[Question] Previously, Codex displayed the context window size after compression, but it now appears to show an unexpected/incorrect value.

[BsoBird]

This issue did not occur with versions 0.1.70–0.1.80. After upgrading to the latest version, the long context window size in Codex appears to display a strange number.

My conversation has actually taken up 20M of space, but the current conversation window displays: 1315/950k

I tried continuing the conversation, and the window statistics then became: 1313.2/950

Why is this happening?

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Thanks for the report — reproducing the reasoning on my side now. The display is used/total: total (950k) comes from the model context-window registry, and used comes from the token usage in the API response. I'm checking what changed in usage handling between 0.1.18 and the latest build; will report findings here shortly.

[BsoBird]

bili proxy log:

2026-09-08T10:48:31.027Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] nudge idle: usage=7% (74941/1050000), growth=3133/22500 (ref=71808, interval=50000), pendingT1=58775/50000 model=gpt-5.6-terra, reason="growth 3133 < floor 22500, ready: T1 58775, blocked: T1 (cadence)"
2026-09-08T10:48:31.060Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] host usage backfill armed: +1234768 tok (forwarded view is folded); host usage will report the uncompressed baseline
2026-09-08T10:48:31.063Z [info] forward POST → http://<private-host>:8787/v1/responses
2026-09-08T10:48:35.089Z [warn] [tag-echo] stripped model-emitted render tag: <acp tokens="52" type="text">
2026-09-08T10:48:42.253Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] [acp-usage] round 1 input=76895 cached=70315 (cache hit 91%)
2026-09-08T10:48:42.351Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] dropped 8 whitespace-only message item(s) before projection (flattened-turn artifact)
2026-09-08T10:48:42.442Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] processTurn: 152 msgs, renderTags=text-only, 21 text tagged, 0 tool tagged (should be 0 with text-only)
2026-09-08T10:48:42.442Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] nudge idle: usage=7% (76895/1050000), growth=5087/22500 (ref=71808, interval=50000), pendingT1=59997/50000 model=gpt-5.6-terra, reason="growth 5087 < floor 22500, ready: T1 59997, blocked: T1 (cadence)"
2026-09-08T10:48:42.470Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] host usage backfill armed: +1234757 tok (forwarded view is folded); host usage will report the uncompressed baseline
2026-09-08T10:48:42.476Z [info] forward POST → http://<private-host>:8787/v1/responses

[BsoBird]

in old version, i think this context window show same with nudge idle,example: usage=7%
but now seems not

[BsoBird]

-- I'm checking what changed in usage handling between 0.1.18
not 0.1.18, is 0.1.70~0.1.80

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Thanks for the correction — 0.1.70–0.1.80 is the key data point, and it lines up exactly: the change that broke this is the host usage backfill (issue #408), which shipped in v0.1.88.

What it does (your log shows it directly):

• nudge idle: usage=7% (74941/1050000) — the proxy's internal ledger sees the real post-compress size: ~75k tokens.
• host usage backfill armed: +1234768 tok (forwarded view is folded) — the proxy then adds the folded-out tokens back into the input_tokens it reports to the client, so Codex receives ~75k + 1.23M ≈ 1315k instead of 75k.
• Codex divides that by its own window (950k) → 1315/950k (138%).

So the "used" number is now the uncompressed baseline (what your context would be without bili), not the post-compress size the old versions reported — which is why it no longer matches the nudge's usage=7%.

I'm now checking how this interacts with Codex's native auto-compact (that was the stated reason to keep the backfill for Codex, unlike pi/omp which got exempted in #590/#623) and will report the fix direction here.

[BsoBird]

so,which value should we use?
currently,when i append chat,metrics can decrement with no acp compress trigger.
seems strange

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Recommendation: Codex should get the post-fold (actually-forwarded) usage — the pre-0.1.88 behavior. I verified the mechanics end-to-end before answering, here's why:

Why the current number is wrong for Codex

1. It's a virtual number. The backfilled input_tokens (≈1315k) is the uncompressed baseline — what your context would be without bili. The request actually forwarded to the model is ~75k (your log: usage=7% (74941/1050000)). A client that displays used/window can only be sane if used is what the model really received.
2. It's unstable — explains your "decrements with no compress" observation. The reported value = real upstream usage (post-fold) + a character-based estimate of the folded-out tokens (est(original) − est(processed)). The estimate drifts turn to turn (different content mix, different folded ranges), so the sum can go down on a plain append (1315 → 1313.2) with no compression at all. The post-fold value has no such drift: it only moves when the model actually consumes more/less, and drops only when a compress really happens.
3. It exceeds the window → 138%, which looks broken and can trip client-side overflow heuristics.
4. It doesn't even drive Codex's auto-compact, so it currently buys nothing. I checked codex-rs: the compaction decision (context_window.rs → token_limit_reached) keys off usage.total_tokens, and bili's backfill only patches input_tokens — total_tokens passes through un-backfilled. So today the baseline number produces the broken display and no functional benefit.

Precedent: pi (#590, showed 302.7%) and omp (#623, showed 205%) were already exempted from this exact backfill for this exact >100%-footer symptom. Codex is the last host still receiving it.

What stays intact if we exempt Codex

• The uncompressed baseline remains visible where it belongs: the /acp panel (host baseline line, hostContextTokens) and the proxy log.
• All safety nets: nudge (proactive) → preflight (hard limit, judged on the real post-fold payload) → Codex's own native auto-compact, which will now see the true post-fold number and fire at 90% of its window — the correct backstop timing (that's the design from REQ: codex 压缩预算协调 + 匿名代理模式拦截伪造（根本解决方案） #321: ACP compresses first at ~55%×W, Codex native compaction backstops only when the real request gets near-full).

The change (small, targeted): in armHostUsageCredit (src/server.ts:1819), extend the pi/omp exemption to Codex, detected by the same UA signal the native-compact interception already uses (codex_cli_rs/ / codex_exec/) — covers both bili codex and plain Codex pointed at the proxy. Plus a regression test mirroring the #590/#623 ones.

One secondary finding while digging: your Codex shows a 950k window, but the proxy's nudge uses 1050k (the models.dev registry value for gpt-5.6-terra). Are you running bili codex (the launcher should hand your config.toml's model_context_window to the proxy) or plain Codex with base_url pointed at the proxy? Your log suggests the latter (nudge uses the registry value). If it matters to you, we can reconcile that too — but it doesn't change the recommendation above.

Shall I go ahead and implement the Codex exemption (PR)?

[BsoBird]

okay,go ahead

[BsoBird]

tips: i use Option 2 — URL change (/bili/ prefix)

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Thanks — that confirms the setup: plain Codex pointed at the proxy via the /bili/ URL prefix (not the bili codex launcher). That also explains the second discrepancy: with the launcher, your config.toml's window would be handed to the proxy, but with a URL change the proxy falls back to the models.dev registry value (1050k for gpt-5.6-terra) while Codex keeps its own 950k — cosmetic for the nudge math, but worth knowing.

Good news: the fix I'm implementing covers exactly your setup — the exemption is detected by Codex's User-Agent (the same signal the native-compact interception uses), so it works for plain Codex via URL change, bili codex, and anything else that identifies as Codex. No launcher or header involvement. I'll post the PR here once tests pass.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Done — implemented per the recommendation above. PR: #647

What changed: armHostUsageCredit (src/server.ts) now exempts Codex, same gate as pi (#590) and omp (#623). Codex gets the post-fold (actually-forwarded) usage — the pre-0.1.88 behavior — instead of the #408 uncompressed-baseline backfill.

• Detection is by User-Agent (codex_cli_rs/, codex_exec/) — the same signal native-compact interception uses. This covers your setup (plain codex + /bili/ URL) and bili codex; the x-bili-plugin header can't be used because launcher sessions register as mcp (shared with claude).
• The backfilled number was wrong for codex on four counts: it's virtual (the model really receives the folded ~75k), it drifts turn-to-turn (which is why your metric decremented with no compress firing), it exceeded the window (1315/950k = 138%), and it drove nothing — codex's auto-compact keys off total_tokens, which the backfill never touches.
• Unchanged: ACP compression still runs on the real number (nudge → preflight), codex's native auto-compact stays the backstop at 90% of its window, and the uncompressed baseline is still visible in the /acp panel and the proxy log.

Tests: two new regression tests on the responses wire with a real kernel fold — codex UA reports the folded usage, and an identical non-codex client still gets the #408 backfill (control). typecheck clean, full suite green (one pre-existing environment-dependent launcher failure on clean master, unrelated), build OK.

One side note, now mostly moot: your codex shows a 950k window while the proxy's nudge line uses 1050k (the models.dev registry value for gpt-5.6-terra). With plain codex + /bili/ URL there's no window handoff, so each side uses its own table. After this fix the numerator is the real post-fold usage, so the two denominators no longer produce a >100% display — but if you want the numbers to agree exactly, setting model_context_window in codex's config.toml to the same value the proxy uses will align them.

The fix ships with the next release; until then, BILI_CODEX_COMPACT=pass doesn't help here (that's the native-compact kill-switch, unrelated) — upgrading after the release is the path.

[BsoBird]

now i try use patched version
bili --version
0.1.95-pr.649.467
still has problem 1408.5/950

log

2026-09-08T13:11:03.877Z [info] forward POST → http://<private-host>:8787/v1/responses
2026-09-08T13:11:12.548Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] [acp-usage] round 1 input=79000 cached=72231 (cache hit 91%)
2026-09-08T13:11:12.620Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] dropped 8 whitespace-only message item(s) before projection (flattened-turn artifact)
2026-09-08T13:11:12.686Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] processTurn: 114 msgs, renderTags=text-only, 16 text tagged, 0 tool tagged (should be 0 with text-only)
2026-09-08T13:11:12.686Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] nudge idle: usage=8% (79000/1050000), growth=63031/22500 (ref=15969, interval=50000), pendingT1=43311/50000 model=gpt-5.6-terra, reason="max compressible 43311 < threshold 50000"
2026-09-08T13:11:12.711Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] host usage backfill armed: +1325970 tok (forwarded view is folded); host usage will report the uncompressed baseline
2026-09-08T13:11:12.714Z [info] forward POST → http://<private-host>:8787/v1/responses
2026-09-08T13:11:32.536Z [info] [01a07bd1-6d97-7be3-8df4-504f9a78eab8] processTurn: 16 msgs, renderTags=text-only, 13 text tagged, 0 tool tagged (should be 0 with text-only)
2026-09-08T13:11:32.536Z [info] [01a07bd1-6d97-7be3-8df4-504f9a78eab8] nudge idle: usage=2% (17175/1050000), growth=-2883/22500 (ref=20058, interval=50000), pendingT1=0/50000 model=gpt-5.6-terra, reason="max compressible 0 < threshold 50000; growth -2883 < floor 22500"
2026-09-08T13:11:32.537Z [info] forward POST → http://<private-host>:8787/v1/responses
2026-09-08T13:11:39.912Z [info] [01a07bd1-6d97-7be3-8df4-504f9a78eab8] [acp-usage] round 1 input=17269 cached=15816 (cache hit 92%)
2026-09-08T13:11:41.220Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] [acp-usage] round 1 input=79352 cached=74208 (cache hit 94%)
2026-09-08T13:11:41.294Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] dropped 8 whitespace-only message item(s) before projection (flattened-turn artifact)
2026-09-08T13:11:41.372Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] processTurn: 117 msgs, renderTags=text-only, 16 text tagged, 0 tool tagged (should be 0 with text-only)
2026-09-08T13:11:41.372Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] nudge idle: usage=8% (79352/1050000), growth=63383/22500 (ref=15969, interval=50000), pendingT1=43311/50000 model=gpt-5.6-terra, reason="max compressible 43311 < threshold 50000"
2026-09-08T13:11:41.398Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] host usage backfill armed: +1325970 tok (forwarded view is folded); host usage will report the uncompressed baseline
2026-09-08T13:11:41.401Z [info] forward POST → http://<private-host>:8787/v1/responses
2026-09-08T13:11:57.411Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] [acp-usage] round 1 input=80769 cached=75181 (cache hit 93%)
2026-09-08T13:11:57.493Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] dropped 8 whitespace-only message item(s) before projection (flattened-turn artifact)
2026-09-08T13:11:57.578Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] processTurn: 120 msgs, renderTags=text-only, 16 text tagged, 0 tool tagged (should be 0 with text-only)
2026-09-08T13:11:57.578Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] nudge idle: usage=8% (80769/1050000), growth=64800/22500 (ref=15969, interval=50000), pendingT1=45883/50000 model=gpt-5.6-terra, reason="max compressible 45883 < threshold 50000"
2026-09-08T13:11:57.600Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] host usage backfill armed: +1325970 tok (forwarded view is folded); host usage will report the uncompressed baseline
2026-09-08T13:11:57.603Z [info] forward POST → http://<private-host>:8787/v1/responses
2026-09-08T13:12:02.385Z [warn] [tag-echo] stripped model-emitted render tag: <acp tokens="488" type="text">
2026-09-08T13:12:12.843Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] [acp-usage] round 1 input=82532 cached=76154 (cache hit 92%)

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 You tested the wrong build — that's why it still shows 1408.5/950.

0.1.95-pr.649.467 is the prerelease for PR #649, which is a different fix (a hostUsageCredit: "off" config switch for zcode, issue #648). I checked its bundled code — its armHostUsageCredit has the config switch but no codex check, so the backfill still fires for your codex session (exactly the host usage backfill armed: +1325970 tok line in your log).

The codex fix is PR #647, and its prerelease is already published:

npm install -g billion-context@pr-647
bili --version   # → 0.1.95-pr.647.464

I verified 0.1.95-pr.647.464 contains the fix — its armHostUsageCredit has if (isCodexClient(headers)) return; (the codex UA exemption).

What to expect after switching to pr-647:

• The host usage backfill armed: +N tok line should disappear for your codex sessions (it's the smoking gun — if you still see it, the fix isn't active).
• Codex's context metric should drop from 1408/950 to the real post-fold usage — in your log that's ~79k ([acp-usage] round 1 input=79000), i.e. roughly 79/950, not 148%.

Two notes:

1. The pr-649 build could also silence it if you set hostUsageCredit: "off" in the bili config — but that's a manual global opt-out (affects every host), not the codex-specific auto-exemption. pr-647 is the targeted fix.
2. Heads-up for the merge: PR fix #645: suppress #408 host-usage backfill for codex (report folded usage) #647 (codex auto-exempt by UA) and PR fix #648: add hostUsageCredit switch to opt out of #408 host-usage backfill #649 (manual hostUsageCredit switch) both touch armHostUsageCredit with different signatures, so they'll need a small integration when both land. They're complementary, not conflicting in intent.

[BsoBird]

seems not okay

2026-09-08T13:34:11.517Z [info] forward POST → http://<private-host>:8787/v1/responses
2026-09-08T13:34:17.754Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] [acp-usage] round 1 input=42719 cached=40146 (cache hit 94%)
2026-09-08T13:34:17.831Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] dropped 8 whitespace-only message item(s) before projection (flattened-turn artifact)
2026-09-08T13:34:17.886Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] processTurn: 52 msgs, renderTags=text-only, 7 text tagged, 0 tool tagged (should be 0 with text-only)
2026-09-08T13:34:17.886Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] nudge idle: usage=4% (42719/1050000), growth=16692/22500 (ref=26027, interval=50000), pendingT1=12966/50000 model=gpt-5.6-terra, reason="max compressible 12966 < threshold 50000; growth 16692 < floor 22500"
2026-09-08T13:34:17.907Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] host usage backfill armed: +1376387 tok (forwarded view is folded); host usage will report the uncompressed baseline
2026-09-08T13:34:17.909Z [info] forward POST → http://<private-host>:8787/v1/responses
2026-09-08T13:34:38.197Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] [acp-usage] round 1 input=42964 cached=40146 (cache hit 93%)
2026-09-08T13:34:38.282Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] dropped 8 whitespace-only message item(s) before projection (flattened-turn artifact)
2026-09-08T13:34:38.337Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] processTurn: 55 msgs, renderTags=text-only, 7 text tagged, 0 tool tagged (should be 0 with text-only)
2026-09-08T13:34:38.337Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] nudge idle: usage=4% (42964/1050000), growth=16937/22500 (ref=26027, interval=50000), pendingT1=15628/50000 model=gpt-5.6-terra, reason="max compressible 15628 < threshold 50000; growth 16937 < floor 22500"
2026-09-08T13:34:38.357Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] host usage backfill armed: +1376387 tok (forwarded view is folded); host usage will report the uncompressed baseline
2026-09-08T13:34:38.359Z [info] forward POST → http://<private-host>:8787/v1/responses
2026-09-08T13:35:10.513Z [info] [01a038c1-35e8-7811-b975-4ceae655e242] [acp-usage] round 1 input=45595 cached=41818 (cache hit 92%)
2026-09-08T13:35:35.466Z [info] [update] throttled — last checked 170s ago, retry in 9s
2026-09-08T13:38:35.483Z [info] [update] checking npm registry for billion-context (last check 350s ago)…
2026-09-08T13:38:36.771Z [info] [update] new version found: 0.1.95-pr.647.464 → 0.1.95, downloading…
2026-09-08T13:38:36.772Z [warn] [update] install failed: install dir not writable: /usr/lib/node_modules/billion-context. Will retry next cycle.
2026-09-08T13:41:35.500Z [info] [update] checking npm registry for billion-context (last check 180s ago)…
2026-09-08T13:41:37.571Z [info] [update] new version found: 0.1.95-pr.647.464 → 0.1.95, downloading…
2026-09-08T13:41:37.572Z [warn] [update] install failed: install dir not writable: /usr/lib/node_modules/billion-context. Will retry next cycle.

still nor correct: 1422.0/950

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 The backfill is still firing in this log (host usage backfill armed: +1376387 tok), which means the running bili process is not actually executing pr-647's code — even though pr-647 is now on disk.

Your own log proves the live process is old:

[update] new version found: 0.1.95-pr.647.464 → 0.1.95, downloading…
[update] install failed: install dir not writable: /usr/lib/node_modules/billion-context

• The on-disk version is 0.1.95-pr.647.464 — the updater read it from /usr/lib/node_modules/billion-context/package.json, so the install did land.
• But the live process is trying to "upgrade" it to 0.1.95. pr-647's own updater would never do that. Its version compare looks only at the numeric 0.1.95 parts (equal → not newer), so it would not attempt this upgrade. A process that thinks 0.1.95 beats 0.1.95-pr.647.464 is running older bili code — it was started before the pr-647 install and never restarted.
• It can't self-heal because /usr/lib/node_modules/billion-context isn't writable by the running user.

So: pr-647 is on disk, but the live process is still the old (buggy) code.

Fix — restart bili so it loads pr-647 from disk:

1. Stop the running bili process.
2. Start it again.
3. Check the log: the host usage backfill armed: +N tok line should be gone for your codex sessions, and Codex's metric should drop from 1422/950 to the real post-fold value (~43/950, per your [acp-usage] input=42719).

Two caveats:

• Don't let it "upgrade" to 0.1.95. That stable build does not contain the codex fix (it predates PR fix #645: suppress #408 host-usage backfill for codex (report folded usage) #647). pr-647 on disk is the correct build — just restart into it. (The install dir not writable error is why the auto-updater can't touch it; that's a permissions issue on your machine, separate from this bug.)
• If, after a clean restart on a confirmed pr-647, the host usage backfill armed line still appears, then it's a codex UA-detection miss in your specific setup — send me the exact User-Agent value your codex sends (not sensitive) and I'll track it down. But I expect the restart resolves it.