llmposter 0.5.0

Drop-in mock server for OpenAI, Anthropic & Gemini APIs — library or standalone CLI. SSE streaming, tool calling, OAuth2, failure injection, streaming chaos, stateful scenarios, request capture, hot-reload, response templating, record/replay (VCR). Test LLM apps without burning tokens.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403

//! Record-mode request path: forward the request upstream, persist
//! extractable 2xx responses as fixtures, and relay the upstream
//! response to the client.

use std::collections::HashMap;
use std::sync::Arc;

use axum::body::Body;
use axum::http::{header, Response, StatusCode};
use axum::response::IntoResponse;

use crate::format::Provider;
use crate::handler::{push_captured, ProviderHandler};
use crate::server::{AppState, RequestOutcome};

use super::{extract_for, reassemble_for, OpenAiEndpoint, Recorder};

/// Upstream RESPONSE headers copied onto the relayed response when
/// present. `retry-after` keeps client backoff logic working on
/// passed-through 429s (and pre-empts the middleware's `retry-after: 60`
/// fallback, which uses `or_insert`); the request-id pair keeps upstream
/// responses correlatable.
const RELAY_RESPONSE_HEADERS: &[&str] = &["retry-after", "x-request-id", "request-id"];

/// Upstream RESPONSE header PREFIXES relayed by lowercase prefix match —
/// the OpenAI and Anthropic rate-limit families, so client throttling
/// logic sees the real upstream budget instead of llmposter's mock values.
const RELAY_RESPONSE_HEADER_PREFIXES: &[&str] = &["x-ratelimit-", "anthropic-ratelimit-"];

/// Cap on the stream tee's capture buffer — symmetric with the
/// request-side `DefaultBodyLimit` (16 MB, server.rs). The cap bounds
/// only the RECORDING: exceeding it abandons the capture (buffer freed)
/// while the relay to a connected client continues untouched. It also
/// bounds the post-client-disconnect salvage drain against an endlessly
/// streaming upstream.
const SALVAGE_BUFFER_CAP: usize = 16 * 1024 * 1024;

/// `true` when growing the capture buffer by `chunk_len` would blow
/// [`SALVAGE_BUFFER_CAP`] — the tee abandons the recording (never the
/// relay) at that point.
fn exceeds_capture_cap(buf_len: usize, chunk_len: usize) -> bool {
    buf_len.saturating_add(chunk_len) > SALVAGE_BUFFER_CAP
}

/// What [`forward_and_relay`] needs to know about the endpoint being
/// forwarded — built from a [`ProviderHandler`] for the five generic
/// routes, or inline for the standalone embeddings handler.
struct ForwardTarget<'a> {
    path: String,
    query: Option<String>,
    provider: Provider,
    /// OpenAI-endpoint discriminator for extraction; non-OpenAI providers
    /// never consult it.
    endpoint: OpenAiEndpoint,
    is_streaming: bool,
    /// Whether a streaming response is SSE (`true` for everything except
    /// Gemini's default JSON-array `streamGenerateContent` — SSE only
    /// with `?alt=sse`). Only consulted when `is_streaming` is true.
    sse: bool,
    /// Provider-shaped error body for the 502 exits.
    error_body: &'a (dyn Fn(u16, &str) -> String + Sync),
}

/// Record-mode path for the five [`ProviderHandler`] routes. See
/// [`forward_and_relay`] for the forward/persist/relay semantics.
#[allow(clippy::too_many_arguments)] // mirrors ProviderHandler's allow — the args are the request
pub(crate) async fn record_and_respond(
    recorder: Arc<Recorder>,
    handler: &dyn ProviderHandler,
    state: &Arc<AppState>,
    headers: &HashMap<String, String>,
    body: String,
    json_body: &serde_json::Value,
    model: &str,
    user_message: &str,
) -> Response<Body> {
    let (path, query) = handler.forward_path_and_query();
    forward_and_relay(
        recorder,
        ForwardTarget {
            endpoint: OpenAiEndpoint::from_path(&path),
            path,
            query,
            provider: handler.provider(),
            is_streaming: handler.is_streaming(json_body),
            sse: handler.streaming_is_sse(),
            error_body: &|status, msg| handler.build_error_body(status, msg),
        },
        state,
        headers,
        body,
        model,
        user_message,
    )
    .await
}

/// Record-mode path for the standalone embeddings handler (which does
/// not implement [`ProviderHandler`]): fixed `/v1/embeddings` path,
/// OpenAI provider and error shape, never streaming.
pub(crate) async fn record_and_respond_embeddings(
    recorder: Arc<Recorder>,
    state: &Arc<AppState>,
    headers: &HashMap<String, String>,
    body: String,
    model: &str,
    user_message: &str,
) -> Response<Body> {
    forward_and_relay(
        recorder,
        ForwardTarget {
            path: "/v1/embeddings".to_string(),
            query: None,
            provider: Provider::OpenAI,
            endpoint: OpenAiEndpoint::Embeddings,
            is_streaming: false,
            sse: true, // never consulted — embeddings never streams
            error_body: &|status, msg| crate::failure::build_error_body(status, msg),
        },
        state,
        headers,
        body,
        model,
        user_message,
    )
    .await
}

/// Forward the request upstream and relay the response.
///
/// - 2xx non-streaming: extract → redact → dedupe → persist to the
///   cassette AND the live fixture set, synchronously, before responding.
/// - 2xx SSE streaming: teed to the client chunk-by-chunk while a spawned
///   task buffers the frames; the task reassembles and persists once the
///   stream ends cleanly, so the client never waits on the recording.
/// - 2xx Gemini JSON-array streaming (no `?alt=sse`): buffered
///   passthrough, UNRECORDED — out of capture scope by spec.
/// - Non-2xx: passed through unrecorded (a 429 must not be immortalized).
/// - Transport errors: 502 naming the upstream base — NEVER echoing any
///   header or auth material.
async fn forward_and_relay(
    recorder: Arc<Recorder>,
    target: ForwardTarget<'_>,
    state: &Arc<AppState>,
    headers: &HashMap<String, String>,
    body: String,
    model: &str,
    user_message: &str,
) -> Response<Body> {
    let ForwardTarget {
        path,
        query,
        provider,
        endpoint,
        is_streaming,
        sse,
        error_body,
    } = target;
    let capture_body = body.clone();
    let upstream_base = recorder.upstream_base(provider).to_string();

    // Shared 502 exit for both transport failures below. `err` is a
    // reqwest error stripped of its URL (`without_url`) so a Gemini
    // `?key=...` can never leak into logs or the response body.
    let bad_gateway = |err: reqwest::Error, capture_body: String| -> Response<Body> {
        let err = err.without_url();
        eprintln!(
            "[llmposter] ERROR: record-mode forward to {} failed: {}",
            upstream_base, err
        );
        push_captured(
            state,
            "POST",
            &path,
            capture_body,
            RequestOutcome::Recorded,
            None,
            502,
        );
        (
            StatusCode::BAD_GATEWAY,
            [(header::CONTENT_TYPE, "application/json")],
            error_body(
                502,
                &format!("upstream {} unreachable: {}", upstream_base, err),
            ),
        )
            .into_response()
    };

    let upstream = match recorder
        .forward(provider, &path, query.as_deref(), headers, body)
        .await
    {
        Ok(resp) => resp,
        Err(e) => return bad_gateway(e, capture_body),
    };

    // reqwest and axum share the `http` crate, so the upstream's typed
    // status and header values relay verbatim — no fallible re-parse.
    let status_code = upstream.status();
    let status = status_code.as_u16();
    let content_type = upstream
        .headers()
        .get(header::CONTENT_TYPE)
        .cloned()
        .unwrap_or_else(|| axum::http::HeaderValue::from_static("application/json"));
    // `HeaderName::as_str()` is always lowercase, so exact names and
    // prefix matches below are already case-insensitive.
    let relay_headers: Vec<(axum::http::HeaderName, axum::http::HeaderValue)> = upstream
        .headers()
        .iter()
        .filter_map(|(name, value)| {
            let lower = name.as_str();
            let keep = RELAY_RESPONSE_HEADERS.contains(&lower)
                || RELAY_RESPONSE_HEADER_PREFIXES
                    .iter()
                    .any(|prefix| lower.starts_with(prefix));
            keep.then(|| (name.clone(), value.clone()))
        })
        .collect();

    let is_2xx = (200..300).contains(&status);

    // --- 2xx SSE stream: tee upstream chunks to the client while a
    // spawned task buffers them; reassemble + persist once the stream
    // ends cleanly. The response returns immediately — the recording
    // happens strictly after the last byte reaches the buffer.
    if is_2xx && is_streaming && sse {
        let (tx, rx) = tokio::sync::mpsc::channel::<Result<axum::body::Bytes, std::io::Error>>(32);
        let state = Arc::clone(state);
        let model = model.to_string();
        let user_message = user_message.to_string();
        let task_path = path;
        // Deliberately unowned: the task may still be finishing a salvage
        // persist after MockServer::drop; runtime teardown bounds it in
        // tests.
        tokio::spawn(async move {
            let mut stream = Box::pin(upstream.bytes_stream());
            let mut buf: Vec<u8> = Vec::new();
            let mut clean = true;
            let mut recording = true;
            let mut client_connected = true;
            while let Some(item) = tokio_stream::StreamExt::next(&mut stream).await {
                match item {
                    Ok(chunk) => {
                        if recording {
                            if exceeds_capture_cap(buf.len(), chunk.len()) {
                                // Abandon the RECORDING only — the relay
                                // to a connected client continues below.
                                eprintln!(
                                    "[llmposter] record: POST {} — stream exceeded the \
                                     {} MiB capture cap — passed through, not recorded",
                                    task_path,
                                    SALVAGE_BUFFER_CAP / (1024 * 1024)
                                );
                                recording = false;
                                buf.clear();
                                buf.shrink_to_fit();
                            } else {
                                buf.extend_from_slice(&chunk);
                            }
                        }
                        if client_connected && tx.send(Ok(chunk)).await.is_err() {
                            // Client hung up — keep draining so the
                            // recording can still complete.
                            client_connected = false;
                        }
                        if !client_connected && !recording {
                            break; // nothing left to relay OR record
                        }
                    }
                    Err(e) => {
                        // `without_url` so a Gemini `?key=...` never leaks.
                        eprintln!(
                            "[llmposter] ERROR: record-mode upstream stream for POST {} \
                             failed mid-stream: {}",
                            task_path,
                            e.without_url()
                        );
                        clean = false;
                        // Surface a REAL transport error to the client —
                        // without it, dropping tx would read as a
                        // clean-looking end of a chunked response.
                        let _ = tx
                            .send(Err(std::io::Error::other("upstream stream failed")))
                            .await;
                        break;
                    }
                }
            }
            drop(tx);
            if clean && recording {
                let body = String::from_utf8_lossy(&buf);
                match reassemble_for(provider, endpoint, &body, &model, &user_message) {
                    Some(rec) => recorder.persist(rec, &state).await,
                    None => eprintln!(
                        "[llmposter] record: POST {} (model='{}') — stream had no \
                         extractable content or no completion sentinel — \
                         passed through, not recorded",
                        task_path, model
                    ),
                }
            }
            // ALWAYS pushed from the task — the capture log records what
            // the client actually received, clean or truncated.
            push_captured(
                &state,
                "POST",
                &task_path,
                capture_body,
                RequestOutcome::Recorded,
                None,
                status,
            );
        });

        let stream = tokio_stream::wrappers::ReceiverStream::new(rx);
        let mut resp = relay_response(
            status_code,
            content_type,
            &relay_headers,
            Body::from_stream(stream),
        );
        resp.headers_mut().insert(
            header::CACHE_CONTROL,
            axum::http::HeaderValue::from_static("no-cache"),
        );
        return resp;
    }

    // Exact-byte passthrough for everything else.
    let bytes = match upstream.bytes().await {
        Ok(b) => b,
        Err(e) => return bad_gateway(e, capture_body),
    };

    if is_2xx && !is_streaming {
        match serde_json::from_slice::<serde_json::Value>(&bytes) {
            Ok(value) => match extract_for(provider, endpoint, &value, model, user_message) {
                Some(rec) => recorder.persist(rec, state).await,
                None => eprintln!(
                    "[llmposter] record: POST {} (model='{}') — no extractable \
                         content in upstream response — passed through, not recorded",
                    path, model
                ),
            },
            Err(_) => eprintln!(
                "[llmposter] record: upstream 2xx body was not JSON — \
                 passed through, not recorded"
            ),
        }
    }
    // Non-2xx and Gemini JSON-array streams (is_streaming && !sse) fall
    // through here unrecorded.

    push_captured(
        state,
        "POST",
        &path,
        capture_body,
        RequestOutcome::Recorded,
        None,
        status,
    );

    relay_response(status_code, content_type, &relay_headers, Body::from(bytes))
}

/// Infallibly assemble the relayed response. Every input is a typed
/// `http` value that came off the parsed upstream response, so there is
/// no fallible builder step — and no unreachable error arm to maintain.
fn relay_response(
    status: StatusCode,
    content_type: axum::http::HeaderValue,
    relay_headers: &[(axum::http::HeaderName, axum::http::HeaderValue)],
    body: Body,
) -> Response<Body> {
    let mut resp = Response::new(body);
    *resp.status_mut() = status;
    resp.headers_mut()
        .insert(header::CONTENT_TYPE, content_type);
    for (name, value) in relay_headers {
        resp.headers_mut().append(name.clone(), value.clone());
    }
    resp
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn should_trip_capture_cap_only_when_buffer_would_exceed_it() {
        // At the cap exactly is fine; one byte past it trips.
        assert!(!exceeds_capture_cap(0, SALVAGE_BUFFER_CAP));
        assert!(!exceeds_capture_cap(SALVAGE_BUFFER_CAP - 5, 5));
        assert!(exceeds_capture_cap(SALVAGE_BUFFER_CAP - 5, 6));
        assert!(exceeds_capture_cap(SALVAGE_BUFFER_CAP, 1));
        // Saturating add — no overflow panic on absurd lengths.
        assert!(exceeds_capture_cap(usize::MAX, usize::MAX));
    }
}