fix(server): scope the Claude context meter to the parent session

[SamGu-NRX]

Fixes #5942

Problem

The parent thread's Context Window meter counted tokens spent by its subagents, and could measure them against the largest context window in the agent tree instead of the session's own. A workflow with background agents drove the parent bar toward 100% while its own transcript stayed small.

Two people beyond the reporter confirmed it on the issue, one noting it "caused me to unnecessarily compact my session because I wasn't paying attention to how the parent context got so full." The meter is the signal for deciding whether to compact or start a fresh thread, so a wrong reading costs real work.

Evidence from a real session

A 13-turn Claude session (Fable, 1M window) that spawned 15 Luna scouts, read from its own provider event log:

claude/system/task_progress   subagent_type: theo-luna-context-scout
  child   usage.total_tokens : 141875
  parent  usedTokens         : 141875   <- the child's number, verbatim
          maxTokens          : 1000000

Twelve parent meter events in that thread came from child task_progress, and in all twelve the parent's usedTokens equals the child's total_tokens exactly. Each is followed by a drop of about 10,000 when the parent's next assistant snapshot arrives, which is the sawtooth described in #4650: the bar climbs on someone else's spend, then corrects.

The same log shows the second half of this PR is not hypothetical either. modelUsage at turn end carries both claude-fable-5[1m] at 1,000,000 and gpt-5.6-luna at 200,000, so a maximum across that map is a coin flip on which window the meter reports.

Fix

Both paths are in ClaudeAdapter.ts:

• normalizeClaudeTaskProgressTokenUsage folded each child's task_progress total into the parent's used count via Math.max. A child's tokens are spent in the child's own window, so they now advance totalProcessedTokens only. A follow-up commit answers review feedback: the running total is held back until it exceeds the parent's own usage (a child's first small tick would otherwise render as "3,000 used of 2,000 total"), and the floor-not-sum choice is stated at the accumulation site.
• Because child tokens now flow only into that running total, completeTurn had to stop replacing it with the parent's own smaller figure, which dropped the field from the snapshot entirely. Total processed is cumulative thread work, so it keeps the larger value. Both reviewers of this PR flagged this independently; it is included because the first change is what makes it reachable.
• Where turn completion falls back to result modelUsage for the window, it took the maximum across every entry, so one 1M subagent widened a 200k parent's denominator. It now prefers the session model's own entry. When getContextUsage() answers, its maxTokens still wins and neither change applies.

That lookup needs the id the SDK reports is actually serving the session, which is not always the selected one: sessions started without an explicit selection have none recorded, and a refusal retry swaps the model "persistent for the session" per the SDK's own docs. system/init and model_refusal_fallback therefore record it in a separate observedApiModelId, whose value matches how modelUsage is keyed, [1m] suffix included. It is deliberately not currentApiModelId: that field mirrors the last id passed to setModel, and sendTurn compares the user's selection against it, so writing a fallback there would make the next turn re-send the refused model. Its own test pins that.

One consequence, pinned by its own test: subagent usage arriving before the parent has reported any usage no longer emits a parent meter event, because there is no baseline to attach a running total to. Previously that event was built from the child's numbers, which is the bug itself. The meter is briefly empty rather than wrong, and the child's tokens still ride on task.progress.

Verification

vp test run src/provider/Layers/ClaudeAdapter.test.ts   87 passed (87)
tsgo --noEmit                                           no new errors
vp fmt                                                  clean

Nine regression tests, each verified failing on main and passing here:

• keeps subagent tokens out of the parent context meter (#5942) — the parent's own 4,200 survives a child's 900,000; before, usedTokens became 900,000.
• measures the meter against the session model's window, not a subagent's — a 200k session model alongside a 1M subagent reports 200,000; before, 1,000,000.
• uses the init model's window when no model was explicitly selected — covers the default-model path, which the second fix would otherwise miss.
• follows a persistent refusal fallback to the model that ran — after a refusal swap, reports the fallback's 1,000,000; before, the rejected model's 200,000.
• does not re-send a refused model on the next turn with the same selection — the second turn makes no setModel call at all. Caught by review; verified failing against the first version of this change.
• measures the window against the model a mid-thread switch selected — after switching from a 1M model to a 200k one, reports 200,000. Also caught by review on the commit that introduced the field.
• carries a subagent's running total into the parent's next snapshot — the child's 480,000 reaches totalProcessedTokens while the parent's usedTokens stays 3,000.
• keeps a subagent's running total when the parent turn completes — with a child at 900,000 and the parent's own result at 4,200, the final snapshot reports both; before, totalProcessedTokens disappeared entirely.
• holds the running total until it exceeds the parent's own usage — a child's 2,000 tick under a 3,000-token parent emits nothing; the 5,000 tick that follows lands. Verified failing without the guard.

One existing test, preserves oversized Claude result totals after task progress snapshots are recorded, emitted its task progress before the parent had any usage, so the new guard discarded it and the assertion passed from the following result alone. It now establishes a parent baseline first, so the task total it names is actually retained.

Three existing tests asserted the old behavior. Two counted an event that is no longer emitted; the third expected a child's 190,000 as the parent's usedTokens and now expects the parent's own result. A fourth was renamed to state the contract directly.

Overlap

• fix(claude): report subagent model and effort #7287 (fix(claude): report subagent model and effort) changes which model a subagent is labeled with; this changes whose tokens the parent meter counts and against which window. Different issues (Agents panel shows the parent session's model/effort for every subagent instead of the subagent's own #7281 vs [Bug]: Workflow/background agent tokens inflate main agent Context Window clock (parent should only show own context) #5942). I applied fix(claude): report subagent model and effort #7287 and this PR's source together on main: they typecheck, and the suite passes at 85 tests once the three tests above are reconciled. Only the shared test file collides. If fix(claude): report subagent model and effort #7287 lands first I will rebase within a day.
• fix: make Claude 200k context window selection take effect #8409 touches this file for the context window a session runs at; no overlapping hunks.
• fix(server): refresh Claude context meter after compact #7249 (size:XL, +2790) rewrites compaction and session lifecycle and edits the same token-normalization functions, so it is the one open PR with hunk-level conflict here. It covers the accounting half of this bug but leaves the modelUsage window selection untouched, so neither PR makes the other unnecessary. Whichever lands first, the other needs its accounting hunks reconciled; happy to rebase this one within a day.
• fix(claude): subagent usage no longer fills the context meter #6586 proposed deleting the task-progress usage path outright. It no longer applies to main.
• [Bug]: context meter ratchets up and never reflects /compact — task_progress writes cumulative tokens into usedTokens via Math.max #4650 turned out to be two bugs, and this PR fixes one of them. A commenter there traced the ratcheting meter to the same Math.max in the task-progress path this PR removes, and the issue was retitled accordingly; @t3dotgg linked [Bug]: Workflow/background agent tokens inflate main agent Context Window clock (parent should only show own context) #5942 as the canonical report for that half yesterday, and a fresh capture on 0.0.35 today notes the practical harm — the inflated meter convinces people autocompact is broken, so they compact sessions that are nowhere near full. The other half — compact_boundary emitting nothing when post_tokens is absent, so /compact never visibly resets the meter — is untouched here and is what fix(server): refresh Claude context meter after compact #7249 addresses.
• [Bug]: Codex Agents panel counts inherited parent history in each child token total #5793 is the same class of bug in CodexAdapter.ts (child totals carrying inherited parent history). Different file; not addressed here.
• [Bug]: Claude context meter jumps to 100% at end of turn - completeTurn falls back to cumulative session usage from result.usage #8594, filed today, is the third distinct fault in this meter: when getContextUsage() misses its one-second budget, completeTurn falls back to result.usage, which is cumulative for the session, and the clamp turns that into a flat 100%. Its author proposes preferring the last message_delta reading over the cumulative one. That is a different branch of completeTurn from either hunk here — this PR changes which modelUsage entry supplies the window, not which usage record is chosen — so the two are compatible, and [Bug]: Claude context meter jumps to 100% at end of turn - completeTurn falls back to cumulative session usage from result.usage #8594's fix would still want this one to pick the right denominator.

Checklist

• This PR is small and focused
• I explained what changed and why
• Screenshots for UI changes — no client files changed; the existing meter renders corrected server data
• Video for animation/interaction changes — not applicable

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 1822f7ba-3de6-494a-9d6d-7344a5d6b805

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

_{Comment @coderabbitai help to get the list of available commands.}

[macroscopeapp]

Approvability

Verdict: Approved at 50b5d09

Macroscope's review found this PR approvable — This is a contained Claude adapter bug fix that corrects parent token accounting and context-window selection while preserving existing model-selection behavior. Production changes are localized and accompanied by targeted regression tests, with no product-default, schema, security, billing, or deployment changes.

^{You can add or adjust custom eligibility rules. Learn more.}

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 1ef15f8. Configure here.}

Dismissing prior approval to re-evaluate 50b5d09