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