fix(#4985): statusline meter reads 100% where Claude Code auto-compacts, and the full window when auto-compact is off

[0xdhx]

Fix PR

Using the wrong template?
— Enhancement: use enhancement.md
— Feature: use feature.md

---

Linked Issue

Required. This PR will be auto-closed if no valid issue link is found.

Fixes #4985
Fixes #4844

The linked issue must have the confirmed-bug label. If it doesn't, ask a maintainer to confirm the bug before continuing.

Supersedes #4959, at a maintainer's request: trek-e, 2026-10-07 asked for one PR that supersedes #4959 and carries the statusline fix.

Credit. The first seven commits are @tanji-dg's #4959 (the #4844 fix), carried with their authorship and rebased onto next. Their content is unchanged; the first one's subject now takes the fix(#4844): form. Their changeset is kept and credits the contribution. The 17 commits after them are the #4985 fix built on top, which reuses #4959's isAutoCompactDisabled for the disabled case, and the changes from the 2026-10-09 review.

The review findings on #4959 are addressed in the carried commits. #4959's head d21d3d1 (pushed 2026-10-07 02:41Z) addressed all three findings from the 2026-10-07 01:23Z review:

• DISABLE_AUTO_COMPACT accepted values. Read from the Claude Code 2.1.292 binary: the second operand of Boolean(Le(DISABLE_COMPACT) || a.DISABLE_AUTO_COMPACT) is the env registry's D.bool() field, which runs the same ["1","true","yes","on"] helper. So both keys share one truthy set. The source is stated beside the constants and in the test.
• tests/perf-317-context-monitor-fs.test.cjs. It sent the invented context_window.total_tokens; it now sends context_window_size.
• The PR-body count. Remaining 16.5 renders 84% (round(83.5)), not 83%.

Two of #4844's acceptance criteria are superseded here, on purpose. #4844's AC2 (enabled behaviour unchanged) and AC3 (the bridge stays raw) were written with the enabled-case math out of scope. #4985 is that scope, and its fix changes both: the enabled case now measures against the real threshold, and the bridge moves to the threshold scale (option a). #4844's AC1 holds. With auto-compact disabled the meter shows the used share of the model window, which is round(100 − remaining) with no buffer (raw 40 reads 40; at the old buffer boundary, raw 83.5 reads 84).

Maintainer confirmation needed: the bridge scale. Option (a) reverses #2451's raw-bridge contract and #4844's AC3. @tanji-dg agreed to the supersede; that is not a maintainer's agreement to the scale change, and this PR should not merge without one.

---

What was broken

Two halves of the same meter block in hooks/gsd-statusline.js:

• bug(statusline): context meter subtracts the auto-compact buffer even when autoCompactEnabled is false #4844, auto-compact off. The meter always subtracted the auto-compact buffer, so with compaction disabled the bar read high and pinned at 100% while context remained (raw 40% used displayed as 48%).
• bug(statusline): context meter's 100% is not the auto-compact point — reads absent context_window.total_tokens, ignores autoCompactWindow #4985, auto-compact on. The meter's 100% was not where Claude Code compacts. It read a payload field Claude Code does not send (context_window.total_tokens, so the window was always the 1M fallback), ignored the autoCompactWindow setting /autocompact saves, and put 100% at the window itself instead of the threshold. With /autocompact 650k on a 1M model it read 74% at the moment of compaction, and gsd-context-monitor.js stayed silent because the bridge's remaining_percentage was on the model-window scale (38% remaining when compaction fired).

What this fix does

The meter measures usage against the token count at which Claude Code auto-compacts, resolved the way Claude Code resolves it, so it reaches 100% there (rounded: within half a percent of it). When compaction is off, the denominator is the model window. The bridge counts down to the same point, so the context monitor warns before compaction.

• Window: CLAUDE_CODE_AUTO_COMPACT_WINDOW (raised to 100k, capped at 1M; an unparseable or non-positive value is ignored), else the autoCompactWindow setting, else context_window_size. Always capped at the model window.
• Threshold: window − min(maxOutput, 20000) − 13000, lowered to floor(effective × pct/100) by CLAUDE_AUTOCOMPACT_PCT_OVERRIDE. A statusline cannot see maxOutput, so the hook takes the 20,000 cap; on a model whose output limit is below 20,000, Claude Code compacts slightly later than the meter's 100%. /context shows the 33k as "Autocompact buffer". The hardcoded 16.5% this replaces is 33,000 / 200,000, so it was right only on a 200K window with nothing configured.
• Usage: current_usage.input_tokens + cache_creation_input_tokens + cache_read_input_tokens, the sum /context reports, measured against the threshold. This avoids re-scaling remaining_percentage, which is already rounded against the model window and loses up to 5,000 tokens on a 1M model.
• Per-model setting (Claude Code 2.1.292). /autocompact now saves the window under modelSettings.<canonical model name>.autoCompactWindow, not the top-level key earlier versions wrote. Reading only the top-level key would leave this issue's headline case unfixed on a current Claude Code. The hook mirrors the binary's merge (lFt() / bv()):
  • Settings files are read lowest precedence first. Each file's per-model entries layer over the lower files' entries.
  • A file that sets a top-level autoCompactWindow replaces the default and drops the lower files' per-model entries.
  • The session model's own entry wins over the default. The model is the payload's model.id, canonicalized across its dated, [1m], Bedrock and Vertex spellings. Canonicalizing is idempotent (a key is its own key), and a key longer than 256 characters once lowercased matches no model, which bounds the work a hostile settings key can cause.
  • "auto" falls back to the model window.
• Settings files are untrusted input. The project's .claude/ is repository content, and the statusline reads the settings files on renders that draw the meter (unless an env var already settles it). A file is read only when it is a regular file of at most 1 MiB. It is opened non-blocking and checked through its own descriptor, so a FIFO is skipped instead of blocking the render. The read stops one byte past the checked size, and a file that grew or shrank after the check is treated as absent. A symlink is followed, since a settings file linked in from a dotfiles repo is ordinary.
• Bridge, option (a) from the bug(statusline): context meter's 100% is not the auto-compact point — reads absent context_window.total_tokens, ignores autoCompactWindow #4985 thread. remaining_percentage / used_pct are on the threshold scale, and the bridge gains used_tokens / threshold_tokens. The monitor gates on remaining_percentage. On the raw scale its 35% / 25% sit at a different distance from compaction for every configured window, and with a window well below the model's (the 650k-on-1M case) compaction fires before the warning does. On the threshold scale they measure runway to compaction whatever the window. The /context agreement fix(#2315): respect review.default_reviewers in bare convergence invocation #2451 protected moves to the token counts, which the monitor's message now quotes: CONTEXT WARNING: Usage at 65% (401k of 617k usable tokens). The monitor's 35% / 25% are unchanged; they now measure runway to compaction, per the triage scope. The token-floor idea from the issue stays out of this PR.

Root cause

• bug(statusline): context meter subtracts the auto-compact buffer even when autoCompactEnabled is false #4844. The buffer subtraction was never gated on whether auto-compaction was on.
• bug(statusline): context meter's 100% is not the auto-compact point — reads absent context_window.total_tokens, ignores autoCompactWindow #4985.
  • The meter read an absent field.
  • It recognized the window only through the env var.
  • It treated the window as the compaction point.
  • The bridge stayed raw, which kept the monitor on a different scale from the compaction point.

Testing

How I verified the fix

The three cases from #4985's repro script, run against this branch's hooks/:

Case	Compaction threshold	Before	After
1. autoCompactWindow: 650000 in settings, 1M model, 616,927 used	617,000	74%, monitor silent	100%, CONTEXT CRITICAL: Usage at 100% (617k of 617k usable tokens)
2. CLAUDE_CODE_AUTO_COMPACT_WINDOW=150000, 200K model, 100,000 used	117,000	100%	85%
3. CLAUDE_CODE_AUTO_COMPACT_WINDOW=650000, 1M model, 616,927 used	617,000	95%	100%

Claude Code 2.1.292's resolution, re-read from the shipped binary (minified names):

• YV(e,n) is the threshold: e − 13000, or min(floor(e × pct/100), e − 13000) under the pct override.
• U3 is the effective window: window − min(maxOutputTokens, 20000).
• bv resolves the window in this order:
  1. env (≥ 100k, capped at the model window)
  2. settings: byModel[settingsKey], else the default. "auto" falls through.
  3. clientdata
  4. experiment
  5. model-default table
  6. model window
• The compaction trigger is YV. The newer eLe (a precompute buffer fraction) only feeds the earlier summary-precompute check.

Suites at this head (d50621d0b), run locally on Linux with HOME sandboxed and CLAUDE_CONFIG_DIR cleared:

• tests/gsd-statusline.test.cjs and tests/perf-317-context-monitor-fs.test.cjs: 394/394
• CI's 12 tree-level required jobs, each run with its CI command, all pass: changeset-lint, docs-lint, default-flip-documentation, Detect test scope, plugin-validate, security, lint-tests (npm run lint:ci), the targeted lane (926/926), the QA ratchet (162/162), and the three unit shards, whose first also runs the integration, security, install and slow suites (44,246 tests, 0 failed, 45 skipped, all platform- or install-gated and none in this PR's files)
• Not reproducible locally, so left to CI: the Windows and macOS conformance legs, the packed-tarball smoke installs, the merged-shard coverage gate, Stryker, and the PR-metadata checks

Regression test added?

• Yes — added a test that would have caught this bug
• No — explain why:

The new #4985 tests fail on the pre-fix hooks:

• Threshold resolution: precedence, validation and caps, CLAUDE_AUTOCOMPACT_PCT_OVERRIDE, and the per-model modelSettings merge.
• The bridge scale.
• The monitor quoting token counts.
• An end-to-end hook run with model.id in the payload. 617k (100%) is due. On next's hooks the meter reads 74%; before the per-model fix (13ccef5f8) the window resolved to 967k and the meter read 64%.

From the 2026-10-09 review:

• Boundaries at limit−1, limit and limit+1 for the window bounds (100k and 1M, in the setting and the env var), the pct override (0, 1, 100, 101) and the threshold's positive edge.
• fast-check properties for the env parser, canonicalModelKey (idempotent, spelling-invariant), the threshold (never falls as the window grows, never above the window less the 33,000 buffer, only ever lowered by the pct override), the meter (used + remaining = 100, its percentage within 0.5 below the threshold and exactly 100 at or above it) and the settings merge, which is checked against an independent reference model rather than the production helpers.
• Hostile settings: a FIFO, a file over 1 MiB, a file that grows or shrinks after the size check, and an over-long model key.

The #2451 assertions in gsd-statusline.test.cjs and perf-317-context-monitor-fs.test.cjs are rewritten, not dropped. The bridge's used_tokens must equal what /context reports as used, and used_pct + remaining_percentage = 100.

One intermediate commit is red at its own tree. a710ed88e moves the bridge to the threshold scale, and five of perf-317's #2451 raw-bridge assertions fail there until the next commit, 0af4c9fcc, moves them to the new scale. A squash merge makes this moot; I can restructure the history if you'd rather keep the commits.

Platforms tested

• macOS
• Windows (including backslash path handling)
• Linux
• N/A (not platform-specific)

Run locally on Linux only. Both changed test files are on CI's Windows (platform-conformance-tier.generated.cjs) and macOS (macos-conformance-tier.generated.cjs) conformance lists, so those legs run them (the FIFO case skips on Windows); their result is CI's, not a local run.

Runtimes tested

• Claude Code
• Antigravity
• OpenCode
• Other: ___
• N/A (not runtime-specific)

---

Checklist

• Issue linked above with Fixes #NNN — PR will be auto-closed if missing
• Linked issue has the confirmed-bug label
• Fix is scoped to the reported bug — no unrelated changes included
• Regression test added (or explained why not)
• All existing tests pass (npm test)
• .changeset/ fragment added if this is a user-facing fix (npm run changeset -- --type Fixed --pr <NNN> --body "...") — or no-changelog label applied
• No unnecessary dependencies added

Breaking changes

None to an API. A reviewer will see these behaviour changes in the test diff:

• No-window default. At 50% of a 1M window the meter reads 52%, not 60%: 100% is now 967k, not the 16.5% share.
• Bridge scale. used_pct / remaining_percentage are on the threshold scale (option a), and the bridge gains used_tokens / threshold_tokens. The context monitor is the bridge's only runtime reader in this repo (two test files read it too).
• Payload fallbacks.
  • A payload without current_usage recovers tokens from the percentage.
  • A payload without context_window_size falls back to Claude Code's own percentage, unscaled.
• Project settings are read from workspace.project_dir (Claude Code's launch directory), not current_dir.
• isAutoCompactDisabled's settings walk moved into a shared reader. It now also skips a settings file that is not a regular file, is over 1 MiB, or changes size while it is read.

Stated limits. These fall back to the model window because a statusline cannot see them:

• the --autocompact launch flag
• managed-policy and --settings files
• Claude Code's clientdata, experiment and per-model default windows, which is what "auto" resolves to
• autoCompactEnabled stored in the legacy global config (~/.claude.json), which 2.1.292 still reads as a fallback source; the disabled check reads settings files only

Domain census. Settings files are read in this order, highest precedence first:

1. project settings.local.json
2. project settings.json
3. user settings.local.json
4. user settings.json

The window merge walks them in reverse. Not reached: managed-policy and --settings files.

[tanji-dg]

Thanks for carrying this forward, @0xdhx. For the record, d21d3d1 (02:38Z) and my 02:47Z comment on #4959 already addressed trek-e's 01:23Z findings; happy for #5252 to proceed as the superseding PR.

[trek-e]

Automated maintainer review of head 87d7efa7b7. Zero-tolerance policy: every finding at any severity must be resolved before approval.

Review PR #5252 — fix(#4985): statusline meter reads 100% where Claude Code auto-compacts, and the full window when auto-compact is off

Head 87d7efa (fresh gh pr view); mergeStateStatus BEHIND, mergeable MERGEABLE, not draft, reviewDecision REVIEW_REQUIRED.
CI: only non-success rollup entries are SKIPPED (auto-merge, inert CI, emitted-baseline publish); no failing checks.

Summary

Carries tanji-dg #4959 (#4844, 7 commits, authorship kept) plus 0xdhx #4985 fix: resolveAutoCompactThreshold / contextMeter / per-model autoCompactWindow merge in hooks/gsd-statusline.js; bridge moved to threshold scale (+used_tokens/threshold_tokens); monitor message quotes tokens. Root cause addressed at the right site (meter block + bridge writer). tanji-dg comment (data): consents to supersede; no effect on verdict.

Acceptance checklist

#4844 (confirmed-bug) AC:

1. disabled => used == round(100-remaining) at mid and at 16.5 boundary: MET (isAutoCompactDisabled -> limit=model window; tests added).
2. enabled/default unchanged incl. custom-window override: NOT MET by design (default now 52% not 60% at 50%/1M). PR body declares superseded; #4985 scope requires it.
3. bridge stays raw: NOT MET by design (option (a)). "Out of scope" in #4844 triage lists the bridge raw contract. #4985 triage allows scaling the monitor thresholds. Option (a) was proposed by author; maintainer never replied (author said "no reply needed"). Judgment: maintainer must confirm.
   #4985 (confirmed-bug) AC:
4. settings-only autoCompactWindow => 100% at real compaction point: MET (650k/1M, 616,927 -> 100%, test asserts threshold_tokens 617000).
5. monitor can warn/critical before compaction at any window size: MET (bridge on threshold scale; end-to-end monitor test).
   Out of scope respected: token-floor idea excluded.
   Scope: two issues, one PR, at maintainer request (trek-e 2026-10-07 on #4959). OK vs RULESET.PR-SCOPE.

Findings

1. MAJOR | deterministic | tests/gsd-statusline.test.cjs (new describe blocks, ~L'context meter resolves...') | RULESET.TESTS.property-based-testing: parseAutoCompactWindowEnv, canonicalModelKey, readAutoCompactWindowSetting merge and resolveAutoCompactThreshold (parser + budget/threshold) have no fast-check property. Fix: add property suite (e.g. threshold monotone in window, <= model window - 33k, used+remaining==100, canonicalModelKey idempotent and spelling-invariant, env parse never throws / result in [100k,1M] or null).
2. MAJOR | deterministic | same tests | RULESET.TESTS.boundary-coverage / .fixtures: no limit-1/limit/limit+1 for window bounds. Existing cases use 50_000, 1_500_000, 5_000_000; none at 99_999/100_000/100_001 (setting + env) or 999_999/1_000_000/1_000_001 (setting + env); pct override not at 0/1/100/101; threshold>0 edge untested. Fix: add boundary table for isAutoCompactWindowValue and parseAutoCompactWindowEnv, and pct override edges.
3. MINOR | judgment | hooks/gsd-statusline.js readSettingsFile / claudeSettingsPaths (project_dir/.claude/settings*.json): hostile-input rule (CONTRIBUTING.md "Security and prompt-injection surfaces", user-controlled files hostile). Settings are read with unbounded fs.readFileSync, no size cap, no lstat/symlink/non-regular-file rejection; a repo shipping .claude/settings.json as a FIFO or huge file stalls/bloats the per-turn statusline. Read is also now done for up to 4 files per render (extra reads vs prior single walk, still bounded in count). Partly inherited from #4959 but this PR now expands the surface (project_dir). Fix: lstat regular-file check + size cap (e.g. 1 MiB) -> treat as absent.
4. MINOR | judgment | PR body/maintainer gate: option (a) bridge-scale change (reverses #2451 contract and #4844 AC3) has no explicit maintainer approval on #4985; surface in chat for confirmation before merge.
5. NIT | deterministic | docs/context-monitor.md ~L37: one prose line far exceeds the wrapped width of neighbours ("lowered further when CLAUDE_AUTOCOMPACT_PCT_OVERRIDE sets an earlier point. With auto-compact off, it is the model\nwindow." ragged wrap). Reflow.
6. NIT | judgment | PR template: "All existing tests pass" unchecked, Platforms Linux-only (Windows path handling for settings paths untested locally; CI matrix covers). Check or explain.
7. NIT | deterministic | carried commit ee7ceb1 subject fix(statusline): ... lacks (#4844) scope form type(#issue):; carried unchanged by request; later commits conform. PR title conforms.
   Not findings (verified): no source-grep tests (bridge read via JSON.parse of SUT-written tmp file; RULESET.TESTS.no-source-grep applies to source paths only), no wall-clock assertions (Date.now only in unique ids), try/finally only inside module-level helper runBoundaryHook.

Blast radius

codegraph index points at the main checkout (stale vs PR; contextMeter not found) so codegraph_impact was not usable for new symbols. By inspection: callers confined to runStatusline in hooks/gsd-statusline.js; bridge file reader is hooks/gsd-context-monitor.js only (tests/security.test.cjs, install-minimal-hooks.test.cjs, perf-317 reference the path). New module.exports: isAutoCompactDisabled, AUTO_COMPACT_DISABLE_ENV_KEYS, AUTO_COMPACT_ENV_TRUTHY, resolveAutoCompactThreshold, contextMeter (additive). Behavior change: bridge used_pct/remaining_percentage scale (consumers: monitor only; stale pre-upgrade bridge files lack token fields and are handled). Monitor thresholds 35/25 now mean runway to compaction. Risk: low-medium, 2 hook files.

Co-change gaps

• Changesets: two present, type Fixed, bold — format OK, pr: 5252 backfilled. Fine.
• docs: Fixed type => docs exempt; PR updated docs/context-monitor.md, FEATURES.md (REQ-HOOK-07 mirrored in docs/features/hook-system.md). Parity OK.
• docs/features/context-window-monitoring.md and docs/ARCHITECTURE.md only mention bridge path; no stale scale claims found. Translated docs (ja/zh/ko/pt) not updated: not required.
• No hooks/dist copies; no workflows/references added => INVENTORY/manifest N/A; no gsd-core/workflows growth => no Emitted-Drift-Ack-Growth trailer needed. CONTEXT.md: no new vocabulary/module.
• Closing keywords: "Fixes #4985 / Fixes #4844" present. Branch fix/4985-... conforms (/NNN-slug).

Behind-by / rebase

1 commit behind origin/next. git merge-tree --write-tree HEAD origin/next produced a clean tree (no conflicts). Fix: git fetch origin && git rebase origin/next, push. BEHIND is a CI condition.

Verdict

Changes requested (no Blockers; 2 Major test-rule violations, 2 Minor, 3 Nit). Substance of the fix is sound.

Memtrace Evidence

Memtrace quota was exhausted for this review run, so the graph-backed checks were NOT executed; this is stated rather than reported as passed:

• get_impact: not run (quota). Blast radius above comes from codegraph impact/callers where present.
• get_symbol_context: not run (quota). Callers/callees were read via codegraph where present.
• recall_decision: not run (quota). The recorded-decision gate was NOT evaluated; this is not a finding that no decision governs the change.
• find_code_review_issues: not run (quota). Deterministic AST/YAML/cross-module pass did not execute; every finding above is judgment-sourced unless marked deterministic.

[0xdhx]

Thanks for the review. All seven findings are addressed below; the branch is rebased onto next (b747aca5b) and the head is now d50621d0b.

Findings → commits

1. MAJOR — property-based tests. Writing the properties exposed a real defect: canonicalModelKey was not idempotent (x[1m][1m] keyed as x[1m]). 383215ba8 fixes it and adds the idempotence property; cc125e505 adds the rest, for the env parser, canonicalModelKey's spelling invariance, the threshold, the meter and the settings merge. d50621d0b makes two of the properties stricter: the settings merge is now checked against an independent reference model of Claude Code's merge, which calls no production helper, and the meter property checks the percentage itself (within 0.5 below the threshold, exactly 100 at or above it).
2. MAJOR — limit−1 / limit / limit+1. 42cf89233 adds the boundary table: 99,999 / 100,000 / 100,001 and 999,999 / 1,000,000 / 1,000,001 for the setting (top-level and per-model) and the env var, the pct override at −1 / 0 / 1 / 100 / 101, and the threshold's positive edge at 32,999 / 33,000 / 33,001. At pct 100 the override is always above the threshold it is min'd with, so 100 and 101 read the same; the test says so rather than asserting an edge that cannot be observed.
3. MINOR — hostile settings files. fce0e58dc reads a settings file only when it is a regular file of at most 1 MiB, opened non-blocking and checked through its own descriptor, so a FIFO is skipped instead of blocking the render. A symlink is followed, since a dotfiles-linked settings.json is ordinary; rejecting symlinks via lstat would break that. 28681b591 closes the gap between the size check and the read: a file that grows or shrinks in between is treated as absent. Our own cross-review of this round then found one more cost under the cap: a 524 KB settings file whose model key repeats [1m] 131,072 times made key normalisation take 14 to 17 s, well past the hook's 4 s timeout. 7622f1b0f bounds it: a key over 256 characters once lowercased matches no model, and a run of [1m] suffixes is stripped in one pass.
4. MINOR — the bridge scale needs a maintainer's yes. Agreed, and it is not established yet: @tanji-dg agreed to the supersede, which is a different question. Option (a) moves used_pct / remaining_percentage onto the threshold scale, reversing fix(#2315): respect review.default_reviewers in bare convergence invocation #2451's raw-bridge contract and bug(statusline): context meter subtracts the auto-compact buffer even when autoCompactEnabled is false #4844's AC3; the token counts (used_tokens / threshold_tokens) carry the /context agreement instead: used_tokens is exact when the payload has current_usage, and estimated from the percentage when it doesn't. The PR body now says it should not merge without your confirmation. If you'd rather keep the bridge raw, I can keep the raw fields and add threshold-scale ones beside them for the monitor to read.
5. NIT — docs reflow. d33f8dab6 reflows the docs/context-monitor.md paragraph; no word changed.
6. NIT — test checkbox and platforms. The body now reports every suite npm test runs, plus each tree-level required job, at d50621d0b, and checks the box. It states Linux as the only local platform: both changed test files are on CI's Windows and macOS conformance lists, so CI's Windows and macOS legs run them; I have not run either platform locally.
7. NIT — carried commit subject. The first carried commit's subject is now fix(#4844): …; content and @tanji-dg's authorship are unchanged.

Rebase. 16 carried commits replayed onto b747aca5b with no conflicts; the base range touched none of this PR's files.

One intermediate commit is red at its own tree. a710ed88e moves the bridge to the threshold scale, and five of perf-317's #2451 raw-bridge assertions fail there until 0af4c9fcc moves them to the new scale. A squash merge makes it moot; I can restructure the history if you'd prefer to keep the commits.

Reversion controls. Restoring the old pathname settings reader fails the FIFO and size tests; the earlier descriptor reader without the re-check, and a variant that accepts a short read, fail the grow/shrink test; six off-by-one mutants of the window, pct and threshold bounds each fail the boundary table; the pre-7622f1b0f key function fails the hostile-key test; and dropping lower files' entries, mis-assigning a dated alias, or a meter reporting used = 0 each fail a property. Two mutants survive these controls. One is the pct upper edge from finding 2 (pct ≤ 100 → < 100), which is not observable. The other strips one [1m] per pass instead of a run, which gives identical output for every string under the 256-character cap. It is not caught by output, only by its work: counting replace calls tells the two apart (10 against 15 on x[1m][1m]). I haven't added a test that pins the call count, since the cap already bounds the cost.

Tests at d50621d0b (local, Linux, HOME sandboxed, CLAUDE_CONFIG_DIR cleared): the two changed test files 394/394; the full suite across CI's unit shards plus integration, security, install and slow, 44,246 tests, 0 failed. CI on d50621d0b as of 16:51Z, after the body edit: 42 check-run registrations, all terminal, 39 SUCCESS and 3 SKIPPED, none failed.

[trek-e]

Review: fix(#4985) statusline meter at the real auto-compact threshold

Summary

All seven findings from the 2026-10-09 review are resolved at head d50621d0be. CI at this exact SHA is green (newest run of every check is success, none pending, Required tests success).

Classification & gate compliance

Bug-fix track. #4985 and #4844 are confirmed-bug, both have Fixes, the changesets are Fixed with pr: 5252, docs and FEATURES are mirrored. No shipped workflow growth, so no emitted-drift trailer.

Root-cause verdict

The meter now resolves the real threshold. The bridge writes threshold-scale used_pct and remaining_percentage plus used_tokens and threshold_tokens.

Maintainer decisions recorded

• Option (a), the bridge scale change that reverses the #2451 raw-bridge contract and #4844 AC3, is approved. The additive-fields alternative is not needed.
• No history rewrite is needed for a710ed88e, which is red at its own tree. The merge is a squash.
• Following a symlinked settings file is accepted. Reading through the descriptor with a regular-file check and a 1 MiB cap meets the hostile-input goal, which was never to break dotfile-linked settings.

Blast radius

hooks/gsd-statusline.js and hooks/gsd-context-monitor.js plus their tests and docs. Blast radius was derived by diff reading and grep.

Co-change completeness

Docs, FEATURES, hook-system.md REQ-HOOK-07 and changesets are all in the diff.

Findings

None.

Verdict

Approved. The branch is behind next, so it will be brought up to date server-side, re-verified green, and merged.

Memtrace Evidence

The Memtrace MCP server was unavailable for this review (it failed to connect: memtrace not found in PATH). None of the four checks below ran, so none is a clean graph result.

• get_impact: not run. Blast radius comes from the diff and grep.
• get_symbol_context: not run. Callers were checked by grep.
• recall_decision: not run. This does not show that no decision governs the change. The #2451 and #4844 contract was checked against the PR threads, and the maintainer ruled on it.
• find_code_review_issues: not run. A manual diff review by an isolated reviewer stood in for it, and CI (lint, Required tests) is green.