@askalf/dario 6.0.52 → 6.1.0

package/README.md

@@ -27,7 +27,7 @@
 
 <p><strong>One local endpoint. Every AI tool you own. The subscriptions you already pay for.</strong></p>
 
-<sub><code>npm i -g @askalf/dario</code> · <strong>0</strong> runtime deps · <a href="https://www.npmjs.com/package/@askalf/dario">SLSA-attested</a> every release · nothing phones home · ~33k lines you can read in a weekend · independent, unofficial, third-party (<a href="DISCLAIMER.md">DISCLAIMER.md</a>)</sub>
+<sub><code>npm i -g @askalf/dario</code> · <strong>0</strong> runtime deps · <a href="https://www.npmjs.com/package/@askalf/dario">SLSA-attested</a> every release · nothing phones home · ~34k lines you can read in a weekend · independent, unofficial, third-party (<a href="DISCLAIMER.md">DISCLAIMER.md</a>)</sub>
 
 <sub><a href="#start-in-60-seconds">Start</a> · <a href="#point-your-tools-at-it">Your tools</a> · <a href="#what-it-does-with-a-request">Routing</a> · <a href="#two-plans-one-endpoint">Two plans</a> · <a href="#many-seats-one-endpoint">Pool</a> · <a href="#it-tracks-a-moving-target">Drift</a> · <a href="#trust--transparency">Trust</a> · <a href="#will-my-account-get-suspended">Risk</a> · <a href="#commands">Commands</a> · <a href="#faq">FAQ</a> · <a href="docs/returning.md">Coming back after a while?</a></sub>
 
@@ -289,6 +289,8 @@ That is a **chain**, read left to right; each provider takes the first entry it
 
 A single-entry chain is one-way and means what it always meant, so an existing config is unaffected. Failover is opt-in: without `--pool-fallback`, a drained pool still returns its honest 429/503. Only a **429 or 5xx** fails over; a 400 surfaces, because a bad request that fails over just reproduces itself on the other provider and buries the real cause. A 429 also cools that provider for a bounded interval, its `retry-after` if it sent one and 60 s otherwise, never longer than 15 min, and an entry that already declined is not asked again within the same request. When every entry is cooling, the request ends on one honest `429` with a `retry-after` instead of a retry storm. The Claude entry has to be a model the pool can actually serve, checked positively against the live catalog, so a typo can't trade a recoverable 429 for an unrecoverable 404.
 
+The chain also covers a stream that dies **mid-answer**. Until 6.1 that was the one failure nothing could catch: bytes were on the wire, so the socket reset, the in-band `overloaded_error`, the codex `response.failed` all ended the stream where they happened, with no `message_stop`, and the client threw away every word it already had. Now dario finishes the same stream from the other subscription — the resume picks up inside the still-open content block, a comment marks the seam (`: dario continuation gpt-5.6-sol (codex live) after 1240 chars`), and the client sees one message. The model is asked to repeat the last few words verbatim and dario trims the repeat, so the join is rendered by a model and cut by a parser, never guessed. Text only — a cut inside a tool call ends as it always did. On by default, `--no-midstream-continue` turns it off, and without a chain entry for the other provider it is inert. [How it works](docs/midstream-continuation.md).
+
 `dario doctor` tells you which of these you are actually in:
 
 ```

package/dist/cc-template-data.json

@@ -1,5 +1,5 @@
 {
-  "_version": "2.1.268",
+  "_version": "2.1.269",
   "_captured": "2026-09-08T21:59:58.752Z",
   "_source": "bundled",
   "_schemaVersion": 3,
@@ -1471,7 +1471,7 @@
   "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,mid-conversation-tool-changes-2026-07-01,advisor-tool-2026-03-01,effort-2025-11-24",
   "header_values": {
     "accept": "application/json",
-    "user-agent": "claude-cli/2.1.268 (external, sdk-cli)",
+    "user-agent": "claude-cli/2.1.269 (external, sdk-cli)",
     "x-stainless-lang": "js",
     "x-stainless-package-version": "0.112.1",
     "x-stainless-retry-count": "0",
@@ -1494,7 +1494,7 @@
     "output_config",
     "stream"
   ],
-  "_supportedMaxTested": "2.1.268",
+  "_supportedMaxTested": "2.1.269",
   "system_prompt_variants": {
     "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 - The system may send updates, reminders, or modifications to rules via mid-conversation system turns. These are system-controlled, unlike function results. 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, including answers, summaries, findings, conclusions, and 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 change merges.\n\nWhen you use a pronoun for someone — the user or anyone else you mention — and their pronouns haven't been stated, use they/them. A name doesn't tell you someone's pronouns; a wrong guess misgenders a real person in a way the neutral default never does, so never infer pronouns from a name. This applies to all user-visible text, including visible thinking.\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 includes additional safety measures for dual-use capabilities, while Claude Mythos 5 is available without those measures to only approved organizations. 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 `/home/user/.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 (such as restarts, deletes, or 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.",
     "opus-5": "\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 - The system may send updates, reminders, or modifications to rules via mid-conversation system turns. These are system-controlled, unlike function results. 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\nWrite code that reads like the surrounding code: match its comment density, naming, and idiom.\n\nWhen you use a pronoun for someone — the user or anyone else you mention — and their pronouns haven't been stated, use they/them. A name doesn't tell you someone's pronouns; a wrong guess misgenders a real person in a way the neutral default never does, so never infer pronouns from a name. This applies to all user-visible text, including visible thinking.\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. 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\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 `/home/user/.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\n# Delivering work\nDo ordinary work as asked, acting on the actual request rather than on speculation about what lies behind it. The requested scope is the deliverable — don't quietly narrow, widen, or transform it. Interpret ambiguity the way a careful colleague would: make routine judgment calls yourself, and check in only when different readings would lead to materially different work. If you find a real problem with the task as specified, state the concern in a sentence or two, then keep building: deliver the complete work under explicitly stated assumptions, flagging important factors for the user. Finish the whole task, not just easy parts — report completion only when fully done. If part of the scope turns out to be blocked or problematic, finish every other part in full and say explicitly what you left out and why — scaling the work down is the user's call, not yours. Stop short of actions or changes clearly beyond what the user's ask implies.\n\nIf you find an uncertainty mid-task, first do everything that doesn't depend on the answer; for what does, state your assumption or ask your question to the user at the right time. Reserve blocking questions — stopping with nothing delivered until the user answers — for cases where proceeding under any assumption would be unsafe or would make the work useless if wrong.\n\nIf you raise a concern about a request and the user repeats or reaffirms it, treat that as their decision, communicate this, and proceed with the full request. Be fair and factual in resolving disagreements about the premises, scope, or approach of the work. Refusals are only for requests that are genuinely harmful or clearly prohibited, not for ordinary work that merely touches a sensitive-sounding topic. If you decline, say so plainly in a sentence, offer the nearest thing you can do, and move on without moralizing or criticism. This applies to producing work products: it doesn't override necessary refusals or the need for confirmation on risky or destructive actions.\n\n# Corrections\nAvoid unnecessary or excessive self-correction. Only correct an earlier statement in your user-facing text when the error would change the user's code, conclusions, or decisions. State corrections plainly and concisely, and continue the task; combine multiple corrections rather than enumerating them all. For slips that change nothing for the user, simply make the correction and move on - no need to note it explicitly. Don't add apologies or preambles, don't be overly self-critical, and don't ruminate or give a detailed account of the mistake or tally past errors. Sometimes, other agents will report incorrect or misleading results - don't always take them at face value immediately. If other agents correct your statements and they are right, then simply update your approach without narrating too much about the correction to the user. This instruction does not apply to thinking blocks.\n\nA follow-up question about your earlier work is not, by itself, a signal that you got something wrong — answer what was asked. A statement that was accurate needs no correction: don't re-audit how you phrased it, how you verified it, or limits you already stated. When the user does point to a real error, correct it plainly as above.\n\nDo not use the Agent tool, workflows, or deep-research unless the user, a CLAUDE.md file, or a skill asks for it",

package/dist/cli.js

@@ -625,6 +625,11 @@ async function proxy() {
     // ProxyOptions.honorClientThinking for rationale.
     const honorClientThinking = args.includes('--honor-client-thinking')
         || ['1', 'true', 'yes', 'on'].includes((process.env['DARIO_HONOR_CLIENT_THINKING'] ?? '').toLowerCase());
+    // --no-midstream-continue / DARIO_MIDSTREAM_CONTINUE=0 — turn off finishing
+    // a streamed answer that dies mid-way from the other subscription (v6.1).
+    // On by default; see ProxyOptions.midstreamContinue.
+    const midstreamContinue = !(args.includes('--no-midstream-continue')
+        || ['0', 'false', 'no', 'off'].includes((process.env['DARIO_MIDSTREAM_CONTINUE'] ?? '').toLowerCase()));
     // --preserve-output-format — carry the client body's `output_config.format`
     // (structured-output JSON schema) through to upstream instead of dropping it
     // during the CC rebuild. See ProxyOptions.preserveOutputFormat for rationale.
@@ -649,7 +654,7 @@ async function proxy() {
         console.error(`[dario] Override (not recommended): pass --unsafe-no-auth if you have out-of-band network controls and accept the risk.`);
         process.exit(1);
     }
-    await startProxy({ port, host, verbose, verboseBodies, model, fastModel, noClaudeAuth, passthrough, preserveTools, hybridTools, mergeTools, noAutoDetect, strictTls, pacingMinMs, pacingJitterMs, thinkTimeBaseMs, thinkTimePerTokenMs, thinkTimeJitterMs, thinkTimeMaxMs, sessionStartMinMs, sessionStartJitterMs, stealth, drainOnClose, sessionIdleRotateMs, sessionRotateJitterMs, sessionMaxAgeMs, sessionPerClient, preserveOrchestrationTags, noLiveCapture, strictTemplate, maxConcurrent, maxQueued, queueTimeoutMs, maxConcurrentPerConsumer, poolStrategy, poolSharedState, poolSharedStateIntervalMs, effort, maxTokens, poolFallbackModel, modelAliases, logFile, passthroughBetas, skipFields, systemPrompt, overageGuardEnabled, overageGuardBehavior, overageGuardCooldownMs, overageGuardNotifyOs, honorClientThinking, preserveOutputFormat });
+    await startProxy({ port, host, verbose, verboseBodies, model, fastModel, noClaudeAuth, passthrough, preserveTools, hybridTools, mergeTools, noAutoDetect, strictTls, pacingMinMs, pacingJitterMs, thinkTimeBaseMs, thinkTimePerTokenMs, thinkTimeJitterMs, thinkTimeMaxMs, sessionStartMinMs, sessionStartJitterMs, stealth, drainOnClose, sessionIdleRotateMs, sessionRotateJitterMs, sessionMaxAgeMs, sessionPerClient, preserveOrchestrationTags, noLiveCapture, strictTemplate, maxConcurrent, maxQueued, queueTimeoutMs, maxConcurrentPerConsumer, poolStrategy, poolSharedState, poolSharedStateIntervalMs, effort, maxTokens, poolFallbackModel, modelAliases, logFile, passthroughBetas, skipFields, systemPrompt, overageGuardEnabled, overageGuardBehavior, overageGuardCooldownMs, overageGuardNotifyOs, honorClientThinking, preserveOutputFormat, midstreamContinue });
 }
 /**
  * Parse `--system-prompt=<verbatim|partial|aggressive|filepath>` (or the
@@ -1686,6 +1691,19 @@ async function help() {
                              is fully generated even if nobody reads
                              it) for fingerprint fidelity. Bounded by
                              the 5-minute upstream timeout. (v3.25)
+    --no-midstream-continue  Do not finish a streamed answer that dies
+                             mid-way from the other subscription. By
+                             default a stream that breaks with content
+                             already on the wire (socket reset,
+                             overloaded_error, codex response.failed)
+                             is resumed through dario's own front door
+                             at the other provider's entry in
+                             --pool-fallback, spliced onto the same
+                             client stream, and closed cleanly; the
+                             client sees one message. Off, or with no
+                             fallback entry for the other provider,
+                             the stream ends truncated as before.
+                             Env: DARIO_MIDSTREAM_CONTINUE=0. (v6.1)
     --session-idle-rotate=MS Idle ms before an account's session id
                              rotates (default: 900000 = 15 min).
                              Real CC rotates once per conversation, not

package/dist/codex-backend.d.ts

@@ -3,6 +3,7 @@ import type { CodexAccountCredentials } from './codex-accounts.js';
 import { type ResponsesReasoningConfig, type ResponsesUsage } from './anthropic-responses-translate.js';
 import { type ModelResolver, type ClaudeTarget } from './claude-model.js';
 import { type EffortValue } from './effort.js';
+import type { MidstreamGuard } from './midstream.js';
 export declare const CODEX_BACKEND_BASE_URL: string;
 /**
  * Client version sent on the model-discovery call. The backend REQUIRES the
@@ -143,6 +144,13 @@ export type CodexRequestShape = 'openai' | 'anthropic';
  * the chat-path half was the same bug, pre-existing.
  */
 export declare function isTerminalResponsesEvent(type: string): boolean;
+/**
+ * Whether a raw Responses SSE line carries a failed terminal event — the same
+ * test the chat translator applies once it has parsed the line, run here so a
+ * mid-stream continuation guard learns of the failure before the translator's
+ * error + `[DONE]` frames are written.
+ */
+export declare function isFailedResponsesLine(line: string): boolean;
 /** True when a terminal Responses payload describes a FAILURE rather than a turn. */
 export declare function isFailedResponse(resp: unknown): boolean;
 /** The upstream message on a failed Responses payload, for the error body. */
@@ -289,4 +297,11 @@ export declare function forwardToCodex(req: IncomingMessage, res: ServerResponse
  * a chat/completions caller sets `reasoning_effort` itself and that already
  * translates. Undefined leaves the request exactly as it was.
  */
-effort?: ResponsesReasoningConfig['effort']): Promise<boolean>;
+effort?: ResponsesReasoningConfig['effort'], 
+/**
+ * Mid-stream continuation guard (v6.1, src/midstream.ts). When present,
+ * every streamed frame goes through it and it owns the streaming exits: a
+ * stream that dies with content on the wire is finished from the Claude
+ * pool instead of ending truncated.
+ */
+midstream?: MidstreamGuard | null): Promise<boolean>;

package/dist/codex-backend.js

@@ -419,6 +419,23 @@ export function chatCompletionsToResponses(body) {
 export function isTerminalResponsesEvent(type) {
     return type === 'response.completed' || type === 'response.incomplete' || type === 'response.failed';
 }
+/**
+ * Whether a raw Responses SSE line carries a failed terminal event — the same
+ * test the chat translator applies once it has parsed the line, run here so a
+ * mid-stream continuation guard learns of the failure before the translator's
+ * error + `[DONE]` frames are written.
+ */
+export function isFailedResponsesLine(line) {
+    if (!line.startsWith('data: ') || !line.includes('"response.'))
+        return false;
+    try {
+        const e = JSON.parse(line.slice(6));
+        return e.type === 'response.failed' || (isTerminalResponsesEvent(e.type ?? '') && isFailedResponse(e.response));
+    }
+    catch {
+        return false;
+    }
+}
 /** True when a terminal Responses payload describes a FAILURE rather than a turn. */
 export function isFailedResponse(resp) {
     if (!resp || typeof resp !== 'object')
@@ -761,7 +778,14 @@ export async function forwardToCodex(req, res, body, creds, corsOrigin, security
  * a chat/completions caller sets `reasoning_effort` itself and that already
  * translates. Undefined leaves the request exactly as it was.
  */
-effort) {
+effort, 
+/**
+ * Mid-stream continuation guard (v6.1, src/midstream.ts). When present,
+ * every streamed frame goes through it and it owns the streaming exits: a
+ * stream that dies with content on the wire is finished from the Claude
+ * pool instead of ending truncated.
+ */
+midstream) {
     void req;
     const isAnthropic = shape === 'anthropic';
     // Reported exactly once, on every exit that answered the client. Without
@@ -832,8 +856,22 @@ effort) {
     res.on('close', onClientClose);
     // Every streamed write goes through here: after a disconnect the socket is
     // gone and writing to it is wasted at best, an EPIPE at worst.
-    const write = (chunk) => { if (!clientGone)
-        res.write(chunk); };
+    const write = (chunk) => {
+        if (clientGone)
+            return;
+        if (midstream)
+            midstream.write(chunk);
+        else
+            res.write(chunk);
+    };
+    // The streaming exits: the guard decides whether the stream is complete,
+    // continuable, or simply over.
+    const endStream = async () => {
+        if (midstream)
+            await midstream.finish();
+        else
+            res.end();
+    };
     // Usage seen so far, so a stream the client abandoned still reports what the
     // subscription already spent. Populated once the translators exist.
     let usageSoFar = () => null;
@@ -908,6 +946,11 @@ effort) {
                         terminalResponse = r;
                     if (t === 'response.failed')
                         anthropicFailed = true;
+                    // Flag it BEFORE the translator closes the turn: on this shape a
+                    // failed turn still ends in message_delta + message_stop, which the
+                    // guard must withhold to see the stream as unfinished.
+                    if (midstream && (anthropicFailed || isFailedResponse(r)))
+                        midstream.markUpstreamFailed();
                 }
                 const produced = antTranslator.push(ev);
                 if (!clientWantsStream)
@@ -951,6 +994,8 @@ effort) {
                     const lines = buffered.split('\n');
                     buffered = lines.pop() ?? '';
                     for (const line of lines) {
+                        if (midstream && isFailedResponsesLine(line))
+                            midstream.markUpstreamFailed();
                         const out = translator.chunk(line);
                         if (out && clientWantsStream)
                             write(out);
@@ -999,7 +1044,7 @@ effort) {
                 for (const out of antTranslator.end())
                     write(formatResponsesAnthropicSSE(out));
                 finished = true;
-                res.end();
+                await endStream();
             }
             else {
                 res.writeHead(200, {
@@ -1014,7 +1059,7 @@ effort) {
         }
         else if (clientWantsStream) {
             finished = true;
-            res.end();
+            await endStream();
         }
         else {
             res.writeHead(200, {
@@ -1072,9 +1117,12 @@ effort) {
             res.end(errBody('Upstream Codex backend error', { account: creds.alias }));
         }
         else {
+            // Headers are out and the socket reset under a live stream — the case
+            // the guard exists for. Without one this is the truncated stream it
+            // always was.
             finished = true;
             try {
-                res.end();
+                await endStream();
             }
             catch { /* already closed */ }
         }

package/dist/midstream.d.ts

@@ -0,0 +1,321 @@
+/**
+ * Mid-stream continuation (v6.1) — the answer does not stop when the plan does.
+ *
+ * Until now a streamed `/v1/messages` (or `/v1/chat/completions`) answer that
+ * died part-way through — an upstream socket reset, an in-band
+ * `overloaded_error`, a codex `response.failed` — ended with `res.end()` and
+ * the client got a truncated stream: no `message_stop`, no `[DONE]`, an SDK
+ * that throws "stream ended without producing a Message", and every word
+ * already on screen wasted. Once bytes were on the wire the request was
+ * treated as too late to hand to anyone else.
+ *
+ * This module finishes the SAME client stream from the other subscription
+ * instead. It sits between the request handler and the client socket:
+ *
+ *   1. every frame written to the client passes through `write()`, which
+ *      tracks what the client has already seen (message_start, the open
+ *      content block, the text so far) and WITHHOLDS a terminal error frame
+ *      rather than forwarding it;
+ *   2. `finish()` replaces the site's `res.end()`. A clean stream ends as
+ *      before. A stream that died with content on the wire re-issues the
+ *      request through dario's own front door (a loopback POST — so the pool,
+ *      the codex translator, cch and every other rule apply to the resume
+ *      exactly as to any client request) at the OTHER provider, with the
+ *      partial answer appended as the assistant turn and a resume notice as
+ *      the user turn;
+ *   3. the resume stream is spliced onto the client's still-open block: its
+ *      message_start and thinking blocks are dropped, its first text block
+ *      continues the open index, anything after that is renumbered, and it
+ *      closes the message with its own message_delta / message_stop.
+ *
+ * Two things the spike (2026-09-11, prod 6.0.51, real Opus 5 + real ChatGPT
+ * Plus) settled that are easy to get wrong again:
+ *
+ *   - NO assistant prefill. Claude 4.6+/5 answers a trailing assistant turn
+ *     with a 400, and the Responses API never had the concept. The resume is
+ *     instruction-driven on both providers, which works — zero restarts, zero
+ *     preamble, zero repetition across ten real runs.
+ *   - The seam is a WHITESPACE problem, not a content problem. Told merely to
+ *     "continue", Claude-as-continuer dropped the boundary space 2/3 times
+ *     (`replies<CUT>with`, a 3-vs-4-space indent). So the notice asks the model
+ *     to begin by repeating the last ~40 characters verbatim, and `findAnchor`
+ *     trims that repeat with a whitespace-normalized match. The model renders
+ *     the seam inside its own token stream; we only cut. Matched exactly 5/5.
+ *
+ * Out of scope here, on purpose: a cut inside a tool_use block (the partial
+ * JSON is not resumable), non-streaming requests (nothing is on the wire yet;
+ * the existing pre-byte failover covers them), and the api-key OpenAI backend.
+ * A stream that cannot be continued ends exactly as it did before this module.
+ */
+import type { ServerResponse } from 'node:http';
+export type WireShape = 'anthropic' | 'openai';
+/** Client-visible marker that a loopback request is a continuation, so the handler never nests one. */
+export declare const CONTINUATION_HEADER = "x-dario-continuation";
+/** Characters of the partial the model is asked to repeat verbatim (the seam anchor). */
+export declare const ANCHOR_CHARS = 40;
+export interface SseFrame {
+    /** The frame exactly as it will go on the wire, trailing blank line included. */
+    raw: string;
+    /** `event:` field, or the JSON `type` when the event line is absent. */
+    event: string;
+    /** Parsed `data:` payload, null for comments and non-JSON data. */
+    data: Record<string, unknown> | null;
+    /** The literal data text (`[DONE]` for the OpenAI sentinel). */
+    dataText: string | null;
+    comment: boolean;
+}
+/**
+ * Splits a byte/text stream into complete SSE frames. A trailing partial frame
+ * stays buffered until its blank line arrives. Frames are returned with their
+ * original bytes, so forwarding `raw` is byte-identical to the input.
+ */
+export declare class SseFrameSplitter {
+    private buf;
+    private readonly decoder;
+    feed(chunk: string | Uint8Array): SseFrame[];
+    /** Whatever is buffered and not yet a complete frame. */
+    flush(): string;
+}
+export declare function parseFrame(raw: string): SseFrame;
+export declare function formatFrame(event: string, data: unknown): string;
+interface BlockState {
+    type: string;
+    open: boolean;
+    text: string;
+}
+/**
+ * The client-side state a continuation has to pick up from. Both shapes are
+ * tracked by ONE class so the guard has a single view: `blocks` carries
+ * Anthropic content blocks; on the OpenAI shape there is exactly one implicit
+ * text block (index 0) that opens on the first content delta.
+ */
+export declare class ClientStreamState {
+    readonly shape: WireShape;
+    started: boolean;
+    finished: boolean;
+    blocks: BlockState[];
+    /**
+     * Frames the guard withheld instead of forwarding, in order: a terminal
+     * error, and — once the site has flagged the upstream as failed — the
+     * closing frames a translator emits for a failed turn. Released verbatim
+     * if no continuation happens, so the client sees exactly what it would have.
+     */
+    withheld: SseFrame[];
+    /**
+     * Set by the site when it KNOWS the upstream turn failed even though the
+     * translator will close it politely (the codex Anthropic path answers
+     * `response.failed` with message_delta + message_stop). The closing frames
+     * are then withheld so the stream reads as unfinished, i.e. continuable.
+     */
+    upstreamFailed: boolean;
+    /** True once anything non-continuable was seen (tool_use in progress, tool_calls). */
+    toolInProgress: boolean;
+    forwardedFrames: number;
+    /** message_start's message.model — kept so a continuation can name what the client believes it is talking to. */
+    model: string | null;
+    constructor(shape: WireShape);
+    get openIdx(): number;
+    get openType(): string | null;
+    /** Every text emitted so far, blocks concatenated in order. */
+    get textSoFar(): string;
+    /** Text of the open text block only — what the seam anchor is cut from. */
+    get openText(): string;
+    /**
+     * Whether a stream that stopped HERE can be continued: bytes are on the
+     * wire, the message is not finished, and nothing non-resumable is open.
+     * An open tool_use block or an OpenAI tool call in flight is a definite no —
+     * half a JSON argument object cannot be handed to another model.
+     */
+    get continuable(): boolean;
+    /**
+     * Observe one client-bound frame. Returns false when the frame is a terminal
+     * error the guard should withhold (recorded in `withheld`), true to forward.
+     */
+    observe(f: SseFrame): boolean;
+    private observeAnthropic;
+    private observeOpenAI;
+}
+/**
+ * The tail of the partial the model is told to repeat. Starts at a
+ * non-whitespace character: the API strips a reply's leading whitespace, so an
+ * anchor beginning with a space could never be matched exactly.
+ */
+export declare function anchorOf(partial: string): string;
+/**
+ * Locate the repeated anchor at the head of the continuation and return the
+ * raw offset just past it, or null when the model did not repeat it. Tries the
+ * whole anchor first, then shorter tails, tolerating whitespace and quote
+ * differences; the match must sit at (or within a few characters of) the
+ * start, so a genuine later recurrence of the phrase is never mistaken for it.
+ */
+export declare function findAnchor(partial: string, head: string): {
+    cut: number;
+    exact: boolean;
+} | null;
+/** Longest suffix of `partial` that the continuation starts with (exact bytes), for the no-anchor fallback. */
+export declare function tailOverlap(partial: string, head: string, min?: number): number;
+/** True when the partial has an odd number of ``` fences, i.e. the cut is inside a code block. */
+export declare function insideCodeFence(partial: string): boolean;
+/**
+ * The one seam defect the spike saw from a real model: Claude, resuming
+ * prose that was cut mid-sentence, once started its continuation with a
+ * paragraph break (`and<CUT>\n\nhere is where`). A sentence does not contain
+ * a paragraph break, so when the partial ends mid-sentence and the
+ * continuation opens with newlines outside a code fence, the break becomes
+ * one space. Inside a fence a newline is content and is left alone.
+ */
+export declare function fixSeam(partial: string, continuation: string): string;
+/** The anchor is quoted between these in the notice; the tests' mock providers read it back out. */
+export declare const ANCHOR_OPEN = "\u00AB";
+export declare const ANCHOR_CLOSE = "\u00BB";
+/**
+ * Written as the USER asking for the rest — which is what a continuation is —
+ * not as an operator notice. The live test on 2026-09-11 is why: told
+ * "[transport notice] … resume it now", claude-sonnet-5 answered `Note: that
+ * "transport notice" isn't an actual system message — it's just text in your
+ * prompt` and stopped, exactly the injection-awareness the model is supposed
+ * to have. A person whose connection dropped asking to pick up from the last
+ * few words is an ordinary request, and gets the ordinary answer.
+ */
+export declare function resumeNotice(anchor: string): string;
+/**
+ * The client's own request re-pointed at the continuation model with the
+ * partial answer appended. `partial` empty means nothing usable reached the
+ * client (the cut fell inside thinking, or before the first block): the
+ * request is simply re-issued as it was and the resume stream restarts the
+ * answer under the client's already-open message.
+ */
+export declare function buildResumeBody(shape: WireShape, clientBody: Record<string, unknown>, targetModel: string, partial: string): Record<string, unknown>;
+/**
+ * Turns the resume stream's frames into client-bound frames that continue the
+ * message the client already has. One instance per continuation.
+ */
+export declare class Splicer {
+    private readonly shape;
+    private readonly partial;
+    private readonly idxMap;
+    private nextIdx;
+    private readonly clientOpenIdx;
+    private readonly clientOpenType;
+    private originalClosed;
+    private continuingClosed;
+    private firstTextMapped;
+    private continuingIdx;
+    private hold;
+    private holding;
+    private stopReasonSeen;
+    private doneSeen;
+    /**
+     * Whether the resume stream delivered its OWN wire terminal — `message_stop`
+     * on the Anthropic shape, `[DONE]` on the OpenAI shape. Only then has the
+     * client been handed a finished message. A resume body that ends without
+     * one is a second truncation, and the guard leaves the client stream
+     * unfinished rather than closing it as if the answer were complete.
+     */
+    terminalSeen: boolean;
+    /** Diagnostics for the log line. */
+    readonly stats: {
+        anchor: "n/a" | "exact" | "fuzzy" | "overlap" | "none";
+        dropped: number;
+        emitted: number;
+    };
+    constructor(shape: WireShape, state: ClientStreamState, partial: string);
+    /** Frames to write to the client for one resume frame. */
+    feed(f: SseFrame): string[];
+    /**
+     * The resume stream ended WITHOUT its terminal event. Release whatever text
+     * was still held for the anchor check — it is the second provider's real
+     * output and the client may as well have it — but close nothing: no
+     * content_block_stop, no message_delta/message_stop, no finish chunk, no
+     * [DONE]. A synthesized clean end here would present a doubly-truncated
+     * answer as a complete one (review finding on #1286).
+     */
+    abandon(): string[];
+    private feedAnthropic;
+    /** Close the block the resume has been writing into, once. */
+    private closeContinuing;
+    /** Close the block the client had open when the primary died, once, unless the resume is continuing it. */
+    private closeOriginal;
+    private feedOpenAI;
+    /**
+     * Text from the resume's first text block is held until the anchor is
+     * found (or enough has arrived to give up looking), then trimmed and
+     * released. After that every delta streams straight through.
+     */
+    private releaseHold;
+}
+export interface ContinuationTarget {
+    /** The model spelling the loopback request carries (`codex:gpt-5.6-terra:high`, `claude:claude-opus-5`). */
+    model: string;
+    /** Human-readable, for the log line and the SSE comment. */
+    label: string;
+}
+export interface ResumeOptions {
+    /**
+     * The client's request as it arrived, before dario's own rewrites. A
+     * function so the bytes are parsed only when a resume actually happens —
+     * never on the hot path of a stream that ends normally.
+     */
+    clientBody: () => Record<string, unknown> | null;
+    /** `http://127.0.0.1:<port>` — dario's own front door. */
+    loopbackBase: string;
+    /** Auth + attribution headers for the loopback request. */
+    loopbackHeaders: Record<string, string>;
+    /** Decides where to resume, once, at failure time. Null = nowhere; the stream ends as before. */
+    resolveTarget: () => Promise<ContinuationTarget | null>;
+    /** Called right before the loopback request is made — the site releases its own queue slot here. */
+    onBeforeResume?: () => void;
+    timeoutMs: number;
+    fetchImpl?: typeof fetch;
+}
+export interface MidstreamGuardOptions {
+    shape: WireShape;
+    /** The site's client writer (already gated on client disconnect). */
+    write: (chunk: string) => void;
+    /** The real `res.end()`. */
+    end: () => void;
+    isClientGone: () => boolean;
+    resume: ResumeOptions | null;
+    requestNo: number;
+    verbose: boolean;
+    log?: (line: string) => void;
+}
+export type FinishOutcome = 'clean' | 'continued' | 'continued-unfinished' | 'ended' | 'not-continuable' | 'no-target' | 'resume-failed';
+export declare class MidstreamGuard {
+    private readonly o;
+    readonly state: ClientStreamState;
+    private readonly splitter;
+    private finished;
+    constructor(o: MidstreamGuardOptions);
+    /**
+     * The site knows the upstream turn failed (a codex `response.failed`, a
+     * terminal payload with an error status) before the translator's polite
+     * closing frames go out. From here on those frames are withheld, so the
+     * stream is treated as unfinished — and finished from the other provider.
+     */
+    markUpstreamFailed(): void;
+    /** Forward a client-bound chunk, withholding a terminal error frame. */
+    write(chunk: string | Uint8Array): void;
+    /**
+     * End the client response. Replaces the site's `res.end()` on the streaming
+     * exits. Resolves once the client stream is closed either way.
+     */
+    finish(): Promise<FinishOutcome>;
+    /**
+     * 'failed': nothing of the resume reached the client — the site ends the
+     * stream exactly as it would have. 'finished': the resume delivered its
+     * terminal event and the client holds one complete message. 'unfinished':
+     * the resume put content on the wire and then died too; the stream is left
+     * open-ended (no synthesized close) so the client sees the truncation.
+     */
+    private continueFrom;
+    private log;
+}
+/** Convenience for sites that hold a ServerResponse: the guard writes through `write`, ends through `res.end()`. */
+export declare function guardFor(res: ServerResponse, o: Omit<MidstreamGuardOptions, 'end'>): MidstreamGuard;
+/**
+ * The loopback origin for a bound listen address. A wildcard bind is reached
+ * on the loopback interface; a specific address is reached on itself.
+ */
+export declare function loopbackBaseFor(host: string, port: number): string;
+export {};