tj_core/classifier/
agent_sdk.rs

1//! Claude CLI ("agent SDK") classifier backend.
2//!
3//! Runs the locally-installed, already-authenticated `claude` binary in
4//! non-interactive print mode, pinned to Haiku, to classify a chunk *without*
5//! an `ANTHROPIC_API_KEY`. This resurrects the v0.7.x `cli` backend that was
6//! removed in v0.8.0 — but honestly: since **2026-06-15** a headless
7//! `claude -p` run draws from the separate **Agent SDK** monthly credit pool
8//! (~$20 Pro / $100 Max 5x / $200 Max 20x, at API rates), not the interactive
9//! Pro/Max pool. Classification is Haiku-class and tiny (a few hundred tokens
10//! per chunk), so the credit lasts a long time — but it is not strictly free.
11//!
12//! The command execution is abstracted behind [`CommandRunner`] so the parsing
13//! path is unit-testable with a fake; the suite never shells out to `claude`.
14
15use super::{Classifier, ClassifyInput, ClassifyOutput};
16use anyhow::{anyhow, Context};
17use std::process::Command;
18
19/// Default model. `claude --model` accepts the short alias and resolves it to
20/// the current dated id (`claude-haiku-4-5-20251001`). Override with
21/// `TJ_AGENT_SDK_MODEL`.
22pub const DEFAULT_MODEL: &str = "claude-haiku-4-5";
23
24/// Env var stamped onto every spawned classifier `claude -p` subprocess. That
25/// subprocess is a full Claude Code instance, so on startup it re-runs the
26/// user's SessionStart hooks — including `task-journal ingest-hook`, which
27/// would spawn yet another classifier `claude -p`, and so on: an unbounded
28/// fork bomb. `ingest-hook` checks for this marker and no-ops when it is set,
29/// breaking the recursion. The CLI guard and the worker's `env_remove` both
30/// reference this constant so the setter and the checker can never drift
31/// (which is exactly the bug that let the fork bomb through: the guard checked
32/// `TJ_IN_CLASSIFIER` but no spawn site ever set it).
33pub const IN_CLASSIFIER_ENV: &str = "TJ_IN_CLASSIFIER";
34
35/// "Run the classifier command and hand back its raw stdout." The production
36/// impl shells out to `claude`; tests inject a fake returning canned JSON.
37pub trait CommandRunner: Send + Sync {
38    /// Run the classification for `prompt` against `model`, returning the raw
39    /// stdout (the `--output-format json` wrapper) on success.
40    fn run(&self, model: &str, prompt: &str) -> anyhow::Result<String>;
41}
42
43/// Build the base `claude` invocation shared by both runners: print mode, the
44/// pinned model, the JSON envelope, an isolated MCP config, and — critically —
45/// the [`IN_CLASSIFIER_ENV`] recursion marker. The argv runner appends the
46/// prompt as a positional arg; the stdin runner feeds it on stdin. Extracted so
47/// a unit test can assert the marker is present without spawning `claude` (the
48/// missing marker is exactly what let the fork bomb through before).
49fn base_claude_command(model: &str) -> Command {
50    let mut cmd = Command::new("claude");
51    cmd.arg("-p")
52        .arg("--model")
53        .arg(model)
54        .arg("--output-format")
55        .arg("json")
56        .arg("--strict-mcp-config")
57        .env(IN_CLASSIFIER_ENV, "1");
58    cmd
59}
60
61/// Production runner: invokes the local `claude` binary in print mode, pinned
62/// to the given model, asking for the JSON envelope and an isolated MCP config
63/// (`--strict-mcp-config` keeps the project's own MCP servers — including this
64/// very journal — out of the classification subprocess).
65pub struct ClaudeBinaryRunner;
66
67/// Build the error for a non-zero `claude -p` exit. With `--output-format
68/// json` claude reports the real cause (invalid model, usage limit, auth) as
69/// JSON on **stdout**, not stderr — so surface both, capped, or the user just
70/// sees a bare "exit status 1".
71fn claude_exit_error(
72    status: std::process::ExitStatus,
73    stdout: &[u8],
74    stderr: &[u8],
75) -> anyhow::Error {
76    let cap = |b: &[u8]| {
77        let s = String::from_utf8_lossy(b);
78        let s = s.trim().to_string();
79        if s.chars().count() > 600 {
80            format!("{}…", s.chars().take(600).collect::<String>())
81        } else {
82            s
83        }
84    };
85    let out = cap(stdout);
86    let err = cap(stderr);
87    let detail = match (out.is_empty(), err.is_empty()) {
88        (true, true) => "(no output)".to_string(),
89        (false, true) => out,
90        (true, false) => err,
91        (false, false) => format!("{err} | stdout: {out}"),
92    };
93    anyhow!("`claude -p` exited with {status}: {detail}")
94}
95
96impl CommandRunner for ClaudeBinaryRunner {
97    fn run(&self, model: &str, prompt: &str) -> anyhow::Result<String> {
98        let output = base_claude_command(model)
99            .arg(prompt)
100            .output()
101            .context("failed to spawn `claude` (is Claude Code installed and on PATH?)")?;
102        if !output.status.success() {
103            return Err(claude_exit_error(
104                output.status,
105                &output.stdout,
106                &output.stderr,
107            ));
108        }
109        Ok(String::from_utf8_lossy(&output.stdout).into_owned())
110    }
111}
112
113/// Like [`ClaudeBinaryRunner`] but feeds the prompt on **stdin** instead of as
114/// an argv argument. Use for large prompts (e.g. a whole session transcript in
115/// dream backfill) that would otherwise blow the per-argument size limit
116/// (`E2BIG`, ~128 KiB on Linux). `claude -p` with no positional prompt reads
117/// the prompt from stdin.
118pub struct ClaudeBinaryStdinRunner;
119
120impl CommandRunner for ClaudeBinaryStdinRunner {
121    fn run(&self, model: &str, prompt: &str) -> anyhow::Result<String> {
122        use std::io::Write;
123        use std::process::Stdio;
124        let mut child = base_claude_command(model)
125            .stdin(Stdio::piped())
126            .stdout(Stdio::piped())
127            .stderr(Stdio::piped())
128            .spawn()
129            .context("failed to spawn `claude` (is Claude Code installed and on PATH?)")?;
130        // Write the prompt, then drop the handle to close stdin so `claude`
131        // sees EOF and starts working.
132        child
133            .stdin
134            .take()
135            .context("claude stdin was not captured")?
136            .write_all(prompt.as_bytes())
137            .context("failed to write prompt to claude stdin")?;
138        let output = child
139            .wait_with_output()
140            .context("failed to wait for `claude`")?;
141        if !output.status.success() {
142            return Err(claude_exit_error(
143                output.status,
144                &output.stdout,
145                &output.stderr,
146            ));
147        }
148        Ok(String::from_utf8_lossy(&output.stdout).into_owned())
149    }
150}
151
152pub struct ClaudeCliClassifier {
153    model: String,
154    runner: Box<dyn CommandRunner>,
155}
156
157impl ClaudeCliClassifier {
158    /// Build from environment. Returns `None` unless a `claude` binary is on
159    /// PATH (probed with `claude --version`) — the caller then falls through to
160    /// the next backend. Model comes from `TJ_AGENT_SDK_MODEL`, else Haiku.
161    pub fn from_env() -> Option<Self> {
162        if !claude_on_path() {
163            return None;
164        }
165        let model = std::env::var("TJ_AGENT_SDK_MODEL").unwrap_or_else(|_| DEFAULT_MODEL.into());
166        Some(Self {
167            model,
168            runner: Box::new(ClaudeBinaryRunner),
169        })
170    }
171
172    /// Test/dev constructor: inject a fake runner and an explicit model so the
173    /// parse path can be exercised without a live `claude` login.
174    pub fn with_runner(model: impl Into<String>, runner: Box<dyn CommandRunner>) -> Self {
175        Self {
176            model: model.into(),
177            runner,
178        }
179    }
180}
181
182/// The JSON wrapper emitted by `claude --output-format json`. We only need the
183/// error flag and the `result` string (the model's verdict text); the rest of
184/// the envelope (usage, cost, timings) is ignored.
185#[derive(serde::Deserialize)]
186struct CliEnvelope {
187    #[serde(default)]
188    is_error: bool,
189    #[serde(default)]
190    result: Option<String>,
191    #[serde(default)]
192    subtype: Option<String>,
193}
194
195impl Classifier for ClaudeCliClassifier {
196    fn classify(&self, input: &ClassifyInput) -> anyhow::Result<ClassifyOutput> {
197        let prompt = crate::classifier::prompt::build(input);
198        let verdict = run_claude_json(self.runner.as_ref(), &self.model, &prompt)?;
199        super::parse_verdict(&verdict)
200    }
201}
202
203/// Run `prompt` through the claude CLI (via `runner`) and return the model's
204/// reply text — the `result` field of the `--output-format json` envelope.
205/// Shared by the classifier and the dream agent-sdk backend so the envelope
206/// handling lives in one place.
207pub fn run_claude_json(
208    runner: &dyn CommandRunner,
209    model: &str,
210    prompt: &str,
211) -> anyhow::Result<String> {
212    let stdout = runner.run(model, prompt)?;
213    let envelope: CliEnvelope = serde_json::from_str(stdout.trim()).with_context(|| {
214        format!(
215            "claude --output-format json wrapper parse failed; got: {}",
216            stdout.trim()
217        )
218    })?;
219    if envelope.is_error {
220        return Err(anyhow!(
221            "claude reported an error (subtype={})",
222            envelope.subtype.as_deref().unwrap_or("unknown")
223        ));
224    }
225    envelope
226        .result
227        .ok_or_else(|| anyhow!("claude json wrapper had no `result` field"))
228}
229
230/// Probe whether `claude` resolves on PATH and runs. Cheap (`--version` does
231/// no network) and tolerant — any spawn/exec failure means "not available".
232pub fn claude_on_path() -> bool {
233    Command::new("claude")
234        .arg("--version")
235        .output()
236        .map(|o| o.status.success())
237        .unwrap_or(false)
238}
239
240#[cfg(test)]
241mod tests {
242    use super::*;
243    use crate::classifier::{decide_status, CONFIDENCE_THRESHOLD};
244    use crate::event::{EventStatus, EventType};
245
246    /// Fake runner: returns canned stdout, ignoring model/prompt. Captures the
247    /// model it was asked for so tests can assert the pin.
248    struct FakeRunner {
249        canned: String,
250        seen_model: std::sync::Mutex<Option<String>>,
251    }
252
253    impl FakeRunner {
254        fn new(canned: impl Into<String>) -> Self {
255            Self {
256                canned: canned.into(),
257                seen_model: std::sync::Mutex::new(None),
258            }
259        }
260    }
261
262    impl CommandRunner for FakeRunner {
263        fn run(&self, model: &str, _prompt: &str) -> anyhow::Result<String> {
264            *self.seen_model.lock().unwrap() = Some(model.to_string());
265            Ok(self.canned.clone())
266        }
267    }
268
269    fn input() -> ClassifyInput {
270        ClassifyInput {
271            text: "We adopted Rust for the journal core.".into(),
272            author_hint: "assistant".into(),
273            recent_tasks: vec![],
274        }
275    }
276
277    fn envelope(result_json: &str) -> String {
278        serde_json::json!({
279            "type": "result",
280            "subtype": "success",
281            "is_error": false,
282            "result": result_json,
283        })
284        .to_string()
285    }
286
287    #[test]
288    fn base_command_carries_recursion_marker() {
289        use std::ffi::OsStr;
290        // The tj-cli ingest-hook guard short-circuits on this exact var; if the
291        // const and the spawn site ever drift, the fork bomb returns.
292        assert_eq!(IN_CLASSIFIER_ENV, "TJ_IN_CLASSIFIER");
293        let cmd = base_claude_command("claude-haiku-4-5");
294        let marker = cmd
295            .get_envs()
296            .any(|(k, v)| k == OsStr::new(IN_CLASSIFIER_ENV) && v == Some(OsStr::new("1")));
297        assert!(
298            marker,
299            "every spawned `claude -p` must set {IN_CLASSIFIER_ENV}=1 to break ingest-hook recursion"
300        );
301    }
302
303    #[test]
304    fn parses_canned_verdict_into_classify_output() {
305        let verdict = r#"{"event_type":"decision","task_id_guess":"tj-x","confidence":0.93,"evidence_strength":null,"suggested_text":"Adopt Rust."}"#;
306        let c = ClaudeCliClassifier::with_runner(
307            DEFAULT_MODEL,
308            Box::new(FakeRunner::new(envelope(verdict))),
309        );
310        let out = c.classify(&input()).unwrap();
311        assert_eq!(out.event_type, EventType::Decision);
312        assert_eq!(out.task_id_guess.as_deref(), Some("tj-x"));
313        assert!((out.confidence - 0.93).abs() < 1e-6);
314        // 0.93 >= 0.85 → confirmed.
315        assert_eq!(decide_status(out.confidence), EventStatus::Confirmed);
316    }
317
318    /// Adapter so a test can keep an `Arc` handle to inspect the runner after
319    /// it is boxed into the classifier.
320    struct ArcRunner(std::sync::Arc<FakeRunner>);
321    impl CommandRunner for ArcRunner {
322        fn run(&self, model: &str, prompt: &str) -> anyhow::Result<String> {
323            self.0.run(model, prompt)
324        }
325    }
326
327    #[test]
328    fn pins_the_configured_model() {
329        let verdict = r#"{"event_type":"finding","task_id_guess":null,"confidence":0.9,"evidence_strength":null,"suggested_text":"x"}"#;
330        let captured = std::sync::Arc::new(FakeRunner::new(envelope(verdict)));
331        let c = ClaudeCliClassifier::with_runner(
332            "claude-haiku-4-5",
333            Box::new(ArcRunner(captured.clone())),
334        );
335        let _ = c.classify(&input()).unwrap();
336        assert_eq!(
337            captured.seen_model.lock().unwrap().as_deref(),
338            Some("claude-haiku-4-5"),
339            "classifier must pin the model it was constructed with"
340        );
341    }
342
343    #[test]
344    fn decide_status_at_the_0_85_threshold() {
345        for (conf, expect) in [
346            (0.85_f64, EventStatus::Confirmed),
347            (0.84_f64, EventStatus::Suggested),
348        ] {
349            let verdict = format!(
350                r#"{{"event_type":"evidence","task_id_guess":null,"confidence":{conf},"evidence_strength":"strong","suggested_text":"t"}}"#
351            );
352            let c = ClaudeCliClassifier::with_runner(
353                DEFAULT_MODEL,
354                Box::new(FakeRunner::new(envelope(&verdict))),
355            );
356            let out = c.classify(&input()).unwrap();
357            assert!((out.confidence - conf).abs() < 1e-6);
358            assert_eq!(decide_status(out.confidence), expect);
359            assert_eq!(CONFIDENCE_THRESHOLD, 0.85);
360        }
361    }
362
363    #[test]
364    fn tolerates_code_fence_wrapped_verdict() {
365        let verdict = "```json\n{\"event_type\":\"rejection\",\"task_id_guess\":null,\"confidence\":0.88,\"evidence_strength\":null,\"suggested_text\":\"won't work\"}\n```";
366        let c = ClaudeCliClassifier::with_runner(
367            DEFAULT_MODEL,
368            Box::new(FakeRunner::new(envelope(verdict))),
369        );
370        let out = c.classify(&input()).unwrap();
371        assert_eq!(out.event_type, EventType::Rejection);
372    }
373
374    #[test]
375    fn errors_when_claude_reports_is_error() {
376        let canned = serde_json::json!({
377            "type": "result",
378            "subtype": "error_during_execution",
379            "is_error": true,
380            "result": null,
381        })
382        .to_string();
383        let c = ClaudeCliClassifier::with_runner(DEFAULT_MODEL, Box::new(FakeRunner::new(canned)));
384        let err = c.classify(&input()).unwrap_err();
385        assert!(format!("{err}").contains("error"), "got: {err}");
386    }
387}
tj_core/classifier/agent_sdk.rs

tj_core/classifier/
agent_sdk.rs
