Re-sync KITT 96583d5: the rename's last corrections and the 2026-11-01 price table

.agents/core/agents-core.md

@@ -22,7 +22,7 @@
   routes, its subsystem map, and its setup/verification/CI commands.
 -->
 
-> **Who this is for.** Every AI agent working in a repo governed by the
+> **Who this is for.** Every AI agent working in a repo governed by
 > KITT — Codex, Claude Code, Replit, and whatever comes next. It tells
 > you **how to behave and what is non-negotiable**, independent of which
 > product you are working on. The repo's own `AGENTS.md` carries the product
@@ -31,7 +31,7 @@
 
 > **One source of truth for all agents.** These rules are shared. Claude
 > Code's `CLAUDE.md` holds only Claude-specific ceremony and defers here for
-> every cross-agent principle. When a shared rule changes, it changes in the
+> every cross-agent principle. When a shared rule changes, it changes in
 > KITT — never as a forked copy in one agent's own file or private
 > memory.
 

.agents/memory/git-fetch-refspec-limited-to-main-hides-a-pushed-branch.md

@@ -17,7 +17,7 @@ PR is open on it.
 `remote.origin.fetch = +refs/heads/main:refs/remotes/origin/main` — so no
 `refs/remotes/origin/claude/<branch>` ref is ever created or updated, and
 every tool that compares the branch to its remote-tracking ref sees nothing
-there. Measured twice on 2026-10-11 (kitt, then DojoOS), both clones
+there. Measured twice on 2026-10-11 (ai-handbook, then DojoOS), both clones
 set up by the session harness.
 
 **Fix.** Widen the refspec and fetch the branch once; never re-push to

.agents/reviews/.gitignore

@@ -15,7 +15,7 @@
 # the same hole remains for the plan loop.)
 #
 # WHY THERE ARE TWO COPIES. Git does not follow a symlinked .gitignore, so a
-# pointer is impossible and a real file at each root is the only option. The
+# pointer is impossible and a real file at each root is the only option.
 # KITT mirrors this one at .agents/reviews/.gitignore, and
 # scripts/check-root-wiring.mjs compares the pattern lines so the duplication
 # is enforced rather than merely intended. Edit THIS copy; the root follows.

docs/ai-context/token-pricing.md

@@ -15,7 +15,7 @@ published!"*).
 
 ## Anthropic — Claude API list prices
 
-Read 2026-10-12 from
+Read 2026-11-01 from
 <https://platform.claude.com/docs/en/about-claude/pricing> (the `.md` form of
 the page). Standard tier; USD per million tokens. Rows quoted as the page
 printed them, whitespace collapsed and the cache-read column's footnote
@@ -43,20 +43,37 @@ write, cache read, output.
 
 ## OpenAI — the Codex reviewer (Astra)
 
-Read 2026-10-12 from <https://developers.openai.com/api/docs/pricing>
-(`openai.com/api/pricing` answered 403 to this container's fetch). Standard
-tier, prompts up to 272K tokens; USD per million tokens. Row quoted as the
-page printed it: input, cached input, cache write, output, then the four
-long-context columns:
+Read 2026-11-01 from <https://developers.openai.com/api/docs/pricing>, the
+rendered HTML fetched directly (`openai.com/api/pricing` answered 403 to this
+container's fetch on 2026-10-12). Standard tab; USD per million tokens.
+Header and row as the page's text-token table prints them:
 
 ```
-| gpt-6-astra | $10.00 | $1.00 | $12.50 | $50.00 | $20.00 | $2.00 | $25.00 | $75.00 |
+| Model | Input | Output | Cached input | 30m write | 12h write |
+| gpt-6-astra | $10.00 | $50.00 | $1.00 | $12.50 | $25.00 |
 ```
 
-| Model | Input | Cached input | Cache write | Output |
-|---|---|---|---|---|
-| `gpt-6-astra` (the `strongestCodex` pin), prompt ≤ 272K | $10 | $1 | $12.50 | $50 |
-| `gpt-6-astra`, prompt > 272K (the long-context columns) | $20 | $2 | $25 | $75 |
+The page no longer prints long-context figures for this model. The long-context
+row below is **derived, not quoted**, from the rule the `gpt-6-astra` model
+page (<https://developers.openai.com/api/docs/models/gpt-6-astra>) prints,
+read the same day: "Prompts with more than 272K input tokens are priced at 2x
+input and cache rates and 1.5x output for the full request." The derived
+figures for input, cached input, the 30-minute write and output match the
+long-context columns the pricing page printed on 2026-10-12.
+
+| Model | Input | Cached input | Cache write 30m | Cache write 12h | Output |
+|---|---|---|---|---|---|
+| `gpt-6-astra` (the `strongestCodex` pin), prompt ≤ 272K | $10 | $1 | $12.50 | $25 | $50 |
+| `gpt-6-astra`, prompt > 272K (derived: 2x input and cache, 1.5x output) | $20 | $2 | $25 | $50 | $75 |
+
+Until 2026-11-01 the page printed one cache-write column, at the figure it
+now prints as the 30-minute write, and the model page still prints that
+single figure; the 12-hour write is new, and every other short-context
+figure is unchanged from the 2026-10-12 read. **The CLI's cache-write bucket
+does not say which retention it bought, so it is priced at the 12-hour rate**
+— the higher of the two, which overstates and never understates, the side
+the long-context and Haiku rules below take for the same reason. Every probe
+so far has read that bucket as 0, so the choice has not yet moved a figure.
 
 ### Pricing an Astra run — per bucket, never by the proxy
 
@@ -87,7 +104,7 @@ Three readings behind that formula, each settled rather than assumed:
 - **The usage is the turn's cumulative total, but the 272K threshold is per
   request** — "Prompts with more than 272K input tokens are priced at 2x input
   and cache rates and 1.5x output for the full request" (the `gpt-6-astra`
-  model page, read 2026-10-12), and the CLI's `turn.completed` is built from
+  model page, read 2026-10-12 and again, unchanged, 2026-11-01), and the CLI's `turn.completed` is built from
   the turn's running `total`. A turn of several tool-using requests can sum
   past 272K with every request short. So: **`input_tokens` at or under
   272,000 is priced exactly from the short row** — no single request can

scripts/__tests__/map.unit.test.mjs

@@ -2,7 +2,7 @@
 //
 // The map tool's pure functions, on bare Node: no compiler, no sync, no
 // fixture repository. This file ships with the payload, so it must run
-// wherever it lands; the compiler-backed behaviour is proven by the
+// wherever it lands; the compiler-backed behaviour is proven by
 // KITT's own integration suite, which never ships.
 import test from "node:test";
 import assert from "node:assert/strict";

scripts/machinery.mjs

@@ -249,7 +249,7 @@ export const PAID_RUN_LEDGER_REPO = "TheAnswerManIsHere/KITT";
  * equal to the pin; `review-proxy.mjs` reads the declared effort to say, in an
  * assessment's header, what the role actually runs at. Those are the two ends
  * of one fact, and a second copy of the reader is a second thing to drift --
- * the failure this whole area is about. KITT-only checker can import
+ * the failure this whole area is about. The KITT-only checker can import
  * payload code; payload code cannot import the checker, so the shared copy
  * lives here. (Both assessors, #131 round 1.)
  *

scripts/sweep-scope.mjs

@@ -347,7 +347,7 @@ export function plan({ root, spec, workers = DEFAULT_WORKERS, include = [] }) {
   const fullSet = new Set();
   for (const g of spec.readInFull ?? []) {
     // A glob is payload-relative like `home`, but an `--include`d file is
-    // repo-relative and outside the payload; match both spellings so the
+    // repo-relative and outside the payload; match both spellings so
     // KITT can name a non-payload document for a full read.
     const asPayload = globToRegExp(toRepoPath(prefix, g));
     const asRepo = globToRegExp(g.replace(/^\.\//, ""));