magi/run.rs
1//! Run state: what happened, where it is stored, and how a run is resumed.
2//!
3//! Every node writes its result into [`RunState`] and the whole struct is
4//! flushed to `run.json` before the next node starts. That is what makes a run
5//! resumable: a competition can take an hour, and dying in review round four
6//! should not throw away three implementations, nine judge reads and a
7//! deliberation.
8//!
9//! Patches and raw agent transcripts are *not* in `run.json` — they live beside
10//! it under `artifacts/`, so the state file stays small enough to read by hand.
11use std::collections::BTreeMap;
12use std::path::{Path, PathBuf};
13
14use anyhow::{Context as _, Result, bail};
15use jiff::{Timestamp, Zoned};
16use serde::{Deserialize, Serialize};
17
18use crate::agent::SeatState;
19use crate::blind::Leak;
20use crate::config::{Config, MergeMode};
21use crate::verdict::{Finding, Rejection, ReviewVote, Severity};
22
23/// On-disk format version. Bumped when a field changes meaning, so a resumed
24/// run never half-reads a state file written by a different magi.
25///
26/// 2: added `RunStatus::Stalled`, `RunState::quota` (rate-limit losses), and
27/// the quorum fields on `Tally`. `RunState::load` already fails loudly and
28/// clearly on a schema mismatch; an old `run.json` from schema 1 now says so
29/// instead of silently half-reading.
30///
31/// 3: added `RunState::judge_skipped`. A solo candidate makes `judge` write
32/// only an event, leaving `judgements` empty forever — indistinguishable from
33/// "not yet judged" on every later reentry, which is what let `judge` re-run
34/// on a finished run and clobber its status back to `Judging`. The flag is
35/// the missing record of the fact that judging was skipped on purpose.
36///
37/// Also 3: a single-viable-candidate tally records `Tally::judges` as `0` and
38/// fills `Tally::uncontested`, instead of leaving the full roster size sitting
39/// next to a panel that never sat. A schema-2 record keeps reading as "0 of 3
40/// judges present" forever, because a tally is computed once and never
41/// recomputed on resume; the bump keeps that stale reading from being mixed
42/// with the new meaning.
43///
44/// 4: added `ReviewRound::progressed`. `graph::STAGNANT_LIMIT` counts
45/// consecutive rounds with `progressed == false` to decide whether the
46/// review loop should give up early, and a schema-3 record's default
47/// `false` would misreport a round that, at the time, actually committed a
48/// real diff — the field simply did not exist yet to say so. Without the
49/// bump, resuming an old multi-round review could spuriously trip the
50/// stagnation check on rounds that were never stagnant.
51///
52/// 5: added `RunStatus::Landing`. A run inside [`crate::land`]'s post-merge
53/// loop used to carry whatever status `merge` set before calling it forward
54/// unchanged - `Merged`, even while still watching CI or waiting on the
55/// owner's approval - which is also the one status [`RunStatus::resumable`]
56/// treats as finished. A daemon that gave this run's slot back to poll
57/// something else while an approval was outstanding, or one that simply
58/// crashed mid-land, had no way to tell "still landing" from "actually
59/// merged" and would either restart the whole competition or leave the run
60/// stuck reading as done. A schema-4 record has no notion of `Landing` at
61/// all, so this is a meaning a resumed old run cannot be guessed into rather
62/// than a value it can default to - hence the bump, not a `#[serde(default)]`.
63///
64/// 6: a deferred e2e is represented by an empty outcome list plus
65/// `ReviewRound::e2e_deferred`. Schema 5 treated that same empty list as an
66/// unconfigured, successful check, so schema-5 records are migrated with the
67/// old (not-deferred) meaning while older binaries reject schema-6 records.
68///
69/// 7: added `RunState::gate_ran`. An empty `RunState::gate` used to carry two
70/// meanings at once — "never attempted, or the last attempt was
71/// resource-blocked" (`graph::Runner::gate`'s retry case) and "attempted,
72/// zero commands configured, vacuously passed" (a repo with no
73/// `verify.gate`) — and nothing told them apart. `graph::Runner::merge`
74/// therefore read the second case as the first and refused forever: a
75/// review-only run with no gate commands configured reached `Gating` and
76/// then could never leave it. A schema-6 record's non-empty `gate` is
77/// migrated to `gate_ran = true` (a recorded attempt, real or historical,
78/// should not be spent again); an empty one migrates to `gate_ran = false`
79/// and is simply re-attempted by the next `gate()` call, which self-heals
80/// instantly for the zero-commands case.
81///
82/// 8: `ReviewRound::verified_head` used to be `None` for the overwhelming
83/// majority of rounds — every ordinary round that ran e2e against its own
84/// `head` in the main review loop never set it at all, leaving only the
85/// rare catch-up-on-a-different-commit case populated. A reader (a review
86/// prompt, `magi show`, the web UI) had no field to ask "which commit did
87/// this round's `e2e` actually check" and fell back to assuming it was
88/// always `head`, which is also what let a stale round's red output get
89/// quoted to a later round's reviewers as if it were about their patch, not
90/// an earlier one (see `ReviewRound::verification_summary`, which now exists
91/// so nowhere else has to guess). `verified_head` is now set whenever `e2e`
92/// held a real attempt (`E2eStatus::Passed`/`Failed`), always naming the
93/// commit actually checked instead of only the divergent case, and
94/// `ReviewRound::verified_at` is new alongside it. A schema-7 round's own
95/// unconditional main-loop check was always against `head` whether or not
96/// this field said so, so a `None` with a non-empty `e2e` migrates to
97/// `Some(head)` — a reconstruction of a fact that was always true, not a
98/// guess. `verified_at` has no historical value to reconstruct and stays
99/// `None`, which reads through `verification_summary` as "checked at:
100/// unknown" — an honest gap, not a fabricated time.
101/// 9: added [`RunState::operator_fixes`] — one record per `magi fix`
102/// invocation, routing specific, already-recorded findings to a fixer as a
103/// targeted, out-of-band fix outside the normal round sequence. Kept in a
104/// channel of its own rather than folded into [`ReviewRound`], because a
105/// reviewer's own severity and vote (copied verbatim onto
106/// [`OperatorFixFinding`]) must never be rewritten to look like the operator
107/// manufactured a blocking verdict — see `graph::Runner::fix_selected`. A
108/// schema-8 record has no operator-fix history at all, and
109/// `#[serde(default)]` reads an empty list as exactly that: "none happened",
110/// not an unknown gap. Nothing about an existing field's meaning changes.
111///
112/// 10: added `RunStatus::VerifiedNoop` and `Candidate::verified_noop`. Before
113/// this, an implementer that correctly concluded (with evidence) that a
114/// task's request was already satisfied elsewhere had no way to say so: the
115/// run ended the same way as one where every candidate simply failed to
116/// write anything — `after_implement` bailing with "no candidate produced a
117/// change; nothing to judge" and the run settling as a plain `Failed`. That
118/// conflated two very different facts (investigation run 391f's audit is
119/// what surfaced it: two attempts that had, correctly, found their fix
120/// already on `main`). A schema-9 record has no notion of either the new
121/// status or field, so a `VerifiedNoop` value is a meaning that cannot be
122/// reconstructed from an old record — hence the bump, not a
123/// `#[serde(default)]` for the status. `Candidate::verified_noop` alone
124/// *does* default-read as `None` on an old record, which is the honest
125/// reading: a run written before this schema never made the claim.
126///
127/// The report task 391f itself was raised from also named `6c5e`, `8df3` and
128/// `e9ce` as three more tasks whose implement wave ended the same
129/// diff-zero way, and the investigation traced all three — they do not
130/// share one cause.
131///
132/// `6c5e` and `8df3` are the same already-landed pattern as `391f`, not a
133/// coincidence: all three were re-queued together by a same-day audit of
134/// `done`-but-unlanded magi tasks (queue talk `20260912-115153-7216`,
135/// 2026-09-12 02:51–04:24), which found 17 magi tasks marked `done` with no
136/// merge to show for it and re-queued 16 of them, `6c5e` (a fix for the
137/// owner's `magi ask --thread` back-and-forth) and `8df3` (release
138/// automation) included. A second, same-day audit (talk
139/// `20260912-222053-07fe`, 13:20–13:36) then found 12 of those re-queued
140/// tasks — `391f`, `6c5e` and `8df3` among them — already merged by another
141/// route, and the owner had them deleted (`magi task rm`); `391f` alone
142/// survived because a daemon still held its run at the moment of deletion,
143/// which is the only reason any record of this group still exists to audit.
144/// Quoted directly from that second audit's own turn (talk `07fe`, so this
145/// reads without needing access to that talk store), naming both by id:
146///
147/// > 12件がマージ済み(対応不要)、3件が未実装(妥当)、2件が部分実装(要確認)でした。
148/// > **マージ済み → hold/rmを推奨:** 6c5e, 1ddc, fcf5, e25b, cea2, 391f, 3202,
149/// > b0a1, 5365, af85, 9f26, 8df3
150///
151/// — followed by the owner answering "削除!" and the agent confirming "11件
152/// 削除完了。391f はいま実行中のdaemonが掴んでいて削除できませんでした."
153/// `git log` independently confirms both fixes: the ask-back feature `6c5e`
154/// wanted landed as `f0df474` ("let the owner ask back on a question...",
155/// #93) on 2026-09-06, and the release-bump automation `8df3` wanted landed
156/// as `61005dd`/`bedd925` (open a release-bump PR on merge) on 2026-09-07
157/// and `116fcdc` (proportional version bump, #108) on 2026-09-08 — all
158/// before the 09-12 requeue. No run record survives the deletion for either
159/// task, so this schema's evidence is the audit transcript plus the
160/// independently re-checked `git log`, not a `run.json`.
161///
162/// `e9ce` is not that pattern at all, and is the reason the adoption guard
163/// below is all-or-nothing rather than "any candidate said so": its task
164/// asked an implementer to merge the real repository's `main` and cut a
165/// GitHub release — a destructive, out-of-worktree operation `AGENTS.md`
166/// names explicitly as not something to hand to an unattended candidate.
167/// Both of its runs (`20260912-053352-49ad`, `20260912-062629-bab1`)
168/// correctly refused, filed `magi ask` (questions `6196`, `6c9a`), and ended
169/// with an empty diff only because no answer arrived before the implement
170/// node's timeout — `49ad` looped `magi ask --wait` in the foreground for
171/// roughly 50 minutes as instructed before the timeout cut it off; `bab1`
172/// ended its turn moments after filing its question without ever actually
173/// blocking on the wait, a separate protocol slip this schema change does
174/// not attempt to fix. `49ad`'s own `candidates[0].summary` (quoted here
175/// because both records predate schema 10 and, separately, predate a
176/// still-unrelated struct change that already makes today's `magi show`
177/// refuse to parse either of them — `unknown field 'planner'` — so this is
178/// read straight from `run.json` on disk, not through that command):
179///
180/// > タスクの内容(READY 状態の run を実リポジトリの main に `merge --no-ff`
181/// > する、GitHub Release を作る)を精査した結果、これは全てこのワーカーの
182/// > worktree の外にある実リポジトリと GitHub 上の共有状態に対する不可逆な
183/// > 操作であり […] 私自身の運用ルール「Work only inside this worktree.
184/// > Nothing outside it is yours.」と正面から矛盾すると判断しました。
185///
186/// Neither candidate's reply carries the
187/// `NO CHANGE NEEDED` marker below, so both runs correctly stay `Failed`
188/// under this schema, not `VerifiedNoop`: a run blocked on an unanswered
189/// authorization question is not a verified no-op, and reading the two
190/// alike is exactly the misclassification the guard's per-candidate and
191/// whole-run conditions exist to refuse.
192///
193/// `RunStatus::Superseded` (added without a bump, still schema 10) is
194/// additive in the ordinary sense — an old build reading a run written under
195/// a newer one only had `RunStatus` variants to worry about before this, and
196/// a record already on disk never contained the new variant to begin with —
197/// but it is not additive in the sense every earlier schema bump on this
198/// constant was: `daemon::supersede_prior_runs` now rewrites a `Blocked` or
199/// `Stalled` run's `status` field *after* it was first written, once a later
200/// attempt at the same task lands. Any tooling outside magi that reads
201/// `run.json` and assumes a terminal `status` is permanent once set — a
202/// monitoring script polling for `blocked`, say — needs to know that
203/// `superseded` is where some of those records now go instead of staying put
204/// forever; see that function's own doc for exactly when.
205///
206/// Schema 11 adds [`RunState::released_to`] / [`RunState::released_branches`]:
207/// a run whose worktree a later attempt at the same task took over
208/// (`crate::handover`) can no longer be resumed from where it was. The fields
209/// are additive, but an older build would happily resume such a run into a
210/// worktree that no longer exists, which is the one thing the bump is for.
211pub const SCHEMA: u32 = 11;
212
213/// Where a run got to.
214#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
215#[serde(rename_all = "snake_case")]
216pub enum RunStatus {
217 /// Worktrees being prepared.
218 Prep,
219 /// Candidates being implemented.
220 Implementing,
221 /// Judges ranking blind.
222 Judging,
223 /// Judges deliberating after a split.
224 Deliberating,
225 /// Final votes being collected privately.
226 Voting,
227 /// Winner in the review + verification loop.
228 Reviewing,
229 /// Gate commands running. Transient: `graph::Runner::gate` and
230 /// `graph::Runner::merge` always move a run on from here, whether or not
231 /// any gate commands are configured — see [`RunState::gate_status`] and
232 /// `SCHEMA`'s doc for schema 7, which fixed a repo with an empty
233 /// `verify.gate` stranding a review-only run in `Gating` forever.
234 Gating,
235 /// Inside [`crate::land`]'s post-merge loop: watching CI, running a fix
236 /// round, rebasing onto a moved base, or waiting on the owner's merge
237 /// approval. A run parked here while an approval is outstanding has
238 /// handed its daemon slot back — see [`crate::daemon`] — and resumes
239 /// through exactly this status, not a fresh competition.
240 Landing,
241 /// Winner merged.
242 Merged,
243 /// Winner passed the gate; merge was not requested.
244 Ready,
245 /// The judgement did not gather enough judges (e.g. rate limiting took out
246 /// seats), so the verdict is not trustworthy. The run stopped and kept its
247 /// work so it can be resumed or folded — it must never be confused with a
248 /// healthy `Ready`.
249 Stalled,
250 /// Review rounds exhausted with findings still open, or the gate failed.
251 Blocked,
252 /// The graph could not complete.
253 Failed,
254 /// Every candidate wrote nothing, and every one of them said why in a way
255 /// that survived [`crate::graph::Runner`]'s adoption guard: a clean CLI
256 /// exit, an actually-empty tree, no command left with an unconfirmed
257 /// exit status, and non-empty evidence. Distinct from `Failed` on
258 /// purpose — see `SCHEMA`'s doc for schema 10 — because the two read
259 /// identically to an operator glancing at a card ("nothing happened")
260 /// while meaning opposite things: one is an agent that could not do the
261 /// work, the other is an agent that checked and the work was already
262 /// done. Settles the task through [`crate::queue::Task::handed_off`], not
263 /// [`crate::queue::Task::fail`]: a human still has to look — the claim is
264 /// unverified by magi itself — and `Held` (not `Failed`-and-requeued)
265 /// means nothing retries the task unattended on the same unconfirmed
266 /// claim while that look is pending.
267 VerifiedNoop,
268 /// A later attempt at the same task already finished the job — see
269 /// `crate::queue::Task::superseded_attempts` — so this run's own
270 /// `Blocked`/`Stalled` no longer means anyone has to look at it. Written
271 /// only over one of those two statuses, only once the task itself is
272 /// `crate::queue::TaskStatus::Done`, and only onto an attempt that comes
273 /// before the one that succeeded. Unlike them, not [`Self::resumable`]:
274 /// there is nothing left to resume towards, the task already has its
275 /// answer, and a fold is free to clean this run's worktree away without
276 /// waiting on a human to confirm that first.
277 Superseded,
278}
279
280impl RunStatus {
281 /// Is this a terminal state?
282 pub fn done(self) -> bool {
283 matches!(
284 self,
285 Self::Merged
286 | Self::Ready
287 | Self::Stalled
288 | Self::Blocked
289 | Self::Failed
290 | Self::VerifiedNoop
291 | Self::Superseded
292 )
293 }
294
295 /// The name this status is written and shown under, matching the
296 /// `snake_case` serde spelling so a log line, an error message and the
297 /// JSON a phone reads all say the same word.
298 pub fn as_str(self) -> &'static str {
299 match self {
300 Self::Prep => "prep",
301 Self::Implementing => "implementing",
302 Self::Judging => "judging",
303 Self::Deliberating => "deliberating",
304 Self::Voting => "voting",
305 Self::Reviewing => "reviewing",
306 Self::Gating => "gating",
307 Self::Landing => "landing",
308 Self::Merged => "merged",
309 Self::Ready => "ready",
310 Self::Stalled => "stalled",
311 Self::Blocked => "blocked",
312 Self::Failed => "failed",
313 Self::VerifiedNoop => "verified_noop",
314 Self::Superseded => "superseded",
315 }
316 }
317
318 /// Label for a human-facing listing or report — the same word as
319 /// [`Self::as_str`] except where the machine spelling would read harsher
320 /// than the state actually is. `VerifiedNoop` is the one case: its own
321 /// `as_str` exists for logs, JSON and event messages, none of which
322 /// should quietly grow a second vocabulary, but a bare "verified_noop" in
323 /// a report reads like an error code, not the qualified, evidence-backed
324 /// claim it actually is.
325 pub fn display_label(self) -> &'static str {
326 match self {
327 Self::VerifiedNoop => "agent-verified no-op",
328 other => other.as_str(),
329 }
330 }
331
332 /// Can this run be carried on from where it stopped?
333 ///
334 /// Everything except a finished run and a failed one. `execute` skips
335 /// nodes already recorded, so re-entering is cheap wherever the run
336 /// stopped, and the alternative is always a fresh competition against
337 /// work that already exists.
338 ///
339 /// - `Stalled` re-asks only the seats whose absence collapsed the panel,
340 /// keeping the candidates that were already paid for.
341 /// - `Blocked` re-enters the review loop against a branch that is built.
342 /// - **A non-terminal status** means the run was interrupted: a parked
343 /// run waiting for its upgrade, or one whose daemon was killed. This
344 /// used to be excluded, which left run 4043 stuck at `reviewing` with
345 /// the deck telling the operator it could not be resumed - the one
346 /// state where resuming is the only sensible answer.
347 ///
348 /// `Failed` does not qualify: the graph could not complete and there is
349 /// no established point to continue from. Nor does a finished run, whose
350 /// answer is a new competition. Nor does `VerifiedNoop`: every candidate
351 /// already agreed nothing belongs in this worktree, and resuming would
352 /// only re-ask the same question — the answer is for a human to check
353 /// the evidence, not for the graph to run again. Nor does `Superseded`:
354 /// a later attempt at the same task already finished it, so there is
355 /// nothing left this run's own answer could still contribute.
356 ///
357 /// Whether anything is *already* driving the run is a separate question,
358 /// answered by `daemon::is_working_on` at the callers that need it.
359 pub fn resumable(self) -> bool {
360 !matches!(
361 self,
362 Self::Merged | Self::Ready | Self::Failed | Self::VerifiedNoop | Self::Superseded
363 )
364 }
365}
366
367/// One candidate implementation.
368#[derive(Debug, Clone, Serialize, Deserialize)]
369pub struct Candidate {
370 /// Position in the implementer list.
371 pub index: usize,
372 /// Blind label as presented to judges.
373 pub label: char,
374 /// Which agent wrote it. Recorded for the stats tables, never shown to a
375 /// judge.
376 pub agent: String,
377 /// Branch, named after the label so judges can inspect it without learning
378 /// the author.
379 pub branch: String,
380 /// Worktree path.
381 pub worktree: PathBuf,
382 /// Sanitized author summary.
383 #[serde(default)]
384 pub summary: String,
385 /// `git diff --stat`.
386 #[serde(default)]
387 pub stat: String,
388 /// Files touched.
389 #[serde(default)]
390 pub files: usize,
391 /// Commits ahead of base.
392 #[serde(default)]
393 pub commits: usize,
394 /// True when the agent produced no change at all.
395 #[serde(default)]
396 pub empty: bool,
397 /// Why this candidate is not in the running.
398 #[serde(default)]
399 pub failed: Option<String>,
400 /// The evidence this candidate gave for writing no change on purpose —
401 /// the `NO CHANGE NEEDED:` marker `prompt::implement`'s reply format
402 /// documents, verbatim. `Some` only when [`crate::graph`]'s adoption
403 /// guard accepted the claim: the CLI exited cleanly, the tree really is
404 /// empty, no command in the reply was left with an unconfirmed exit
405 /// status, and the evidence itself is non-empty. A candidate that wrote
406 /// nothing and said nothing about why — the ordinary empty loss — always
407 /// reads `None` here, same as one written before schema 10 ever existed.
408 #[serde(default)]
409 pub verified_noop: Option<String>,
410 /// Wall-clock time for the implementation.
411 #[serde(default)]
412 pub duration_ms: u64,
413 /// Whether the worktree has been folded away.
414 #[serde(default)]
415 pub folded: bool,
416}
417
418impl Candidate {
419 /// Can this candidate be judged?
420 pub fn viable(&self) -> bool {
421 self.failed.is_none() && !self.empty
422 }
423}
424
425/// One judge's independent ranking.
426#[derive(Debug, Clone, Serialize, Deserialize)]
427pub struct Judgement {
428 /// Judge seat number, 1-based.
429 pub judge: usize,
430 /// Seat key.
431 pub seat: String,
432 /// Agent occupying the seat.
433 pub agent: String,
434 /// Best-first labels.
435 #[serde(default)]
436 pub ranking: Vec<char>,
437 /// Per-label justification.
438 #[serde(default)]
439 pub reasons: BTreeMap<String, String>,
440 /// Self-reported confidence.
441 #[serde(default)]
442 pub confidence: Option<u8>,
443 /// Order the candidates were presented in, as candidate indices.
444 #[serde(default)]
445 pub order: Vec<usize>,
446 /// Why this judge has no ranking.
447 #[serde(default)]
448 pub failed: Option<String>,
449 /// Wall-clock time.
450 #[serde(default)]
451 pub duration_ms: u64,
452}
453
454/// One judge's turn in a deliberation round.
455#[derive(Debug, Clone, Serialize, Deserialize)]
456pub struct DeliberationTurn {
457 /// Judge seat number, 1-based.
458 pub judge: usize,
459 /// Agent occupying the seat.
460 pub agent: String,
461 /// The argument, as written.
462 pub body: String,
463 /// Where the judge stood at the end of the turn.
464 #[serde(default)]
465 pub tentative: Option<char>,
466}
467
468/// A deliberation round.
469#[derive(Debug, Clone, Serialize, Deserialize)]
470pub struct DeliberationRound {
471 /// 1-based round number.
472 pub round: usize,
473 /// Turns, in the order they were taken.
474 pub turns: Vec<DeliberationTurn>,
475}
476
477/// A final vote, collected privately.
478#[derive(Debug, Clone, Serialize, Deserialize)]
479pub struct VoteRecord {
480 /// Judge seat number, 1-based.
481 pub judge: usize,
482 /// Agent occupying the seat.
483 pub agent: String,
484 /// The vote.
485 #[serde(default)]
486 pub vote: Option<char>,
487 /// Why.
488 #[serde(default)]
489 pub reason: String,
490 /// Did this judge move from its initial first choice?
491 #[serde(default)]
492 pub changed: bool,
493}
494
495/// A seat that was taken out by a CLI rate limit / quota, recorded so a run
496/// whose panel collapsed does not masquerade as a healthy one.
497#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
498pub struct QuotaLoss {
499 /// Seat key, e.g. `judge-1` or `review-2`.
500 pub seat: String,
501 /// Node that was running, e.g. `judge`, `vote`, `review`.
502 pub node: String,
503 /// When the CLI reported the limit.
504 pub at: Timestamp,
505 /// Reset hint if the CLI printed one, free text.
506 #[serde(default)]
507 pub reset: Option<String>,
508}
509
510/// A newly created lockfile of a package manager the directory does not use,
511/// which a rescue commit left untracked instead of committing. Recorded so the
512/// omission is visible: the file may be one the task really wanted.
513#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
514pub struct Withheld {
515 /// Repo-relative path.
516 pub path: String,
517 /// The manager the file belongs to.
518 pub manager: String,
519 /// The tracked file (or missing manifest) that made it foreign.
520 pub kept_by: String,
521 /// Node whose rescue commit withheld it.
522 pub node: String,
523 /// When.
524 pub at: Timestamp,
525}
526
527/// The mechanical count.
528#[derive(Debug, Clone, Serialize, Deserialize)]
529pub struct Tally {
530 /// First-choice votes per label.
531 pub first_choice: BTreeMap<char, usize>,
532 /// Borda points from the initial rankings, used only to break a tie.
533 pub borda: BTreeMap<char, usize>,
534 /// The winning label.
535 pub winner: char,
536 /// How many judges produced a usable ranking. A panel of one is not a
537 /// consensus and must not be reported as a split.
538 #[serde(default)]
539 pub rankings: usize,
540 /// Did every judge's *initial* first choice agree?
541 pub unanimous_initial: bool,
542 /// Was deliberation run?
543 pub deliberated: bool,
544 /// Judges who moved between their initial ranking and their final vote.
545 pub changed_votes: usize,
546 /// Did the final votes agree?
547 pub unanimous_final: bool,
548 /// How the tie was broken, when it had to be.
549 #[serde(default)]
550 pub tie_break: Option<String>,
551 /// Configured judge count — the size of the full panel. `0` when no
552 /// panel was asked (see `uncontested`), not the roster size a panel that
553 /// never sat would have had.
554 #[serde(default)]
555 pub judges: usize,
556 /// Judges who actually contributed to the decision (not taken out by a
557 /// rate limit and producing a usable rank or vote).
558 #[serde(default)]
559 pub present: usize,
560 /// How many judges are required for a trustworthy verdict. Chosen as a
561 /// strict majority (`judges / 2 + 1`): a verdict backed by a minority must
562 /// never be presented as a healthy one, while a bare majority is still
563 /// real signal. A one-candidate run needs no quorum.
564 #[serde(default)]
565 pub quorum: usize,
566 /// `present >= quorum`, or no quorum was required.
567 #[serde(default)]
568 pub met_quorum: bool,
569 /// Why no panel was asked, when none was: a single viable candidate, or
570 /// a review-only run that never competed. `None` when judges actually
571 /// ranked and voted — including when too few of them survived to reach
572 /// quorum, which is a collapse and must keep reading as one.
573 #[serde(default)]
574 pub uncontested: Option<String>,
575}
576
577/// One reviewer's report in a round.
578#[derive(Debug, Clone, Serialize, Deserialize)]
579pub struct ReviewRecord {
580 /// Reviewer seat number, 1-based.
581 pub reviewer: usize,
582 /// Agent occupying the seat.
583 pub agent: String,
584 /// Reviewer prose.
585 #[serde(default)]
586 pub summary: String,
587 /// Findings, with magi-assigned ids.
588 #[serde(default)]
589 pub findings: Vec<Finding>,
590 /// This seat's initial vote. `None` on a record predating votes, exactly
591 /// like a round that genuinely had none cast — never a stand-in for a
592 /// vote that was lost.
593 #[serde(default)]
594 pub vote: Option<ReviewVote>,
595 /// Why this reviewer produced nothing.
596 #[serde(default)]
597 pub failed: Option<String>,
598 /// Wall-clock time.
599 #[serde(default)]
600 pub duration_ms: u64,
601 /// How many times this seat was asked before it settled — 0 for a first
602 /// answer, N after N nudges (`ask_json_wave`'s retry loop). Read this
603 /// together with [`Self::failed`], never `failed` alone: `failed: Some(_)`
604 /// with `attempts == 0` is a seat that never answered at all, while
605 /// `failed: None` with `attempts > 0` is one that only came back after a
606 /// nudge — recovered, not silent — and the two must not look the same in
607 /// history. A record written before this field existed defaults to `0`,
608 /// which under-reports a pre-existing retry rather than inventing one;
609 /// see `ask_json_wave`'s own doc for where this is filled in.
610 #[serde(default)]
611 pub attempts: usize,
612}
613
614/// One seat's revote during a round's reconsideration (see
615/// [`ReviewRound::reconsideration`]).
616#[derive(Debug, Clone, Serialize, Deserialize)]
617pub struct ReviewRevoteRecord {
618 /// Reviewer seat number, 1-based.
619 pub reviewer: usize,
620 /// Agent occupying the seat.
621 pub agent: String,
622 /// The revote. `None` when the seat did not answer.
623 #[serde(default)]
624 pub vote: Option<ReviewVote>,
625 /// Why.
626 #[serde(default)]
627 pub reason: String,
628 /// Why this seat produced no revote.
629 #[serde(default)]
630 pub failed: Option<String>,
631}
632
633/// One fix round spent on a failing `verify.gate`: which fixer, what it did to
634/// the tree, and the gate output it was shown.
635#[derive(Debug, Clone, Serialize, Deserialize)]
636pub struct GateFixRecord {
637 /// Agent that applied the fix.
638 pub agent: String,
639 /// The gate output the fixer was handed, tail-truncated.
640 pub failed: Vec<CommandOutcome>,
641 /// What the fixer said it changed.
642 #[serde(default)]
643 pub notes: String,
644 /// Did the fix produce a commit?
645 #[serde(default)]
646 pub committed: bool,
647 /// Why the fix step produced nothing.
648 #[serde(default)]
649 pub error: Option<String>,
650}
651
652/// The fixer's response to a round.
653#[derive(Debug, Clone, Serialize, Deserialize)]
654pub struct FixRecord {
655 /// Agent that applied the fixes.
656 pub agent: String,
657 /// Finding ids acted on.
658 #[serde(default)]
659 pub addressed: Vec<String>,
660 /// Findings declined, with reasons.
661 #[serde(default)]
662 pub rejected: Vec<Rejection>,
663 /// What changed.
664 #[serde(default)]
665 pub notes: String,
666 /// Did the fix produce a commit?
667 #[serde(default)]
668 pub committed: bool,
669 /// Why the fix step produced nothing.
670 #[serde(default)]
671 pub failed: Option<String>,
672 /// Wall-clock time.
673 #[serde(default)]
674 pub duration_ms: u64,
675 /// How the fixer's own seat was made to answer when its CLI turn ended
676 /// cleanly but without an addressed/rejected report — see
677 /// [`graph::Runner::continue_fix_report`]. `None` for a record written
678 /// before this existed, which must read as "unknown", not as
679 /// [`ContinuationOutcome::NotNeeded`]: an old run really may have hit
680 /// this exact gap and simply had no mechanism to say so.
681 #[serde(default)]
682 pub continuation: Option<ContinuationRecord>,
683}
684
685/// How a node recovered — or failed to recover — a structured report after
686/// the CLI's own turn ended cleanly (a usable, non-empty, exit-0 reply)
687/// without it. A clean CLI turn is not the same fact as the node's own work
688/// being done — see the `fix` node's `continue_fix_report`, which is what
689/// produces this.
690#[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq)]
691pub enum ContinuationOutcome {
692 /// The first reply already carried the report; nothing was resumed.
693 NotNeeded,
694 /// A follow-up call in the same session recovered the report.
695 Resumed,
696 /// The continuation budget was spent without ever recovering it.
697 Exhausted,
698 /// A continuation attempt hit the CLI's rate limit; not retried further
699 /// — a quota fails the same way again immediately.
700 QuotaLost,
701 /// No session was left to resume into, so nothing was attempted.
702 NoSession,
703}
704
705/// Cost and outcome of one node's attempt to recover a missing report by
706/// resuming its own seat.
707#[derive(Debug, Clone, Copy, Serialize, Deserialize)]
708pub struct ContinuationRecord {
709 /// Follow-up calls made to the same seat. `0` when the outcome is
710 /// [`ContinuationOutcome::NotNeeded`] or [`ContinuationOutcome::NoSession`].
711 pub attempts: usize,
712 /// Wall-clock time spent on those follow-up calls, summed — not counting
713 /// the original call whose reply this is recovering from.
714 pub cumulative_wait_ms: u64,
715 /// What ended the loop.
716 pub outcome: ContinuationOutcome,
717}
718
719impl ContinuationRecord {
720 /// The report was already there on the first try.
721 pub fn not_needed() -> Self {
722 Self {
723 attempts: 0,
724 cumulative_wait_ms: 0,
725 outcome: ContinuationOutcome::NotNeeded,
726 }
727 }
728}
729
730/// What happened to one operator-selected finding after the fixer ran, as
731/// part of an [`OperatorFixRequest`].
732#[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize)]
733#[serde(rename_all = "snake_case")]
734pub enum OperatorFixOutcome {
735 /// The request has not run yet, or never got far enough to report.
736 #[default]
737 Pending,
738 /// The fixer's adoption report named this finding as addressed.
739 Addressed,
740 /// The fixer's adoption report declined it, with an argument.
741 Rejected {
742 /// The fixer's own reason.
743 why: String,
744 },
745 /// The fixer never delivered a usable adoption report at all — a
746 /// dropped stream, a quota hit, or a continuation budget spent without
747 /// recovering one (see `graph::Runner::continue_fix_report`). Distinct
748 /// from `Rejected`, which needs an argument this never produced, and
749 /// never written back as "addressed" or silently left `Pending` — a
750 /// gap in the report is its own outcome, not evidence either way about
751 /// the finding.
752 Unreported,
753}
754
755/// One finding an operator selected for [`OperatorFixRequest`], with the
756/// provenance a reviewer originally gave it, copied here verbatim.
757///
758/// Severity and vote are snapshots, never recomputed and never treated as
759/// blocking just because an operator picked the finding — only
760/// [`Severity::blocks`] on the original [`ReviewRecord`] decides that. This
761/// type exists so an operator's selection is an auditable *addition* to the
762/// record, not a rewrite of what a reviewer actually said.
763#[derive(Debug, Clone, Serialize, Deserialize)]
764pub struct OperatorFixFinding {
765 /// Finding id, e.g. `R2-1-3`.
766 pub id: String,
767 /// Severity as the reviewer recorded it.
768 pub severity: Severity,
769 /// The reviewer seat's overall vote for the round this finding came
770 /// from, if one was cast.
771 #[serde(default)]
772 pub reviewer_vote: Option<ReviewVote>,
773 /// Review round the finding was raised in.
774 pub round: usize,
775 /// That round's own head — the commit the finding was actually raised
776 /// against, used for the freshness check against the branch's current
777 /// head at request time.
778 pub round_head: String,
779 /// Reviewer seat number, 1-based.
780 pub reviewer: usize,
781 /// Agent occupying that seat.
782 pub agent: String,
783 /// File the finding concerns.
784 #[serde(default)]
785 pub file: Option<String>,
786 /// Line the finding concerns.
787 #[serde(default)]
788 pub line: Option<u32>,
789 /// One-line summary.
790 pub title: String,
791 /// The argument.
792 #[serde(default)]
793 pub detail: String,
794 /// What happened to this finding after the fixer ran.
795 #[serde(default)]
796 pub outcome: OperatorFixOutcome,
797}
798
799/// One `magi fix` invocation: the operator's own record of which
800/// already-recorded findings they routed to a fixer, why, and what came
801/// back. See [`SCHEMA`]'s doc for schema 9 on why this is a channel of its
802/// own rather than a field on [`ReviewRound`].
803#[derive(Debug, Clone, Serialize, Deserialize)]
804pub struct OperatorFixRequest {
805 /// When `magi fix` was invoked.
806 pub requested_at: Timestamp,
807 /// The operator's own reasoning. Required and never empty at the CLI —
808 /// the audit trail this feature exists for.
809 pub reason: String,
810 /// The findings selected, each with its own provenance and outcome.
811 pub findings: Vec<OperatorFixFinding>,
812 /// The branch's head at the moment this request started executing.
813 pub head_at_request: String,
814 /// Did the operator pass `--allow-stale`?
815 pub allow_stale: bool,
816 /// Did any selected finding's own `round_head` differ from
817 /// `head_at_request`? Kept distinct from `allow_stale` — flipping that
818 /// flag does not retroactively make a request that was actually fresh
819 /// read as stale, or the reverse.
820 pub stale: bool,
821 /// The fixer's own attempt, once dispatched.
822 #[serde(default)]
823 pub fix: Option<FixRecord>,
824 /// Head after the fixer's commit, when it produced one.
825 #[serde(default)]
826 pub result_head: Option<String>,
827 /// The review-only run opened to re-verify the change, when one was
828 /// actually committed. `None` when nothing changed, so there was
829 /// nothing new to re-review — never left implicit as "not gotten to
830 /// yet".
831 #[serde(default)]
832 pub follow_up_review_run: Option<String>,
833}
834
835/// A command a seat's own CLI reported running, kept for `magi show` and for
836/// telling "this seat's turn ended" apart from "the process it started is
837/// done" — see `agent::CommandEvidence`, which is the only source this is
838/// ever built from. Never something magi polled or supervised; a command the
839/// CLI never reported finishing (or a CLI this crate has no adapter for at
840/// all) simply has no entry here, which must read as "unknown", not as
841/// "nothing ran".
842#[derive(Debug, Clone, Serialize, Deserialize)]
843pub struct JobRecord {
844 /// Graph node the seat belongs to, e.g. `"implement"`, `"fix"`.
845 pub node: String,
846 /// Review round this job belongs to, for a `"review"`/`"fix"` node —
847 /// `None` for every other node, where rounds do not apply, and for every
848 /// record written before this was tracked. Lets a reader ask "what did
849 /// this seat itself actually run this round", distinct from and never
850 /// substituted for magi's own recorded `ReviewRound::e2e` — an absent
851 /// entry here means unobserved, not that nothing ran (see this type's
852 /// own doc).
853 #[serde(default)]
854 pub round: Option<usize>,
855 /// Seat key, e.g. `"impl-A"`.
856 pub seat: String,
857 /// The CLI's own id for this command.
858 pub id: String,
859 /// The command itself, as the CLI reported it.
860 pub description: String,
861 /// When this evidence was captured — the moment this seat's reply
862 /// carrying it was read, not the command's own start time, which no
863 /// adapter here currently has. A lower bound on staleness only.
864 pub checked_at: Timestamp,
865 /// What the CLI reported for it.
866 pub status: JobStatus,
867 /// Exit code the CLI reported.
868 pub exit_code: Option<i32>,
869 /// Tail of the command's own output, when reported.
870 #[serde(default)]
871 pub result_summary: String,
872 /// Which CLI/event stream this came from, e.g. `"codex"`.
873 pub source: String,
874}
875
876/// What a [`JobRecord`]'s own CLI reported for it. There is no `Running`
877/// variant: nothing here is ever polled live, so "still running" and
878/// "finished but never reported" are the same absence of evidence, not a
879/// state this type can name — see [`JobRecord`]'s own doc.
880#[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq)]
881pub enum JobStatus {
882 /// The command's own reported exit code was `0`.
883 Completed,
884 /// The command's own reported exit code was non-zero.
885 Failed,
886 /// The CLI reported this command but not a readable exit code.
887 Unknown,
888}
889
890/// Outcome of one shell command.
891#[derive(Debug, Clone, Serialize, Deserialize)]
892pub struct CommandOutcome {
893 /// The command, as configured.
894 pub command: String,
895 /// Exit code, `None` on timeout or signal.
896 pub code: Option<i32>,
897 /// Tail of the combined output, for the report and the fix prompt.
898 #[serde(default)]
899 pub output_tail: String,
900 /// Wall-clock time.
901 #[serde(default)]
902 pub duration_ms: u64,
903 /// Set only by magi itself, never inferred from `output_tail`: the
904 /// configured command was never actually run because a resource it
905 /// needs — right now, only the shared build cache's lease or the
906 /// freshness check that must precede using it — was not available
907 /// within budget. Distinct from an ordinary failure or timeout (both of
908 /// which *did* run something and are evidence about the patch); this is
909 /// evidence about the machine, and must never be read as a verdict on
910 /// the tree it named. `#[serde(default)]` so every record written
911 /// before this field existed keeps reading as `false` — exactly what it
912 /// was.
913 #[serde(default)]
914 pub resource_blocked: bool,
915}
916
917/// Substrings that mark a Cargo/rustc/link failure: the toolchain could not
918/// produce a binary to run at all, as opposed to producing one that ran and
919/// failed. A Windows link race against a shared `CARGO_TARGET_DIR` (see
920/// AGENTS.md, "Running magi on magi") looks exactly like a red command
921/// otherwise, and a run has concluded `Blocked` on nothing but that race.
922const BUILD_FAILURE_MARKERS: &[&str] = &[
923 "error: could not compile",
924 "error: linking with",
925 "LINK : fatal error",
926 "fatal error LNK",
927];
928
929impl CommandOutcome {
930 /// Did it pass?
931 pub fn ok(&self) -> bool {
932 self.code == Some(0)
933 }
934
935 /// Did this command fail because the code could not be built or linked,
936 /// rather than because it ran and produced a wrong result? A failure here
937 /// is not a verdict on the patch under review.
938 pub fn build_failed(&self) -> bool {
939 !self.ok()
940 && BUILD_FAILURE_MARKERS
941 .iter()
942 .any(|m| self.output_tail.contains(m))
943 }
944}
945
946/// One review + verify + fix round.
947#[derive(Debug, Clone, Serialize, Deserialize)]
948pub struct ReviewRound {
949 /// 1-based round number.
950 pub round: usize,
951 /// Commit the round reviewed.
952 pub head: String,
953 /// The commit `e2e` was actually attempted against. Set whenever an
954 /// attempt was dispatched (`e2e_status()` reads `Passed`, `Failed`, or
955 /// `ResourceBlocked`), naming that commit even when it equals `head` —
956 /// never left implicit, because an implicit "must have been `head`" is
957 /// exactly what let a later round quote an earlier round's result
958 /// without saying which commit it came from. A resource-blocked attempt
959 /// still targeted a specific commit even though no command finished, and
960 /// leaving that unrecorded is exactly what made a *fresh* blocked
961 /// attempt read the same as an untracked one from before schema 8.
962 /// `None` only when nothing was attempted at all (`NotConfigured`,
963 /// `Deferred`). See `SCHEMA`'s doc for schema 8 for why this broadened
964 /// from only the catch-up-on-a-different-commit case.
965 #[serde(default)]
966 pub verified_head: Option<String>,
967 /// When the attempt behind `verified_head` actually ran. `None` on
968 /// every record written before schema 8, and on a round where nothing
969 /// ran —
970 /// both read as "unknown", not as "now" or "never asked".
971 #[serde(default)]
972 pub verified_at: Option<Timestamp>,
973 /// Reviewer reports.
974 pub reviews: Vec<ReviewRecord>,
975 /// E2E command outcomes for this round.
976 #[serde(default)]
977 pub e2e: Vec<CommandOutcome>,
978 /// True when the first verify attempt this round could not build or
979 /// link, and `e2e` above holds a second attempt run before concluding.
980 /// A run must never be decided on a red it could not tell from an
981 /// unrelated build race.
982 #[serde(default)]
983 pub verify_retried: bool,
984 /// True when `e2e` was intentionally left empty this round: the round
985 /// already had blocking findings and another round was available, so
986 /// `graph::Runner::review_loop` sent the fixer straight at them instead
987 /// of spending a full verify run on a head it already knew would need
988 /// another fix. Distinct from an `e2e` that is simply empty because
989 /// `verify.e2e` has no commands configured — `e2e.is_empty()` alone
990 /// cannot tell those apart, and conflating them is exactly how a
991 /// deferred check would get painted green. A record written before this
992 /// field existed defaults to `false`, which is the truth for it: every
993 /// round used to run e2e unconditionally.
994 #[serde(default)]
995 pub e2e_deferred: bool,
996 /// Why `e2e` was deferred, set only when [`Self::e2e_deferred`] is true.
997 /// Carried to the fixer's prompt and shown in the report so "deferred"
998 /// never reads as silence.
999 #[serde(default)]
1000 pub e2e_defer_reason: Option<String>,
1001 /// Fixer response, absent when the round was already clean.
1002 #[serde(default)]
1003 pub fix: Option<FixRecord>,
1004 /// Findings that hold the merge.
1005 #[serde(default)]
1006 pub blocking: usize,
1007 /// Reviewer seats that answered (did not time out, crash, or return
1008 /// something unparsable).
1009 #[serde(default)]
1010 pub answered: usize,
1011 /// Reviewer seats the round expected an answer from — normally
1012 /// `graph.reviewers`, but recorded per round so a config change between
1013 /// runs never has to be inferred from history.
1014 #[serde(default)]
1015 pub expected: usize,
1016 /// Round ended with no blocking findings and green verification, judged
1017 /// against the seats that answered. See [`Self::incomplete`] for whether
1018 /// that verdict is missing input.
1019 #[serde(default)]
1020 pub clean: bool,
1021 /// Did the tree actually move against `base` this round, comparing the
1022 /// diff after the fix to the diff the reviewers saw at the start of the
1023 /// round?
1024 ///
1025 /// Never derived from the fixer's own `addressed`/`rejected` count: that
1026 /// self-report has been caught lying twice on this workload (runs `b455`
1027 /// and `6218`, both of which committed a real, substantial diff while
1028 /// reporting `0 addressed`). `git` does not lie about whether the tree
1029 /// changed, so this is what `graph::Runner::review_loop` counts rounds of
1030 /// no progress against. Absent on a round with no fix attempt (already
1031 /// clean, or the round the budget ran out on), where it defaults to
1032 /// `false` and is not consulted.
1033 #[serde(default)]
1034 pub progressed: bool,
1035 /// Did the seats' initial votes ([`ReviewRecord::vote`]) disagree?
1036 #[serde(default)]
1037 pub vote_split: bool,
1038 /// One round of revoting, run only when `vote_split`: each seat that cast
1039 /// an initial vote reads every seat's findings and votes, then revotes.
1040 /// Empty when the initial votes already agreed, the same as a solo
1041 /// candidate leaving `deliberation` empty.
1042 #[serde(default)]
1043 pub reconsideration: Vec<ReviewRevoteRecord>,
1044 /// The round's verdict: the most cautious vote among the seats that
1045 /// answered, using each seat's revote where reconsideration ran and its
1046 /// initial vote otherwise. `None` when no seat produced a usable vote —
1047 /// including every record written before votes existed, which is the
1048 /// truth for those rounds, not a gap in this one.
1049 #[serde(default)]
1050 pub verdict: Option<ReviewVote>,
1051}
1052
1053impl ReviewRound {
1054 /// Did at least one reviewer seat fail to answer this round?
1055 pub fn incomplete(&self) -> bool {
1056 self.answered < self.expected
1057 }
1058
1059 /// The honest state of this round's e2e leg.
1060 ///
1061 /// Never derive this from `e2e.is_empty()` alone anywhere else in the
1062 /// codebase — `NotConfigured` and `Deferred` both leave it empty, and
1063 /// only this method (backed by [`Self::e2e_deferred`]) tells them apart.
1064 /// A resource-blocked attempt is checked first and ahead of both: `e2e`
1065 /// is non-empty for it too, but `CommandOutcome::resource_blocked` says
1066 /// no command actually ran, and reading that as `Failed` is exactly how
1067 /// shared build-cache contention gets misreported as a verdict on the
1068 /// patch (see `CommandOutcome::resource_blocked`'s own doc).
1069 pub fn e2e_status(&self) -> E2eStatus {
1070 if self.e2e.iter().any(|o| o.resource_blocked) {
1071 E2eStatus::ResourceBlocked
1072 } else if !self.e2e.is_empty() {
1073 if self.e2e.iter().all(CommandOutcome::ok) {
1074 E2eStatus::Passed
1075 } else {
1076 E2eStatus::Failed
1077 }
1078 } else if self.e2e_deferred {
1079 E2eStatus::Deferred
1080 } else {
1081 E2eStatus::NotConfigured
1082 }
1083 }
1084
1085 /// Facts about this round's verification leg, judged against
1086 /// `current_head` — the commit whoever is asking is actually looking at
1087 /// right now. `None` when there is nothing worth surfacing: no
1088 /// `verify.e2e` configured, or the round's own check came back green (a
1089 /// passing result needs no skepticism attached to it, and an unread
1090 /// `None` is exactly what keeps a quiet round quiet instead of padding
1091 /// every prompt with "everything was fine").
1092 ///
1093 /// This is the single place that turns `e2e`/`e2e_deferred`/
1094 /// `verified_head`/`verified_at` into text. Every prompt and report that
1095 /// shows a round's verification result must build its wording from this,
1096 /// not re-derive its own summary at the call site — a hand-rolled
1097 /// version at one more place is exactly how "an old red read as today's
1098 /// answer" comes back through a different door (see the incident this
1099 /// type exists to prevent, recorded alongside `SCHEMA`'s doc for schema
1100 /// 8).
1101 pub fn verification_summary(&self, current_head: &str) -> Option<VerificationSummary> {
1102 let status = self.e2e_status();
1103 if matches!(status, E2eStatus::NotConfigured | E2eStatus::Passed) {
1104 return None;
1105 }
1106 let commit = match &self.verified_head {
1107 Some(h) if h == current_head => {
1108 format!("commit {} (this is the head being looked at now)", short(h))
1109 }
1110 Some(h) => format!("commit {} (an earlier head, since superseded)", short(h)),
1111 None => "commit unknown (no command finished checking one)".to_owned(),
1112 };
1113 let checked_at = match self.verified_at {
1114 Some(t) => format!("checked at {t}"),
1115 None => "checked at: unknown (recorded before this was tracked)".to_owned(),
1116 };
1117 let result = match status {
1118 E2eStatus::NotConfigured | E2eStatus::Passed => unreachable!("checked above"),
1119 E2eStatus::Failed => "result: FAILED".to_owned(),
1120 E2eStatus::Deferred => format!(
1121 "result: not run this round yet — deferred to the fixer{}. Not passed, not \
1122 failed.",
1123 self.e2e_defer_reason
1124 .as_deref()
1125 .map(|why| format!(" ({why})"))
1126 .unwrap_or_default()
1127 ),
1128 E2eStatus::ResourceBlocked => "result: could not run — the shared build cache was \
1129 not available. This is evidence about the machine, \
1130 not about the patch."
1131 .to_owned(),
1132 };
1133 let label = format!("round {}, {commit}, {checked_at}\n{result}", self.round);
1134 // `Failed` names the command that actually ran and failed;
1135 // `ResourceBlocked` names the operation magi was waiting on (or the
1136 // freshness check it could not confirm) — `command` still says what
1137 // was attempted even though nothing finished, and leaving it out is
1138 // exactly how a reviewer or fixer lost the one thing this leg *can*
1139 // still tell them: what was being checked, not whether it passed.
1140 let tail = matches!(status, E2eStatus::Failed | E2eStatus::ResourceBlocked).then(|| {
1141 self.e2e
1142 .iter()
1143 .filter(|o| !o.ok())
1144 .map(|o| format!("$ {}\n{}\n", o.command, o.output_tail))
1145 .collect::<String>()
1146 });
1147 Some(VerificationSummary { label, tail })
1148 }
1149}
1150
1151/// The honest state of a round's e2e leg. See [`ReviewRound::e2e_status`].
1152#[derive(Debug, Clone, Copy, PartialEq, Eq)]
1153pub enum E2eStatus {
1154 /// `verify.e2e` has no commands configured.
1155 NotConfigured,
1156 /// Skipped this round on purpose: blocking findings already required a
1157 /// fix, so the round went straight to the fixer instead of spending a
1158 /// full verify run on a head it already knew would need another pass.
1159 Deferred,
1160 /// Ran, and every command exited 0.
1161 Passed,
1162 /// Ran, and at least one command did not exit 0.
1163 Failed,
1164 /// Magi could not even get a command to run — the shared build cache's
1165 /// lease or freshness check was not available within budget. Evidence
1166 /// about the machine, never a verdict on the tree it named; must not be
1167 /// shown or counted the same as [`Self::Failed`].
1168 ResourceBlocked,
1169}
1170
1171/// [`ReviewRound::verification_summary`]'s output: the facts, pre-worded, for
1172/// a prompt or report to place under its own heading. Kept as two pieces
1173/// rather than one pre-joined string so a caller that wants to insert its own
1174/// note between the label and the raw command tail (see `prompt::review`) can
1175/// do so without re-parsing text back apart.
1176#[derive(Debug, Clone)]
1177pub struct VerificationSummary {
1178 /// Round, commit, freshness and result — always present.
1179 pub label: String,
1180 /// Raw `$ command` / output tail, present for `result: FAILED` and for
1181 /// a resource-blocked attempt (naming the operation magi was waiting on,
1182 /// even though nothing finished) — absent for every other result, which
1183 /// has nothing to add past the label.
1184 pub tail: Option<String>,
1185}
1186
1187/// The honest state of a run's final gate. See [`RunState::gate_status`].
1188#[derive(Debug, Clone, Copy, PartialEq, Eq)]
1189pub enum GateStatus {
1190 /// Never attempted, or the last attempt was resource-blocked (the shared
1191 /// build cache could not be acquired or confirmed fresh in time) and
1192 /// needs a retry.
1193 NotRun,
1194 /// Ran with zero commands configured (`verify.gate` is empty) and
1195 /// therefore vacuously passed — there was nothing to check.
1196 PassedWithNoCommands,
1197 /// Ran one or more commands, and every one of them exited 0.
1198 Passed,
1199 /// Ran one or more commands, and at least one did not exit 0.
1200 Failed,
1201}
1202
1203impl GateStatus {
1204 /// May a run in this state proceed to merge?
1205 pub fn ok(self) -> bool {
1206 matches!(self, Self::PassedWithNoCommands | Self::Passed)
1207 }
1208}
1209
1210/// What happened to the winning branch.
1211#[derive(Debug, Clone, Serialize, Deserialize)]
1212pub struct MergeOutcome {
1213 /// Requested mode.
1214 pub mode: MergeMode,
1215 /// Did it land?
1216 pub ok: bool,
1217 /// Command output, or the command the operator should run.
1218 #[serde(default)]
1219 pub detail: String,
1220}
1221
1222/// A seat currently mid-answer: a prompt was sent and no reply has landed yet.
1223///
1224/// This is not the whole story of "is it alive" — a daemon killed mid-wave
1225/// leaves its last wave's entries here forever, since nothing ran to clear
1226/// them. A reader must cross-check a live daemon's heartbeat
1227/// (`daemon::is_working_on`) before trusting one of these as "still running"
1228/// rather than "abandoned". [`RunState::clear_active`] is what keeps that
1229/// leftover from surviving into the next attempt at this run: `execute` calls
1230/// it before doing anything else, so a resumed run never carries a stale
1231/// entry into its own report before the next wave repopulates it.
1232///
1233/// Deliberately carries no agent id: an implementer's agent is no secret, but
1234/// a judge or reviewer seat is blind (`SeatState::key` is keyed by seat, never
1235/// agent, for exactly this reason), and this struct has no way to tell which
1236/// kind of seat it describes. The seat key alone — already in the map this
1237/// lives under — is what every caller needs to say which seat is running.
1238///
1239/// The same map also carries entries for shell-command work that runs
1240/// outside any seat — `verify.e2e`, `verify.gate` — keyed by the task's own
1241/// name (`"e2e"`, `"gate"`) rather than a seat key. [`Self::task`] is `Some`
1242/// only for those; it is how a reader tells the two kinds of entry apart
1243/// without a second map, a second route, or a second SSE reason to poll for
1244/// — see [`RunState::seats_active`] / [`RunState::tasks_active`] for the
1245/// accessors that split them back apart. A task entry is exactly as blind as
1246/// a seat entry: no agent runs it, so there is nothing to leak, and
1247/// [`Self::command`] carries only the shell command being run, never
1248/// anything about who is running it.
1249#[derive(Debug, Clone, Serialize, Deserialize)]
1250pub struct ActiveSeat {
1251 /// Node the seat is answering for, e.g. `implement`, `judge`, `review`.
1252 /// For a task entry, the node the command list runs under (`verify`,
1253 /// `gate`).
1254 pub node: String,
1255 /// When this attempt — or, for a task entry, this one command — was
1256 /// started. A task entry's timer resets at every command boundary,
1257 /// because `verify.e2e` / `verify.gate` apply their timeout per command,
1258 /// not once across the whole list — see [`RunState::task_command`].
1259 pub started_at: Timestamp,
1260 /// The wall-clock budget for this attempt (a seat) or this one command
1261 /// (a task entry).
1262 pub timeout_secs: u64,
1263 /// 0 for the first ask, N for the Nth nudge or resume. For a task entry,
1264 /// 0 for the first pass over the command list, N for the Nth retry (see
1265 /// `run_e2e_with_retry`'s build/link retry).
1266 #[serde(default)]
1267 pub attempt: usize,
1268 /// `None` for a seat; `Some("e2e")` / `Some("gate")` for a running
1269 /// command-list task. This is the type tag that lets both kinds of entry
1270 /// share one map without a task ever being mistaken for a (blind) seat —
1271 /// see this struct's own doc. Omitted from JSON when absent (the common,
1272 /// seat case), rather than written out as a literal `null` on every one
1273 /// of a run's seat entries.
1274 #[serde(default, skip_serializing_if = "Option::is_none")]
1275 pub task: Option<String>,
1276 /// The command currently running, task entries only. Never set on a
1277 /// seat entry — a seat has no command, only a prompt, and a prompt is
1278 /// not safe to show mid-run (see this struct's blindness note).
1279 #[serde(default, skip_serializing_if = "Option::is_none")]
1280 pub command: Option<String>,
1281 /// 1-based position of [`Self::command`] within the task's command list.
1282 #[serde(default, skip_serializing_if = "Option::is_none")]
1283 pub index: Option<usize>,
1284 /// Number of commands in the task's list.
1285 #[serde(default, skip_serializing_if = "Option::is_none")]
1286 pub total: Option<usize>,
1287}
1288
1289impl ActiveSeat {
1290 /// Seconds since this attempt was sent.
1291 #[must_use]
1292 pub fn elapsed_secs(&self, now: Timestamp) -> i64 {
1293 (now.as_second() - self.started_at.as_second()).max(0)
1294 }
1295
1296 /// Seconds left before this attempt's own timeout fires, floored at zero
1297 /// rather than going negative once the CLI has overrun its budget.
1298 #[must_use]
1299 pub fn remaining_secs(&self, now: Timestamp) -> i64 {
1300 (self.timeout_secs as i64 - self.elapsed_secs(now)).max(0)
1301 }
1302}
1303
1304/// Whether a process is provably still driving a run, provably not, or
1305/// neither — see [`RunState::liveness`]. Serialized as a lowercase string
1306/// (`"live"` / `"dead"` / `"unknown"`) rather than a bool: a bool has no room
1307/// for "could not tell", and folding that case into either `true` or `false`
1308/// is exactly the wrong call for a display an operator uses to decide
1309/// whether to wait or to act — see the schema-10 field doc on
1310/// [`RunState::driver_pid`] for the report it used to produce instead.
1311#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
1312#[serde(rename_all = "lowercase")]
1313pub enum Liveness {
1314 /// Proven: a daemon's heartbeat claims the run, or `driver_pid` answers
1315 /// alive under the same identity (`driver_started_at`) this run recorded
1316 /// for it.
1317 Live,
1318 /// Proven: no daemon claim, and either `driver_pid` answers dead outright
1319 /// or it answers alive under a *different* identity than recorded — a
1320 /// pid the OS has since handed to an unrelated process is exactly as
1321 /// good as proof the original driver is gone (see
1322 /// [`RunState::driver_started_at`]'s own doc).
1323 Dead,
1324 /// Neither proven — no daemon claim, and either no `driver_pid` to ask,
1325 /// the platform could not answer for it, or a live pid with nothing (or
1326 /// nothing queryable) to corroborate its identity against. Never treated
1327 /// as `Dead`: see [`RunState::liveness_with`].
1328 Unknown,
1329}
1330
1331/// A timestamped note about a node.
1332#[derive(Debug, Clone, Serialize, Deserialize)]
1333pub struct Event {
1334 /// When.
1335 pub at: Timestamp,
1336 /// Node name.
1337 pub node: String,
1338 /// What happened.
1339 pub message: String,
1340}
1341
1342/// How far the winner's tree trailed the landing base, last time it was
1343/// checked, and what came of trying to close that gap.
1344///
1345/// Set by `graph::Runner::sync_to_base`, which runs before the review loop and
1346/// again before the gate: verifying against a tree that does not yet contain
1347/// the base's tip answers "green on the commit this run branched from", not
1348/// "green on what is about to land", and a merge on that answer can revert
1349/// whatever landed elsewhere while the run was thinking.
1350#[derive(Debug, Clone, Serialize, Deserialize)]
1351pub struct BaseSync {
1352 /// `<remote>/<base>` tip the tree was last checked against.
1353 pub tip: String,
1354 /// Commits `tip` was ahead of the tree at that check, before any rebase
1355 /// this round tried to close the gap. Zero means the tree already
1356 /// contained `tip`.
1357 pub behind: usize,
1358 /// Rebase attempts spent so far this run, bounded by
1359 /// `graph::BASE_SYNC_ROUNDS`.
1360 pub attempts: usize,
1361 /// What git said, if the most recent rebase attempt conflicted or could
1362 /// not be pushed. `Some` here is what makes a `Blocked` run read as
1363 /// "stopped on the base, not on review or the gate" - the rebase is not
1364 /// retried again while this is set; a person has to look.
1365 #[serde(default)]
1366 pub conflict: Option<String>,
1367}
1368
1369/// What the land loop saw last time it looked at the pull request.
1370///
1371/// Strings for `state` and `checks` on purpose: they are `gh`'s vocabulary, and
1372/// pinning them into an enum here would mean a new GitHub check conclusion
1373/// turns a readable status into a deserialisation error on a run someone is
1374/// trying to look at.
1375#[derive(Debug, Clone, Serialize, Deserialize)]
1376pub struct PrRecord {
1377 /// Pull request url.
1378 pub url: String,
1379 /// Pull request number.
1380 pub number: u64,
1381 /// `open`, `merged` or `closed`.
1382 pub state: String,
1383 /// `pending`, `green`, `red` or `unknown`.
1384 pub checks: String,
1385 /// Land round, 1-based, or 0 before the first fix.
1386 pub round: usize,
1387 /// Land round budget.
1388 pub rounds: usize,
1389 /// Checks that were red when the pull request merged anyway (the forge
1390 /// said they do not block it). Empty for a green merge, and for any run
1391 /// recorded before this field existed - so empty is not proof nothing was
1392 /// red.
1393 #[serde(default, skip_serializing_if = "Vec::is_empty")]
1394 pub red_at_merge: Vec<String>,
1395}
1396
1397/// What the post-merge release-bump step did, and whether it needs a human.
1398///
1399/// Recorded next to the `bump` events rather than instead of them: the events
1400/// are the narrative, this is what `magi show`, the status word and the
1401/// notification read without parsing prose. `status` stays `Merged` - the
1402/// merge did happen - so [`RunState::needs_attention`] is the one place that
1403/// says a merged run is not fully done.
1404#[derive(Debug, Clone, Default, PartialEq, Eq, Serialize, Deserialize)]
1405pub struct ReleaseBump {
1406 /// The release pull request, when one was opened.
1407 #[serde(default)]
1408 pub pr_url: Option<String>,
1409 /// The version it releases.
1410 #[serde(default)]
1411 pub version: Option<String>,
1412 /// Did enabling automerge on `pr_url` succeed?
1413 #[serde(default)]
1414 pub automerge_enabled: bool,
1415 /// GitHub refused automerge because CI had already finished (the pull
1416 /// request was in clean status), so magi merged it directly. Kept apart
1417 /// from `automerge_enabled`, which stays the plain fact it says it is.
1418 #[serde(default)]
1419 pub merged_directly: bool,
1420 /// What went wrong, verbatim from the tool that said it.
1421 #[serde(default)]
1422 pub problem: Option<String>,
1423 /// What the operator has to do about it. `Some` is what makes a merged run
1424 /// read as needing attention.
1425 #[serde(default)]
1426 pub action_required: Option<String>,
1427}
1428
1429/// The whole run.
1430#[derive(Debug, Clone, Serialize, Deserialize)]
1431pub struct RunState {
1432 /// On-disk format version.
1433 pub schema: u32,
1434 /// Run id, e.g. `20260830-153012-a1b2`.
1435 pub id: String,
1436 /// Repository the run operates on.
1437 pub repo: PathBuf,
1438 /// Branch the run started from.
1439 pub base_branch: String,
1440 /// Commit the run started from.
1441 pub base_commit: String,
1442 /// The task, verbatim.
1443 pub instruction: String,
1444 /// When the run was created.
1445 pub created_at: Timestamp,
1446 /// Last state flush.
1447 pub updated_at: Timestamp,
1448 /// Current status.
1449 pub status: RunStatus,
1450 /// Seed for labels and session ids.
1451 pub seed: u64,
1452 /// Config snapshot, so a resumed run behaves like the original.
1453 pub config: Config,
1454 /// Did this run take a reference on `extensions.worktreeConfig` being on
1455 /// (see [`crate::git::acquire_worktree_config`])? If so, cleanup releases
1456 /// it - which only actually turns the setting back off once every other
1457 /// run sharing this repository has released its own reference too.
1458 #[serde(default)]
1459 pub enabled_worktree_config: bool,
1460 /// Candidates.
1461 #[serde(default)]
1462 pub candidates: Vec<Candidate>,
1463 /// Initial blind rankings.
1464 #[serde(default)]
1465 pub judgements: Vec<Judgement>,
1466 /// `judge` decided a solo candidate needs no panel and only logged it.
1467 ///
1468 /// `judgements` stays empty in that case — nothing to distinguish from
1469 /// "not yet judged" — so this is the record that makes the skip
1470 /// idempotent: without it, every reentry re-ran `judge`, re-logged the
1471 /// same event, and rewrote `status` to `Judging` over whatever a later
1472 /// node had already concluded.
1473 #[serde(default)]
1474 pub judge_skipped: bool,
1475 /// Deliberation, if it happened.
1476 #[serde(default)]
1477 pub deliberation: Vec<DeliberationRound>,
1478 /// Private final votes.
1479 #[serde(default)]
1480 pub votes: Vec<VoteRecord>,
1481 /// The count.
1482 #[serde(default)]
1483 pub tally: Option<Tally>,
1484 /// Review rounds.
1485 #[serde(default)]
1486 pub reviews: Vec<ReviewRound>,
1487 /// Final gate.
1488 ///
1489 /// Never derive whether the gate has run from `gate.is_empty()` alone —
1490 /// use [`Self::gate_status`] instead. An empty list is ambiguous on its
1491 /// own: it is what an unattempted gate looks like, what a
1492 /// resource-blocked attempt leaves behind (see `graph::Runner::gate`'s
1493 /// own doc), and also what a repo with no `verify.gate` commands
1494 /// configured produces once it *has* run. [`Self::gate_ran`] is what
1495 /// tells the third case apart from the first two.
1496 #[serde(default)]
1497 pub gate: Vec<CommandOutcome>,
1498 /// Did `gate()` actually record an attempt — zero commands configured
1499 /// and vacuously passed, or one or more commands that ran to
1500 /// completion — as opposed to never having run, or having last hit a
1501 /// resource-blocked retry?
1502 ///
1503 /// `gate.is_empty()` cannot tell those apart by itself: a repo with no
1504 /// `verify.gate` commands leaves `gate` empty exactly like an
1505 /// unattempted or resource-blocked one does, and reading that empty list
1506 /// as "not yet run" is what stranded a review-only run in
1507 /// `RunStatus::Gating` forever on such a repo — see `SCHEMA`'s doc for
1508 /// schema 7. A record written before this field existed defaults to
1509 /// `false` and is migrated in [`migrate_schema`].
1510 #[serde(default)]
1511 pub gate_ran: bool,
1512 /// Fix rounds a failing gate has already spent, oldest first. Persisted
1513 /// right after each fixer call so a resumed run spends only what is left
1514 /// of `graph.gate_fix_rounds` instead of starting the budget over.
1515 #[serde(default)]
1516 pub gate_fixes: Vec<GateFixRecord>,
1517 /// Outcomes of the `verify.pre_gate` commands from the latest time they
1518 /// ran on the winner. Informational only: a failure here never blocks the
1519 /// run, the gate stays the single arbiter. Overwritten on each pass.
1520 #[serde(default)]
1521 pub pre_gate: Vec<CommandOutcome>,
1522 /// The latest commit a `pre_gate` pass made, if any pass left changes.
1523 #[serde(default)]
1524 pub pre_gate_commit: Option<String>,
1525 /// Merge outcome.
1526 #[serde(default)]
1527 pub merge: Option<MergeOutcome>,
1528 /// Vendor tokens seen in judged material.
1529 #[serde(default)]
1530 pub leaks: Vec<Leak>,
1531 /// Seats lost to a CLI rate limit / quota, in the order they hit.
1532 #[serde(default)]
1533 pub quota: Vec<QuotaLoss>,
1534 /// Stray foreign lockfiles a rescue commit left out, one entry per path.
1535 #[serde(default)]
1536 pub withheld: Vec<Withheld>,
1537 /// Parked at a node boundary, waiting to be resumed.
1538 ///
1539 /// A run that is neither finished nor being worked on is otherwise
1540 /// indistinguishable from one whose daemon was killed, and the two want
1541 /// opposite things from an operator: the first is expected to be resumed,
1542 /// the second is a leftover. Cleared by the resume that carries it on.
1543 #[serde(default)]
1544 pub parked: bool,
1545 /// The run that took this run's worktree over, when one did — see
1546 /// [`crate::handover`]. `Some` means the worktree is gone on purpose and
1547 /// the run can no longer be resumed from it ([`Self::released`]).
1548 #[serde(default)]
1549 pub released_to: Option<String>,
1550 /// Branches that outlived this run's released worktree and now belong to
1551 /// the run in [`Self::released_to`] (and to its pull request). A later
1552 /// fold must leave them alone.
1553 #[serde(default)]
1554 pub released_branches: Vec<String>,
1555 /// Per-seat conversation state.
1556 #[serde(default)]
1557 pub seats: BTreeMap<String, SeatState>,
1558 /// Seats currently mid-answer, keyed by seat.
1559 ///
1560 /// An entry exists from the moment a prompt is sent until a reply (of any
1561 /// kind — success, failure, quota, drop) comes back, so its keys are
1562 /// exactly "who hasn't answered yet" for whichever node populated it. See
1563 /// [`ActiveSeat`] for why a reader still has to check a live daemon
1564 /// before trusting one of these as "running" rather than "abandoned".
1565 #[serde(default)]
1566 pub active: BTreeMap<String, ActiveSeat>,
1567 /// Process id of whichever `execute()` call last drove this run —
1568 /// written at the very top of that method, the same place
1569 /// [`Self::clear_active`] runs, so a fresh reentry always overwrites the
1570 /// pid a previous, possibly-dead process left behind.
1571 ///
1572 /// A daemon-claimed run already has a stronger signal
1573 /// (`daemon::is_working_on`), but a `magi run` / `magi review` typed
1574 /// straight into a terminal claims nothing there — before this field
1575 /// existed, [`report::active_seats`] had no way to tell that run apart
1576 /// from one a killed process abandoned, and printed the same "no live
1577 /// daemon claims this run" warning over a run that was, in fact, still
1578 /// answering. See [`Liveness`] for how this and the daemon claim combine.
1579 #[serde(default)]
1580 pub driver_pid: Option<u32>,
1581 /// The OS-reported moment [`Self::driver_pid`] started, recorded in the
1582 /// same breath as the pid itself — an opaque marker
1583 /// (`crate::proc::process_started_at`), compared only for equality.
1584 ///
1585 /// A pid alone never proves a live process is *this run's* driver: pids
1586 /// get reused, sometimes within minutes on a busy machine, and a killed
1587 /// manual `magi run` whose pid a later, wholly unrelated process happens
1588 /// to receive would otherwise read back as `Liveness::Live` from that
1589 /// coincidence alone. [`Self::liveness`] re-queries the current holder
1590 /// of `driver_pid` and requires this marker to still match before
1591 /// trusting a live answer — a mismatch means a different process now
1592 /// answers to that number, and no marker to compare (an old run, or a
1593 /// platform this build could not ask at record time) means neither
1594 /// extreme can be proven.
1595 #[serde(default)]
1596 pub driver_started_at: Option<String>,
1597 /// Last observation of the winner's pull request, when a land loop ran.
1598 ///
1599 /// Persisted rather than derived from the event log because the phone asks
1600 /// two questions about a run that has opened a PR - how are its checks and
1601 /// which round is it on - and parsing prose out of events to answer them
1602 /// would break the first time an event message was reworded.
1603 #[serde(default)]
1604 pub pr: Option<PrRecord>,
1605 /// The post-merge release-bump step's outcome, when it got as far as
1606 /// having one to report. See [`ReleaseBump`].
1607 #[serde(default)]
1608 pub release_bump: Option<ReleaseBump>,
1609 /// The last look at how far the winner's tree trailed the landing base,
1610 /// and the rebase(s) tried to close that gap. `None` until the tree has a
1611 /// winner to check.
1612 #[serde(default)]
1613 pub base_sync: Option<BaseSync>,
1614 /// The design-deliberation stage's output, when `[graph] advise` ran it:
1615 /// one record per advisor seat, plus the synthesis blended into the
1616 /// implementer's prompt. `None` when the stage is off, has not run yet,
1617 /// or could not even resolve its seats - see
1618 /// [`crate::graph::Runner::advise`].
1619 #[serde(default)]
1620 pub advice: Option<crate::advise::Advice>,
1621 /// Whether the design-deliberation stage has already been attempted this
1622 /// run, whatever it produced. The idempotency marker `Runner::advise`
1623 /// checks on reentry, the same role [`Self::judge_skipped`] plays for
1624 /// `judge` - without it a resumed run whose stage failed (a misconfigured
1625 /// `[roles] advisors`, every seat quota'd) would re-run it, and re-spend
1626 /// the agent calls, on every single reentry before `implement`.
1627 #[serde(default)]
1628 pub advise_attempted: bool,
1629 /// Node log.
1630 #[serde(default)]
1631 pub events: Vec<Event>,
1632 /// Commands seats' own CLIs reported running, across every node — see
1633 /// [`JobRecord`]. Populated in [`crate::graph::wave`] as each seat
1634 /// answers, so a resumed run keeps what earlier waves already collected
1635 /// rather than losing it to a reentry. Empty on a record written before
1636 /// this existed, or wherever no adapter reads structured job events for
1637 /// the backend a seat used — both read as "no evidence", not "nothing
1638 /// ran".
1639 #[serde(default)]
1640 pub jobs: Vec<JobRecord>,
1641 /// Operator-triggered targeted fixes — see [`OperatorFixRequest`] and
1642 /// `SCHEMA`'s doc for schema 9. Empty on every record written before
1643 /// this existed, which reads correctly as "no operator fix ever
1644 /// requested".
1645 #[serde(default)]
1646 pub operator_fixes: Vec<OperatorFixRequest>,
1647}
1648
1649impl RunState {
1650 /// A fresh run.
1651 pub fn new(
1652 repo: PathBuf,
1653 base_branch: String,
1654 base_commit: String,
1655 instruction: String,
1656 config: Config,
1657 ) -> Self {
1658 let now = Timestamp::now();
1659 let seed = config.blind.seed.unwrap_or_else(crate::rng::entropy);
1660 Self {
1661 schema: SCHEMA,
1662 id: new_id(),
1663 repo,
1664 base_branch,
1665 base_commit,
1666 instruction,
1667 created_at: now,
1668 updated_at: now,
1669 status: RunStatus::Prep,
1670 seed,
1671 config,
1672 enabled_worktree_config: false,
1673 candidates: Vec::new(),
1674 judgements: Vec::new(),
1675 judge_skipped: false,
1676 deliberation: Vec::new(),
1677 votes: Vec::new(),
1678 tally: None,
1679 reviews: Vec::new(),
1680 gate: Vec::new(),
1681 gate_ran: false,
1682 gate_fixes: Vec::new(),
1683 pre_gate: Vec::new(),
1684 pre_gate_commit: None,
1685 merge: None,
1686 leaks: Vec::new(),
1687 quota: Vec::new(),
1688 withheld: Vec::new(),
1689 parked: false,
1690 released_to: None,
1691 released_branches: Vec::new(),
1692 seats: BTreeMap::new(),
1693 active: BTreeMap::new(),
1694 driver_pid: None,
1695 driver_started_at: None,
1696 pr: None,
1697 release_bump: None,
1698 base_sync: None,
1699 advice: None,
1700 advise_attempted: false,
1701 events: Vec::new(),
1702 jobs: Vec::new(),
1703 operator_fixes: Vec::new(),
1704 }
1705 }
1706
1707 /// Directory holding this run's state and artifacts.
1708 pub fn dir(&self) -> PathBuf {
1709 run_dir(&self.id)
1710 }
1711
1712 /// Short form used in branch names and reports.
1713 pub fn short(&self) -> &str {
1714 short_of(&self.id)
1715 }
1716
1717 /// Branch name for a label.
1718 pub fn branch_for(&self, label: char) -> String {
1719 format!("magi/{}/{}", self.short(), label)
1720 }
1721
1722 /// Root of this run's worktrees.
1723 pub fn worktree_root(&self) -> PathBuf {
1724 self.config
1725 .graph
1726 .worktree_root
1727 .clone()
1728 .unwrap_or_else(default_worktree_root)
1729 .join(self.short())
1730 }
1731
1732 /// Record lockfiles a rescue commit withheld; a path already recorded by an
1733 /// earlier round is not repeated.
1734 pub fn note_withheld(&mut self, node: &str, strays: &[crate::git::Stray]) {
1735 for s in strays {
1736 if self.withheld.iter().any(|w| w.path == s.path) {
1737 continue;
1738 }
1739 self.event(
1740 node,
1741 format!(
1742 "withheld stray lockfile {} ({}; the directory uses {})",
1743 s.path, s.manager, s.kept_by
1744 ),
1745 );
1746 self.withheld.push(Withheld {
1747 path: s.path.clone(),
1748 manager: s.manager.clone(),
1749 kept_by: s.kept_by.clone(),
1750 node: node.to_owned(),
1751 at: Timestamp::now(),
1752 });
1753 }
1754 }
1755
1756 /// Note something in the run log and on the tracing stream.
1757 pub fn event(&mut self, node: &str, message: impl Into<String>) {
1758 let message = message.into();
1759 tracing::info!(node, "{message}");
1760 self.events.push(Event {
1761 at: Timestamp::now(),
1762 node: node.to_owned(),
1763 message,
1764 });
1765 }
1766
1767 /// The honest state of the final gate.
1768 ///
1769 /// Never derive this from `gate.is_empty()` alone anywhere else in the
1770 /// codebase — `NotRun` and `PassedWithNoCommands` both leave `gate`
1771 /// empty, and only this method (backed by [`Self::gate_ran`]) tells them
1772 /// apart. See `SCHEMA`'s doc for schema 7 for what conflating them used
1773 /// to do.
1774 pub fn gate_status(&self) -> GateStatus {
1775 if !self.gate_ran {
1776 GateStatus::NotRun
1777 } else if self.gate.is_empty() {
1778 GateStatus::PassedWithNoCommands
1779 } else if self.gate.iter().all(CommandOutcome::ok) {
1780 GateStatus::Passed
1781 } else {
1782 GateStatus::Failed
1783 }
1784 }
1785
1786 /// Record that `seat` was just sent a prompt for `node`, with the given
1787 /// wall-clock budget. `attempt` is 0 for the first ask and N for the Nth
1788 /// nudge or resume, purely for display — it does not change how the seat
1789 /// is treated.
1790 pub fn seat_started(
1791 &mut self,
1792 node: &str,
1793 seat: &str,
1794 timeout: std::time::Duration,
1795 attempt: usize,
1796 ) {
1797 self.active.insert(
1798 seat.to_owned(),
1799 ActiveSeat {
1800 node: node.to_owned(),
1801 started_at: Timestamp::now(),
1802 timeout_secs: timeout.as_secs(),
1803 attempt,
1804 task: None,
1805 command: None,
1806 index: None,
1807 total: None,
1808 },
1809 );
1810 }
1811
1812 /// Record that `seat` has answered, whatever the answer was.
1813 pub fn seat_finished(&mut self, seat: &str) {
1814 self.active.remove(seat);
1815 }
1816
1817 /// Record that `task` (`"e2e"` or `"gate"` — a shell-command list run
1818 /// outside any seat) has just started `command`, the `index`-th of
1819 /// `total`. Called at every command boundary, not once for the whole
1820 /// list: `verify.e2e` / `verify.gate` apply `timeout` per command, so
1821 /// this is the only way a reader can tell "how long is left" for
1822 /// whichever command is actually running right now, rather than a stale
1823 /// budget left over from the first one.
1824 #[allow(clippy::too_many_arguments)]
1825 pub fn task_command(
1826 &mut self,
1827 task: &str,
1828 node: &str,
1829 attempt: usize,
1830 command: &str,
1831 index: usize,
1832 total: usize,
1833 timeout: std::time::Duration,
1834 ) {
1835 self.active.insert(
1836 task.to_owned(),
1837 ActiveSeat {
1838 node: node.to_owned(),
1839 started_at: Timestamp::now(),
1840 timeout_secs: timeout.as_secs(),
1841 attempt,
1842 task: Some(task.to_owned()),
1843 command: Some(command.to_owned()),
1844 index: Some(index),
1845 total: Some(total),
1846 },
1847 );
1848 }
1849
1850 /// Record that `task` has finished its whole command list for this
1851 /// attempt.
1852 pub fn task_finished(&mut self, task: &str) {
1853 self.active.remove(task);
1854 }
1855
1856 /// The seats — never task entries — currently mid-answer. What
1857 /// `report::active_seats` and the phone's "who has not answered yet"
1858 /// note need: a seat's identifier is safe to show ([`ActiveSeat`]'s doc),
1859 /// so nothing here filters anything out beyond the type tag itself.
1860 pub fn seats_active(&self) -> impl Iterator<Item = (&String, &ActiveSeat)> {
1861 self.active.iter().filter(|(_, a)| a.task.is_none())
1862 }
1863
1864 /// The command-list tasks — never seat entries — currently running.
1865 /// Counterpart to [`Self::seats_active`]; see [`ActiveSeat::task`] for
1866 /// the tag both read.
1867 pub fn tasks_active(&self) -> impl Iterator<Item = (&String, &ActiveSeat)> {
1868 self.active.iter().filter(|(_, a)| a.task.is_some())
1869 }
1870
1871 /// Drop every seat this state still lists as answering, reporting whether
1872 /// anything was dropped.
1873 ///
1874 /// Called first thing in `execute`, on every entry — fresh, resumed, or
1875 /// recovering a stall — because an entry here only means something while
1876 /// the process that wrote it is still asking that seat something. A
1877 /// process killed mid-wave leaves its last batch of seats here with
1878 /// nobody left to clear them, and the next process to touch this run must
1879 /// not let that leftover read as "still going" before it has asked
1880 /// anyone anything.
1881 pub fn clear_active(&mut self) -> bool {
1882 if self.active.is_empty() {
1883 return false;
1884 }
1885 self.active.clear();
1886 true
1887 }
1888
1889 /// Does every seat this run still lists as [`Self::active`] sit past its
1890 /// own [`ActiveSeat::timeout_secs`]? `false` when nothing is active at
1891 /// all — an empty map is not evidence of anything overrunning.
1892 ///
1893 /// This alone is not proof the run is dead: a seat's own attempt can
1894 /// legitimately run a little past its budget while the process driving it
1895 /// is still tearing the attempt down. Every caller pairs this with its own
1896 /// `!live` reading (`daemon::is_working_on`) before treating the run as
1897 /// abandoned — this module cannot check that itself without depending on
1898 /// `crate::daemon`, and callers already have to ask that question anyway.
1899 #[must_use]
1900 pub fn active_all_overrun(&self, now: Timestamp) -> bool {
1901 !self.active.is_empty()
1902 && self
1903 .active
1904 .values()
1905 .all(|a| a.elapsed_secs(now) > a.timeout_secs as i64)
1906 }
1907
1908 /// Whether a process is actually still driving this run, given whether a
1909 /// daemon's heartbeat claims it and process-liveness/identity queries for
1910 /// [`Self::driver_pid`].
1911 ///
1912 /// A daemon claim wins outright when present — it is the stronger,
1913 /// independently-heartbeating signal. Absent that (every manual `magi
1914 /// run` / `magi review`, and every daemon-driven run whose daemon has
1915 /// since exited cleanly), `driver_pid` is asked directly. A live answer
1916 /// alone is not enough to trust, though: pids get reused, so `identity`
1917 /// re-queries whoever currently holds that pid and the result must still
1918 /// match [`Self::driver_started_at`] — the marker recorded at the same
1919 /// moment `driver_pid` was — before this reads `Live`. A mismatch means
1920 /// a *different* process now answers to that number, which is exactly as
1921 /// good as proof the original driver is gone, so that reads `Dead`; no
1922 /// marker to compare against (an old run, or a platform this build could
1923 /// not ask at record time) or a `None` from either query, and this
1924 /// cannot tell either way, so it reads [`Liveness::Unknown`] — never
1925 /// guessed as [`Liveness::Dead`] out of mere silence. A display that
1926 /// guessed "dead" out of missing information would be exactly the
1927 /// mtime-and-task-manager guessing this type exists to replace.
1928 ///
1929 /// Kept generic over `query` and `identity` so a test can inject answers
1930 /// without spawning a real process query — production code goes through
1931 /// [`Self::liveness`], which supplies [`crate::proc::pid_status`] and
1932 /// [`crate::proc::process_started_at`].
1933 #[must_use]
1934 pub fn liveness_with<F, G>(&self, daemon_claims: bool, query: F, identity: G) -> Liveness
1935 where
1936 F: FnOnce(u32) -> Option<bool>,
1937 G: FnOnce(u32) -> Option<String>,
1938 {
1939 if daemon_claims {
1940 return Liveness::Live;
1941 }
1942 let Some(pid) = self.driver_pid else {
1943 return Liveness::Unknown;
1944 };
1945 match query(pid) {
1946 Some(false) => Liveness::Dead,
1947 None => Liveness::Unknown,
1948 Some(true) => match (&self.driver_started_at, identity(pid)) {
1949 (Some(recorded), Some(current)) if *recorded == current => Liveness::Live,
1950 (Some(_), Some(_)) => Liveness::Dead,
1951 _ => Liveness::Unknown,
1952 },
1953 }
1954 }
1955
1956 /// [`Self::liveness_with`], backed by the real process-liveness and
1957 /// identity queries.
1958 #[must_use]
1959 pub fn liveness(&self, daemon_claims: bool) -> Liveness {
1960 self.liveness_with(
1961 daemon_claims,
1962 crate::proc::pid_status,
1963 crate::proc::process_started_at,
1964 )
1965 }
1966
1967 /// Clear every seat this run still lists as active and fail it, unless it
1968 /// had already reached a terminal status some other way.
1969 ///
1970 /// Callers must already have proven this run is dead — [`Self::active_all_overrun`]
1971 /// plus their own `!live` reading — before calling this; it does not
1972 /// check either itself. Unlike [`Self::clear_active`] (dropping a resumed
1973 /// run's own stale wave before repopulating it, called unconditionally at
1974 /// the top of every `execute()`), this is a verdict: a run left this way
1975 /// has nothing left to repopulate the wave, ever, and must stop reading as
1976 /// `implementing` (or whichever node) forever.
1977 pub fn abandon(&mut self, by: &str) {
1978 let seats: Vec<String> = self.active.keys().cloned().collect();
1979 self.clear_active();
1980 if !self.status.done() {
1981 self.status = RunStatus::Failed;
1982 }
1983 self.event(
1984 by,
1985 format!(
1986 "abandoned: seat(s) {} left behind by a killed process, past their own \
1987 timeout with no live daemon claiming this run",
1988 seats.join(", ")
1989 ),
1990 );
1991 }
1992
1993 /// Flush to `run.json`, atomically, under the process-global [`home`].
1994 pub fn save(&mut self) -> Result<()> {
1995 let home = home();
1996 self.save_under(&home)
1997 }
1998
1999 /// [`Self::save`], rooted at an explicit `home` instead of the
2000 /// process-global one.
2001 ///
2002 /// For a caller that was already handed its own `home` explicitly — a
2003 /// housekeeping pass, mainly, for the same reason `Queue::at` and the
2004 /// daemon status path are parameters rather than resolved here (see
2005 /// `daemon::drive`'s own doc) — falling through to the global would write
2006 /// back through whichever directory some *other* process or test pinned
2007 /// into that `OnceLock` first, not the one this call was actually handed.
2008 pub fn save_under(&mut self, home: &Path) -> Result<()> {
2009 self.updated_at = Timestamp::now();
2010 let dir = home.join("runs").join(&self.id);
2011 std::fs::create_dir_all(&dir).with_context(|| format!("create {}", dir.display()))?;
2012 let body = serde_json::to_string_pretty(self).context("serialize run state")?;
2013 let tmp = dir.join("run.json.tmp");
2014 std::fs::write(&tmp, &body).with_context(|| format!("write {}", tmp.display()))?;
2015 std::fs::rename(&tmp, dir.join("run.json")).with_context(|| "replace run.json")?;
2016 Ok(())
2017 }
2018
2019 /// Load a run by id or unambiguous id prefix.
2020 pub fn load(id: &str) -> Result<Self> {
2021 let resolved = resolve_id(id)?;
2022 let path = run_dir(&resolved).join("run.json");
2023 let body =
2024 std::fs::read_to_string(&path).with_context(|| format!("read {}", path.display()))?;
2025 let state: Self =
2026 serde_json::from_str(&body).with_context(|| format!("parse {}", path.display()))?;
2027 migrate_schema(state)
2028 }
2029
2030 /// [`Self::load`], rooted at an explicit `home` instead of the
2031 /// process-global one, and for a full id rather than a prefix — a caller
2032 /// with its own `home` already has the exact id (from a `RunState` it
2033 /// already read, or from `Task::runs`) and has no `runs_root` to search
2034 /// for a prefix against in the first place. Same reasoning as
2035 /// [`Self::save_under`]: a caller holding its own `home` explicitly must
2036 /// not read back through whichever directory some *other* process or
2037 /// test pinned into the global [`home`] `OnceLock` first.
2038 pub fn load_under(id: &str, home: &Path) -> Result<Self> {
2039 let path = home.join("runs").join(id).join("run.json");
2040 let body =
2041 std::fs::read_to_string(&path).with_context(|| format!("read {}", path.display()))?;
2042 let state: Self =
2043 serde_json::from_str(&body).with_context(|| format!("parse {}", path.display()))?;
2044 migrate_schema(state)
2045 }
2046}
2047
2048fn migrate_schema(mut state: RunState) -> Result<RunState> {
2049 // Schema 5 predates deferred e2e. Its empty e2e lists therefore mean
2050 // "not configured", never "deferred"; serde's field defaults retain
2051 // exactly that representation while this migration permits resumes.
2052 if state.schema == 5 {
2053 state.schema = 6;
2054 }
2055 // Schema 6 predates `gate_ran` and could not tell "never attempted or
2056 // resource-blocked" apart from "ran with zero commands configured" — see
2057 // `SCHEMA`'s doc for schema 7. A non-empty `gate` is a real recorded
2058 // attempt either way, so it is trusted as `gate_ran = true` rather than
2059 // spent again; an empty one is simply handed back to the next `gate()`
2060 // call, which re-attempts it and, for the zero-commands case, resolves
2061 // instantly.
2062 if state.schema == 6 {
2063 state.gate_ran = !state.gate.is_empty();
2064 state.schema = 7;
2065 }
2066 // Schema 7's main review loop always checked `e2e` against the round's
2067 // own `head`, it just never wrote that fact into `verified_head` unless
2068 // a catch-up run had checked a *different* commit — see `SCHEMA`'s doc
2069 // for schema 8. Reconstructing `Some(head)` for a round whose `e2e` held
2070 // a real attempt restores a fact that was always true; `verified_at` has
2071 // no historical value to recover and stays `None`.
2072 if state.schema == 7 {
2073 for round in &mut state.reviews {
2074 if round.verified_head.is_none()
2075 && matches!(round.e2e_status(), E2eStatus::Passed | E2eStatus::Failed)
2076 {
2077 round.verified_head = Some(round.head.clone());
2078 }
2079 }
2080 state.schema = 8;
2081 }
2082 // Schema 8 predates `operator_fixes`. There is nothing to reconstruct —
2083 // an old run simply never had one requested — so `#[serde(default)]`
2084 // already left it as the correct empty `Vec`; this only advances the
2085 // version number.
2086 if state.schema == 8 {
2087 state.schema = 9;
2088 }
2089 // Schema 9 predates `RunStatus::VerifiedNoop` and
2090 // `Candidate::verified_noop`. Nothing to reconstruct: an old record never
2091 // made the claim, `#[serde(default)]` already reads `verified_noop` as
2092 // `None` on every candidate, and a `VerifiedNoop` status cannot appear in
2093 // a schema-9 record at all — see `SCHEMA`'s doc for schema 10. This only
2094 // advances the version number.
2095 if state.schema == 9 {
2096 state.schema = 10;
2097 }
2098 // Schema 10 predates `released_to`: no old run ever had its worktree
2099 // handed over, and `#[serde(default)]` reads exactly that. Only the
2100 // version number advances.
2101 if state.schema == 10 {
2102 state.schema = SCHEMA;
2103 }
2104 if state.schema != SCHEMA {
2105 bail!(
2106 "run {} was written by a different magi (schema {}, this build \
2107 speaks {SCHEMA})",
2108 state.id,
2109 state.schema
2110 );
2111 }
2112 Ok(state)
2113}
2114
2115impl RunState {
2116 /// Whether a later attempt took this run's worktree over, so a resume has
2117 /// nothing left to continue from.
2118 pub fn released(&self) -> bool {
2119 self.released_to.is_some()
2120 }
2121
2122 /// The winning candidate, once the tally has run.
2123 pub fn winner(&self) -> Option<&Candidate> {
2124 let label = self.tally.as_ref()?.winner;
2125 self.candidates.iter().find(|c| c.label == label)
2126 }
2127
2128 /// Candidates eligible for judging.
2129 pub fn viable(&self) -> Vec<&Candidate> {
2130 self.candidates.iter().filter(|c| c.viable()).collect()
2131 }
2132
2133 /// Did every candidate write nothing, and every one of them back it with
2134 /// evidence [`crate::graph`]'s adoption guard accepted?
2135 ///
2136 /// All-or-nothing on purpose: one candidate declaring `NO CHANGE NEEDED`
2137 /// while another simply failed to produce anything is not agreement, it
2138 /// is one candidate's unverified claim next to an ordinary loss, and the
2139 /// run must still read as the `Failed` it is. Only ever meaningful when
2140 /// [`Self::viable`] is already empty — a run with any real patch to judge
2141 /// never reaches the caller that asks this.
2142 pub fn all_candidates_verified_noop(&self) -> bool {
2143 !self.candidates.is_empty()
2144 && self
2145 .candidates
2146 .iter()
2147 .all(|c| c.empty && c.verified_noop.is_some())
2148 }
2149
2150 /// Findings still open when the review loop stopped trying: the last
2151 /// round's, exactly when that round was not clean. Empty on a run that
2152 /// never reviewed, or whose last round was clean.
2153 ///
2154 /// This is the last round's findings regardless of what the fixer claims
2155 /// to have addressed in that same round: a round that stopped the loop
2156 /// (round budget spent, or no tree progress for
2157 /// [`crate::graph::STAGNANT_LIMIT`] rounds) never had a *following* round
2158 /// to confirm the fix actually landed, and the self-reported adoption
2159 /// count is not trusted for that judgement either — see
2160 /// [`ReviewRound::progressed`].
2161 pub fn open_findings(&self) -> Vec<&Finding> {
2162 match self.reviews.last() {
2163 Some(r) if !r.clean => r
2164 .reviews
2165 .iter()
2166 .flat_map(|rec| rec.findings.iter())
2167 .collect(),
2168 _ => Vec::new(),
2169 }
2170 }
2171
2172 /// Every finding raised in the most recent review round, regardless of
2173 /// that round's own severity mix — unlike [`Self::open_findings`], not
2174 /// filtered to a round that was not clean. This is the pool `magi fix`
2175 /// reports as available to pick from: a round can conclude clean (no
2176 /// finding blocked merge) while still carrying minor findings nobody
2177 /// has acted on.
2178 pub fn last_round_findings(&self) -> Vec<&Finding> {
2179 self.reviews
2180 .last()
2181 .into_iter()
2182 .flat_map(|r| r.reviews.iter())
2183 .flat_map(|rec| rec.findings.iter())
2184 .collect()
2185 }
2186
2187 /// Look up a finding by id anywhere in this run's review history,
2188 /// together with the round and reviewer record that raised it — the
2189 /// provenance `magi fix` snapshots onto [`OperatorFixFinding`].
2190 pub fn finding(&self, id: &str) -> Option<(&ReviewRound, &ReviewRecord, &Finding)> {
2191 self.reviews.iter().find_map(|round| {
2192 round.reviews.iter().find_map(|rec| {
2193 rec.findings
2194 .iter()
2195 .find(|f| f.id == id)
2196 .map(|f| (round, rec, f))
2197 })
2198 })
2199 }
2200
2201 /// Did this run reach a mergeable status (`Ready` or `Merged`) with
2202 /// review findings still open?
2203 ///
2204 /// That combination is the point of the review hand-off: the review
2205 /// round budget (or an unproductive round, see [`ReviewRound::progressed`])
2206 /// was spent while gate and e2e stayed green, so the run was handed off
2207 /// rather than blocked — but the findings did not disappear, and whoever
2208 /// reads the result should be told they are still there.
2209 pub fn handed_off_with_open_findings(&self) -> bool {
2210 matches!(self.status, RunStatus::Ready | RunStatus::Merged)
2211 && self.reviews.last().is_some_and(|r| !r.clean)
2212 }
2213
2214 /// Reached `Ready` because `[merge] mode = "none"` left it there by
2215 /// design, never to be picked up by the PR-polling merge watcher — as
2216 /// opposed to a `Ready` that is still a plausible landing candidate (a
2217 /// PR closed without merging, or a re-entry onto an already-concluded
2218 /// node). Both leave `status` at `Ready`; only this one leaves the
2219 /// winning branch permanently unwatched, which is what a caller needs to
2220 /// know before labelling the run in a listing.
2221 pub fn unmerged_by_design(&self) -> bool {
2222 self.status == RunStatus::Ready
2223 && self
2224 .merge
2225 .as_ref()
2226 .is_some_and(|m| m.mode == MergeMode::None)
2227 }
2228
2229 /// Did a step after the merge leave something only a human can finish?
2230 ///
2231 /// Derived, not a status: the merge happened, so `status` stays `Merged`
2232 /// and everything that branches on it keeps working. Display code asks
2233 /// this so a merged run with an unmerged release PR does not read as
2234 /// plain green.
2235 pub fn needs_attention(&self) -> bool {
2236 self.status == RunStatus::Merged
2237 && self
2238 .release_bump
2239 .as_ref()
2240 .is_some_and(|b| b.action_required.is_some())
2241 }
2242
2243 /// Local-time creation stamp for reports.
2244 pub fn created_local(&self) -> String {
2245 self.created_at
2246 .to_zoned(jiff::tz::TimeZone::system())
2247 .strftime("%Y-%m-%d %H:%M:%S")
2248 .to_string()
2249 }
2250
2251 /// Assert that this run is safe to delete.
2252 ///
2253 /// Refuses a run a live daemon is working on, and refuses any run whose
2254 /// candidate worktrees and branches have not been folded away with `magi
2255 /// fold`. The fold requirement is the real protection: it is what makes
2256 /// "delete" mean "remove a record" rather than "throw away a worktree
2257 /// somebody may still be editing".
2258 ///
2259 /// `in_flight` has to come from the caller, because a run's own status
2260 /// cannot answer the question. A daemon killed mid-run leaves its status at
2261 /// `implementing` forever, and a guard that trusted that would make every
2262 /// interrupted run permanently undeletable - the operator's only recourse
2263 /// being to edit `run.json` by hand, which is exactly the sort of thing
2264 /// this command exists to avoid. The queue already treats an orphaned
2265 /// `.lock` from a `SIGKILL`ed daemon the same way; this is that rule for
2266 /// runs.
2267 pub fn ensure_can_delete(&self, in_flight: bool) -> Result<()> {
2268 if in_flight {
2269 bail!(
2270 "run {} is being worked on by a live daemon right now",
2271 self.short()
2272 );
2273 }
2274 if self.candidates.iter().any(|c| !c.folded) {
2275 bail!(
2276 "run {} has unfolded candidates; fold first with `magi fold`",
2277 self.short()
2278 );
2279 }
2280 Ok(())
2281 }
2282}
2283
2284/// The short form of a commit, for a label a human or an LLM reads.
2285fn short(commit: &str) -> String {
2286 commit.chars().take(7).collect()
2287}
2288
2289/// The short form of a run id: the trailing block after the last `-`.
2290///
2291/// A free function as well as [`RunState::short`], because callers that have
2292/// only an id - an error message, a daemon status, a route handler - were
2293/// otherwise reimplementing the split, and two spellings of "short id" is one
2294/// rename away from branch names that no longer match their run.
2295pub fn short_of(id: &str) -> &str {
2296 id.split('-').next_back().unwrap_or(id)
2297}
2298
2299/// Where magi keeps its runs.
2300///
2301/// `MAGI_HOME` overrides the default, and [`set_home`] overrides both — which
2302/// is what lets the integration tests drive a whole graph without writing into
2303/// the operator's real history.
2304///
2305/// In a unit test build (`cfg(test)`), falling through to the real
2306/// `<data_local>/magi` is not a fallback worth having: it is exactly how
2307/// three broken fixture runs ended up in the operator's actual history and
2308/// were counted as `unreadable` by the deck. A test that reaches this point
2309/// forgot to call [`set_home`] (or set `MAGI_HOME`) - that is a bug in the
2310/// test, not a case to serve, so it panics instead of writing anywhere.
2311pub fn home() -> PathBuf {
2312 resolve_home(HOME.get().cloned(), std::env::var_os("MAGI_HOME"))
2313}
2314
2315/// The decision `home` makes, taking its two overrides as plain values
2316/// instead of reading the `OnceLock` and the environment itself.
2317///
2318/// Pulled out so the `cfg(test)` panic is asserted directly against a
2319/// `None, None` input, rather than racing every other unit test in the
2320/// binary for who touches the process-global `HOME` first.
2321fn resolve_home(pinned: Option<PathBuf>, magi_home_env: Option<std::ffi::OsString>) -> PathBuf {
2322 if let Some(dir) = pinned {
2323 return dir;
2324 }
2325 if let Some(dir) = magi_home_env {
2326 return PathBuf::from(dir);
2327 }
2328 #[cfg(test)]
2329 {
2330 panic!(
2331 "run::home() was reached in a test without run::set_home() or \
2332 MAGI_HOME; this would write into the operator's real \
2333 <data_local>/magi. Call `run::set_home(temp_dir)` before any \
2334 code path that touches a RunState."
2335 );
2336 }
2337 #[cfg(not(test))]
2338 {
2339 dirs::data_local_dir()
2340 .unwrap_or_else(|| PathBuf::from("."))
2341 .join("magi")
2342 }
2343}
2344
2345/// [`home`] without the test panic: `None` in a test that never pinned a
2346/// home, so best-effort writers (see [`crate::notices::raise`]) skip the write
2347/// instead of touching the operator's real state or aborting the test.
2348pub fn try_home() -> Option<PathBuf> {
2349 if cfg!(test) {
2350 HOME.get()
2351 .cloned()
2352 .or_else(|| std::env::var_os("MAGI_HOME").map(PathBuf::from))
2353 } else {
2354 Some(home())
2355 }
2356}
2357
2358/// Pin the run home for this process. The first call wins.
2359pub fn set_home(dir: PathBuf) {
2360 let _ = HOME.set(dir);
2361}
2362
2363static HOME: std::sync::OnceLock<PathBuf> = std::sync::OnceLock::new();
2364
2365/// `<home>/runs`.
2366pub fn runs_root() -> PathBuf {
2367 home().join("runs")
2368}
2369
2370/// The worktree root a run uses when the config sets none: `~/wt/magi`.
2371///
2372/// One definition of the default, so the folder the janitor folds and the
2373/// folder the health view sizes cannot drift apart: a run with no configured
2374/// [`crate::config::Graph::worktree_root`] lays its worktrees exactly here.
2375pub fn default_worktree_root() -> PathBuf {
2376 dirs::home_dir()
2377 .unwrap_or_else(|| PathBuf::from("."))
2378 .join("wt")
2379 .join("magi")
2380}
2381
2382/// Directory for one run id.
2383pub fn run_dir(id: &str) -> PathBuf {
2384 runs_root().join(id)
2385}
2386
2387/// Every run id on disk, newest first.
2388///
2389/// A directory is a run because of its **name**, not because it holds a
2390/// readable `run.json`. A run whose very first save lost the machine's last
2391/// free bytes leaves `<id>/run.json.tmp` and nothing else, and filtering on
2392/// `run.json` made that run invisible everywhere: not in `magi list`, not in
2393/// `runs_unreadable`, not on the phone, so nothing could report it and no
2394/// route could clear it. `88c0` sat like that for two days. Unreadable is
2395/// counted, never hidden - the readers already say why each one cannot be
2396/// read, and `fold_unreadable` is how a record like this leaves.
2397pub fn list_ids() -> Vec<String> {
2398 let mut ids: Vec<String> = std::fs::read_dir(runs_root())
2399 .into_iter()
2400 .flatten()
2401 .flatten()
2402 .filter(|e| e.path().is_dir())
2403 .map(|e| e.file_name().to_string_lossy().into_owned())
2404 .filter(|name| is_run_id(name))
2405 .collect();
2406 // Ids start with a sortable timestamp.
2407 ids.sort_unstable_by(|a, b| b.cmp(a));
2408 ids
2409}
2410
2411/// Does `name` have the shape [`new_id`] mints: `YYYYMMDD-HHMMSS-xxxx`?
2412///
2413/// The test for "this directory is a run", so a stray folder under
2414/// `<home>/runs` is not reported as a broken run.
2415///
2416/// The tag is checked for length and for being alphanumeric, not for being
2417/// hex: real ids are hex, but fixtures across this crate name runs
2418/// `...-dead` / `...-gone` / `...-once`, and a predicate that disowned those
2419/// would be asserting the fixtures' spelling rather than the shape.
2420pub fn is_run_id(name: &str) -> bool {
2421 let mut parts = name.split('-');
2422 let (Some(day), Some(time), Some(tag), None) =
2423 (parts.next(), parts.next(), parts.next(), parts.next())
2424 else {
2425 return false;
2426 };
2427 day.len() == 8
2428 && day.bytes().all(|b| b.is_ascii_digit())
2429 && time.len() == 6
2430 && time.bytes().all(|b| b.is_ascii_digit())
2431 && tag.len() == 4
2432 && tag.bytes().all(|b| b.is_ascii_alphanumeric())
2433}
2434
2435/// Expand an id prefix to exactly one run id.
2436pub fn resolve_id(prefix: &str) -> Result<String> {
2437 // A whole id names its directory, readable state or not: the run whose
2438 // `run.json` never landed still has to be reachable by `magi show` and
2439 // by the fold route, which is the only way its record ever leaves.
2440 if is_run_id(prefix) && run_dir(prefix).is_dir() {
2441 return Ok(prefix.to_owned());
2442 }
2443 let hits: Vec<String> = list_ids()
2444 .into_iter()
2445 .filter(|id| id.starts_with(prefix) || id.ends_with(prefix))
2446 .collect();
2447 match hits.len() {
2448 1 => Ok(hits.into_iter().next().expect("exactly one hit")),
2449 0 => bail!("no run matches `{prefix}`"),
2450 _ => bail!(
2451 "`{prefix}` matches {} runs: {}",
2452 hits.len(),
2453 hits.join(", ")
2454 ),
2455 }
2456}
2457
2458/// The most recent run, if any.
2459pub fn latest_id() -> Option<String> {
2460 list_ids().into_iter().next()
2461}
2462
2463/// `YYYYMMDD-HHMMSS-xxxx`, sortable and short enough for a branch name.
2464///
2465/// The four hex digits are fresh entropy, **not** `blind.seed`. They were the
2466/// seed, and a pinned seed then made the whole id a function of the second it
2467/// started in: two runs a second apart were distinguishable, two in the same
2468/// second were not. Everything keyed on the id collided with them - the run
2469/// directory, `artifacts/`, and the candidate worktrees under
2470/// `wt/magi/<short>/`.
2471///
2472/// `tests/common` pins the seed on purpose, so its integration tests all share
2473/// one suffix. On Windows the suite is slow enough that the seconds differ and
2474/// nothing showed; on Linux `graph_dropped_stream`'s three tests run inside
2475/// 16s, so two of them shared a run directory and the second read an artifact
2476/// the first had written (`impl-B-resume.out`) - a failure that looked like the
2477/// resume logic misbehaving and was really two runs in one directory.
2478///
2479/// A seed exists to make the *blind* decisions reproducible: label assignment
2480/// and per-judge presentation order. It was never meant to name the run, and
2481/// `RunState::seed` still carries it for what it is for.
2482fn new_id() -> String {
2483 let stamp = Zoned::now().strftime("%Y%m%d-%H%M%S");
2484 let entropy = crate::rng::entropy();
2485 format!("{stamp}-{:04x}", (entropy ^ (entropy >> 32)) & 0xffff)
2486}
2487
2488/// Keep the last `max` bytes of `text`, on a line boundary.
2489pub fn tail(text: &str, max: usize) -> String {
2490 if text.len() <= max {
2491 return text.to_owned();
2492 }
2493 let mut cut = text.len() - max;
2494 while cut < text.len() && !text.is_char_boundary(cut) {
2495 cut += 1;
2496 }
2497 let slice = &text[cut..];
2498 let start = slice.find('\n').map_or(0, |i| i + 1);
2499 format!(
2500 "[... {} earlier bytes omitted ...]\n{}",
2501 cut,
2502 &slice[start..]
2503 )
2504}
2505
2506/// Path of a run artifact.
2507pub fn artifact_path(run: &RunState, name: &str) -> PathBuf {
2508 run.dir().join("artifacts").join(name)
2509}
2510
2511/// Write an artifact, creating the directory if needed.
2512pub fn write_artifact(run: &RunState, name: &str, body: &str) -> Result<PathBuf> {
2513 let path = artifact_path(run, name);
2514 if let Some(parent) = path.parent() {
2515 std::fs::create_dir_all(parent).with_context(|| format!("create {}", parent.display()))?;
2516 }
2517 std::fs::write(&path, body).with_context(|| format!("write {}", path.display()))?;
2518 Ok(path)
2519}
2520
2521/// Read an artifact back, e.g. a stored patch on resume.
2522pub fn read_artifact(run: &RunState, name: &str) -> Option<String> {
2523 std::fs::read_to_string(artifact_path(run, name)).ok()
2524}
2525
2526#[cfg(test)]
2527mod tests {
2528 use super::*;
2529
2530 fn state() -> RunState {
2531 RunState::new(
2532 PathBuf::from("/repo"),
2533 "main".to_owned(),
2534 "abc1234def".to_owned(),
2535 "add retries".to_owned(),
2536 Config::default(),
2537 )
2538 }
2539
2540 #[test]
2541 fn superseded_is_terminal_but_not_resumable() {
2542 assert!(RunStatus::Superseded.done());
2543 assert!(!RunStatus::Superseded.resumable());
2544 assert_eq!(RunStatus::Superseded.as_str(), "superseded");
2545 }
2546
2547 #[test]
2548 fn load_under_reads_back_exactly_what_save_under_wrote_at_an_explicit_home() {
2549 // Both rooted at an explicit `home` rather than the process-global
2550 // one - a caller with its own `home` (a test fixture, a housekeeping
2551 // pass) must round-trip through exactly that directory, never
2552 // through whichever home some other test in the same binary pinned
2553 // into the global `OnceLock` first.
2554 let dir = tempfile::tempdir().unwrap();
2555 let mut original = state();
2556 original.id = "20260101-000000-load".to_owned();
2557 original.status = RunStatus::Blocked;
2558 original.save_under(dir.path()).unwrap();
2559
2560 let reloaded = RunState::load_under(&original.id, dir.path()).unwrap();
2561 assert_eq!(reloaded.id, original.id);
2562 assert_eq!(reloaded.status, RunStatus::Blocked);
2563
2564 assert!(
2565 RunState::load_under("20260101-000000-none", dir.path()).is_err(),
2566 "an id with nothing saved under this home must not silently read something else"
2567 );
2568 }
2569
2570 #[test]
2571 fn resolve_home_prefers_the_pin_then_the_env_var() {
2572 let pinned = PathBuf::from("/pinned");
2573 assert_eq!(
2574 resolve_home(Some(pinned.clone()), Some("/env".into())),
2575 pinned,
2576 "a pin wins even over MAGI_HOME"
2577 );
2578 assert_eq!(
2579 resolve_home(None, Some("/env".into())),
2580 PathBuf::from("/env")
2581 );
2582 }
2583
2584 #[test]
2585 #[should_panic(expected = "run::set_home()")]
2586 fn resolve_home_refuses_to_fall_back_to_the_operators_real_home() {
2587 // Neither override present is exactly the state a test reaches by
2588 // forgetting `set_home`/`MAGI_HOME` - the accident that put three
2589 // broken fixture runs into the operator's real history. Asserted
2590 // against the pure decision directly, not `home()` itself, because
2591 // `HOME` is a process-wide `OnceLock` another test may have already
2592 // set - this must not depend on test execution order.
2593 resolve_home(None, None);
2594 }
2595
2596 #[test]
2597 fn a_run_is_named_by_shape_so_a_state_less_directory_is_still_a_run() {
2598 // The shape `new_id` mints. A directory answering to it is a run even
2599 // with no readable `run.json`: that is how a save that ran out of
2600 // disk stays visible instead of vanishing from every listing.
2601 assert!(is_run_id(&new_id()));
2602 assert!(is_run_id("20260904-014540-88c0"));
2603 // Not runs: a stray folder, a truncated id, a non-hex tag, and an id
2604 // with an extra segment (a worktree label, say).
2605 assert!(!is_run_id("scratch"));
2606 assert!(!is_run_id("20260904-014540"));
2607 assert!(!is_run_id("20260904-014540-88c0f"));
2608 assert!(!is_run_id("2026090x-014540-88c0"));
2609 assert!(!is_run_id("20260904-014540-88c0-A"));
2610 }
2611
2612 #[test]
2613 fn ids_are_sortable_and_short_suffixed() {
2614 let s = state();
2615 let parts: Vec<&str> = s.id.split('-').collect();
2616 assert_eq!(parts.len(), 3);
2617 assert_eq!(parts[0].len(), 8);
2618 assert_eq!(parts[1].len(), 6);
2619 assert_eq!(parts[2].len(), 4);
2620 assert_eq!(s.short(), parts[2]);
2621 }
2622
2623 #[test]
2624 fn branch_names_carry_the_label_not_the_author() {
2625 let s = state();
2626 let b = s.branch_for('B');
2627 assert_eq!(b, format!("magi/{}/B", s.short()));
2628 assert!(!b.contains("claude"));
2629 }
2630
2631 /// A pinned seed reproduces the blind decisions. It must **not** reproduce
2632 /// the run's identity.
2633 ///
2634 /// `assert_eq!(a.short(), b.short())` used to stand where the last
2635 /// assertion is now, and it was pinning the defect: with the id's suffix
2636 /// derived from the seed, two runs started in the same second were the
2637 /// same run as far as the filesystem was concerned - one directory, one
2638 /// `artifacts/`, one set of candidate worktrees. `tests/common` pins a
2639 /// seed for every integration test, so on Linux, where the suite is fast,
2640 /// two tests in `graph_dropped_stream` shared a directory and one read the
2641 /// other's artifact.
2642 #[test]
2643 fn a_pinned_seed_is_reproducible_but_never_the_run_id() {
2644 let mut cfg = Config::default();
2645 cfg.blind.seed = Some(1234);
2646 let a = RunState::new(
2647 PathBuf::from("/r"),
2648 "main".to_owned(),
2649 "c".to_owned(),
2650 "t".to_owned(),
2651 cfg.clone(),
2652 );
2653 let b = RunState::new(
2654 PathBuf::from("/r"),
2655 "main".to_owned(),
2656 "c".to_owned(),
2657 "t".to_owned(),
2658 cfg,
2659 );
2660 // What the seed is for: the same shuffles, run after run.
2661 assert_eq!(a.seed, 1234);
2662 assert_eq!(a.seed, b.seed);
2663 // What it is not for. Two runs are two runs, in the same second or
2664 // not, and everything keyed on the id depends on that.
2665 assert_ne!(
2666 a.id, b.id,
2667 "two runs sharing an id share a directory, artifacts and worktrees"
2668 );
2669 }
2670
2671 #[test]
2672 fn status_terminality() {
2673 assert!(RunStatus::Merged.done());
2674 assert!(RunStatus::Blocked.done());
2675 assert!(!RunStatus::Reviewing.done());
2676 }
2677
2678 fn overrun_seat(now: Timestamp, elapsed_secs: i64, timeout_secs: u64) -> ActiveSeat {
2679 ActiveSeat {
2680 node: "implement".to_owned(),
2681 started_at: now - jiff::SignedDuration::new(elapsed_secs, 0),
2682 timeout_secs,
2683 attempt: 0,
2684 task: None,
2685 command: None,
2686 index: None,
2687 total: None,
2688 }
2689 }
2690
2691 #[test]
2692 fn active_all_overrun_requires_every_seat_past_its_own_timeout() {
2693 let mut s = state();
2694 let now = Timestamp::now();
2695 assert!(
2696 !s.active_all_overrun(now),
2697 "nothing active is not evidence of anything"
2698 );
2699
2700 s.active
2701 .insert("impl-A".to_owned(), overrun_seat(now, 21_000, 3_600));
2702 assert!(
2703 s.active_all_overrun(now),
2704 "21000s elapsed against a 3600s budget"
2705 );
2706
2707 // A seat still well within its own budget means the run is not
2708 // provably dead, however far its sibling has overrun.
2709 s.active
2710 .insert("impl-B".to_owned(), overrun_seat(now, 0, 3_600));
2711 assert!(!s.active_all_overrun(now));
2712 }
2713
2714 /// A daemon claim wins outright, whatever `driver_pid` or either query
2715 /// says — the stronger, independently-heartbeating signal. Neither query
2716 /// closure is even called: a daemon claim short-circuits before either
2717 /// one, which panicking closures here prove.
2718 #[test]
2719 fn liveness_reads_live_from_a_daemon_claim_alone() {
2720 let mut s = state();
2721 s.driver_pid = None;
2722 assert_eq!(
2723 s.liveness_with(
2724 true,
2725 |_| panic!("a daemon claim needs no pid query"),
2726 |_| panic!("a daemon claim needs no identity query")
2727 ),
2728 Liveness::Live,
2729 "a daemon claim needs no pid to back it up"
2730 );
2731 }
2732
2733 /// The gap `driver_pid` closes: no daemon claim (every manual `magi run`
2734 /// / `magi review`), but the recorded pid answers alive *and* the
2735 /// process currently holding it still carries the same start-time
2736 /// marker this run recorded — proof it is genuinely the same process,
2737 /// not merely the same number.
2738 #[test]
2739 fn liveness_reads_live_from_a_confirmed_pid_with_a_matching_identity() {
2740 let mut s = state();
2741 s.driver_pid = Some(4242);
2742 s.driver_started_at = Some("2026-09-22T10:00:00Z".to_owned());
2743 assert_eq!(
2744 s.liveness_with(
2745 false,
2746 |pid| {
2747 assert_eq!(pid, 4242);
2748 Some(true)
2749 },
2750 |pid| {
2751 assert_eq!(pid, 4242);
2752 Some("2026-09-22T10:00:00Z".to_owned())
2753 }
2754 ),
2755 Liveness::Live
2756 );
2757 }
2758
2759 /// No daemon claim and the recorded pid confirmed gone by the OS itself
2760 /// — dead outright, and the identity query is never even reached (a
2761 /// panicking closure proves it), since there is nothing left to
2762 /// corroborate.
2763 #[test]
2764 fn liveness_reads_dead_from_a_confirmed_dead_pid() {
2765 let mut s = state();
2766 s.driver_pid = Some(4242);
2767 assert_eq!(
2768 s.liveness_with(
2769 false,
2770 |_| Some(false),
2771 |_| panic!("a confirmed-dead pid needs no identity query")
2772 ),
2773 Liveness::Dead
2774 );
2775 }
2776
2777 /// The gap this task's review round exists to close: a killed manual
2778 /// run's pid gets handed to a wholly unrelated later process. `pid_status`
2779 /// alone would read that as `Live` — the reused pid really is alive —
2780 /// but the process now holding it started at a different moment than the
2781 /// one this run recorded, so this must read `Dead`, not `Live`: a
2782 /// mismatch is exactly as good as proof the original driver is gone.
2783 #[test]
2784 fn liveness_reads_dead_when_a_live_pid_no_longer_matches_the_recorded_start_time() {
2785 let mut s = state();
2786 s.driver_pid = Some(4242);
2787 s.driver_started_at = Some("2026-09-22T10:00:00Z".to_owned());
2788 assert_eq!(
2789 s.liveness_with(
2790 false,
2791 |_| Some(true),
2792 |_| Some("2026-09-22T11:30:00Z".to_owned())
2793 ),
2794 Liveness::Dead,
2795 "the pid is alive, but under a different process than the one this run recorded"
2796 );
2797 }
2798
2799 /// Missing information never collapses to `Dead`: an old run with no
2800 /// `driver_pid` at all, a `driver_pid` this build could not query, a live
2801 /// pid with no recorded start time to compare (an even older run, before
2802 /// that field existed), and a live pid whose current identity this build
2803 /// could not re-query, all read as `Unknown` — never a guess in either
2804 /// direction.
2805 #[test]
2806 fn liveness_never_guesses_out_of_missing_information() {
2807 let mut s = state();
2808 s.driver_pid = None;
2809 assert_eq!(
2810 s.liveness_with(
2811 false,
2812 |_| panic!("no pid to query"),
2813 |_| panic!("no pid to query")
2814 ),
2815 Liveness::Unknown,
2816 "no driver_pid recorded at all — an old run predating this field"
2817 );
2818
2819 s.driver_pid = Some(4242);
2820 assert_eq!(
2821 s.liveness_with(false, |_| None, |_| panic!("inconclusive already")),
2822 Liveness::Unknown,
2823 "a pid to ask, but the platform could not answer for it"
2824 );
2825
2826 s.driver_started_at = None;
2827 assert_eq!(
2828 s.liveness_with(false, |_| Some(true), |_| Some("anything".to_owned())),
2829 Liveness::Unknown,
2830 "a live pid, but no recorded marker to corroborate it against — an old run \
2831 predating `driver_started_at`"
2832 );
2833
2834 s.driver_started_at = Some("2026-09-22T10:00:00Z".to_owned());
2835 assert_eq!(
2836 s.liveness_with(false, |_| Some(true), |_| None),
2837 Liveness::Unknown,
2838 "a live pid and a recorded marker, but the identity re-query itself failed"
2839 );
2840 }
2841
2842 #[test]
2843 fn abandon_clears_active_and_fails_a_non_terminal_run() {
2844 let mut s = state();
2845 s.status = RunStatus::Implementing;
2846 let now = Timestamp::now();
2847 s.active
2848 .insert("impl-A".to_owned(), overrun_seat(now, 21_000, 3_600));
2849
2850 s.abandon("daemon");
2851
2852 assert!(s.active.is_empty());
2853 assert_eq!(s.status, RunStatus::Failed);
2854 assert!(
2855 s.events
2856 .last()
2857 .expect("an event was logged")
2858 .message
2859 .contains("impl-A"),
2860 "the event names the abandoned seat"
2861 );
2862 }
2863
2864 #[test]
2865 fn abandon_never_overwrites_a_status_already_terminal() {
2866 let mut s = state();
2867 s.status = RunStatus::Ready;
2868 let now = Timestamp::now();
2869 s.active
2870 .insert("impl-A".to_owned(), overrun_seat(now, 21_000, 3_600));
2871
2872 s.abandon("daemon");
2873
2874 assert!(s.active.is_empty());
2875 assert_eq!(
2876 s.status,
2877 RunStatus::Ready,
2878 "a run already done must not be relabelled Failed"
2879 );
2880 }
2881
2882 #[test]
2883 fn candidate_viability_excludes_empty_and_failed() {
2884 let mut c = Candidate {
2885 index: 0,
2886 label: 'A',
2887 agent: "a".to_owned(),
2888 branch: "b".to_owned(),
2889 worktree: PathBuf::from("/w"),
2890 summary: String::new(),
2891 stat: String::new(),
2892 files: 1,
2893 commits: 1,
2894 empty: false,
2895 failed: None,
2896 verified_noop: None,
2897 duration_ms: 0,
2898 folded: false,
2899 };
2900 assert!(c.viable());
2901 c.empty = true;
2902 assert!(!c.viable());
2903 c.empty = false;
2904 c.failed = Some("timeout".to_owned());
2905 assert!(!c.viable());
2906 }
2907
2908 #[test]
2909 fn build_failure_is_distinguished_from_a_failing_test() {
2910 let link_race = CommandOutcome {
2911 command: "cargo test".to_owned(),
2912 code: Some(1),
2913 output_tail: "LINK : fatal error LNK1104: cannot open file \
2914 'graph_dirty_tree-71d4dc8e.exe'\n\
2915 error: could not compile `magi-cli` (test \"graph_dirty_tree\")"
2916 .to_owned(),
2917 duration_ms: 500,
2918 resource_blocked: false,
2919 };
2920 assert!(!link_race.ok());
2921 assert!(link_race.build_failed());
2922
2923 let failing_test = CommandOutcome {
2924 command: "cargo test".to_owned(),
2925 code: Some(101),
2926 output_tail: "thread 'it_works' panicked at 'assertion failed'".to_owned(),
2927 duration_ms: 500,
2928 resource_blocked: false,
2929 };
2930 assert!(!failing_test.ok());
2931 assert!(
2932 !failing_test.build_failed(),
2933 "a real test failure must not be classed as a build failure"
2934 );
2935
2936 let passing = CommandOutcome {
2937 command: "cargo test".to_owned(),
2938 code: Some(0),
2939 output_tail: String::new(),
2940 duration_ms: 500,
2941 resource_blocked: false,
2942 };
2943 assert!(passing.ok());
2944 assert!(!passing.build_failed());
2945 }
2946
2947 #[test]
2948 fn tail_keeps_the_end_on_a_line_boundary() {
2949 let text = (0..100).map(|i| format!("line {i}\n")).collect::<String>();
2950 let t = tail(&text, 40);
2951 assert!(t.starts_with("[..."));
2952 assert!(t.ends_with("line 99\n"));
2953 assert!(t.len() < 120);
2954 assert_eq!(tail("short", 40), "short");
2955 }
2956
2957 #[test]
2958 fn tail_survives_multibyte_cuts() {
2959 let text = "あ".repeat(50);
2960 let t = tail(&text, 10);
2961 assert!(t.contains("earlier bytes omitted"));
2962 assert!(t.ends_with('あ'));
2963 }
2964
2965 fn finding(id: &str, severity: crate::verdict::Severity) -> crate::verdict::Finding {
2966 crate::verdict::Finding {
2967 id: id.to_owned(),
2968 severity,
2969 file: None,
2970 line: None,
2971 title: "x".to_owned(),
2972 detail: String::new(),
2973 }
2974 }
2975
2976 fn round(clean: bool, findings: Vec<crate::verdict::Finding>) -> ReviewRound {
2977 ReviewRound {
2978 round: 1,
2979 head: "h".to_owned(),
2980 verified_head: None,
2981 verified_at: None,
2982 reviews: vec![ReviewRecord {
2983 attempts: 0,
2984 reviewer: 1,
2985 agent: "a".to_owned(),
2986 summary: String::new(),
2987 findings,
2988 vote: None,
2989 failed: None,
2990 duration_ms: 0,
2991 }],
2992 e2e: Vec::new(),
2993 verify_retried: false,
2994 e2e_deferred: false,
2995 e2e_defer_reason: None,
2996 fix: None,
2997 blocking: 0,
2998 answered: 1,
2999 expected: 1,
3000 clean,
3001 progressed: false,
3002 vote_split: false,
3003 reconsideration: Vec::new(),
3004 verdict: None,
3005 }
3006 }
3007
3008 #[test]
3009 fn e2e_status_tells_deferred_apart_from_not_configured() {
3010 let mut r = round(false, Vec::new());
3011 assert_eq!(r.e2e_status(), E2eStatus::NotConfigured);
3012
3013 r.e2e_deferred = true;
3014 assert_eq!(
3015 r.e2e_status(),
3016 E2eStatus::Deferred,
3017 "an empty e2e must not read as unconfigured once it was deferred on purpose"
3018 );
3019
3020 r.e2e = vec![CommandOutcome {
3021 command: "test".to_owned(),
3022 code: Some(0),
3023 output_tail: String::new(),
3024 duration_ms: 0,
3025 resource_blocked: false,
3026 }];
3027 assert_eq!(
3028 r.e2e_status(),
3029 E2eStatus::Passed,
3030 "a round with real outcomes is never read as deferred, even if the flag is still set"
3031 );
3032 }
3033
3034 #[test]
3035 fn e2e_status_reports_a_real_failure_as_failed_not_deferred() {
3036 let mut r = round(false, Vec::new());
3037 r.e2e = vec![CommandOutcome {
3038 command: "test".to_owned(),
3039 code: Some(1),
3040 output_tail: "boom".to_owned(),
3041 duration_ms: 0,
3042 resource_blocked: false,
3043 }];
3044 assert_eq!(r.e2e_status(), E2eStatus::Failed);
3045 }
3046
3047 #[test]
3048 fn e2e_status_never_reads_a_resource_block_as_a_failure() {
3049 // The exact shape of contention on the shared build cache: `e2e`
3050 // holds one outcome, and it is `resource_blocked`, never a command
3051 // that actually ran and produced a red exit code.
3052 let mut r = round(false, Vec::new());
3053 r.e2e = vec![CommandOutcome {
3054 command: "(waiting for the shared build cache)".to_owned(),
3055 code: None,
3056 output_tail: "contended".to_owned(),
3057 duration_ms: 0,
3058 resource_blocked: true,
3059 }];
3060 assert_eq!(
3061 r.e2e_status(),
3062 E2eStatus::ResourceBlocked,
3063 "magi's own inability to get a command to run must not read as a verdict on the \
3064 patch"
3065 );
3066 }
3067
3068 #[test]
3069 fn verification_summary_is_silent_when_there_is_nothing_worth_saying() {
3070 let mut r = round(true, Vec::new());
3071 assert!(
3072 r.verification_summary("h").is_none(),
3073 "no verify.e2e configured: nothing to surface"
3074 );
3075 r.e2e = vec![CommandOutcome {
3076 command: "test".to_owned(),
3077 code: Some(0),
3078 output_tail: String::new(),
3079 duration_ms: 0,
3080 resource_blocked: false,
3081 }];
3082 assert!(
3083 r.verification_summary("h").is_none(),
3084 "a green result needs no skepticism attached to it"
3085 );
3086 }
3087
3088 #[test]
3089 fn verification_summary_tells_the_current_head_apart_from_an_earlier_one() {
3090 let mut r = round(false, Vec::new());
3091 r.head = "h1".to_owned();
3092 r.e2e = vec![CommandOutcome {
3093 command: "test".to_owned(),
3094 code: Some(1),
3095 output_tail: "boom".to_owned(),
3096 duration_ms: 0,
3097 resource_blocked: false,
3098 }];
3099 r.verified_head = Some("h1".to_owned());
3100 r.verified_at = Some(Timestamp::now());
3101
3102 let fresh = r.verification_summary("h1").expect("a failure is surfaced");
3103 assert!(
3104 fresh.label.contains("this is the head being looked at now"),
3105 "{}",
3106 fresh.label
3107 );
3108 assert_eq!(fresh.tail.as_deref(), Some("$ test\nboom\n"));
3109
3110 let stale = r.verification_summary("h2").expect("still surfaced");
3111 assert!(
3112 stale.label.contains("an earlier head, since superseded"),
3113 "a result about a different commit than the one being looked at now must say so, \
3114 not read as current: {}",
3115 stale.label
3116 );
3117 }
3118
3119 #[test]
3120 fn verification_summary_marks_a_resource_block_and_a_deferral_distinctly_from_a_failure() {
3121 let mut r = round(false, Vec::new());
3122 r.e2e = vec![CommandOutcome {
3123 command: "(waiting for the shared build cache)".to_owned(),
3124 code: None,
3125 output_tail: "contended".to_owned(),
3126 duration_ms: 0,
3127 resource_blocked: true,
3128 }];
3129 let blocked = r
3130 .verification_summary("h")
3131 .expect("a resource block is still surfaced, never silent");
3132 assert!(blocked.label.contains("could not run"));
3133 // No command actually ran, but which operation was attempted is
3134 // still a fact worth showing — never silent past the label either.
3135 let tail = blocked
3136 .tail
3137 .expect("the attempted operation is still named");
3138 assert!(tail.contains("(waiting for the shared build cache)"));
3139 assert!(tail.contains("contended"));
3140
3141 let mut d = round(false, Vec::new());
3142 d.e2e_deferred = true;
3143 d.e2e_defer_reason = Some("2 blocking finding(s) already required a fix".to_owned());
3144 let deferred = d.verification_summary("h").expect("deferred is surfaced");
3145 assert!(deferred.label.contains("deferred to the fixer"));
3146 assert!(deferred.label.contains("2 blocking finding(s)"));
3147 assert!(deferred.tail.is_none());
3148 }
3149
3150 #[test]
3151 fn verification_summary_says_unknown_rather_than_guessing_a_time_or_a_commit() {
3152 let mut r = round(false, Vec::new());
3153 r.e2e = vec![CommandOutcome {
3154 command: "test".to_owned(),
3155 code: Some(1),
3156 output_tail: "boom".to_owned(),
3157 duration_ms: 0,
3158 resource_blocked: false,
3159 }];
3160 // verified_head/verified_at left at their default `None` — exactly
3161 // the shape a schema-7 round with no reconstructable timestamp has.
3162 let summary = r.verification_summary("h").expect("a failure is surfaced");
3163 assert!(summary.label.contains("commit unknown"));
3164 assert!(summary.label.contains("checked at: unknown"));
3165 }
3166
3167 #[test]
3168 fn gate_status_tells_not_run_apart_from_passed_with_no_commands() {
3169 let mut s = state();
3170 assert_eq!(s.gate_status(), GateStatus::NotRun);
3171
3172 s.gate_ran = true;
3173 assert_eq!(
3174 s.gate_status(),
3175 GateStatus::PassedWithNoCommands,
3176 "an empty gate must read as a real pass once gate_ran says it actually ran"
3177 );
3178
3179 s.gate = vec![CommandOutcome {
3180 command: "cargo make check".to_owned(),
3181 code: Some(0),
3182 output_tail: String::new(),
3183 duration_ms: 0,
3184 resource_blocked: false,
3185 }];
3186 assert_eq!(s.gate_status(), GateStatus::Passed);
3187
3188 s.gate[0].code = Some(1);
3189 assert_eq!(s.gate_status(), GateStatus::Failed);
3190
3191 s.gate_ran = false;
3192 assert_eq!(
3193 s.gate_status(),
3194 GateStatus::NotRun,
3195 "gate_ran false must win even over a non-empty gate left from a stale record"
3196 );
3197 }
3198
3199 #[test]
3200 fn open_findings_is_empty_when_the_last_round_was_clean() {
3201 let mut s = state();
3202 s.reviews = vec![round(
3203 true,
3204 vec![finding("R1-1-1", crate::verdict::Severity::Minor)],
3205 )];
3206 assert!(s.open_findings().is_empty());
3207 }
3208
3209 #[test]
3210 fn open_findings_reads_the_last_non_clean_round() {
3211 let mut s = state();
3212 s.reviews = vec![round(
3213 false,
3214 vec![finding("R1-1-1", crate::verdict::Severity::Major)],
3215 )];
3216 let open = s.open_findings();
3217 assert_eq!(open.len(), 1);
3218 assert_eq!(open[0].id, "R1-1-1");
3219 }
3220
3221 #[test]
3222 fn handed_off_with_open_findings_needs_a_mergeable_status_and_an_open_round() {
3223 let mut s = state();
3224 s.reviews = vec![round(
3225 false,
3226 vec![finding("R1-1-1", crate::verdict::Severity::Major)],
3227 )];
3228
3229 s.status = RunStatus::Blocked;
3230 assert!(
3231 !s.handed_off_with_open_findings(),
3232 "a blocked run is not a hand-off"
3233 );
3234
3235 s.status = RunStatus::Ready;
3236 assert!(s.handed_off_with_open_findings());
3237
3238 s.reviews = vec![round(true, Vec::new())];
3239 assert!(
3240 !s.handed_off_with_open_findings(),
3241 "a clean last round has nothing to hand off"
3242 );
3243 }
3244
3245 #[test]
3246 fn unmerged_by_design_is_only_ready_reached_via_merge_mode_none() {
3247 let mut s = state();
3248
3249 s.status = RunStatus::Ready;
3250 assert!(
3251 !s.unmerged_by_design(),
3252 "no merge outcome recorded at all must not be flagged"
3253 );
3254
3255 s.merge = Some(MergeOutcome {
3256 mode: MergeMode::None,
3257 ok: true,
3258 detail: "git merge --no-ff magi/x/A".to_owned(),
3259 });
3260 assert!(
3261 s.unmerged_by_design(),
3262 "Ready reached through mode none is the case this exists to flag"
3263 );
3264
3265 // A PR closed without merging also leaves `status` at `Ready`, but
3266 // through `mode = "pr"` — a run that may still have been landable by
3267 // a person watching the PR, unlike the honest mode-none no-op.
3268 s.merge = Some(MergeOutcome {
3269 mode: MergeMode::Pr,
3270 ok: false,
3271 detail: "https://example.com/pr/1 was closed without merging".to_owned(),
3272 });
3273 assert!(
3274 !s.unmerged_by_design(),
3275 "a closed pull request is a different Ready and must not be relabelled"
3276 );
3277
3278 // Same signal must not fire before the run actually got there.
3279 s.status = RunStatus::Gating;
3280 s.merge = Some(MergeOutcome {
3281 mode: MergeMode::None,
3282 ok: true,
3283 detail: "git merge --no-ff magi/x/A".to_owned(),
3284 });
3285 assert!(
3286 !s.unmerged_by_design(),
3287 "status must actually be Ready, not merely have a stale mode-none merge record"
3288 );
3289 }
3290
3291 #[test]
3292 fn state_round_trips_through_json() {
3293 let s = state();
3294 let body = serde_json::to_string(&s).unwrap();
3295 let back: RunState = serde_json::from_str(&body).unwrap();
3296 assert_eq!(back.id, s.id);
3297 assert_eq!(back.instruction, "add retries");
3298 assert_eq!(back.status, RunStatus::Prep);
3299 }
3300
3301 #[test]
3302 fn a_round_recorded_before_e2e_deferral_existed_still_loads() {
3303 // Exactly the shape a pre-existing `run.json` has for a round: no
3304 // `e2e_deferred`, no `e2e_defer_reason`. Every round used to run e2e
3305 // unconditionally, so the honest reading of an old record's silence
3306 // on this is "it was not deferred" — `false`/`None`, not a load
3307 // failure and not a schema bump (see the `SCHEMA` doc comment: a
3308 // purely additive field whose absence has one unambiguous meaning
3309 // does not need one).
3310 let body = r#"{
3311 "round": 1,
3312 "head": "deadbeef",
3313 "reviews": [],
3314 "e2e": [],
3315 "verify_retried": false,
3316 "fix": null,
3317 "blocking": 0,
3318 "answered": 1,
3319 "expected": 1,
3320 "clean": true
3321 }"#;
3322 let r: ReviewRound = serde_json::from_str(body).expect("an old-shaped round must load");
3323 assert!(!r.e2e_deferred);
3324 assert!(r.e2e_defer_reason.is_none());
3325 assert_eq!(r.e2e_status(), E2eStatus::NotConfigured);
3326 }
3327
3328 #[test]
3329 fn schema_five_state_migrates_old_empty_e2e_and_legacy_verify_budget() {
3330 let mut value = serde_json::to_value(state()).expect("serialize state");
3331 let object = value.as_object_mut().expect("state object");
3332 object.insert("schema".to_owned(), serde_json::json!(5));
3333 let graph = object["config"]["graph"]
3334 .as_object_mut()
3335 .expect("graph object");
3336 graph.insert("timeout_review".to_owned(), serde_json::json!(3600));
3337 graph.remove("timeout_verify");
3338 let review = object["reviews"].as_array_mut().expect("reviews");
3339 review.push(serde_json::json!({
3340 "round": 1, "head": "old", "reviews": [], "e2e": [],
3341 "verify_retried": false, "blocking": 0, "answered": 1,
3342 "expected": 1, "clean": true
3343 }));
3344 let old: RunState = serde_json::from_value(value).expect("schema-5 shape parses");
3345 let migrated = migrate_schema(old).expect("schema 5 migrates");
3346 assert_eq!(migrated.schema, SCHEMA);
3347 assert_eq!(migrated.config.graph.verify_timeout(), 3600);
3348 assert_eq!(migrated.reviews[0].e2e_status(), E2eStatus::NotConfigured);
3349 }
3350
3351 #[test]
3352 fn schema_six_state_with_a_recorded_gate_migrates_to_gate_ran_true() {
3353 let mut value = serde_json::to_value(state()).expect("serialize state");
3354 let object = value.as_object_mut().expect("state object");
3355 object.insert("schema".to_owned(), serde_json::json!(6));
3356 object.insert(
3357 "gate".to_owned(),
3358 serde_json::json!([{
3359 "command": "cargo make check",
3360 "code": 0,
3361 "output_tail": "",
3362 "duration_ms": 0,
3363 "resource_blocked": false
3364 }]),
3365 );
3366 let old: RunState = serde_json::from_value(value).expect("schema-6 shape parses");
3367 let migrated = migrate_schema(old).expect("schema 6 migrates");
3368 assert_eq!(migrated.schema, SCHEMA);
3369 assert!(
3370 migrated.gate_ran,
3371 "a non-empty recorded gate is a real attempt, not an unrun one"
3372 );
3373 assert_eq!(migrated.gate_status(), GateStatus::Passed);
3374 }
3375
3376 #[test]
3377 fn schema_six_state_with_an_empty_gate_migrates_to_gate_ran_false_and_is_retried() {
3378 // The exact shape of the stuck `shoka` run this schema bump fixes:
3379 // `verify.gate` empty, `gate` empty, schema 6. It must come back as
3380 // "not yet run" so the next `gate()` call re-attempts it — and for a
3381 // repo with no gate commands configured, that resolves instantly to
3382 // `PassedWithNoCommands` instead of staying stuck forever.
3383 let mut value = serde_json::to_value(state()).expect("serialize state");
3384 let object = value.as_object_mut().expect("state object");
3385 object.insert("schema".to_owned(), serde_json::json!(6));
3386 object.insert("gate".to_owned(), serde_json::json!([]));
3387 let old: RunState = serde_json::from_value(value).expect("schema-6 shape parses");
3388 let migrated = migrate_schema(old).expect("schema 6 migrates");
3389 assert_eq!(migrated.schema, SCHEMA);
3390 assert!(
3391 !migrated.gate_ran,
3392 "an empty gate on schema 6 is ambiguous and must be treated as unrun"
3393 );
3394 assert_eq!(migrated.gate_status(), GateStatus::NotRun);
3395 }
3396
3397 #[test]
3398 fn schema_seven_state_reconstructs_verified_head_for_a_round_that_actually_ran_e2e() {
3399 // Schema 7's main review loop always checked `e2e` against the
3400 // round's own `head` — it just never wrote that into `verified_head`
3401 // unless a catch-up run had checked a *different* commit. Migrating
3402 // to schema 8 restores that always-true fact instead of leaving a
3403 // reader to assume it.
3404 let mut value = serde_json::to_value(state()).expect("serialize state");
3405 let object = value.as_object_mut().expect("state object");
3406 object.insert("schema".to_owned(), serde_json::json!(7));
3407 let reviews = object["reviews"].as_array_mut().expect("reviews");
3408 reviews.push(serde_json::json!({
3409 "round": 1, "head": "deadbeef", "reviews": [],
3410 "e2e": [{
3411 "command": "cargo test", "code": 0, "output_tail": "",
3412 "duration_ms": 0, "resource_blocked": false
3413 }],
3414 "verify_retried": false, "blocking": 0, "answered": 1,
3415 "expected": 1, "clean": true
3416 }));
3417 let old: RunState = serde_json::from_value(value).expect("schema-7 shape parses");
3418 let migrated = migrate_schema(old).expect("schema 7 migrates");
3419 assert_eq!(migrated.schema, SCHEMA);
3420 assert_eq!(
3421 migrated.reviews[0].verified_head.as_deref(),
3422 Some("deadbeef"),
3423 "a schema-7 round's main-loop e2e was always against its own head, even though the \
3424 field never said so"
3425 );
3426 assert!(
3427 migrated.reviews[0].verified_at.is_none(),
3428 "no historical timestamp exists to reconstruct; unknown stays unknown, not a \
3429 guessed 'now'"
3430 );
3431 }
3432
3433 #[test]
3434 fn schema_seven_state_leaves_a_deferred_round_with_no_verified_head() {
3435 let mut value = serde_json::to_value(state()).expect("serialize state");
3436 let object = value.as_object_mut().expect("state object");
3437 object.insert("schema".to_owned(), serde_json::json!(7));
3438 let reviews = object["reviews"].as_array_mut().expect("reviews");
3439 reviews.push(serde_json::json!({
3440 "round": 1, "head": "deadbeef", "reviews": [],
3441 "e2e": [], "e2e_deferred": true,
3442 "verify_retried": false, "blocking": 1, "answered": 1,
3443 "expected": 1, "clean": false
3444 }));
3445 let old: RunState = serde_json::from_value(value).expect("schema-7 shape parses");
3446 let migrated = migrate_schema(old).expect("schema 7 migrates");
3447 assert!(
3448 migrated.reviews[0].verified_head.is_none(),
3449 "a deferred round never ran e2e; there is nothing to reconstruct"
3450 );
3451 }
3452
3453 #[test]
3454 fn schema_six_serialization_is_rejected_by_a_schema_five_reader() {
3455 let body = serde_json::to_value(state()).expect("serialize state");
3456 assert_eq!(body["schema"], serde_json::json!(SCHEMA));
3457 assert_ne!(body["schema"], serde_json::json!(5));
3458 }
3459
3460 #[test]
3461 fn seat_started_and_finished_track_who_has_not_answered_yet() {
3462 let mut s = state();
3463 s.seat_started("judge", "judge-1", std::time::Duration::from_secs(60), 0);
3464 s.seat_started("judge", "judge-2", std::time::Duration::from_secs(60), 0);
3465 assert_eq!(s.active.len(), 2, "both seats are still out");
3466
3467 s.seat_finished("judge-1");
3468 assert_eq!(
3469 s.active.keys().collect::<Vec<_>>(),
3470 vec!["judge-2"],
3471 "only the seat that answered drops out; judge-2 is still waited on"
3472 );
3473 }
3474
3475 /// `seats_active` / `tasks_active` are the accessors report/web read
3476 /// instead of `active` directly, so neither ever counts the other kind of
3477 /// entry as a seat — a `verify.e2e` task must never inflate a quorum or
3478 /// seat count, and a seat must never show up in a task listing.
3479 #[test]
3480 fn seats_active_and_tasks_active_never_cross_over() {
3481 let mut s = state();
3482 s.seat_started("judge", "judge-1", std::time::Duration::from_secs(60), 0);
3483 s.task_command(
3484 "e2e",
3485 "verify",
3486 0,
3487 "cargo test",
3488 1,
3489 2,
3490 std::time::Duration::from_secs(600),
3491 );
3492
3493 assert_eq!(
3494 s.seats_active()
3495 .map(|(k, _)| k.as_str())
3496 .collect::<Vec<_>>(),
3497 vec!["judge-1"]
3498 );
3499 assert_eq!(
3500 s.tasks_active()
3501 .map(|(k, _)| k.as_str())
3502 .collect::<Vec<_>>(),
3503 vec!["e2e"]
3504 );
3505
3506 // A command boundary updates the same entry in place — still one
3507 // task, never a second one accumulating alongside it.
3508 s.task_command(
3509 "e2e",
3510 "verify",
3511 0,
3512 "cargo clippy",
3513 2,
3514 2,
3515 std::time::Duration::from_secs(600),
3516 );
3517 assert_eq!(s.tasks_active().count(), 1);
3518 assert_eq!(s.active["e2e"].command.as_deref(), Some("cargo clippy"));
3519
3520 s.task_finished("e2e");
3521 assert!(s.tasks_active().next().is_none());
3522 assert_eq!(
3523 s.seats_active()
3524 .map(|(k, _)| k.as_str())
3525 .collect::<Vec<_>>(),
3526 vec!["judge-1"],
3527 "clearing the task must not touch the seat entry"
3528 );
3529 }
3530
3531 #[test]
3532 fn a_retry_is_recorded_as_a_later_attempt_on_the_same_seat() {
3533 let mut s = state();
3534 s.seat_started("review", "review-2", std::time::Duration::from_secs(30), 0);
3535 s.seat_finished("review-2");
3536 // A nudge re-asks the same seat; attempt says this is not the first
3537 // time, which is the only trace a nudge otherwise leaves behind.
3538 s.seat_started("review", "review-2", std::time::Duration::from_secs(30), 1);
3539 assert_eq!(s.active["review-2"].attempt, 1);
3540 }
3541
3542 #[test]
3543 fn active_seat_reports_elapsed_and_remaining_time() {
3544 let now = Timestamp::now();
3545 let started = now - jiff::SignedDuration::from_secs(30);
3546 let seat = ActiveSeat {
3547 node: "judge".to_owned(),
3548 started_at: started,
3549 timeout_secs: 100,
3550 attempt: 0,
3551 task: None,
3552 command: None,
3553 index: None,
3554 total: None,
3555 };
3556 assert_eq!(seat.elapsed_secs(now), 30);
3557 assert_eq!(seat.remaining_secs(now), 70);
3558 }
3559
3560 #[test]
3561 fn remaining_time_never_goes_negative_past_the_timeout() {
3562 // `agy`'s own print-timeout occasionally overruns by a hair before the
3563 // kill lands; a naive subtraction would print a negative "time left".
3564 let now = Timestamp::now();
3565 let started = now - jiff::SignedDuration::from_secs(200);
3566 let seat = ActiveSeat {
3567 node: "implement".to_owned(),
3568 started_at: started,
3569 timeout_secs: 100,
3570 attempt: 1,
3571 task: None,
3572 command: None,
3573 index: None,
3574 total: None,
3575 };
3576 assert_eq!(seat.remaining_secs(now), 0);
3577 }
3578
3579 #[test]
3580 fn clear_active_drops_stale_seats_and_reports_whether_it_did() {
3581 let mut s = state();
3582 assert!(!s.clear_active(), "nothing to clear on a fresh run");
3583 s.seat_started(
3584 "implement",
3585 "impl-B",
3586 std::time::Duration::from_secs(3600),
3587 0,
3588 );
3589 assert!(s.clear_active(), "a leftover entry is reported as cleared");
3590 assert!(s.active.is_empty());
3591 }
3592
3593 #[test]
3594 fn active_seat_carries_nothing_that_could_be_read_as_output_bytes() {
3595 // `agy` prints exactly one JSON object, at the very end (see
3596 // `agent::dropped_stream`'s doc comment) — a seat can sit at zero
3597 // captured bytes for its whole timeout while working normally. So
3598 // `ActiveSeat` records only the wall-clock facts (when it started,
3599 // its budget, which attempt), never a byte count, which is what
3600 // keeps a reader from being able to build "0 bytes => dead" out of
3601 // it even by accident.
3602 let seat = ActiveSeat {
3603 node: "implement".to_owned(),
3604 started_at: Timestamp::now(),
3605 timeout_secs: 60,
3606 attempt: 0,
3607 task: None,
3608 command: None,
3609 index: None,
3610 total: None,
3611 };
3612 let value = serde_json::to_value(&seat).unwrap();
3613 let keys: std::collections::BTreeSet<String> =
3614 value.as_object().unwrap().keys().cloned().collect();
3615 assert_eq!(
3616 keys,
3617 std::collections::BTreeSet::from([
3618 "node".to_owned(),
3619 "started_at".to_owned(),
3620 "timeout_secs".to_owned(),
3621 "attempt".to_owned(),
3622 ]),
3623 "a byte count here would be a lever to declare a silent-but-healthy seat dead, and \
3624 the task-only fields must stay absent (not null) on an ordinary seat entry"
3625 );
3626 }
3627
3628 /// The same guarantee as
3629 /// [`active_seat_carries_nothing_that_could_be_read_as_output_bytes`],
3630 /// extended to a task entry: `verify.e2e` / `verify.gate` are exactly as
3631 /// silent as `agy` between commands, so a running command-list task must
3632 /// never carry anything a reader could mistake for output-byte evidence
3633 /// either.
3634 #[test]
3635 fn task_active_seat_carries_nothing_that_could_be_read_as_output_bytes() {
3636 let seat = ActiveSeat {
3637 node: "verify".to_owned(),
3638 started_at: Timestamp::now(),
3639 timeout_secs: 600,
3640 attempt: 0,
3641 task: Some("e2e".to_owned()),
3642 command: Some("cargo test".to_owned()),
3643 index: Some(1),
3644 total: Some(3),
3645 };
3646 let value = serde_json::to_value(&seat).unwrap();
3647 let keys: std::collections::BTreeSet<String> =
3648 value.as_object().unwrap().keys().cloned().collect();
3649 assert_eq!(
3650 keys,
3651 std::collections::BTreeSet::from([
3652 "node".to_owned(),
3653 "started_at".to_owned(),
3654 "timeout_secs".to_owned(),
3655 "attempt".to_owned(),
3656 "task".to_owned(),
3657 "command".to_owned(),
3658 "index".to_owned(),
3659 "total".to_owned(),
3660 ]),
3661 );
3662 }
3663
3664 #[test]
3665 fn an_old_run_json_without_active_seats_still_loads() {
3666 // Schema did not bump for this field: an already-written run.json
3667 // simply lacks the key, and `#[serde(default)]` must fill it in
3668 // rather than fail the whole read.
3669 let s = state();
3670 let mut value = serde_json::to_value(&s).unwrap();
3671 value.as_object_mut().unwrap().remove("active");
3672 let back: RunState = serde_json::from_value(value).unwrap();
3673 assert!(back.active.is_empty());
3674 assert_eq!(back.schema, SCHEMA);
3675 }
3676
3677 #[test]
3678 fn an_old_run_json_without_jobs_still_loads() {
3679 // No schema bump for this field either, for the same reason: an
3680 // empty `jobs` list on an old record means exactly what it always
3681 // meant for that record — no adapter existed yet to report one —
3682 // and `#[serde(default)]` fills it in rather than failing the read.
3683 let s = state();
3684 let mut value = serde_json::to_value(&s).unwrap();
3685 value.as_object_mut().unwrap().remove("jobs");
3686 let back: RunState = serde_json::from_value(value).unwrap();
3687 assert!(back.jobs.is_empty());
3688 assert_eq!(back.schema, SCHEMA);
3689 }
3690
3691 #[test]
3692 fn ensure_can_delete_guards_live_and_unfolded_runs() {
3693 let mut s = state();
3694 // 1. A daemon is working on it right now.
3695 s.status = RunStatus::Prep;
3696 let err = s.ensure_can_delete(true).unwrap_err().to_string();
3697 assert!(err.contains("live daemon"), "{err}");
3698
3699 // 2. The same unfinished run with no daemon behind it is a leftover
3700 // from a killed process, and deletable. Without this an interrupted
3701 // run could never be removed: its status stays `prep` forever.
3702 assert!(s.ensure_can_delete(false).is_ok());
3703
3704 // 3. Unfolded candidates are refused either way — that is the guard
3705 // that stops a delete from discarding a worktree.
3706 s.status = RunStatus::Merged;
3707 s.candidates.push(Candidate {
3708 index: 0,
3709 label: 'A',
3710 agent: "a".to_owned(),
3711 branch: "b".to_owned(),
3712 worktree: PathBuf::from("/w"),
3713 summary: String::new(),
3714 stat: String::new(),
3715 files: 1,
3716 commits: 1,
3717 empty: false,
3718 failed: None,
3719 verified_noop: None,
3720 duration_ms: 0,
3721 folded: false,
3722 });
3723 let err = s.ensure_can_delete(false).unwrap_err().to_string();
3724 assert!(
3725 err.contains("magi fold"),
3726 "error must suggest `magi fold`: {err}"
3727 );
3728
3729 // 4. Folded and nobody working on it.
3730 s.candidates[0].folded = true;
3731 assert!(s.ensure_can_delete(false).is_ok());
3732 }
3733}