fix(context-guard): read dumb past 500000 tokens on a 1M window, not 250000

.changes/context-guard/fix-context-guard-1m-band-2f868982.md

@@ -0,0 +1,11 @@
+---
+bump: minor
+---
+
+### Changed
+
+- **The 1M-window token band reads `dumb` past 500000 tokens, not 250000 ([#6644](https://github.com/melodic-software/claude-code-plugins/issues/6644)).** At 26% of a 1M window the shipped bands read `dumb (3 of 3)` and sessions stopped work on it. The `smart` edge stays at 128000. The reader contract records the new basis: Google's GraphWalks BFS figures for Opus 5.5 and Fable 5.1, and Context Arena's 8-needle MRCR, where both Claude rows first score under one half at 512K. The 200k row and the percentage bands are unchanged; `zones.json` still overrides.
+
+### Fixed
+
+- **Band provenance figures.** The reader contract gave Claude Sonnet 5's 8-needle MRCR as at least 0.957 through 256K; the live table reads 0.529 at 128K and 0.522 at 256K. It also said no per-length data exists for Opus 5.5, but Google publishes its GraphWalks BFS scores (90.6% up to 128K, 66.8% from 256K to 1M).

.changes/plugin-quality/fix-context-guard-1m-band-a79971d9.md

@@ -0,0 +1,7 @@
+---
+bump: patch
+---
+
+### Changed
+
+- **The bundled zone resolver's 1M-window token band reads `dumb` past 500000 tokens, not 250000 ([#6644](https://github.com/melodic-software/claude-code-plugins/issues/6644)),** synced from context-guard's reader contract.

lib/context-zone.fixtures.mjs

@@ -38,8 +38,9 @@ export default [
   {"name":"tokens 200000+1 in 200000","sid":"s1","snapshot":{"captured_at":"@now-30","session_id":"s1","context_window":{"used_percentage":null,"remaining_percentage":80,"current_usage":{"input_tokens":100},"total_input_tokens":200000,"total_output_tokens":1,"context_window_size":200000},"cli_version":"2.1.200"},"zones":null,"word":"unknown","notices":[]},
   {"name":"tokens 118000+10000 in 1000000","sid":"s1","snapshot":{"captured_at":"@now-30","session_id":"s1","context_window":{"used_percentage":null,"remaining_percentage":80,"current_usage":{"input_tokens":100},"total_input_tokens":118000,"total_output_tokens":10000,"context_window_size":1000000},"cli_version":"2.1.200"},"zones":null,"word":"smart","notices":[]},
   {"name":"tokens 128000+1 in 1000000","sid":"s1","snapshot":{"captured_at":"@now-30","session_id":"s1","context_window":{"used_percentage":null,"remaining_percentage":80,"current_usage":{"input_tokens":100},"total_input_tokens":128000,"total_output_tokens":1,"context_window_size":1000000},"cli_version":"2.1.200"},"zones":null,"word":"acceptable","notices":[]},
-  {"name":"tokens 249000+1000 in 1000000","sid":"s1","snapshot":{"captured_at":"@now-30","session_id":"s1","context_window":{"used_percentage":null,"remaining_percentage":80,"current_usage":{"input_tokens":100},"total_input_tokens":249000,"total_output_tokens":1000,"context_window_size":1000000},"cli_version":"2.1.200"},"zones":null,"word":"acceptable","notices":[]},
-  {"name":"tokens 250000+1 in 1000000","sid":"s1","snapshot":{"captured_at":"@now-30","session_id":"s1","context_window":{"used_percentage":null,"remaining_percentage":80,"current_usage":{"input_tokens":100},"total_input_tokens":250000,"total_output_tokens":1,"context_window_size":1000000},"cli_version":"2.1.200"},"zones":null,"word":"dumb","notices":[]},
+  {"name":"tokens 260000 in 1000000 (26% used, #6644) is not dumb","sid":"s1","snapshot":{"captured_at":"@now-30","session_id":"s1","context_window":{"used_percentage":26,"remaining_percentage":74,"current_usage":{"input_tokens":100},"total_input_tokens":260000,"total_output_tokens":0,"context_window_size":1000000},"cli_version":"2.1.200"},"zones":null,"word":"acceptable","notices":[]},
+  {"name":"tokens 499000+1000 in 1000000","sid":"s1","snapshot":{"captured_at":"@now-30","session_id":"s1","context_window":{"used_percentage":null,"remaining_percentage":80,"current_usage":{"input_tokens":100},"total_input_tokens":499000,"total_output_tokens":1000,"context_window_size":1000000},"cli_version":"2.1.200"},"zones":null,"word":"acceptable","notices":[]},
+  {"name":"tokens 500000+1 in 1000000","sid":"s1","snapshot":{"captured_at":"@now-30","session_id":"s1","context_window":{"used_percentage":null,"remaining_percentage":80,"current_usage":{"input_tokens":100},"total_input_tokens":500000,"total_output_tokens":1,"context_window_size":1000000},"cli_version":"2.1.200"},"zones":null,"word":"dumb","notices":[]},
   {"name":"tokens 999000+1000 in 1000000","sid":"s1","snapshot":{"captured_at":"@now-30","session_id":"s1","context_window":{"used_percentage":null,"remaining_percentage":80,"current_usage":{"input_tokens":100},"total_input_tokens":999000,"total_output_tokens":1000,"context_window_size":1000000},"cli_version":"2.1.200"},"zones":null,"word":"dumb","notices":[]},
   {"name":"tokens 1000000+1 in 1000000","sid":"s1","snapshot":{"captured_at":"@now-30","session_id":"s1","context_window":{"used_percentage":null,"remaining_percentage":80,"current_usage":{"input_tokens":100},"total_input_tokens":1000000,"total_output_tokens":1,"context_window_size":1000000},"cli_version":"2.1.200"},"zones":null,"word":"unknown","notices":[]},
   {"name":"window 500000 takes the 200000 class","sid":"s1","snapshot":{"captured_at":"@now-30","session_id":"s1","context_window":{"used_percentage":null,"remaining_percentage":80,"current_usage":{"input_tokens":100},"total_input_tokens":110000,"total_output_tokens":0,"context_window_size":500000},"cli_version":"2.1.200"},"zones":null,"word":"acceptable","notices":[]},

lib/context-zone.sh

@@ -53,7 +53,7 @@
 # window-class token bands over occupancy:
 #
 #   window class 200000:  smart ≤ 100000 < acceptable ≤ 150000 < dumb
-#   window class 1000000: smart ≤ 128000 < acceptable ≤ 250000 < dumb
+#   window class 1000000: smart ≤ 128000 < acceptable ≤ 500000 < dumb
 #
 # All are declared judgment defaults, NOT doc- or benchmark-derived
 # constants (anchors and provenance are recorded in the reader contract).
@@ -64,7 +64,7 @@
 #     "acceptable_max_used_percentage": 75,
 #     "token_bands": {
 #       "200000":  { "smart_max_tokens": 100000, "acceptable_max_tokens": 150000 },
-#       "1000000": { "smart_max_tokens": 128000, "acceptable_max_tokens": 250000 }
+#       "1000000": { "smart_max_tokens": 128000, "acceptable_max_tokens": 500000 }
 #     }
 #   }
 #
@@ -92,7 +92,7 @@ DEFAULT_ACCEPTABLE_MAX=75
 # [class, smart_max, acceptable_max] rows, ascending class order. JSON rather
 # than whitespace rows because the bands are handed to the snapshot pass as
 # data, so the comparison happens there instead of in a separate awk process.
-DEFAULT_TOKEN_BANDS='[[200000,100000,150000],[1000000,128000,250000]]'
+DEFAULT_TOKEN_BANDS='[[200000,100000,150000],[1000000,128000,500000]]'
 
 unknown() {
   printf 'unknown\n'
@@ -197,7 +197,7 @@ if [[ -e "$zones" ]]; then
     acceptable_max=${zpct#* }
   fi
   if [[ "$ztb" == "invalid" ]]; then
-    printf 'context-guard: zones.json token_bands malformed — using shipped default token bands (200000:100000/150000, 1000000:128000/250000)\n' >&2
+    printf 'context-guard: zones.json token_bands malformed — using shipped default token bands (200000:100000/150000, 1000000:128000/500000)\n' >&2
   elif [[ "$ztb" != "absent" ]]; then
     token_bands=$ztb
   fi

plugins/context-guard/README.md

@@ -14,7 +14,7 @@ degraded context **before** quality slips, instead of guessing. Four parts:
   worse computable zone wins): percentage bands over `used_percentage` (shipped defaults
   smart ≤ 50 < acceptable ≤ 75 < dumb) and window-class token bands over occupancy
   (`total_input_tokens + total_output_tokens`; shipped defaults 100k/150k on a 200k window,
-  128k/250k on a 1M window). Bands come from the machine-scope
+  128k/500k on a 1M window). Bands come from the machine-scope
   `~/.claude/context-guard/zones.json` when present and valid, else from the shipped defaults.
   Zones say *where you are*; consumers decide *what to do*.
 - **PostCompact marker hook** (`hooks/post-compact-mark.sh`), a settings hook, so it runs where

plugins/context-guard/hooks/context-guard.test.ts

@@ -421,6 +421,13 @@ test('zones.json: an extra threshold gets its approach line and its line once pe
   ])
 })
 
+// On a 1000000-token window with the shipped bands, 26% used (260000 tokens) reads acceptable:
+// the reader contract's 1M row puts the dumb edge past 500000 tokens.
+test('shipped bands, 1000000 window: 26% used announces acceptable, never dumb; 51% announces dumb', async ($, on) => {
+  const { w } = world(on, { window: 1_000_000 })
+  expect(await walk($, w, [5, 26, 51])).toEqual([[], [crossing('acceptable')], [crossing('dumb')]])
+})
+
 // The token edges of a 1000000-token window at 210000 and 410000; the percentage edges there
 // (50% and 75%) sit at 500000 and 750000 tokens, so the token shape decides both crossings.
 const TOKEN_EDGES = { token_bands: { '1000000': { smart_max_tokens: 210_000, acceptable_max_tokens: 410_000 } } }

plugins/context-guard/hooks/zone.ts

@@ -13,7 +13,7 @@ export const DEFAULT_BANDS: Bands = {
   acceptable: 75,
   tokens: [
     [200_000, 100_000, 150_000],
-    [1_000_000, 128_000, 250_000],
+    [1_000_000, 128_000, 500_000],
   ],
 }
 export const RANK: Record<Zone, number> = { smart: 0, acceptable: 1, dumb: 2 }

plugins/context-guard/reference/reader-contract.md

@@ -59,7 +59,7 @@ same triggers. These stamps are probe results with a date, not standing facts.
 - **Default token bands (over occupancy = `total_input_tokens` + `total_output_tokens`, uppers
   inclusive, selected by window class as described under "Occupancy and combination rule"):**
   window class **200000**: `smart` ≤ **100000** < `acceptable` ≤ **150000** < `dumb`;
-  window class **1000000**: `smart` ≤ **128000** < `acceptable` ≤ **250000** < `dumb`.
+  window class **1000000**: `smart` ≤ **128000** < `acceptable` ≤ **500000** < `dumb`.
 - **Token-shape version floor (fixed):** the token shape is computable only when the snapshot's
   `cli_version` is present, purely numeric dotted, and **≥ 2.1.132**, the release from which the
   token fields mean current occupancy rather than cumulative session totals.
@@ -226,28 +226,35 @@ a version field cannot (there is no writer authentication, so `cli_version` is u
 every other snapshot value). The bundled resolver implements both gates.
 
 **Band provenance:** the shipped token bands are **declared judgment**, not benchmark-derived
-constants. No per-length long-context data is published for Claude Opus 5.5, Sonnet 5.5 or Fable
-5.1; their system cards report only an aggregate ProgramBench score. The 1M row's edges, 128000
-and 250000, are anchored on the nearest published data: on Context Arena's 8-needle MRCR (max
-effort, Google tokenizer) Claude Opus 5 scores 0.913 at 128K, 0.656 at 256K and 0.425 at 512K, and
-Claude Sonnet 5 a median of at least 0.957 through 256K and 0.041 at 512K; on GraphWalks BFS, run
-by Google at max reasoning and published with Gemini 4 Argon on 2026-09-30, Fable 5.1 scores 91.4%
-up to 128K and 65.0% from 256K to 1M. The 200k row's `acceptable` edge, 150000, sits at
+constants. Anthropic publishes no per-length long-context scores for Claude Opus 5.5, Sonnet 5.5
+or Fable 5.1: their system cards report one aggregate ProgramBench score each, an agentic task
+whose episodes run up to the full 1M window. The published per-length data is retrieval and graph
+traversal, not agentic work, and no study we found links those scores to agentic task quality. On
+GraphWalks BFS, run by Google at max reasoning and published with Gemini 4 Argon, Opus 5.5 scores
+90.6% up to 128K and 66.8% from 256K to 1M, and Fable 5.1 91.4% and 65.0%. On Context Arena's
+8-needle MRCR (max effort), Claude Opus 5 scores 0.913 at 128K, 0.656 at 256K and 0.425 at 512K,
+and Claude Sonnet 5 0.529, 0.522 and 0.320. The 1M row's `smart` edge, 128000, is the last length
+at which those current-model scores are still near their best. Its `acceptable` edge, 500000,
+sits below 512K, the first MRCR length at which both Claude rows score under one half. Between
+the two, scores sag without a measured cliff, so a reading there is `acceptable`, not `dumb`: an
+`acceptable` edge of 250000 read `dumb` at 26% of a 1M window on no evidence for current models
+or agentic work (#6644). The 200k row's `acceptable` edge, 150000, sits at
 the top of the practitioner consensus of about 125-150K. `zones.json` is the correction path, and
 the numeric agreement of the 200k row's percentage translation with the shipped 50/75 percentage
 defaults is coincidence, not validation.
 
 - **Pointer**: when re-deriving a 1M band edge, fetch live:
   [Context Arena, 8 needles](https://contextarena.ai/api/needle-summary?needles=8), the Claude
   Opus 5 and Claude Sonnet 5 rows; [Gemini models](https://deepmind.google/models/gemini), the
-  Gemini 4 Argon table's GraphWalks BFS rows; the absence of per-length data, the
+  Gemini 4 Argon table's Opus 5.5 and Fable 5.1 GraphWalks BFS rows; the absence of per-length data, the
   [Opus 5.5 system card](https://www-cdn.anthropic.com/fc1b44717c85dc068bc6ba5024219938094694bd/Claude%20Opus%205.5%20System%20Card.pdf)
   section 8.10 and the [Fable 5.1 system card](https://www-cdn.anthropic.com/0339e6a7c5c7b87f5c07798616dc32c215d14235/Claude%20Fable%205.1%20&%20Claude%20Mythos%205.1%20System%20Card.pdf)
   section 8.11. For the 200k row, correlate with [AI Hero, "Smart Zone"](https://www.aihero.dev/ai-coding-dictionary/smart-zone)
   and [Geoffrey Huntley, "Ralph"](https://ghuntley.com/ralph/); no docs page covers where quality
   degrades on a 200K window as of 2026-10-04.
-- **As of**: 2026-10-07
-- **Recheck trigger**: Opus 5.5, Fable 5.1 or Sonnet 5.5 appear on Context Arena; a new or revised
+- **As of**: 2026-10-09
+- **Recheck trigger**: Opus 5.5, Fable 5.1 or Sonnet 5.5 appear on Context Arena; Google's table
+  splits the 256K-1M GraphWalks bin; a new or revised
   Anthropic system card publishes per-length scores; or a docs page starts covering where quality degrades on a 200K window, at which point the
   200k row's pointer moves there.
 
@@ -382,7 +389,7 @@ session itself can know.
 compaction fires at or near the model's context limit; the cases that fire earlier depend on the
 model, the window it runs with, and the environment. For the current thresholds, see
 [Claude Code model config, "Default auto-compact thresholds"](https://code.claude.com/docs/en/model-config#default-auto-compact-thresholds).
-**As of:** 2026-09-30, the one number that section publishes, for native 1M windows, sits above
+**As of:** 2026-10-09, the one number that section publishes for native 1M windows, about 967K tokens, sits above
 the shipped `dumb` band, so it does not disturb the margin that the bands-below-the-trigger rule
 protects, the way a lowered window does. **Recheck trigger:** that section is renamed or removed,
 or publishes a default at or below the `dumb` band's lower edge. `CLAUDE_AUTOCOMPACT_PCT_OVERRIDE`
@@ -451,8 +458,8 @@ hook, so a firing is best read diagnostically: **it means the boundary was misse
 window was managed. Lowering the window moves the trigger, so the bands in `zones.json` must move
 with it, normalized into the percentage shape. A 400000-token window on a 1M-class model puts the
 trigger at **40% of the full window**, which is *inside* the shipped percentage `smart` band
-(≤ 50), so the percentage shape still reads smart when auto-compact fires (the token shape decides
-the zone there). Keeping the percentage bands below that trigger means pulling them under 40, not
+(≤ 50), so the percentage shape still reads smart when auto-compact fires, and the token shape,
+which decides the zone there, reads `acceptable`: neither shipped shape reaches `dumb` first. Keeping the percentage bands below that trigger means pulling them under 40, not
 comparing 400000 against the token bands' occupancy edges, which measure a different quantity.
 
 That diagnostic reading is adopted; the prescription that usually travels with it is not. **Leave
@@ -500,7 +507,7 @@ what the human sees and what consumers decide on. Zones say *where you are*; con
   "acceptable_max_used_percentage": 75,
   "token_bands": {
     "200000": { "smart_max_tokens": 100000, "acceptable_max_tokens": 150000 },
-    "1000000": { "smart_max_tokens": 128000, "acceptable_max_tokens": 250000 }
+    "1000000": { "smart_max_tokens": 128000, "acceptable_max_tokens": 500000 }
   }
 }
 ```

plugins/context-guard/scripts/context-zone.fixtures.mjs

@@ -40,8 +40,9 @@ export default [
   {"name":"tokens 200000+1 in 200000","sid":"s1","snapshot":{"captured_at":"@now-30","session_id":"s1","context_window":{"used_percentage":null,"remaining_percentage":80,"current_usage":{"input_tokens":100},"total_input_tokens":200000,"total_output_tokens":1,"context_window_size":200000},"cli_version":"2.1.200"},"zones":null,"word":"unknown","notices":[]},
   {"name":"tokens 118000+10000 in 1000000","sid":"s1","snapshot":{"captured_at":"@now-30","session_id":"s1","context_window":{"used_percentage":null,"remaining_percentage":80,"current_usage":{"input_tokens":100},"total_input_tokens":118000,"total_output_tokens":10000,"context_window_size":1000000},"cli_version":"2.1.200"},"zones":null,"word":"smart","notices":[]},
   {"name":"tokens 128000+1 in 1000000","sid":"s1","snapshot":{"captured_at":"@now-30","session_id":"s1","context_window":{"used_percentage":null,"remaining_percentage":80,"current_usage":{"input_tokens":100},"total_input_tokens":128000,"total_output_tokens":1,"context_window_size":1000000},"cli_version":"2.1.200"},"zones":null,"word":"acceptable","notices":[]},
-  {"name":"tokens 249000+1000 in 1000000","sid":"s1","snapshot":{"captured_at":"@now-30","session_id":"s1","context_window":{"used_percentage":null,"remaining_percentage":80,"current_usage":{"input_tokens":100},"total_input_tokens":249000,"total_output_tokens":1000,"context_window_size":1000000},"cli_version":"2.1.200"},"zones":null,"word":"acceptable","notices":[]},
-  {"name":"tokens 250000+1 in 1000000","sid":"s1","snapshot":{"captured_at":"@now-30","session_id":"s1","context_window":{"used_percentage":null,"remaining_percentage":80,"current_usage":{"input_tokens":100},"total_input_tokens":250000,"total_output_tokens":1,"context_window_size":1000000},"cli_version":"2.1.200"},"zones":null,"word":"dumb","notices":[]},
+  {"name":"tokens 260000 in 1000000 (26% used, #6644) is not dumb","sid":"s1","snapshot":{"captured_at":"@now-30","session_id":"s1","context_window":{"used_percentage":26,"remaining_percentage":74,"current_usage":{"input_tokens":100},"total_input_tokens":260000,"total_output_tokens":0,"context_window_size":1000000},"cli_version":"2.1.200"},"zones":null,"word":"acceptable","notices":[]},
+  {"name":"tokens 499000+1000 in 1000000","sid":"s1","snapshot":{"captured_at":"@now-30","session_id":"s1","context_window":{"used_percentage":null,"remaining_percentage":80,"current_usage":{"input_tokens":100},"total_input_tokens":499000,"total_output_tokens":1000,"context_window_size":1000000},"cli_version":"2.1.200"},"zones":null,"word":"acceptable","notices":[]},
+  {"name":"tokens 500000+1 in 1000000","sid":"s1","snapshot":{"captured_at":"@now-30","session_id":"s1","context_window":{"used_percentage":null,"remaining_percentage":80,"current_usage":{"input_tokens":100},"total_input_tokens":500000,"total_output_tokens":1,"context_window_size":1000000},"cli_version":"2.1.200"},"zones":null,"word":"dumb","notices":[]},
   {"name":"tokens 999000+1000 in 1000000","sid":"s1","snapshot":{"captured_at":"@now-30","session_id":"s1","context_window":{"used_percentage":null,"remaining_percentage":80,"current_usage":{"input_tokens":100},"total_input_tokens":999000,"total_output_tokens":1000,"context_window_size":1000000},"cli_version":"2.1.200"},"zones":null,"word":"dumb","notices":[]},
   {"name":"tokens 1000000+1 in 1000000","sid":"s1","snapshot":{"captured_at":"@now-30","session_id":"s1","context_window":{"used_percentage":null,"remaining_percentage":80,"current_usage":{"input_tokens":100},"total_input_tokens":1000000,"total_output_tokens":1,"context_window_size":1000000},"cli_version":"2.1.200"},"zones":null,"word":"unknown","notices":[]},
   {"name":"window 500000 takes the 200000 class","sid":"s1","snapshot":{"captured_at":"@now-30","session_id":"s1","context_window":{"used_percentage":null,"remaining_percentage":80,"current_usage":{"input_tokens":100},"total_input_tokens":110000,"total_output_tokens":0,"context_window_size":500000},"cli_version":"2.1.200"},"zones":null,"word":"acceptable","notices":[]},