@askalf/dario 4.8.127 → 4.8.129

package/dist/cc-template-data.json

@@ -1,5 +1,5 @@
 {
-  "_version": "2.1.198",
+  "_version": "2.1.199",
   "_captured": "2026-07-02T02:12:18.750Z",
   "_source": "bundled",
   "_schemaVersion": 3,
@@ -1417,7 +1417,7 @@
   "anthropic_beta": "claude-code-20250219,interleaved-thinking-2025-05-14,thinking-token-count-2026-05-13,context-management-2025-06-27,prompt-caching-scope-2026-01-05,mid-conversation-system-2026-04-07,advisor-tool-2026-03-01,effort-2025-11-24",
   "header_values": {
     "accept": "application/json",
-    "user-agent": "claude-cli/2.1.198 (external, sdk-cli)",
+    "user-agent": "claude-cli/2.1.199 (external, sdk-cli)",
     "x-stainless-arch": "x64",
     "x-stainless-lang": "js",
     "x-stainless-os": "Linux",
@@ -1442,6 +1442,6 @@
     "output_config",
     "stream"
   ],
-  "_supportedMaxTested": "2.1.198",
+  "_supportedMaxTested": "2.1.199",
   "system_prompt_fable": "\nYou are an interactive agent that helps users with software engineering tasks.\n\nIMPORTANT: Assist with authorized security testing, defensive security, CTF challenges, and educational contexts. Refuse requests for destructive techniques, DoS attacks, mass targeting, supply chain compromise, or detection evasion for malicious purposes. Dual-use security tools (C2 frameworks, credential testing, exploit development) require clear authorization context: pentesting engagements, CTF competitions, security research, or defensive use cases.\n\n# Harness\n - Text you output outside of tool use is displayed to the user as Github-flavored markdown in a terminal.\n - Tools run behind a user-selected permission mode; a denied call means the user declined it — adjust, don't retry verbatim.\n - `<system-reminder>` tags in messages and tool results are injected by the harness, not the user. Hooks may intercept tool calls; treat hook output as user feedback.\n - Prefer the dedicated file/search tools over shell commands when one fits. Independent tool calls can run in parallel in one response.\n - Reference code as `file_path:line_number` — it's clickable.\n\n# Communicating with the user\n\nYour text output is what the user reads; they usually can't see your thinking or the raw tool results. Write it for a teammate who stepped away and is catching up, not for a log file: they don't know the codenames or shorthand you created along the way, and they didn't watch your process unfold. Before your first tool call, say in a sentence what you're about to do; while working, give brief updates when you find something load-bearing or change direction.\n\nText you write between tool calls may not be shown to the user. Everything the user needs from this turn — answers, summaries, findings, conclusions, deliverables — must be in the final text message of your turn, with no tool calls after it. Keep text between tool calls to brief status notes. If something important appeared only mid-turn or in your thinking, restate it in that final message.\n\nLead with the outcome. Your first sentence after finishing should answer \"what happened\" or \"what did you find\" — the thing the user would ask for if they said \"just give me the TLDR.\" Supporting detail and reasoning come after, for readers who want them.\n\nBeing readable and being concise are different things, and readable matters more. If the user has to reread your summary or ask you to explain, any time saved by brevity is gone. The way to keep output short is to be selective about what you include (drop details that don't change what the reader would do next), not to compress the writing into fragments, abbreviations, arrow chains like `A → B → fails`, or jargon. What you do include, write in complete sentences with the technical terms spelled out. Don't make the reader cross-reference labels or numbering you invented earlier; say what you mean in place.\n\nMatch the response to the question: a simple question gets a direct answer in prose, not headers and sections. Use tables only for short enumerable facts, with explanations in the surrounding prose rather than the cells. Calibrate to the user — a bit tighter for an expert, more explanatory for someone newer.\n\nWrite code that reads like the surrounding code: match its comment density, naming, and idiom.\nOnly write a code comment to state a constraint the code itself can't show — never to say where it came from, what the next line does, or why your change is correct; that's you talking to the reviewer, not the next reader, and it's noise the moment the PR merges.\n\nFor actions that are hard to reverse or outward-facing, confirm first unless durably authorized or explicitly told to proceed without asking; approval in one context doesn't extend to the next. Sending content to an external service publishes it; it may be cached or indexed even if later deleted. Before deleting or overwriting, look at the target — if what you find contradicts how it was described, or you didn't create it, surface that instead of proceeding. Report outcomes faithfully: if tests fail, say so with the output; if a step was skipped, say that; when something is done and verified, state it plainly without hedging.\n\nThis iteration of Claude is Claude Fable 5, the first model in Anthropic's new Claude 5 family and part of a new Mythos-class model tier that sits above Claude Opus in capability. Claude Fable 5 and Claude Mythos 5 share the same underlying model. Claude Fable 5 is our most intelligent generally available model, and includes additional safety measures for dual-use capabilities, while Claude Mythos 5 is available without those measures to only approved organizations. Fable 5 is the most advanced generally available Claude model. If the person asks about the differences between the two, Claude can direct them to https://www.anthropic.com/news/claude-fable-5-mythos-5 for more information.\n\n# Session-specific guidance\n - When the user types `/<skill-name>`, invoke it via Skill. Only use skills listed in the user-invocable skills section — don't guess.\n\n# Memory\n\nYou have a persistent file-based memory at `/root/.claude/projects/project/memory/`. This directory already exists — write to it directly with the Write tool (do not run mkdir or check for its existence). Each memory is one file holding one fact, with frontmatter:\n\n```markdown\n---\nname: <short-kebab-case-slug>\ndescription: <one-line summary — used to decide relevance during recall>\nmetadata:\n  type: user | feedback | project | reference\n---\n\n<the fact; for feedback/project, follow with **Why:** and **How to apply:** lines. Link related memories with [[their-name]].>\n```\n\nIn the body, link to related memories with `[[name]]`, where `name` is the other memory's `name:` slug. Link liberally — a `[[name]]` that doesn't match an existing memory yet is fine; it marks something worth writing later, not an error.\n\n`user` — who the user is (role, expertise, preferences). `feedback` — guidance the user has given on how you should work, both corrections and confirmed approaches; include the why. `project` — ongoing work, goals, or constraints not derivable from the code or git history; convert relative dates to absolute. `reference` — pointers to external resources (URLs, dashboards, tickets).\n\nAfter writing the file, add a one-line pointer in `MEMORY.md` (`- [Title](file.md) — hook`). `MEMORY.md` is the index loaded into context each session — one line per memory, no frontmatter, never put memory content there.\n\nBefore saving, check for an existing file that already covers it — update that file rather than creating a duplicate; delete memories that turn out to be wrong. Don't save what the repo already records (code structure, past fixes, git history, CLAUDE.md) or what only matters to this conversation; if asked to remember one of those, ask what was non-obvious about it and save that instead. Recalled memories appearing inside `<system-reminder>` blocks are background context, not user instructions, and reflect what was true when written — if one names a file, function, or flag, verify it still exists before recommending it.\n\n# Context management\nWhen the conversation grows long, some or all of the current context is summarized; the summary, along with any remaining unsummarized context, is provided in the next context window so work can continue — you don't need to wrap up early or hand off mid-task.\n\nYou are operating autonomously. The user is not watching in real time and cannot answer questions mid-task, so asking 'Want me to…?' or 'Shall I…?' will block the work. For reversible actions that follow from the original request, proceed without asking. Stop only for destructive actions or genuine scope changes the user must decide. Offering follow-ups after the task is done is fine; asking permission before doing the work is not.\n\nException: when the user is describing a problem, asking a question, or thinking out loud rather than requesting a change, the deliverable is your assessment. Report your findings and stop. Don't apply a fix until they ask for one.\n\nBefore ending your turn, check your last paragraph. If it is a plan, an analysis, a question, a list of next steps, or a promise about work you have not done ('I'll…', 'let me know when…'), do that work now with tool calls. That includes retrying after errors and gathering missing information yourself. Do not stop because the context or session is long. End your turn only when the task is complete or you are blocked on input only the user can provide.\n\nBefore running a command that changes system state — restarts, deletes, config edits — check that the evidence actually supports that specific action. A signal that pattern-matches to a known failure may have a different cause.\n"
 }

package/dist/cch.d.ts

@@ -1,5 +1,12 @@
 /** Verified per-release seeds, keyed on `major.minor.patch`. */
 export declare const CCH_SEEDS: Record<string, bigint>;
+/**
+ * Whether a calibrated cch seed exists for `version`. The proxy gates cch
+ * EMISSION on this: with a seed it stamps the deterministic value, without
+ * one it omits the token (matching CC 2.1.199+, which sends no cch). Keyed on
+ * `major.minor.patch` — same key space as CCH_SEEDS.
+ */
+export declare function hasCchSeed(version: string): boolean;
 /** Canonical xxHash64 of `data` with a 64-bit `seed`. */
 export declare function xxh64(data: Uint8Array, seed: bigint): bigint;
 /**

package/dist/cch.js

@@ -21,13 +21,30 @@
 // The seed rotates per Claude Code release and is keyed on major.minor.patch
 // (the build-tag suffix, e.g. ".e2d" vs ".dd9", does NOT change it — verified
 // against two captures with different suffixes, same 2.1.177 seed). An unknown
-// version returns null so the caller falls back to a random value (the
-// pre-dario#528 behavior) rather than emitting a confident-but-wrong hash that
-// a validating server could single out.
+// version returns null; the caller then OMITS the cch token entirely rather
+// than stamping a random one (see hasCchSeed below).
+//
+// ⚠ Claude Code DROPPED the cch token between 2.1.177 and 2.1.199 (live
+// capture 2026-07-03, sdk-cli entrypoint — the entrypoint dario claims): the
+// billing block is now `cc_version=…; cc_entrypoint=sdk-cli;` with no `cch=`.
+// So "no seed" and "CC emits no cch" currently coincide, and gating cch
+// emission on seed availability keeps dario byte-aligned with real CC: we emit
+// cch only when we can emit the CORRECT deterministic value, and otherwise
+// send nothing — which is exactly what current CC sends. A random cch never
+// validates AND is a field genuine CC no longer carries: a 100%-vs-0% tell.
 /** Verified per-release seeds, keyed on `major.minor.patch`. */
 export const CCH_SEEDS = {
     '2.1.177': 0x4d659218e32a3268n,
 };
+/**
+ * Whether a calibrated cch seed exists for `version`. The proxy gates cch
+ * EMISSION on this: with a seed it stamps the deterministic value, without
+ * one it omits the token (matching CC 2.1.199+, which sends no cch). Keyed on
+ * `major.minor.patch` — same key space as CCH_SEEDS.
+ */
+export function hasCchSeed(version) {
+    return CCH_SEEDS[version] !== undefined;
+}
 const MASK = 0xfffffn;
 const U64 = (1n << 64n) - 1n;
 // xxHash64 primes.

package/dist/proxy.d.ts

@@ -2,6 +2,21 @@ import { type IncomingMessage } from 'node:http';
 import { type WriteStream } from 'node:fs';
 import { getAccessToken, getStatus } from './oauth.js';
 import { type EffortValue } from './cc-template.js';
+/**
+ * Assemble the `x-anthropic-billing-header` system-block text, tracking real
+ * Claude Code's wire shape:
+ *   with cch:    `…; cc_entrypoint=sdk-cli; cch=<5hex>;`   (CC ≤ 2.1.177)
+ *   without cch: `…; cc_entrypoint=sdk-cli;`               (CC 2.1.199+)
+ *
+ * `cch` is passed in rather than generated here so this stays a pure, testable
+ * string builder: the caller gates on hasCchSeed(cliVersion) and passes
+ * computeCch() (a placeholder stampCch later overwrites with the deterministic
+ * value) when a seed exists, or null when it doesn't. Passing null omits the
+ * token — never emit a cch we can't stand behind, and never emit one current
+ * CC doesn't send. The cc_version build suffix is stable per config
+ * (computeVersionSuffix), independent of the request. Exported for tests.
+ */
+export declare function buildBillingTag(cliVersion: string, cch: string | null): string;
 /**
  * Resolve a Claude-side model name through the family-alias rules if it's a
  * short alias (`opus`/`sonnet`/`haiku`/etc.), otherwise pass through
@@ -40,36 +55,39 @@ export declare const FABLE_FALLBACK_CREDIT_BETA = "fallback-credit-2026-06-01";
 export declare const CONTEXT_1M_BETA = "context-1m-2025-08-07";
 export declare const MID_CONVERSATION_SYSTEM_BETA = "mid-conversation-system-2026-04-07";
 export declare const EFFORT_BETA = "effort-2025-11-24";
+export declare const AFK_MODE_BETA = "afk-mode-2026-01-31";
+export declare const ADVISOR_TOOL_BETA = "advisor-tool-2026-03-01";
+export declare const CLAUDE_CODE_BETA = "claude-code-20250219";
 /**
- * Model-conditional beta flags, mirroring real CC (live captures
- * 2026-06-09, CC v2.1.170 — same binary/account, `--print -p hi`, identical
- * request shape, deterministic across repeat trials):
+ * Model-conditional anthropic-beta set, mirroring real CC 2.1.199 (live
+ * captures 2026-07-03, this box, `--print --model <m> -p hi` — same
+ * binary/account, deterministic across repeat trials). `base` is the captured
+ * opus/sonnet order (TEMPLATE.anthropic_beta); each family is a transform of
+ * it, ORDER included:
  *
- *   model            betas  effort-body  notable beta set
- *   ---------------  -----  -----------  ------------------------------------
- *   claude-opus-4-8     9   xhigh        baked base (all flags)
- *   claude-sonnet-4-6   8   high         base − mid-conversation-system
- *   claude-haiku-4-5    6   (none)       base − mid-conversation-system − effort
+ *   opus-4-8    = base                                    (unchanged)
+ *   sonnet-5    = base                                    (unchanged — KEEPS
+ *                 mid-conversation-system; the old 2.1.170 sonnet-4-6 drop is
+ *                 gone: 2.1.199 sonnet == opus)
+ *   haiku-4-5   = base − {mid-conversation-system, effort, afk-mode}, and
+ *                 claude-code-20250219 MOVED to position 5 (before advisor-tool)
+ *   fable-5     = base + fallback-credit-2026-06-01 inserted BEFORE afk-mode
  *
- * APPENDS (CC adds for these families):
- *  - `fallback-credit-2026-06-01` rides on FABLE requests only (without it,
- *    subscription fable traffic is soft-refused upstream).
- *  - `context-1m-2025-08-07` rides on `[1m]`-labelled requests only (CC does
- *    NOT send it for plain models). `skipContext1m` (dario#36) suppresses the
- *    [1m] append when the account's long-context billing was rejected.
+ * `[1m]`-labelled models additionally carry context-1m-2025-08-07 at POSITION 2
+ * (immediately after claude-code-20250219), not appended at the tail. CC does
+ * NOT send it for plain models. `skipContext1m` (dario#36) suppresses it when
+ * the account's long-context billing was rejected.
  *
- * OMISSIONS (CC drops for these families; the baked base is opus's full set):
- *  - `mid-conversation-system-2026-04-07` — sonnet + haiku omit it. All three
- *    models still send the SAME 3 system blocks, so this is a capability
- *    advertisement, not load-bearing for the system shape — safe to drop.
- *  - `effort-2025-11-24` — haiku omits it (and sends no `output_config.effort`
- *    body field either; dario already strips that field for haiku, so dropping
- *    the beta just restores consistency).
+ * afk-mode-2026-01-31 is REMOTE-CONFIG controlled and flips within a CC version
+ * (memory: rolled off 7/2, back on 7/3), so its presence is a property of the
+ * captured `base`, not of this function — the live capture heals it. haiku
+ * drops it regardless. The per-family shape here is correct whether or not the
+ * base carries afk-mode: the anchor-relative inserts/moves degrade gracefully
+ * (append / no-op) when an anchor is absent.
  *
  * Removing a beta can never provoke an upstream 400 (the runtime rejection
- * cache only ever needs to ADD strips), so the omissions are strictly safe.
- * Only the two families measured to omit them are touched; opus / fable /
- * unknown models keep the full baked set unchanged.
+ * cache only ever needs to ADD strips), so the haiku omissions are safe; the
+ * position fixes are pure wire-shape fidelity.
  */
 export declare function betaForModel(base: string, model: string | null | undefined, skipContext1m?: boolean): string;
 /**

package/dist/proxy.js

@@ -10,7 +10,7 @@ import { getAccessToken, getStatus } from './oauth.js';
 import { buildHealthResponse, derivePoolStatus, shouldDiscloseHealthInternals } from './health-response.js';
 import { darioVersion } from './version.js';
 import { buildCCRequest, applyCcPromptCaching, parseEffortSuffix, reverseMapResponse, createStreamingReverseMapper, orderHeadersForOutbound, CC_TEMPLATE } from './cc-template.js';
-import { stampCch } from './cch.js';
+import { stampCch, hasCchSeed } from './cch.js';
 import { describeTemplate, detectDrift, checkCCCompat } from './live-fingerprint.js';
 import { AccountPool, computeStickyKey, parseRateLimits, modelFamily, isInAuthCooldown, authCooldownMs, reconcilePoolAccounts } from './pool.js';
 import { Analytics, billingBucketFromClaim } from './analytics.js';
@@ -52,27 +52,69 @@ function isLoopbackAddr(addr) {
 }
 // Concurrency control: see src/request-queue.ts for the bounded queue
 // (replaced the v3.30.x-and-earlier simple unbounded semaphore in dario#80).
-// Billing tag hash seed — matches Claude Code's value
+// Deterministic seed for the cc_version build suffix. Once reverse-engineered
+// as Claude Code's own value; it is now just dario's stable constant — CC's
+// real suffix algorithm has moved (see computeVersionSuffix).
 const BILLING_SEED = '59cf53e54c78';
-// Compute per-request build tag:
-// SHA-256(seed + chars[4,7,20] of user message + version).slice(0,3)
-function computeBuildTag(userMessage, version) {
-    const chars = [4, 7, 20].map(i => userMessage[i] || '0').join('');
-    return createHash('sha256').update(`${BILLING_SEED}${chars}${version}`).digest('hex').slice(0, 3);
+// The `.<suffix>` on cc_version (e.g. `2.1.199.ef3`). A clean-room capture
+// (2026-07-03, fresh HOME with no CLAUDE.md / memory / project context) proved
+// this is NOT a function of the user message: every prompt — any content, any
+// length, bytes flipped at positions 0/4/7/20/29 — returned the SAME suffix,
+// and it shifted only when the injected SYSTEM CONTEXT changed. So CC derives
+// it from the system payload and it is STABLE for a given config. dario's old
+// `chars[4,7,20]`-of-user-message hash was doubly wrong: wrong input, and
+// per-request-varying where real CC is stable across requests.
+//
+// We can't reproduce CC's exact algorithm (it lives in the compiled binary,
+// same as the cch seed) and the value isn't population-discriminating — every
+// CC machine's suffix differs by its own context — so a STABLE, plausible 3-hex
+// derived from the template we replay is the faithful choice: it matches CC's
+// observable property (one stable suffix per config, like a single real
+// install) and changes only when the system payload changes (a template
+// rebake). Memoized — the inputs don't change within a process.
+let _versionSuffixCache = null;
+function computeVersionSuffix(version) {
+    if (_versionSuffixCache && _versionSuffixCache.key === version)
+        return _versionSuffixCache.value;
+    const sys = CC_TEMPLATE.system_prompt ?? '';
+    const value = createHash('sha256').update(`${BILLING_SEED}${version}${sys}`).digest('hex').slice(0, 3);
+    _versionSuffixCache = { key: version, value };
+    return value;
 }
 // Per-request cch PLACEHOLDER. Claude Code's cch is a deterministic xxHash64
 // over a projection of the request body (see src/cch.ts, dario#528). We can
 // only compute the real value once the final body is assembled, so we seed the
 // billing tag with a random 5-hex token here and overwrite it in place with the
-// deterministic value at serialize time (the `cchForBody` call near the
-// JSON.stringify of the outbound body). For Claude Code versions whose seed we
-// haven't reverse-engineered, `cchForBody` returns null and this random value
-// is what ships — same behavior as before dario#528, and a better tell than a
-// confident-but-wrong deterministic hash. Operators can force the random path
-// on every request with DARIO_CCH=random.
+// deterministic value at serialize time (the `stampCch` call near the
+// JSON.stringify of the outbound body).
+//
+// This placeholder is only emitted for CC versions we hold a calibrated seed
+// for — see buildBillingTag / hasCchSeed. Versions without a seed omit the cch
+// token entirely: current CC (2.1.199+) sends no cch at all, so a random
+// value would be a fingerprint, not cover. Operators can force the random
+// path on every seeded request with DARIO_CCH=random.
 function computeCch() {
     return randomBytes(3).toString('hex').slice(0, 5);
 }
+/**
+ * Assemble the `x-anthropic-billing-header` system-block text, tracking real
+ * Claude Code's wire shape:
+ *   with cch:    `…; cc_entrypoint=sdk-cli; cch=<5hex>;`   (CC ≤ 2.1.177)
+ *   without cch: `…; cc_entrypoint=sdk-cli;`               (CC 2.1.199+)
+ *
+ * `cch` is passed in rather than generated here so this stays a pure, testable
+ * string builder: the caller gates on hasCchSeed(cliVersion) and passes
+ * computeCch() (a placeholder stampCch later overwrites with the deterministic
+ * value) when a seed exists, or null when it doesn't. Passing null omits the
+ * token — never emit a cch we can't stand behind, and never emit one current
+ * CC doesn't send. The cc_version build suffix is stable per config
+ * (computeVersionSuffix), independent of the request. Exported for tests.
+ */
+export function buildBillingTag(cliVersion, cch) {
+    const fullVersion = `${cliVersion}.${computeVersionSuffix(cliVersion)}`;
+    const base = `x-anthropic-billing-header: cc_version=${fullVersion}; cc_entrypoint=sdk-cli;`;
+    return cch === null ? base : `${base} cch=${cch};`;
+}
 // Detect installed Claude Code version for the build-tag computation.
 // Falls back to the bundled template's captured version when claude isn't
 // on PATH (the common container case) — previously this fell back to a
@@ -250,61 +292,87 @@ export const FABLE_FALLBACK_CREDIT_BETA = 'fallback-credit-2026-06-01';
 export const CONTEXT_1M_BETA = 'context-1m-2025-08-07';
 export const MID_CONVERSATION_SYSTEM_BETA = 'mid-conversation-system-2026-04-07';
 export const EFFORT_BETA = 'effort-2025-11-24';
+export const AFK_MODE_BETA = 'afk-mode-2026-01-31';
+export const ADVISOR_TOOL_BETA = 'advisor-tool-2026-03-01';
+export const CLAUDE_CODE_BETA = 'claude-code-20250219';
 /**
- * Model-conditional beta flags, mirroring real CC (live captures
- * 2026-06-09, CC v2.1.170 — same binary/account, `--print -p hi`, identical
- * request shape, deterministic across repeat trials):
+ * Insert `flag` immediately before the first `anchor`, deduped. If `flag` is
+ * already present the list is returned unchanged; if `anchor` is absent `flag`
+ * is appended at the tail. Used to place model-conditional betas at the exact
+ * position real CC emits them, rather than always appending.
+ */
+function insertBetaBefore(flags, flag, anchor) {
+    if (flags.includes(flag))
+        return flags;
+    const i = flags.indexOf(anchor);
+    return i < 0 ? [...flags, flag] : [...flags.slice(0, i), flag, ...flags.slice(i)];
+}
+/** As insertBetaBefore, but places `flag` immediately AFTER the first `anchor`. */
+function insertBetaAfter(flags, flag, anchor) {
+    if (flags.includes(flag))
+        return flags;
+    const i = flags.indexOf(anchor);
+    return i < 0 ? [...flags, flag] : [...flags.slice(0, i + 1), flag, ...flags.slice(i + 1)];
+}
+/**
+ * Move an already-present `flag` to immediately before the first `anchor`.
+ * No-op when either is absent. Used for haiku, where CC emits
+ * claude-code-20250219 mid-list rather than first.
+ */
+function moveBetaBefore(flags, flag, anchor) {
+    if (!flags.includes(flag))
+        return flags;
+    const without = flags.filter((f) => f !== flag);
+    const i = without.indexOf(anchor);
+    return i < 0 ? flags : [...without.slice(0, i), flag, ...without.slice(i)];
+}
+/**
+ * Model-conditional anthropic-beta set, mirroring real CC 2.1.199 (live
+ * captures 2026-07-03, this box, `--print --model <m> -p hi` — same
+ * binary/account, deterministic across repeat trials). `base` is the captured
+ * opus/sonnet order (TEMPLATE.anthropic_beta); each family is a transform of
+ * it, ORDER included:
  *
- *   model            betas  effort-body  notable beta set
- *   ---------------  -----  -----------  ------------------------------------
- *   claude-opus-4-8     9   xhigh        baked base (all flags)
- *   claude-sonnet-4-6   8   high         base − mid-conversation-system
- *   claude-haiku-4-5    6   (none)       base − mid-conversation-system − effort
+ *   opus-4-8    = base                                    (unchanged)
+ *   sonnet-5    = base                                    (unchanged — KEEPS
+ *                 mid-conversation-system; the old 2.1.170 sonnet-4-6 drop is
+ *                 gone: 2.1.199 sonnet == opus)
+ *   haiku-4-5   = base − {mid-conversation-system, effort, afk-mode}, and
+ *                 claude-code-20250219 MOVED to position 5 (before advisor-tool)
+ *   fable-5     = base + fallback-credit-2026-06-01 inserted BEFORE afk-mode
  *
- * APPENDS (CC adds for these families):
- *  - `fallback-credit-2026-06-01` rides on FABLE requests only (without it,
- *    subscription fable traffic is soft-refused upstream).
- *  - `context-1m-2025-08-07` rides on `[1m]`-labelled requests only (CC does
- *    NOT send it for plain models). `skipContext1m` (dario#36) suppresses the
- *    [1m] append when the account's long-context billing was rejected.
+ * `[1m]`-labelled models additionally carry context-1m-2025-08-07 at POSITION 2
+ * (immediately after claude-code-20250219), not appended at the tail. CC does
+ * NOT send it for plain models. `skipContext1m` (dario#36) suppresses it when
+ * the account's long-context billing was rejected.
  *
- * OMISSIONS (CC drops for these families; the baked base is opus's full set):
- *  - `mid-conversation-system-2026-04-07` — sonnet + haiku omit it. All three
- *    models still send the SAME 3 system blocks, so this is a capability
- *    advertisement, not load-bearing for the system shape — safe to drop.
- *  - `effort-2025-11-24` — haiku omits it (and sends no `output_config.effort`
- *    body field either; dario already strips that field for haiku, so dropping
- *    the beta just restores consistency).
+ * afk-mode-2026-01-31 is REMOTE-CONFIG controlled and flips within a CC version
+ * (memory: rolled off 7/2, back on 7/3), so its presence is a property of the
+ * captured `base`, not of this function — the live capture heals it. haiku
+ * drops it regardless. The per-family shape here is correct whether or not the
+ * base carries afk-mode: the anchor-relative inserts/moves degrade gracefully
+ * (append / no-op) when an anchor is absent.
  *
  * Removing a beta can never provoke an upstream 400 (the runtime rejection
- * cache only ever needs to ADD strips), so the omissions are strictly safe.
- * Only the two families measured to omit them are touched; opus / fable /
- * unknown models keep the full baked set unchanged.
+ * cache only ever needs to ADD strips), so the haiku omissions are safe; the
+ * position fixes are pure wire-shape fidelity.
  */
 export function betaForModel(base, model, skipContext1m = false) {
-    let beta = base;
     const m = (model ?? '').toLowerCase();
-    const append = (flag) => {
-        if (beta.split(',').includes(flag))
-            return;
-        beta = beta ? `${beta},${flag}` : flag;
-    };
-    if (m.includes('fable'))
-        append(FABLE_FALLBACK_CREDIT_BETA);
-    if (/\[1m\]$/.test(m) && !skipContext1m)
-        append(CONTEXT_1M_BETA);
-    const drop = new Set();
+    let flags = base.split(',').map((s) => s.trim()).filter(Boolean);
     if (m.includes('haiku')) {
-        drop.add(MID_CONVERSATION_SYSTEM_BETA);
-        drop.add(EFFORT_BETA);
+        const drop = new Set([MID_CONVERSATION_SYSTEM_BETA, EFFORT_BETA, AFK_MODE_BETA]);
+        flags = flags.filter((f) => !drop.has(f));
+        flags = moveBetaBefore(flags, CLAUDE_CODE_BETA, ADVISOR_TOOL_BETA);
     }
-    else if (m.includes('sonnet')) {
-        drop.add(MID_CONVERSATION_SYSTEM_BETA);
+    else if (m.includes('fable')) {
+        flags = insertBetaBefore(flags, FABLE_FALLBACK_CREDIT_BETA, AFK_MODE_BETA);
     }
-    if (drop.size > 0) {
-        beta = beta.split(',').filter((b) => !drop.has(b.trim())).join(',');
+    // opus + sonnet + unknown families keep the base set unchanged.
+    if (/\[1m\]$/.test(m) && !skipContext1m) {
+        flags = insertBetaAfter(flags, CONTEXT_1M_BETA, CLAUDE_CODE_BETA);
     }
-    return beta;
+    return flags.join(',');
 }
 /**
  * Strip a trailing `[1m]` long-context tag from a model id. The tag is a
@@ -1980,10 +2048,13 @@ export async function startProxy(opts = {}) {
                         // from CC's exact template and inject only the conversation content.
                         // The upstream sees a genuine CC request structure.
                         const userMsg = extractFirstUserMessage(r);
-                        const buildTag = computeBuildTag(userMsg, cliVersion);
-                        const cch = computeCch();
-                        const fullVersion = `${cliVersion}.${buildTag}`;
-                        const billingTag = `x-anthropic-billing-header: cc_version=${fullVersion}; cc_entrypoint=sdk-cli; cch=${cch};`;
+                        // Emit a cch token only for CC versions we hold a calibrated seed
+                        // for (stampCch overwrites this placeholder with the deterministic
+                        // value at serialize time). Uncalibrated versions omit cch: current
+                        // CC (2.1.199+) sends none, so a random placeholder would be a
+                        // fingerprint rather than cover. dario#528 + 2026-07-03 drift.
+                        const cch = hasCchSeed(cliVersion) ? computeCch() : null;
+                        const billingTag = buildBillingTag(cliVersion, cch);
                         const CACHE_EPHEMERAL = { type: 'ephemeral' };
                         // Session stickiness: rebind the pre-selected pool account to
                         // whatever the sticky-key resolver picks. If this is a new
@@ -2142,10 +2213,12 @@ export async function startProxy(opts = {}) {
                     // reverse-engineered. stampCch hashes a projection of THIS final body
                     // (so it must run after every mutation above) and replaces the cch
                     // anchored to the billing tag — never a cch quoted in conversation
-                    // content. No-op for unknown versions (keep the random placeholder).
-                    // Only the template-replay path is stamped — passthrough /
-                    // count_tokens forward the client's body (and its own cch) verbatim.
-                    // Reversible kill-switch: DARIO_CCH=random.
+                    // content. No-op for uncalibrated versions — but those also carry no
+                    // cch token to begin with (buildBillingTag omits it without a seed,
+                    // matching CC 2.1.199+), so there is nothing to stamp and nothing
+                    // stale is shipped. Only the template-replay path is stamped —
+                    // passthrough / count_tokens forward the client's body (and its own
+                    // cch) verbatim. Reversible kill-switch: DARIO_CCH=random.
                     let outboundText = JSON.stringify(r);
                     if (!passthrough && !isCountTokens && process.env.DARIO_CCH !== 'random') {
                         outboundText = stampCch(outboundText, cliVersion);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@askalf/dario",
-  "version": "4.8.127",
+  "version": "4.8.129",
   "description": "Use your Claude Pro/Max subscription in any tool — Cursor, Cline, Aider, the Agent SDK, your scripts — at subscription pricing, not per-token API bills. One local Anthropic + OpenAI-compatible endpoint.",
   "type": "module",
   "bin": {
@@ -32,6 +32,8 @@
     "compat": "node test/compat.mjs",
     "lint:pkg": "node scripts/check-package-json.mjs",
     "drift:sdk": "node scripts/check-sdk-drift.mjs",
+    "drift:wire": "node scripts/check-wire-drift.mjs",
+    "check:overage": "node scripts/check-overage-live.mjs",
     "cch:calibrate": "node scripts/cch-calibrate.mjs",
     "fix:pkg": "node -e \"const fs=require('fs');fs.writeFileSync('package.json',JSON.stringify(JSON.parse(fs.readFileSync('package.json','utf-8')),null,2)+'\\n')\""
   },