link
Copy Markdown
Owner

What

Ports headroom's two output-token levers to bili's forward boundary. Output tokens cost ~5x input and are billed the moment they stream out, so the only lever is at request time. Both levers are default off — zero behavior change until a user opts in.

1. Verbosity steering — appends a byte-stable conciseness directive (levels L0-L4, default L2) at the tail of the system prompt. Never prepends (that would bust the prefix cache), skip-if-absent (never fabricates a system carrier the client didn't send), and idempotent (re-applying the same level is a no-op; a level change replaces the block in place). Covers all four wires: Anthropic system, OpenAI system/developer message, Responses instructions, Google systemInstruction.

2. Effort routing — structural turn classification of the final message list (block composition only, no content pattern-matching). On a mechanical continuation turn it lowers an already-present effort field to its floor (reasoning_effort, reasoning.effort, output_config.effort + thinking.budget_tokens->1024, generationConfig.thinkingConfig.thinkingBudget->128). Never injects an effort field the client did not send, never toggles thinking.type; dynamic budgets (-1) and non-mechanical turns are left untouched.

Both run at the forward boundary after compat.roles, so the classifier sees the final wire list; the launcher / native-plugin / /bili/ lanes inherit it automatically, and compress-retry re-sends share the same idempotent transform.

Config

// billion-context.json — default OFF
"outputSteering": { "enabled": true, "verbosityLevel": 2, "effortRouting": true }

A per-provider route entry overlays the global block: "providers": { "<url>": { "outputSteering": { ... } } }.

Files

  • src/output-steering.ts (new): directives, idempotent tail placement, structural classifiers, clamp-only effort lowering, wire-shape detection.
  • src/config.ts: outputSteering in the global + per-route schema.
  • src/server.ts: forward-boundary hook (initial forward + retry re-sends).
  • tests/output-steering.test.ts (new): 23 regression tests across all four wires.

Pre-flight

  • npm run typecheck — clean
  • npm run build — success
  • npm test — 2185 pass / 0 fail / 2 skipped (the skips are the env-gated e2e)
  • The full-client e2e-codex run needs a live Responses-compatible upstream (see tests/e2e/README.md), which isn't available in this environment; its zero-token preflight confirms the codex binary + built dist are ready. Because both levers are inert by default and the entire existing suite passes, the request pipeline is unaffected unless a user enables the feature.

While validating I caught and fixed a bug in this very change: the resolved steering config wasn't defaulted when a ProxyOptions object omitted the new field (hand-built options in the test harnesses), which threw a TypeError on every forward (surfacing as 502s). Fixed by null-coalescing to DEFAULT_OUTPUT_STEERING in forward().

Closes #1093

@github-actions

github-actions Bot commented Sep 21, 2026 •

Copy link
Copy Markdown
Contributor

📦 Built Package Artifact

Branch: 2026-09-21_output-side-compression (1986ffc)

Option A — Install from npm PR tag (recommended)

npm install -g billion-context@pr-1096

Each push to this PR publishes a new version under the pr-1096 npm tag.

Option B — Download artifact

  1. Download the artifact from the Actions run
  2. Extract the tarball and install:
tar xzf billion-context-pr1096.tgz
npm install -g package

Install instructions are refreshed on each push.

ranxianglei pushed a commit that referenced this pull request Sep 21, 2026
…ks (#1093)

Review round on PR #1096: the Responses turn classifier scanned the flat
input list FORWARD for any user signal. Clients re-send full history every
request (codex does), so the original ask was present on every turn and the
classifier returned new_user_ask unconditionally — effort routing could only
fire on delta-only inputs that never occur through bili. The other three
wires classify the LAST message correctly, so this made effort routing dead
on the flagship codex lane while appearing functional in tests (whose
synthetic inputs carry no prior user message).

- classifyResponsesTurn now walks BACKWARD from the tail: tool outputs after
  the last user signal = mechanical continuation; a user signal at the tail
  or with no outputs after it = new_user_ask. Assistant-side items between
  the last user signal and the trailing outputs no longer break the
  continuation (message/function_call/custom_tool_call/local_shell_call/
  apply_patch_call/reasoning — previously custom/local/apply-patch calls
  marked the turn unknown even in the delta case).
- Anthropic + Google classifiers: unrecognized block/part composition now
  returns unknown instead of being silently skipped (all four wires share
  one conservative stance: unrecognized structure => do not classify).
- 6 regression tests incl. the realistic full-history repro; steering suite
  23 -> 29, all green.
- Docs: outputSteering entries in CONFIGURATION.md + CONFIGURATION.zh-CN.md
  (global + per-provider), CHANGELOG [Unreleased] entry.
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 PR review — verified end-to-end, one real bug found & fixed on the branch, otherwise ready for your gate.

What I checked (branch 2026-09-21_output-side-compression, base = current master 9c01dbe, mergeable):

  1. Diff hygiene — clean: the feature commit touches exactly src/output-steering.ts (new), src/config.ts, src/server.ts, tests/output-steering.test.ts; no lockfile/version/unrelated churn, single linear commit.
  2. Design invariants from the Feature: Output-side compression (verbosity steering + effort routing) #1093 discussion — all hold in code:
    • skip-if-absent on all four wires (never fabricates a system carrier; byte-identical no-op when absent);
    • tail-append only, sentinel-wrapped, idempotent (same-level re-apply is byte-stable; a level change replaces the block in place) → prefix-cache safe by construction;
    • clamp-only effort lowering: never injects a field the client didn't send, never toggles thinking.type, dynamic budgets (-1) and error continuations untouched;
    • hook sits after compat.roles (classifier sees the final wire list) and is shared by the initial forward + every compress-retry re-send via wireTransform, so launcher / native-plugin / /bili/ lanes inherit;
    • default OFF is a byte-transparent path (no parse/re-serialize at all when disabled); config parsing falls back to defaults on malformed input rather than clobbering; the self-caught ProxyOptions null-coalesce in forward() is present and is the only read site.
  3. Both compression modes — mode-agnostic by placement: in proxy mode the directive appends after bili's injected compression prompt; in plugin mode only clients that actually send a system carrier are touched (skip-if-absent covers the rest). Idempotency keeps retry re-sends stable.
  4. Pre-flight on the rebased head — typecheck clean · build green · npm test 2185 pass / 0 fail / 2 skipped (env-gated e2e) — matches the claim. Full-client e2e-codex still needs a live Responses upstream (its zero-token preflight was green in CI).

Bug found & fixed (new commit 3d5c121 on the branch): effort routing was dead on the Responses wire in real traffic. The classifier scanned the flat input list forward for any user signal — but codex re-sends full history on every request, so the original ask was present on every turn → classified new_user_ask unconditionally → the lowering never fired. The other three wires classify the last message correctly, which is why the synthetic tests (delta-only inputs with no prior user message) passed. Repro: [user msg, function_call, function_call_output] + reasoning.effort:"high" → expected effort:low, got nothing. Fix: walk the list backward from the tail (tool outputs after the last user signal = mechanical continuation; a user signal at the tail, or with no outputs after it = fresh ask), and treat assistant-side items between the last user signal and the trailing outputs as continuation-neutral — custom_tool_call / local_shell_call / apply_patch_call previously marked even delta-case turns unknown. Also tightened the Anthropic/Google classifiers: unrecognized block/part composition now classifies as unknown instead of being silently skipped, so all four wires share one conservative stance (unrecognized structure ⇒ do not touch). Added 6 regression tests including the realistic full-history repro (steering suite 23→29, all green). Gates re-run after the fix: typecheck clean · 2191 pass / 0 fail / 2 skipped · build green.

Minor observation (report only, no action taken): one transient single-test failure appeared in 1 of 4 full-suite runs in this environment — not reproduced in the other 3 runs, and CI ubuntu 22/24 were green on the original head. That run's log wasn't retained, so the test name is unknown; if it recurs anywhere it's worth hunting per the determinism rule (§7.2).

Docs gap closed: the PR carried no documentation for the new config block — added outputSteering entries to CONFIGURATION.md + CONFIGURATION.zh-CN.md (global + per-provider, including the assumed-128 Google-floor caveat and the whole-block overlay semantics) and a CHANGELOG [Unreleased] entry, in the same commit.

CI is re-running on the updated head; merging stays yours either way (feat ⇒ off the auto-merge path regardless).

中文摘要:审查了输出侧压缩 PR——发现并修复一个真实缺陷:Responses wire 的轮次分类器正向扫描整个 input,而 codex 每轮重发完整历史,原始提问永远在场,导致机械续轮恒被判为新提问、effort routing 在主力 codex 通道上实际失效(合成测试因无历史用户消息而未暴露);已改为从尾部反向扫描并统一四条 wire 的保守分类口径,补 6 个回归测试与中/英配置文档及 CHANGELOG,其余安全不变量(skip-if-absent、幂等尾追加、只降不注、默认关闭零行为变化)逐项核实通过,可以合并。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Follow-up: CI is now fully green on the updated head 3d5c121 — all 8 checks pass (test ubuntu 22/24, test windows 22/24, build-artifact, version-guard, codex MCP overlay, real-codex zero-token e2e). PR state: open, no conflicts, mergeable_state: clean. Nothing left on my side — the merge button is yours whenever you're ready.

@ranxianglei

Copy link
Copy Markdown
Owner Author

Fix follow-up on the earlier review findings (cwst pass): fixed on this branch in b8bf783, gates green (typecheck / 2194 tests / build).

1. minimal effort is no longer raised to low — the lowerEffort guards for openai (reasoning_effort) and responses (reasoning.effort) both add && e !== "minimal" (sub-low values stay untouched; clamp-only invariant restored). Tests: minimal → changed:false on both wires.

2. Compress rounds are exempt from verbosity steering — ApplySteeringOptions { verbosity?: boolean }; applyOutputSteeringJson(..., opts) skips the system-prompt directive when verbosity === false. src/server.ts builds wireTransform(true) for the forward boundary and compressWireTransform(false) for the two compress-loop re-send paths (runCompressLoop / compressLoopResponsesJson) — the L2/L3 "never restate" directive no longer contaminates the summarizer's contract; effort-lowering stays (cheap + harmless). Test: verbosity:false → labels contain effort:low only, no system block.

3. Bad config values now warn — clampVerbosityLevel(v, warn) surfaces through parseOutputSteering warnings → loggerLog("warn", ...) at config load.

Non-blocking notes from the review stay as documented behavior (per-route replace-not-merge; no env override; Google floor 128 assumed).

@ework — please re-review the current head b8bf783 in full (fresh pass over the whole diff, not just these hunks).

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Re-reviewed head b8bf783 in full — fresh pass over the whole diff, not just the three hunks. Verdict: fixes 1 and 2 are correct as implemented; fix 3 shipped with a false-positive bug, which I fixed directly (commit b4df2a1, pushed to this branch).

Fix 1 — minimal effort guard ✓. Both guards verified (reasoning_effort on openai, reasoning.effort on responses: e !== "low" && e !== "minimal"); the two new tests assert changed:false with "minimal" byte-preserved. Anthropic's output_config.effort is deliberately unguarded — "minimal" is not a value on Anthropic's effort scale (low/medium/high), so there is no sub-floor value to protect there. Clamp-only invariant restored.

Fix 2 — compress-round exemption ✓ (one documented coarseness). I traced every request flowing through the loop transform: in runCompressLoop (src/loop/core.ts:262 fetchUpstream, call sites 425/482/529/802/820/845/846) and compressLoopResponsesJson (src/compress-loop-responses.ts:228), ALL re-sends are model-facing continuations of the client's own request — none is a dedicated summarizer call. In proxy mode those bodies carry the client system prompt with bili's injected compression contract, so appending the L2/L3 "never restate code / file contents / diffs / tool output" directive onto it would collide exactly on the post-compress re-send where summary fidelity matters most. Exempting the whole loop is the safe side of that boundary. Coarseness, noted and accepted: nudge-retries etc. also lose the directive mid-request — harmless, it self-heals on the next client turn, and prefix-cache divergence is confined to the small appended tail block (a compression event has just rebuilt the context anyway). A surgical two-transform split inside the loop would touch load-bearing loop internals for marginal gain — not worth it. Also verified the preflight summarizer path (summarizeRange, src/preflight.ts:637) never went through steering at all, so no other contamination route exists.

Fix 3 — bad-config warning ⚠ had a false positive, fixed. As shipped in b8bf783, parseOutputSteering called clampVerbosityLevel(obj.verbosityLevel, warn) unconditionally — so an omitted verbosityLevel (the documented default case, e.g. just {"enabled": true}) hit the out-of-range branch and logged verbosityLevel must be an integer 0-4; got undefined on every config load. That accused valid configs of being malformed (honest-output violation, and contradicts the CONFIGURATION.md entry saying omitted fields fall back to defaults). Fixed with a new exported pure helper resolveVerbosityLevel(v): omitted ⇒ silent L2 default; present-but-invalid (9, -1, 2.5, "high", null) ⇒ fallback plus warning. Three regression tests pin both sides.

Also cleaned: b8bf783 left const wireTransform = makeWireTransform(true) at src/server.ts:3811 with zero consumers (both compress-loop call sites now take compressWireTransform) — removed the dead local and corrected the stale "#552 shared rewrite AND steering" comment to match actual behavior (re-sends carry compat + effort routing, verbosity stays initial-forward-only).

Fresh pass over the rest of the diff: all round-1 invariants still hold — skip-if-absent on all four wires, sentinel/tail/idempotent placement, clamp-only lowering (never injects, never toggles thinking.type, -1 dynamic budgets untouched), hook ordering after compat.roles including retry re-sends, default-OFF byte-transparency, both compression modes reasoned. My round-1 responses-classifier fix (backward scan + neutral item types) and the doc entries are intact.

Gates at b4df2a1: typecheck clean · build success · full suite 2197 pass / 0 fail / 2 skipped (the two env-gated e2e-codex skips, unchanged). CI is running on the new head — early checks green; will confirm here when it finishes.


复审了 b8bf783 的全部改动:fix 1(minimal 不再被抬到 low)与 fix 2(压缩轮免 verbosity 指令)验证无误;fix 3 的告警在字段缺省时误报(合法配置被当成非法并打 warn),已直接在分支上修复(presence-aware 的 resolveVerbosityLevel + 3 个回归测试)并顺手清掉一处死代码,本地三道门全绿,CI 绿后可合并。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 CI follow-up: all 8 checks green on head b4df2a1 (test suites on ubuntu-22.04/24.04 + windows, PR build artifact, version guard, codex overlay #686, real-codex zero-token e2e). PR state: open, mergeable_state: clean, no conflicts with master.

Full re-review of b8bf783 is complete and reported above — the only defect found (false-positive config warning on omitted verbosityLevel) is fixed in b4df2a1 with regression tests; everything else verified intact. Ready for a human to merge whenever you are.


CI 全绿(8/8),PR 无冲突可合并;复审发现的唯一缺陷已在 b4df2a1 修复并带回归测试,等人工合并即可。

@ranxianglei

Copy link
Copy Markdown
Owner Author

审查结论:实现质量过关,建议按 owner 拍板保留为参考实现 + 回归基准,待 acp-kernel#355 落地后收敛为 wire 落位层。

验证(rebased head)

硬边界逐条核对(全部落实 ✅)

  1. 只降不升 — lowerEffort 各 case 只写 floor;minimal / -1 / 已低于 floor 的值一律不动
  2. 客户端没发的字段不注入 — 所有 clamp 都以字段已存在为前提
  3. 不动 thinking.type — 只 clamp budget_tokens
  4. 只加尾部 — 四 wire 全部 tail-append,sentinel 幂等,换档原位替换不累积

与 #1092(锚定)的交互 — 无冲突

锚定在 prepare 内比较的是客户端发来的 fresh 并存为 lastSeen;verbosity 在 forward boundary 之后追加到出站 body,不回流进锚的比较基准,不会产生假 diff。锚头字节 + compress 注入 + steering 尾注三者逐轮字节稳定,前缀缓存兼容。

亮点

  • compress 重发剥离 verbosity 指令(compressWireTransform(false)):L2/L3 的 "never restate code/paths" 会直接对抗压缩提示的逐字保留契约,这个污染点想到了并在正确的层处理
  • Responses 分类器反向扫 input:codex 每请求全历史重发,正向扫永远先撞到原始提问、永远分不出续行轮——反向 + assistant 侧 item 不打断续行,是唯一正确的扫法

残留风险(不阻塞,记录在案)

  1. GOOGLE_MIN_THINKING_BUDGET = 128 是假定值(代码注释自标 "verify per-model")——真实流量上 Google 原生 wire 前需按模型核实,过低的 budget 在部分 Gemini 模型会 400
  2. 换 verbosityLevel 会原位替换 sentinel 块 = 该会话吃一次缓存 miss(已文档化,可接受)
  3. CHANGELOG 写 "29 regression tests",实际 tests/output-steering.test.ts 已有 35 个——纯文档滞后,随手改

定位确认

按 2026-09-22 拍板:轮次分类 / verbosity 档位 / effort 路由的决策逻辑下沉 acp-kernel#355,本 PR 保持打开作为规格与回归基准。届时按标准跨仓流程 bump 内核版本,再删除 src/output-steering.ts 中与内核重复的决策代码,收敛为「wire 落位 + 配置管道」。

ework-agent and others added 5 commits September 23, 2026 16:13
Port headroom's two output-token levers to bili's forward boundary (#1093):

- Verbosity steering: append a byte-stable conciseness directive (L0-L4,
  default L2) at the TAIL of the system prompt on all four wires
  (anthropic system, openai system/developer msg, responses instructions,
  google systemInstruction). Never prepends, never fabricates a system
  carrier when absent (skip-if-absent), idempotent so prefix cache stays hot.
- Effort routing: structural turn classification of the final message list
  (no content pattern-matching); lowers an already-present effort field to its
  floor on mechanical continuations only. Never injects an effort field the
  client did not send, never toggles thinking.type.

Default off; config in billion-context.json (global + per-provider route).
Lives at the forward boundary after compat.roles so launcher/native-plugin/
/bili/ lanes inherit it; compress-retry re-sends share the same transform.

Closes #1093
…ks (#1093)

Review round on PR #1096: the Responses turn classifier scanned the flat
input list FORWARD for any user signal. Clients re-send full history every
request (codex does), so the original ask was present on every turn and the
classifier returned new_user_ask unconditionally — effort routing could only
fire on delta-only inputs that never occur through bili. The other three
wires classify the LAST message correctly, so this made effort routing dead
on the flagship codex lane while appearing functional in tests (whose
synthetic inputs carry no prior user message).

- classifyResponsesTurn now walks BACKWARD from the tail: tool outputs after
  the last user signal = mechanical continuation; a user signal at the tail
  or with no outputs after it = new_user_ask. Assistant-side items between
  the last user signal and the trailing outputs no longer break the
  continuation (message/function_call/custom_tool_call/local_shell_call/
  apply_patch_call/reasoning — previously custom/local/apply-patch calls
  marked the turn unknown even in the delta case).
- Anthropic + Google classifiers: unrecognized block/part composition now
  returns unknown instead of being silently skipped (all four wires share
  one conservative stance: unrecognized structure => do not classify).
- 6 regression tests incl. the realistic full-history repro; steering suite
  23 -> 29, all green.
- Docs: outputSteering entries in CONFIGURATION.md + CONFIGURATION.zh-CN.md
  (global + per-provider), CHANGELOG [Unreleased] entry.
- treat explicit minimal effort as at-floor (clamp-only invariant: never
  raise a client-sent minimal to low on openai/responses)
- compress rounds re-send without the verbosity directive (summary
  fidelity: the never-restate clause contradicts the compress prompt's
  verbatim-preservation contract); effort routing stays
- warn (instead of silently clamping) on out-of-range verbosityLevel
…form (#1093)

b8bf783 made clampVerbosityLevel warn on any non-integer-0..4 value, which
includes undefined: every config that omits verbosityLevel (the documented
default case) logged a false 'must be an integer 0-4' warning at load.
resolveVerbosityLevel now warns only when the field is PRESENT but invalid
(honest output); omitted silently takes L2. Also removes the now-unused
wireTransform(true) local left behind when both compress-loop call sites
moved to compressWireTransform(false), and corrects its stale comment.
@ranxianglei
ranxianglei force-pushed the 2026-09-21_output-side-compression branch from b4df2a1 to 1986ffc Compare September 23, 2026 08:16
@ranxianglei

Copy link
Copy Markdown
Owner Author

决定调整:不走 bili 侧先行实现,直接走内核路线。

决策逻辑(轮次结构分类 / verbosity L0–L4 / effort clamp)全部在 acp-kernel#355 落地并发版,billion-context 只做 wire 侧落位 + 配置管道。

本 PR 关闭重做:等 acp-kernel#355 就绪、按标准跨仓流程 bump 内核版本后重新开 PR,新 PR 只保留:

  • 四 wire 的落位注入(尾部追加、skip-if-absent、sentinel 幂等)
  • effort 字段 clamp 写回(内核产出「是否降档 + 目标值」,bili 只写具体协议字段)
  • outputSteering 配置管道(global + per-provider overlay)
  • 本 PR 的 35 个回归测试大部分可直接复用——断言的是 wire 形状,不依赖决策代码住在哪

本实现的价值不丢:继续作为 acp-kernel#355 的规格参考与回归基准(内核 issue 已引用 PR 链接),审查结论见上一条评论(硬边界逐条核对通过、与 #1092 锚定无交互)。

ranxianglei pushed a commit that referenced this pull request Sep 24, 2026
Add two opt-in output levers at the proxy forward boundary (#1093):

- Verbosity steering: append a deterministic conciseness directive to the
  system-prompt tail (skip-if-absent, sentinel-idempotent, never prepends).
- Effort routing: clamp an already-sent effort field down on mechanical
  continuation turns (never injects an absent field, never raises, never
  toggles thinking.type).

Division of labor (per #1096 rework spec): ALL decision logic — turn
classification, verbosity-level resolution, lower-effort choice — lives in
acp-kernel (decideOutputSteering / resolveOutputSteeringConfig, pinned 0.0.85).
bili's src/output-steering.ts only LANDS those decisions on the four wires
(openai chat / responses / anthropic / google-native); no local cwst/level
classification is duplicated into bili.

Config: `compress.outputSteering` rides the standard three-level compress
cascade (global -> provider -> model) via mergeCompress; resolved + validated
by the kernel's resolveOutputSteeringConfig at the forward boundary. Default
off => byte-for-byte passthrough when disabled (no prefix-cache regression).

Covers both forward paths: the initial string-body forward and compress-retry
re-sends (via wireTransform). Launcher / native-plugin / /bili/ lanes inherit
it automatically since it lives at the shared forward boundary.

35 wire-shape regression tests (tests/output-steering.test.ts).
cskwork pushed a commit to cskwork/billion-context that referenced this pull request Oct 21, 2026
…ks (ranxianglei#1093)

Review round on PR ranxianglei#1096: the Responses turn classifier scanned the flat
input list FORWARD for any user signal. Clients re-send full history every
request (codex does), so the original ask was present on every turn and the
classifier returned new_user_ask unconditionally — effort routing could only
fire on delta-only inputs that never occur through bili. The other three
wires classify the LAST message correctly, so this made effort routing dead
on the flagship codex lane while appearing functional in tests (whose
synthetic inputs carry no prior user message).

- classifyResponsesTurn now walks BACKWARD from the tail: tool outputs after
  the last user signal = mechanical continuation; a user signal at the tail
  or with no outputs after it = new_user_ask. Assistant-side items between
  the last user signal and the trailing outputs no longer break the
  continuation (message/function_call/custom_tool_call/local_shell_call/
  apply_patch_call/reasoning — previously custom/local/apply-patch calls
  marked the turn unknown even in the delta case).
- Anthropic + Google classifiers: unrecognized block/part composition now
  returns unknown instead of being silently skipped (all four wires share
  one conservative stance: unrecognized structure => do not classify).
- 6 regression tests incl. the realistic full-history repro; steering suite
  23 -> 29, all green.
- Docs: outputSteering entries in CONFIGURATION.md + CONFIGURATION.zh-CN.md
  (global + per-provider), CHANGELOG [Unreleased] entry.
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.

Feature: Output-side compression (verbosity steering + effort routing)

1 participant