fix: normalize context usage categories to totalUsed and exclude autocompact buffer row

[lsm]

Normalizes non-free, non-buffer context usage categories so their token counts sum to totalUsed while preserving relative proportions, then recomputes percentages against the final totalCapacity. Recomputes free space as capacity - totalUsed - buffer and removes the autocompact buffer from the breakdown rows (it is rendered as a separate visual zone). Adds a regression test for categories scaled to a larger SDK window.

[lsm]

🤖 Review by glm-5.1[1m] (GLM)

Model: glm-5.1[1m] | Client: NeoKai | Provider: GLM

Recommendation: APPROVE — posted as a comment because GitHub rejects APPROVE/REQUEST_CHANGES from the PR author.

The normalization correctly fixes the reported issue (a single breakdown category exceeding totalUsed). Verified end to end:

• Normalization — non-free, non-buffer usage categories are scaled by totalTokens / sum(usage), with the rounding residual added to the largest category so the values sum exactly to totalUsed while preserving relative proportions. Hand-checked the arithmetic in all three touched cases; sums are exact and no category exceeds totalUsed.
• Buffer excluded — the autocompact category is filtered out of usageEntries, so it can no longer reappear as a usage row; it is rendered via the separate buffer zone driven by autoCompactThreshold.
• Threshold intact — the threshold/reserve derivation was hoisted above the breakdown build but its logic is unchanged (still uses the raw SDK category tokens), so normalization cannot corrupt autoCompactThreshold.
• Free space — recomputed as capacity − totalUsed − buffer, which is consistent with the buffer zone (used + free + buffer = capacity) and fixes the prior inconsistency where synthesized free space ignored the buffer.
• Tests — the regression test asserts all three required invariants (no category > totalUsed, non-free sum = totalUsed, free = capacity − totalUsed − buffer); the ContextUsageBar web tests were correctly left untouched since they use hardcoded fixtures independent of the fetcher.

Verification run:

• context-fetcher.test.ts — 48 pass
• ContextUsageBar.test.tsx — 54 pass
• context-tracker + sdk-message-handler — 94 pass
• typecheck / lint / knip — clean

The lone failure in the 1-core shard (provider-service.test.ts, ENABLE_TOOL_SEARCH env-var leak) is pre-existing and unrelated — context-fetcher.ts touches neither env vars nor provider config.

Zero findings (P0–P3).

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 41143fe81a

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[lsm]

🤖 Review by glm-5.1 (GLM)

Model: glm-5.1 | Client: NeoKai | Provider: GLM

Recommendation: APPROVE

Follow-up review on rebased head 11ab7904c (rebased onto dev tip e39b3d397, which now contains #2244's flaky-test fix). The 2-file effective diff vs dev is unchanged from the prior approval review (#4645684500):

• packages/daemon/src/lib/agent/context-fetcher.ts: the normalization block scales non-free usage categories to sum exactly to totalUsed (preserving relative proportions, rounding error pushed to the largest category). After normalization nonFreeSpaceTokens == totalUsed, so the autocompact-buffer-exclusion invariant free = capacity − totalUsed − buffer holds. Capacity selection, autocompact-threshold conversion, and the [1m] suffix handling are unchanged.
• packages/daemon/tests/unit/1-core/agent/context-fetcher.test.ts: regression test asserts non-free sum = totalUsed, each category ≤ totalUsed, and free = capacity − totalUsed − buffer for the kimi (262144 / 90584) case.

Verdict: zero P0–P3 findings. Code is correct and approved.

Pre-merge status: mergeStateStatus CLEAN, MERGEABLE, 32 checks SUCCESS / 4 SKIPPED (dev-only), zero unresolved review conversations. Proceeding to squash merge into dev.