speccoder_runtime/
attempts.rs

1//! Durable attempt outcomes.
2//!
3//! A passed task is a commit on `speccoder/<feature>`. Every other outcome
4//! — a failed verdict, a stop, an undemonstrated gap, an interrupted run —
5//! is saved as a commit object whose tree is the worktree as the attempt
6//! left it, parented on the branch tip, under
7//! `refs/speccoder/attempts/<feature>/<task>/<run>`, with its outcome in
8//! trailers. The worktree then goes back to where the attempt started, so
9//! the next attempt never builds on (or freezes tests from) unfinished work,
10//! and nothing is lost: `git diff speccoder/<feature> <ref>` shows it.
11//! Git is the durable record; the journal can be replaced across upgrades.
12
13use std::path::Path;
14
15use crate::git::{Git, GitError};
16use crate::worktree::{self, Checkpoint};
17
18/// Where attempt refs live.
19pub const PREFIX: &str = "refs/speccoder/attempts";
20
21/// Field and record separators for `git for-each-ref` output.
22const FS: char = '\u{1f}';
23const RS: char = '\u{1e}';
24
25/// An attempt saved under [`PREFIX`].
26#[derive(Debug, Clone, PartialEq, Eq)]
27pub struct Saved {
28    /// Task ID.
29    pub task: String,
30    /// `blocked`, `verification_failed`, `not_demonstrated`, `interrupted`, …
31    pub outcome: String,
32    /// Why, in one or a few lines.
33    pub reason: String,
34    /// The agentplane run, when there was one.
35    pub run_id: Option<String>,
36    /// The call that awaited approval (`{tool, arguments, reason}`), if any.
37    pub pending: Option<serde_json::Value>,
38    /// Model tokens the attempt used.
39    pub usage: Option<serde_json::Value>,
40    /// The story's scenarios still red when its last task stopped.
41    pub story_open: Vec<String>,
42    /// The task's own test that failed, or that the implementer called
43    /// wrong (a task outside a story): it can be ruled wrong.
44    pub own_test: Option<String>,
45    /// The full ref name.
46    pub reference: String,
47    /// The commit it points at.
48    pub commit: String,
49}
50
51pub(crate) const IDENTITY: [&str; 6] = [
52    "-c",
53    "user.name=speccoder",
54    "-c",
55    "user.email=speccoder@users.noreply.invalid",
56    "-c",
57    "commit.gpgsign=false",
58];
59
60/// What an attempt ended with.
61#[derive(Debug, Clone, Copy)]
62pub struct Record<'a> {
63    /// The agentplane run, when there was one (it names the ref).
64    pub run: Option<&'a str>,
65    /// `blocked`, `verification_failed`, …
66    pub outcome: &'a str,
67    /// Why.
68    pub reason: &'a str,
69    /// The call awaiting approval, if the attempt stopped for one.
70    pub pending: Option<&'a serde_json::Value>,
71    /// Model tokens the attempt used (`cost`).
72    pub usage: Option<&'a serde_json::Value>,
73    /// The story's scenarios still red, when the story's last task stopped.
74    pub story_open: &'a [String],
75    /// The task's own test that failed or was called wrong (outside a story).
76    pub own_test: Option<&'a str>,
77}
78
79/// The `Usage:` trailer speccoder writes on its commits, or nothing.
80#[must_use]
81pub fn usage_trailer(usage: Option<&serde_json::Value>) -> String {
82    let Some(u) = usage.filter(|u| u.is_object()) else {
83        return String::new();
84    };
85    let n = |k: &str| u[k].as_u64().unwrap_or(0);
86    // An external agent that reports its own cost (Claude Code) keeps it.
87    let usd = u["cost_usd"]
88        .as_f64()
89        .map(|c| format!(" usd={c}"))
90        .unwrap_or_default();
91    format!(
92        "Usage: input={} output={} cache_read={} cache_write={}{usd}\n",
93        n("input_tokens"),
94        n("output_tokens"),
95        n("cache_read_tokens"),
96        n("cache_write_tokens")
97    )
98}
99
100/// Reads a `Usage:` trailer value back into a usage object.
101#[must_use]
102pub fn parse_usage(value: &str) -> Option<serde_json::Value> {
103    let mut out = serde_json::Map::new();
104    for part in value.split_whitespace() {
105        let (k, v) = part.split_once('=')?;
106        let key = match k {
107            "input" => "input_tokens",
108            "output" => "output_tokens",
109            "cache_read" => "cache_read_tokens",
110            "cache_write" => "cache_write_tokens",
111            "usd" => {
112                out.insert(
113                    "cost_usd".to_owned(),
114                    serde_json::json!(v.parse::<f64>().ok()?),
115                );
116                continue;
117            }
118            _ => continue,
119        };
120        out.insert(key.to_owned(), serde_json::json!(v.parse::<u64>().ok()?));
121    }
122    (!out.is_empty()).then_some(serde_json::Value::Object(out))
123}
124
125/// Saves the worktree at `dir` as an attempt of `feature/task`.
126pub fn save(
127    dir: &Path,
128    tmp: &Path,
129    feature: &str,
130    task: &str,
131    record: &Record<'_>,
132) -> Result<Saved, GitError> {
133    let Record {
134        run,
135        outcome,
136        reason,
137        pending,
138        usage,
139        story_open,
140        own_test,
141    } = *record;
142    let git = Git::at(dir);
143    let tree = worktree::checkpoint(dir, tmp)?;
144    let head = git.line(&["rev-parse", "--verify", "HEAD^{commit}"])?;
145    // A file whose change holds a value copied from the person's secret
146    // files is left out: a saved attempt is a ref anyone with the
147    // repository sees.
148    let (tree, left_out) = without_secrets(&git, tmp, &head, tree)?;
149    let reason = if left_out.is_empty() {
150        reason.to_owned()
151    } else {
152        format!(
153            "{reason} ({} left out: it held a value copied from your checkout)",
154            left_out.join(", ")
155        )
156    };
157    let reason = reason.as_str();
158    let message = format!(
159        "speccoder attempt: {task} {outcome}\n\n{reason}\n\nSpec-Task: {feature}/{task}\nOutcome: {outcome}\n{}{}{}{}{}",
160        run.map(|r| format!("Run: {r}\n")).unwrap_or_default(),
161        Some(
162            story_open
163                .iter()
164                .filter(|t| t.chars().all(|c| c.is_ascii_alphanumeric() || c == '-'))
165                .map(String::as_str)
166                .collect::<Vec<_>>()
167                .join(", ")
168        )
169        .filter(|t| !t.is_empty())
170        .map(|t| format!("Story-Open: {t}\n"))
171        .unwrap_or_default(),
172        own_test
173            .filter(|p| !p.contains('\n'))
174            .map(|p| format!("Own-Test: {p}\n"))
175            .unwrap_or_default(),
176        usage_trailer(usage),
177        // Compact JSON is one line: a trailer value.
178        pending
179            .map(|p| format!("Pending: {p}\n"))
180            .unwrap_or_default()
181    );
182    let mut args = IDENTITY.to_vec();
183    args.extend(["commit-tree", &tree.0, "-p", &head, "-m", &message]);
184    let commit = git.line(&args)?;
185    let name = run
186        .filter(|r| {
187            r.chars()
188                .all(|c| c.is_ascii_alphanumeric() || c == '_' || c == '-')
189        })
190        .map_or_else(
191            || format!("{outcome}-{}", commit.get(..12).unwrap_or(&commit)),
192            str::to_owned,
193        );
194    let reference = format!("{PREFIX}/{feature}/{task}/{name}");
195    git.run(&["update-ref", &reference, &commit])?;
196    Ok(Saved {
197        task: task.to_owned(),
198        outcome: outcome.to_owned(),
199        reason: reason.to_owned(),
200        run_id: run.map(str::to_owned),
201        pending: pending.cloned(),
202        usage: usage.cloned(),
203        story_open: story_open.to_vec(),
204        own_test: own_test.map(str::to_owned),
205        reference,
206        commit,
207    })
208}
209
210/// `tree` without each file whose change from `head` adds a value
211/// [`crate::secrets`] masks, and those files.
212fn without_secrets(
213    git: &Git,
214    tmp: &Path,
215    head: &str,
216    tree: Checkpoint,
217) -> Result<(Checkpoint, Vec<String>), GitError> {
218    let changed = crate::git::nul_fields(&git.run(&[
219        "diff",
220        "--name-only",
221        "-z",
222        "--no-renames",
223        head,
224        &tree.0,
225    ])?);
226    let held: Vec<String> = changed
227        .into_iter()
228        .filter(|f| {
229            git.run(&["diff", "-U0", "--no-color", head, &tree.0, "--", f])
230                .is_ok_and(|p| crate::include::added_secret(&String::from_utf8_lossy(&p)).is_some())
231        })
232        .collect();
233    if held.is_empty() {
234        return Ok((tree, held));
235    }
236    let index = tmp.join(format!("attempt-index-{}", std::process::id()));
237    let _ = std::fs::create_dir_all(tmp);
238    git.run_with_index(&index, &["read-tree", &tree.0], None)?;
239    let mut rm = vec!["update-index", "--force-remove", "--"];
240    rm.extend(held.iter().map(String::as_str));
241    git.run_with_index(&index, &rm, None)?;
242    let out = git.run_with_index(&index, &["write-tree"], None);
243    let _ = std::fs::remove_file(&index);
244    let kept = String::from_utf8_lossy(&out?).trim().to_owned();
245    Ok((Checkpoint(kept), held))
246}
247
248/// Makes the worktree at `dir` equal to its HEAD commit's tree.
249pub fn restore_head(dir: &Path, tmp: &Path) -> Result<(), GitError> {
250    let tree = Git::at(dir).line(&["rev-parse", "HEAD^{tree}"])?;
251    worktree::restore(dir, tmp, &Checkpoint(tree)).map(|_| ())
252}
253
254/// The saved attempts of `feature`, newest first.
255pub fn list(repo: &Path, feature: &str) -> Vec<Saved> {
256    let format = format!(
257        "--format=%(refname){FS}%(objectname){FS}%(trailers:key=Spec-Task,valueonly){FS}%(trailers:key=Outcome,valueonly){FS}%(trailers:key=Run,valueonly){FS}%(trailers:key=Pending,valueonly){FS}%(trailers:key=Usage,valueonly){FS}%(trailers:key=Story-Open,valueonly){FS}%(trailers:key=Own-Test,valueonly){FS}%(contents:body){RS}"
258    );
259    let out = Git::at(repo)
260        .run(&[
261            "for-each-ref",
262            "--sort=-committerdate",
263            &format,
264            &format!("{PREFIX}/{feature}/"),
265        ])
266        .map(|b| String::from_utf8_lossy(&b).into_owned())
267        .unwrap_or_default();
268    out.split(RS)
269        .filter_map(|record| {
270            let f: Vec<&str> = record.trim_start_matches('\n').split(FS).collect();
271            let [
272                reference,
273                commit,
274                task,
275                outcome,
276                run,
277                pending,
278                usage,
279                story_open,
280                own_test,
281                body,
282            ] = f[..]
283            else {
284                return None;
285            };
286            let task = task.trim().strip_prefix(&format!("{feature}/"))?.to_owned();
287            // The reason is the body before the trailer block.
288            let reason = body
289                .split("\n\nSpec-Task:")
290                .next()
291                .unwrap_or_default()
292                .trim()
293                .to_owned();
294            let run = run.trim();
295            Some(Saved {
296                task,
297                outcome: outcome.trim().to_owned(),
298                reason,
299                run_id: (!run.is_empty()).then(|| run.to_owned()),
300                pending: serde_json::from_str(pending.trim()).ok(),
301                usage: parse_usage(usage.trim()),
302                story_open: story_open
303                    .split(',')
304                    .map(str::trim)
305                    .filter(|t| !t.is_empty())
306                    .map(str::to_owned)
307                    .collect(),
308                own_test: Some(own_test.trim())
309                    .filter(|p| !p.is_empty())
310                    .map(str::to_owned),
311                reference: reference.to_owned(),
312                commit: commit.to_owned(),
313            })
314        })
315        .collect()
316}
317
318#[cfg(test)]
319mod tests {
320    use super::*;
321
322    #[test]
323    fn a_file_holding_a_copied_value_is_left_out_of_a_saved_attempt() {
324        let dir = tempfile::tempdir().unwrap();
325        let r = dir.path();
326        let git = |args: &[&str]| {
327            assert!(
328                std::process::Command::new("git")
329                    .args([
330                        "-c",
331                        "user.name=t",
332                        "-c",
333                        "user.email=t@e",
334                        "-c",
335                        "commit.gpgsign=false"
336                    ])
337                    .args(args)
338                    .current_dir(r)
339                    .status()
340                    .unwrap()
341                    .success()
342            );
343        };
344        std::fs::write(r.join("a.txt"), "a\n").unwrap();
345        git(&["init", "-q"]);
346        git(&["add", "-A"]);
347        git(&["commit", "-qm", "init"]);
348        crate::secrets::register("SNAP_TOKEN", "attempt-snap-Vb3 value");
349        std::fs::write(r.join("a.txt"), "a\nb\n").unwrap();
350        std::fs::write(r.join("snap.txt"), "token: attempt-snap-Vb3 value\n").unwrap();
351        let tmp = tempfile::tempdir().unwrap();
352        let saved = save(
353            r,
354            tmp.path(),
355            "001-x",
356            "T001",
357            &Record {
358                run: None,
359                outcome: "failed",
360                reason: "red",
361                pending: None,
362                usage: None,
363                story_open: &[],
364                own_test: None,
365            },
366        )
367        .unwrap();
368        let g = Git::at(r);
369        let files = g
370            .run(&["ls-tree", "-r", "--name-only", &saved.commit])
371            .unwrap();
372        assert_eq!(String::from_utf8_lossy(&files), "a.txt\n");
373        assert!(
374            saved.reason.contains("snap.txt left out"),
375            "{}",
376            saved.reason
377        );
378    }
379
380    #[test]
381    fn usage_trailers_round_trip_with_a_reported_cost() {
382        let u = serde_json::json!({ "input_tokens": 100, "output_tokens": 5, "cache_read_tokens": 90, "cache_write_tokens": 3, "cost_usd": 0.0125 });
383        let t = usage_trailer(Some(&u));
384        assert_eq!(
385            t,
386            "Usage: input=100 output=5 cache_read=90 cache_write=3 usd=0.0125\n"
387        );
388        let back = parse_usage(t.trim().trim_start_matches("Usage: ")).unwrap_or_default();
389        assert_eq!(back, u);
390        let native = parse_usage("input=1 output=2 cache_read=0 cache_write=0").unwrap_or_default();
391        assert!(native.get("cost_usd").is_none());
392    }
393}
speccoder_runtime/attempts.rs

speccoder_runtime/
attempts.rs
