magi/config.rs
1//! Run configuration: the agent roster, the shape of the graph, and the
2//! blindness / verification policy.
3//!
4//! Layers, lowest precedence first: `<config_dir>/magi/config.toml` (machine),
5//! `.magi/config.toml`, `magi.toml`. With none, built-in defaults apply, the
6//! agent roster derived from the agent CLIs actually installed on this
7//! machine. `--config <path>` replaces the whole stack with that one file.
8//!
9//! **The repository layers are read from `<remote>/<base>`, never from the
10//! working tree.** A primary checkout is often stale or on a detached HEAD, so
11//! a gate read from it can be older than the one the branch under review is
12//! cut from (run ce6e failed on a gate step `origin/main` had already
13//! dropped). `git show <sha>:magi.toml` of the remote-tracking ref's commit is
14//! written to a content-addressed directory under the git common dir
15//! (`<common>/magi/config/<sha>-<hash of repo>/`, temp dir when that is not
16//! writable) and loaded from there.
17//!
18//! `remote` and `base` are themselves in the file, so they are resolved in a
19//! fixed order, first hit wins:
20//!
21//! 1. `[merge] base` / `remote` of the machine layer, rendered as a real load
22//! would (env templates, includes);
23//! 2. the same keys, read minimally (only `[merge]`, nothing deserialized into
24//! `Config`) from the checkout's `magi.toml`, then `.magi/config.toml` -
25//! only to *locate* the ref; an unreadable file is ignored;
26//! 3. `remote` defaults to `origin`, `base` to what
27//! `refs/remotes/<remote>/HEAD` points at.
28//!
29//! [`Config::discover`] is synchronous and reads the remote-tracking ref as it
30//! is (`daemon::fetch_loop` keeps it fresh); [`Config::discover_fetched`] runs
31//! `git fetch` first and is what every path that starts work uses. A failed
32//! fetch, a missing ref or an unresolvable base is an **error**, never a
33//! silent fall back to the working tree or a local branch. Two cases are not
34//! that: an explicit `--config`, and a `repo` that is not a git working tree
35//! at all (no branch exists, so nothing can be stale).
36//!
37//! A repository layer that declares `include = [...]` gets the whole tree of
38//! that commit extracted, so relative includes resolve against the same
39//! commit. The directory is built beside its name and renamed in, so a
40//! concurrent `discover` never reads a half-written layer.
41use std::collections::BTreeMap;
42use std::path::{Path, PathBuf};
43
44use anyhow::{Context as _, Result, bail};
45use serde::{Deserialize, Serialize};
46
47/// Which CLI drives an agent.
48#[derive(Debug, Clone, Copy, PartialEq, Eq, Deserialize, Serialize)]
49#[serde(rename_all = "lowercase")]
50pub enum AgentKind {
51 /// Anthropic Claude Code (`claude -p`).
52 Claude,
53 /// opencode (`opencode run`).
54 Opencode,
55 /// Antigravity CLI (`agy -p`). Gemini CLI is deliberately absent: Google
56 /// retired the standalone client for individual accounts in favour of this
57 /// one, so an adapter for it would be dead code on a live machine.
58 Antigravity,
59 /// OpenAI Codex CLI (`codex exec`). The one roster member with a real
60 /// read-only mode: `--sandbox read-only` is enforced by the CLI, not by
61 /// the prompt.
62 Codex,
63 /// oh-my-pi (`omp -p --mode=json`). Another CLI with no read-only mode of
64 /// its own - `--auto-approve` gates reads and writes together - so a
65 /// read-only seat rests on the prompt and on the worktree discipline the
66 /// other non-codex seats already rely on. It is also the roster member that
67 /// reaches DeepSeek, whose models ship in `omp`'s own catalog.
68 Omp,
69 /// Arbitrary command. The escape hatch, and what the test suite drives.
70 Command,
71}
72
73impl AgentKind {
74 /// Executable that must be on `PATH` for this kind, if any.
75 pub fn program(self) -> Option<&'static str> {
76 match self {
77 Self::Claude => Some("claude"),
78 Self::Opencode => Some("opencode"),
79 Self::Antigravity => Some("agy"),
80 Self::Codex => Some("codex"),
81 Self::Omp => Some("omp"),
82 Self::Command => None,
83 }
84 }
85
86 /// Lowercase name as written in the config file.
87 pub fn as_str(self) -> &'static str {
88 match self {
89 Self::Claude => "claude",
90 Self::Opencode => "opencode",
91 Self::Antigravity => "antigravity",
92 Self::Codex => "codex",
93 Self::Omp => "omp",
94 Self::Command => "command",
95 }
96 }
97
98 /// Every kind the roster can name, in display order.
99 ///
100 /// This is what `magi doctor` lists, and it has to be one list rather than
101 /// the same set typed out again wherever a kind is enumerated. The last
102 /// time it was typed out twice, `omp` was added as a roster member and the
103 /// doctor output went on saying the machine had four CLIs - which reads as
104 /// "that agent is not installed" to the person the command exists for.
105 pub const ALL: [Self; 6] = [
106 Self::Claude,
107 Self::Opencode,
108 Self::Antigravity,
109 Self::Codex,
110 Self::Omp,
111 Self::Command,
112 ];
113}
114
115/// How the prompt reaches the agent process.
116#[derive(Debug, Clone, Copy, PartialEq, Eq, Deserialize, Serialize)]
117#[serde(rename_all = "lowercase")]
118pub enum Delivery {
119 /// Piped on stdin.
120 Stdin,
121 /// Passed as a positional argument. Beware OS command-line limits.
122 Argv,
123 /// Written to a file; the agent is told to read it. No length limit.
124 File,
125}
126
127/// One addressable agent in the roster.
128#[derive(Debug, Clone, Deserialize, Serialize)]
129#[serde(deny_unknown_fields)]
130pub struct AgentSpec {
131 /// Stable identifier used by `[roles]` and by the stats tables.
132 pub id: String,
133 /// Which CLI to drive.
134 pub kind: AgentKind,
135 /// Model passed through to the CLI (`--model` / `-m`). CLI default if unset.
136 #[serde(default)]
137 pub model: Option<String>,
138 /// `kind = "command"` only: argv. Supports `{prompt_file}`, `{cwd}`,
139 /// `{label}`, `{session}` placeholders.
140 #[serde(default)]
141 pub command: Vec<String>,
142 /// Extra arguments appended to the built command line.
143 #[serde(default)]
144 pub extra_args: Vec<String>,
145 /// Extra environment variables for the child process.
146 #[serde(default)]
147 pub env: BTreeMap<String, String>,
148 /// Override the per-kind prompt delivery default.
149 #[serde(default)]
150 pub prompt_delivery: Option<Delivery>,
151}
152
153impl AgentSpec {
154 /// Default prompt delivery for this agent.
155 ///
156 /// `opencode` and `agy` take the prompt as an argument, which on Windows
157 /// caps out around 32 KiB — well under a judging prompt carrying three
158 /// patches — so both get a file instead.
159 pub fn delivery(&self) -> Delivery {
160 self.prompt_delivery.unwrap_or(match self.kind {
161 AgentKind::Claude | AgentKind::Command => Delivery::Stdin,
162 // `codex exec -` reads the prompt from stdin, so the whole
163 // instruction arrives without an argv length limit and without a
164 // tool round-trip to open a file.
165 AgentKind::Codex => Delivery::Stdin,
166 // `omp -p` reads the prompt from stdin too, and a judging prompt
167 // carrying three patches is well past the Windows argv cap, so this
168 // is the only delivery that works for every node.
169 AgentKind::Omp => Delivery::Stdin,
170 AgentKind::Opencode | AgentKind::Antigravity => Delivery::File,
171 })
172 }
173
174 /// Human-facing label, e.g. `opus (claude:opus)`.
175 pub fn display(&self) -> String {
176 match &self.model {
177 Some(m) => format!("{} ({}:{m})", self.id, self.kind.as_str()),
178 None => format!("{} ({})", self.id, self.kind.as_str()),
179 }
180 }
181}
182
183/// A standalone seat's agent: one roster id, or an ordered fallback chain.
184///
185/// The string form is what every existing config uses and serializes back as
186/// a string; the array form tries each id at most once, in order, moving on
187/// when a call ends in quota or fails (see [`crate::agent::pick_chain`]).
188/// An empty string or array is the same as unset.
189#[derive(Debug, Clone, PartialEq, Eq, Deserialize, Serialize)]
190#[serde(untagged)]
191pub enum AgentChoice {
192 /// One agent id.
193 One(String),
194 /// Ordered fallback chain of agent ids.
195 Chain(Vec<String>),
196}
197
198impl AgentChoice {
199 /// The ids named, in order, with blanks dropped. Empty means unset.
200 pub fn ids(&self) -> Vec<&str> {
201 let all: Vec<&str> = match self {
202 Self::One(id) => vec![id.as_str()],
203 Self::Chain(ids) => ids.iter().map(String::as_str).collect(),
204 };
205 all.into_iter()
206 .map(str::trim)
207 .filter(|id| !id.is_empty())
208 .collect()
209 }
210
211 /// The first id named, if any.
212 pub fn primary(&self) -> Option<&str> {
213 self.ids().into_iter().next()
214 }
215}
216
217impl From<&str> for AgentChoice {
218 fn from(id: &str) -> Self {
219 Self::One(id.to_owned())
220 }
221}
222
223/// The first id of an optional choice: what a single-seat caller resolves.
224pub fn primary(choice: Option<&AgentChoice>) -> Option<&str> {
225 choice.and_then(AgentChoice::primary)
226}
227
228/// Explicit role assignment. Empty lists are filled in by
229/// [`Config::resolve_roles`] by rotating the roster.
230#[derive(Debug, Clone, Default, Deserialize, Serialize)]
231#[serde(deny_unknown_fields, default)]
232pub struct Roles {
233 /// Agents that implement the task, one worktree each.
234 pub implementers: Vec<String>,
235 /// Agents that rank the candidates blind.
236 pub judges: Vec<String>,
237 /// Agents that review the winning patch.
238 pub reviewers: Vec<String>,
239 /// Agent that applies review findings. Defaults to the winner's author.
240 ///
241 /// Accepts one id or an ordered array of ids: a fallback chain, see
242 /// [`AgentChoice`]. The review loop, the gate-fix round, `land`'s fix
243 /// rounds and the rebase conflict round move to the next id when one
244 /// errors, hits its quota or answers nothing usable; each id is tried at
245 /// most once per call, and once a fallback answered, the rest of the run
246 /// starts from it (see `crate::fixer`). Unset keeps the winner's author.
247 pub fixer: Option<AgentChoice>,
248 /// Agent that answers the standing chat's turns (`src/talk.rs`).
249 ///
250 /// Unset picks a `claude` seat, else the first runnable agent in roster
251 /// order (see [`crate::agent::pick`]) - which is roster *order*, not a
252 /// judgement about who converses well. Naming one here matters once an
253 /// agent is also sitting as a judge: the chat is opened far more often
254 /// than any single competition, and every open competes with that judge
255 /// seat for the same account's concurrency. A timeout on an ordinary chat
256 /// turn traced to exactly this - `opus` triple-booked as chatter and judge
257 /// - is what this field exists to let an operator break apart.
258 ///
259 /// Accepts one id or an ordered array of ids: a fallback chain, see
260 /// [`AgentChoice`].
261 pub chatter: Option<AgentChoice>,
262 /// Agent that arranges the queue between polls: `crate::conduct`'s single
263 /// seat, called once per cycle to decide a runnable task's `blocked_by`
264 /// and how a stalled or finished task recovers.
265 ///
266 /// The same precedent as [`Self::chatter`] for a seat that stands alone
267 /// rather than rotating through the roster - resolved through
268 /// [`crate::agent::pick`], so unset falls back to its own default order
269 /// (a claude seat, else the first runnable agent) rather than reusing a
270 /// judge or reviewer seat that the conductor's own poll-cycle cadence
271 /// would otherwise compete with for the same account's concurrency.
272 ///
273 /// Accepts a string or a fallback chain, like [`Self::chatter`].
274 pub conductor: Option<AgentChoice>,
275 /// Seats for the design-deliberation stage `graph::Runner::advise` runs
276 /// before `implement`: independent, read-only design proposals gathered
277 /// once a run has a settled task instruction and before any implementer
278 /// touches the repository.
279 ///
280 /// Empty falls back to `judges` rather than to the whole roster: a panel
281 /// trusted to rank patches independently is exactly the panel worth
282 /// asking to sketch a design independently, and an operator who has
283 /// already thought about judge diversity gets advisor diversity for free
284 /// instead of a fourth roster to maintain.
285 pub advisors: Vec<String>,
286 /// Agent that blends [`Self::advisors`]' proposals into the one design
287 /// brief `graph::Runner::synthesize_brief` carries into every
288 /// implementer's prompt.
289 ///
290 /// The same precedent as [`Self::chatter`] and [`Self::conductor`] for a
291 /// seat that stands alone rather than rotating through the roster -
292 /// resolved through [`crate::agent::pick`], so unset falls back to its
293 /// own default order (a claude seat, else the first runnable agent in
294 /// roster order) - exactly today's behavior, unchanged by leaving this
295 /// field out.
296 ///
297 /// Accepts a string or a fallback chain, like [`Self::chatter`].
298 pub synthesizer: Option<AgentChoice>,
299 /// Agent (normally a `kind = "command"` wrapper around Jev) that decides
300 /// whether a pull request title and prose are English, for
301 /// [`crate::github_text`]'s posting gate.
302 ///
303 /// Unset spawns nothing: the gate keeps its built-in heuristics. Set, a
304 /// failed, timed-out or unparsable call also falls back to them, so the
305 /// gate never depends on this seat being reachable. Accepts a string or a
306 /// fallback chain, like [`Self::chatter`].
307 pub language_judge: Option<AgentChoice>,
308}
309
310/// Graph shape and limits.
311#[derive(Debug, Clone, Deserialize, Serialize)]
312#[serde(deny_unknown_fields, default)]
313pub struct Graph {
314 /// Gate GitHub prose before posting. Secret redaction remains mandatory when disabled.
315 pub github_text_guard: bool,
316 /// Parallel implementations of the same task (the number of implementer
317 /// seats; pairs with [`Roles::implementers`]). **One by default.**
318 ///
319 /// Formerly `candidates`, still read as a deprecated alias.
320 ///
321 /// Competition is the thing magi is for, and it is still here - it is just
322 /// no longer what every task buys without being asked. Three days and 13
323 /// runs on this repository, which is the workload these numbers are drawn
324 /// from:
325 ///
326 /// - **0 of 13** competed runs reached a merge. Everything that landed in
327 /// that window went through `magi review` - the cheap half, no
328 /// competition - and passed on the first try.
329 /// - The judges' first choices **split 73% of the time** (8 of 11
330 /// tallies). Candidates that close together make the ranking a weak
331 /// signal for what it costs to produce.
332 /// - One run's own breakdown: implement 60min, judge 40min, fix 40min,
333 /// review 28min, **verify 5min** - and verify is the node that caught a
334 /// defect every reviewer had passed as clean. The cheapest step is the
335 /// one that earns its place every time.
336 ///
337 /// It is not worthless: `oc` won 3 of those tallies against `sonnet`, so a
338 /// single-seat default would have shipped the worse implementation in
339 /// roughly a quarter of them. That is exactly why this is a *default* and
340 /// not a removal - `magi run --candidates N` and a per-task seat count are
341 /// how a task that deserves a competition gets one.
342 ///
343 /// A single-candidate run needs no special case: `Runner::review`'s doc
344 /// records that `execute` already degrades to implement -> review -> gate
345 /// -> merge, because `judge` skips a one-candidate field, `deliberate` has
346 /// no two first choices to reconcile and `vote` returns early.
347 #[serde(alias = "candidates")]
348 pub implementers: usize,
349 /// Independent judges.
350 pub judges: usize,
351 /// Deliberation rounds when the judges' first choices disagree.
352 pub deliberate_rounds: usize,
353 /// Reviewers per review round. **Three by default** - the smallest panel
354 /// a lens cycle (see [`crate::prompt::Lens`]) covers exactly once, so the
355 /// default panel reads the patch for spec compliance, regressions, and
356 /// simplicity without repeating an angle. Review is also the one stage
357 /// [`Self::implementers`]'s doc describes as running on every task
358 /// regardless of competition, which is what makes a panel worth its cost
359 /// here even though `implementers` itself defaults to one.
360 pub reviewers: usize,
361 /// Maximum review+fix rounds before the run is declared blocked.
362 pub review_rounds: usize,
363 /// Fix rounds a failing `verify.gate` gets before the run is blocked. The
364 /// fixer is handed the gate's own output; `0` keeps the old behaviour of
365 /// blocking on the first red gate. Independent of [`Self::review_rounds`]
366 /// so a run that spent its review budget can still repair a one-line gate
367 /// failure.
368 pub gate_fix_rounds: usize,
369 /// Maximum agent processes running at once.
370 pub max_parallel: usize,
371 /// Language for the prose the agents write (`en` / `ja` / any language name).
372 pub language: String,
373 /// Keep one CLI conversation per seat, so a judge remembers its own
374 /// argument across deliberation rounds and the fixer remembers its own
375 /// implementation across review rounds.
376 ///
377 /// Sessions are scoped to a *seat*, never to an agent id: the same model
378 /// sitting as implementer and as judge gets two unrelated conversations,
379 /// which is what keeps blind judging blind.
380 pub sessions: bool,
381 /// Per-node timeouts, seconds.
382 pub timeout_implement: u64,
383 /// Per-node timeouts, seconds.
384 pub timeout_judge: u64,
385 /// Per-node timeouts, seconds.
386 pub timeout_review: u64,
387 /// Timeout for `verify.e2e` and `verify.gate`, seconds. Separate from
388 /// [`Self::timeout_review`] so shrinking a reviewer's budget cannot
389 /// silently shrink a real-machine command's budget too — the two used to
390 /// share `timeout_review`, and turning a slow reviewer down cut the
391 /// timeout `cargo test --all-targets` runs under along with it. When
392 /// omitted, preserves legacy configurations by using
393 /// [`Self::timeout_review`]. Set an explicit value to make verification
394 /// independent of later review-seat budget changes.
395 pub timeout_verify: Option<u64>,
396 /// Per-node timeouts, seconds.
397 pub timeout_fix: u64,
398 /// Wall-clock limit for one turn of [`crate::talk`]'s standing
399 /// conversation, seconds.
400 ///
401 /// An hour: the operator is not watching this turn resolve in real time,
402 /// so the budget can match what the work - reading files, running
403 /// commands, checking their output - actually needs rather than what a
404 /// person waiting on a phone can tolerate.
405 pub timeout_talk: u64,
406 /// Retries for an agent invocation that fails or returns nothing usable.
407 pub retries: usize,
408 /// Root for candidate / judge worktrees. Defaults to `~/wt/magi`.
409 pub worktree_root: Option<PathBuf>,
410 /// After the pull request is open, keep going: watch its checks and
411 /// reviews, run a fix round when they are unhappy, and ask to merge.
412 ///
413 /// On, because stopping at an open pull request left the operator doing
414 /// the watching by hand - six times in the session this was built in - and
415 /// that is the work the loop exists to take. It only engages for
416 /// `merge = "pr"`; every other merge mode ends the run as before.
417 ///
418 /// Turning this on does **not** hand magi the merge button:
419 /// [`Graph::land_approval`] is on too, and nothing merges without an
420 /// explicit answer. Setting both to their non-defaults is the only way to
421 /// get an unattended merge, and it has to be chosen twice.
422 pub land: bool,
423 /// Land rounds - watch, fix, push - before the run is left for a human.
424 pub land_rounds: usize,
425 /// Ask the owner before merging, showing what is about to land.
426 ///
427 /// On, and it is what makes `land` safe to have on: the question carries a
428 /// rendered panel - the diffstat, the patch, the checks, the review
429 /// comments that were addressed, and the subject the squash will use - so
430 /// the decision is made on evidence rather than on trust, from wherever
431 /// the operator happens to be.
432 ///
433 /// Silence is a hold. An unanswered approval never merges, and neither
434 /// does any answer other than the word `merge`.
435 pub land_approval: bool,
436 /// Ask the owner before merging a run whose review loop handed off over a
437 /// contested finding: the last round ended with a Major-or-above finding
438 /// still open *and* at least one reviewer's final vote (after
439 /// reconsideration) was reject. Findings raised in the last round are
440 /// never fixed or re-reviewed, so this is the one case an unattended
441 /// merge would land over a reviewer's explicit objection.
442 ///
443 /// On. Interaction with [`Self::land_approval`]: when that is on, every
444 /// merge is asked about and this key changes nothing. When it is off,
445 /// this key alone decides whether a contested hand-off is asked about
446 /// (the question says why); every other run still merges unattended.
447 /// Read at land time, so turning it off restores the unattended merge at
448 /// once. The pull request is opened either way; only the merge waits.
449 pub hold_contested_merge: bool,
450 /// After a run's pull request merges, file the findings its last review
451 /// round left open as follow-up tasks (`crate::followup`) instead of
452 /// letting them vanish with the merge: every Major-or-above finding, and
453 /// every finding of a seat whose final vote was reject. Other Minor/Nit
454 /// findings are only listed in the pull-request comment. Does not affect
455 /// whether the merge happens. On; best-effort, a failure is a run event.
456 pub file_followups: bool,
457 /// How long to wait for an owner to answer a question before the run is
458 /// abandoned, seconds. A parked run costs nothing, so this is generous;
459 /// it exists so a forgotten question cannot pin a worktree forever.
460 pub answer_timeout: u64,
461 /// What a round does when one or more reviewer seats never answered
462 /// (timeout, crash, unparsable output).
463 pub incomplete_review: IncompleteReviewPolicy,
464 /// Run `verify.e2e` on every round, even one that already has blocking
465 /// findings and another round left to try.
466 ///
467 /// Off by default: a round with a blocking finding and rounds still left
468 /// is going back to the fixer regardless of what `verify.e2e` says, so
469 /// running it first only spends the round's slowest step (minutes, on a
470 /// Rust repo's `cargo test --all-targets`) on a head about to be
471 /// rewritten anyway. `verify.e2e` still runs once a round has no
472 /// blocking findings left (a round cannot go `clean` without it) and the
473 /// final `verify.gate` always runs on the actual tree that would land —
474 /// deferring is about *when* e2e runs mid-loop, never about skipping it.
475 ///
476 /// Set this to restore the old every-round diagnostic behaviour: e2e
477 /// output from a round that still has blocking findings is occasionally
478 /// useful on its own (a runtime failure a reviewer's panel would not
479 /// have caught by reading), and this is the way back to seeing it every
480 /// round instead of only once the panel has nothing left to flag.
481 pub e2e_every_round: bool,
482 /// Run the design-deliberation stage before `implement`: independent
483 /// advisor seats each sketch a design, and (when at least one produced a
484 /// usable proposal) a synthesis blends them into a brief carried in the
485 /// implementer's prompt. See `graph::Runner::advise`.
486 ///
487 /// On by default. A design sketch is a few paragraphs an agent can write
488 /// without touching the repository, where a full implementation is a
489 /// tool loop that re-reads the codebase on every turn - so three
490 /// sketches, gathered once before `implement` starts, cost a fraction of
491 /// a fourth candidate and buy back a form of the same disagreement
492 /// [`Self::implementers`]'s doc describes moving away from being the
493 /// default, on every run rather than only the ones an operator remembers
494 /// to ask for with `--implementers`.
495 pub advise: bool,
496 /// How many independent design proposals the deliberation stage gathers.
497 /// **Three by default** - the same number [`Self::implementers`]'s doc
498 /// names as the point where a fourth judge's first choice stopped
499 /// changing the tally.
500 pub advisors: usize,
501}
502
503impl Default for Graph {
504 fn default() -> Self {
505 Self {
506 github_text_guard: true,
507 implementers: 1,
508 judges: 3,
509 deliberate_rounds: 1,
510 reviewers: 3,
511 review_rounds: 6,
512 gate_fix_rounds: 1,
513 max_parallel: 4,
514 language: "en".to_owned(),
515 sessions: true,
516 timeout_implement: 3600,
517 timeout_judge: 1200,
518 timeout_review: 1200,
519 timeout_verify: None,
520 timeout_fix: 1800,
521 timeout_talk: 3600,
522 retries: 1,
523 worktree_root: None,
524 land: true,
525 land_rounds: 4,
526 land_approval: true,
527 hold_contested_merge: true,
528 file_followups: true,
529 answer_timeout: 86_400,
530 incomplete_review: IncompleteReviewPolicy::Block,
531 e2e_every_round: false,
532 advise: true,
533 advisors: 3,
534 }
535 }
536}
537
538impl Graph {
539 /// Effective machine-command budget. Older configuration files had only
540 /// `timeout_review`, which also governed verification, so absence is a
541 /// compatibility fallback rather than a new 1200-second default.
542 pub fn verify_timeout(&self) -> u64 {
543 self.timeout_verify.unwrap_or(self.timeout_review)
544 }
545}
546
547/// What a review round does when a reviewer seat never answered.
548///
549/// A round where half the panel timed out is not evidence of a clean patch —
550/// it is evidence of nothing. The default refuses to call that clean; `warn`
551/// exists for an operator who would rather keep a flaky seat from stalling
552/// every run, and accepts that the gap is on them to read in the report.
553#[derive(Debug, Clone, Copy, PartialEq, Eq, Deserialize, Serialize)]
554#[serde(rename_all = "lowercase")]
555pub enum IncompleteReviewPolicy {
556 /// A round with a missing seat is never `clean`: with nothing raised to
557 /// fix, the round is re-reviewed instead of gating; with the max rounds
558 /// exhausted, the run is left `Blocked` rather than declared ready.
559 Block,
560 /// A round with a missing seat can still gate as clean, once every seat
561 /// that *did* answer raised nothing blocking and verification is green.
562 /// The record keeps the gap visible (`magi show`, `magi stats`) even
563 /// though the run does not wait on it.
564 Warn,
565}
566
567/// What to do when vendor-identifying text is found in material shown to judges.
568#[derive(Debug, Clone, Copy, PartialEq, Eq, Deserialize, Serialize)]
569#[serde(rename_all = "lowercase")]
570pub enum LeakPolicy {
571 /// Record the leak, show the patch unmodified.
572 Warn,
573 /// Replace the token with `[REDACTED]` in the presented patch.
574 Redact,
575 /// Abort the run.
576 Fail,
577}
578
579/// Blindness policy.
580///
581/// Commit messages and candidate summaries are *always* stripped of
582/// attribution trailers and redacted — that is where signatures actually
583/// appear. [`Blind::on_leak`] governs the patch body only, where blanket
584/// redaction would corrupt the artifact under judgement.
585#[derive(Debug, Clone, Deserialize, Serialize)]
586#[serde(deny_unknown_fields, default)]
587pub struct Blind {
588 /// Install a per-worktree `commit-msg` hook that deletes attribution
589 /// trailers before they can land in a candidate's history.
590 pub commit_msg_hook: bool,
591 /// Literal, case-insensitive substrings. A line containing any of them is
592 /// dropped from commit messages and summaries; the `commit-msg` hook is
593 /// generated from the same list.
594 pub strip_lines: Vec<String>,
595 /// Case-insensitive substrings that identify a vendor or model.
596 pub vendor_tokens: Vec<String>,
597 /// Policy for vendor tokens found in the patch body.
598 pub on_leak: LeakPolicy,
599 /// Seed for label assignment and per-judge presentation order. Derived from
600 /// the run id when unset; set it to make a run reproducible.
601 pub seed: Option<u64>,
602}
603
604impl Default for Blind {
605 fn default() -> Self {
606 Self {
607 commit_msg_hook: true,
608 strip_lines: [
609 "Co-Authored-By:",
610 "Signed-off-by:",
611 "Assisted-by:",
612 "Generated-by:",
613 "Generated with",
614 "\u{1f916}",
615 ]
616 .iter()
617 .map(|s| (*s).to_owned())
618 .collect(),
619 vendor_tokens: [
620 "claude",
621 "anthropic",
622 "codex",
623 "openai",
624 "chatgpt",
625 "gemini",
626 "grok",
627 "xai",
628 "copilot",
629 "opencode",
630 "qoder",
631 "cursor",
632 "\u{1f916}",
633 ]
634 .iter()
635 .map(|s| (*s).to_owned())
636 .collect(),
637 on_leak: LeakPolicy::Warn,
638 seed: None,
639 }
640 }
641}
642
643/// Shell commands that gate the winner.
644#[derive(Debug, Clone, Default, Deserialize, Serialize)]
645#[serde(deny_unknown_fields, default)]
646pub struct Verify {
647 /// Run in the winner's worktree once per review round. Its output is fed
648 /// back to the fixer. This is the "real machine" leg of the review.
649 pub e2e: Vec<String>,
650 /// Final gate. Must all exit 0 before a merge is attempted.
651 pub gate: Vec<String>,
652 /// Mechanical fixers (formatters) run in the winner's worktree just before
653 /// `gate`, after review is clean. Anything they change is committed as one
654 /// neutral-identity commit so the gate and the PR see it. A command that
655 /// fails or times out is a warning, never a failed run: the gate stays the
656 /// single arbiter. Empty (the default) turns the feature off. Runs under
657 /// `graph.verify_timeout()` per command, like the gate. Meant for light
658 /// tools: it is not scanned for `CARGO_TARGET_DIR`, and it should be
659 /// idempotent since a resumed run may repeat it.
660 #[serde(default)]
661 pub pre_gate: Vec<String>,
662 /// Shell used to run the commands above. Defaults to `sh -c`, or
663 /// `cmd /C` when `sh` is not on `PATH`.
664 pub shell: Option<Vec<String>>,
665}
666
667impl Verify {
668 /// The `CARGO_TARGET_DIR=` value of the first rendered command that sets
669 /// one, if any. See [`crate::disk::extract_cargo_target_dir`] for the shape
670 /// this reads back. One rendering is enough - they all set the same
671 /// rendered `{{ vars.cache }}` path via the same shell - and the first e2e
672 /// command is checked before the gate because the e2e rebuilds the crate.
673 pub fn cache_dir(&self) -> Option<PathBuf> {
674 self.e2e
675 .iter()
676 .chain(self.gate.iter())
677 .find_map(|cmd| crate::disk::extract_cargo_target_dir(cmd))
678 }
679}
680
681/// Per-repository worktree preparation, run right after magi creates a worktree.
682#[derive(Debug, Clone, Default, PartialEq, Deserialize, Serialize)]
683#[serde(deny_unknown_fields, default)]
684pub struct Worktree {
685 /// Ordered steps. Declare in one layer only. Empty (the default) changes
686 /// nothing. See [`crate::worktree_setup`].
687 pub setup: Vec<SetupStep>,
688}
689
690/// One setup step: exactly one of `copy` / `run`.
691#[derive(Debug, Clone, Default, PartialEq, Deserialize, Serialize)]
692#[serde(deny_unknown_fields, default)]
693pub struct SetupStep {
694 /// `"src -> dst"` (or just `"path"`): copy from the primary checkout (an
695 /// absolute source is allowed) into the worktree. Never overwrites.
696 pub copy: Option<String>,
697 /// With `copy`: a missing source is skipped instead of an error.
698 pub optional: bool,
699 /// Shell command, run with the worktree as cwd.
700 pub run: Option<String>,
701 /// With `run`: seconds before the command is killed. Defaults to
702 /// `graph.verify_timeout()`.
703 pub timeout_secs: Option<u64>,
704}
705
706impl Worktree {
707 /// Refuse a malformed step at load time rather than in the middle of a run.
708 pub fn validate(&self) -> Result<()> {
709 for (i, step) in self.setup.iter().enumerate() {
710 let n = i + 1;
711 match (&step.copy, &step.run) {
712 (Some(c), None) => {
713 if step.timeout_secs.is_some() {
714 bail!("[worktree] setup step {n}: `timeout_secs` applies to `run` only");
715 }
716 crate::worktree_setup::validate_copy(c)
717 .with_context(|| format!("[worktree] setup step {n}"))?;
718 }
719 (None, Some(r)) => {
720 if r.trim().is_empty() {
721 bail!("[worktree] setup step {n}: `run` is empty");
722 }
723 if step.optional {
724 bail!("[worktree] setup step {n}: `optional` applies to `copy` only");
725 }
726 }
727 _ => bail!("[worktree] setup step {n} must set exactly one of `copy` and `run`"),
728 }
729 }
730 Ok(())
731 }
732}
733
734/// Disk hygiene: how hard magi is allowed to press on the machine's free space.
735///
736/// The numbers below come from one incident, not from theory: a machine with
737/// 951.8 GB free ran a few competitions and plans and best read 6.7 GB free.
738/// Three multi-gigabyte classes of junk accumulated side by side - per-run
739/// worktrees that end as `Merged`/`Ready`/`Failed`, a shared build cache whose
740/// each verify round and each implementation wave recompiles the derived
741/// section of the project into, and the outputs of runs that were removed but
742/// whose folders nobody deleted.
743#[derive(Debug, Clone, PartialEq, Deserialize, Serialize)]
744#[serde(deny_unknown_fields, default)]
745pub struct Disk {
746 /// Free space, in bytes, below which no new run may start: the daemon and
747 /// the `magi run` gate answer with "the disk is full" instead of letting
748 /// the graph fill it the rest of the way. `0` turns the gate off.
749 ///
750 /// Default 8 GiB. The incident ran down to 6.7 GB free of 951.8 GB total
751 /// before anybody noticed; 8 GiB is enough headroom for the compile a fresh
752 /// competition triggers and small enough that a 1 TB disk with 100 GB free
753 /// is nowhere near the threshold.
754 pub min_free_bytes: u64,
755 /// Fold finished runs without being asked. `Merged`, `Ready` and `Failed`
756 /// runs older than [`fold_grace_secs`](Self::fold_grace_secs) have their
757 /// worktrees removed. `0` turns the janitor off.
758 ///
759 /// Default true.
760 pub auto_fold: bool,
761 /// How old a finished run must be before the janitor folds it, seconds.
762 ///
763 /// Default 6 hours. A run that `Ready` at 8am is the operator's answer; a
764 /// run that `Ready` a week ago is worktrees holding a compile each. Six
765 /// hours is long enough that nobody loses an answer in the gap between
766 /// reading the report and starting from it, and short enough that a backlog
767 /// cannot pile up across two nights.
768 pub fold_grace_secs: u64,
769 /// Ceiling for the shared build cache (`CARGO_TARGET_DIR` in the rendered
770 /// verify commands), in bytes. When the janitor runs and the cache is over
771 /// it, files are dropped oldest-first until it is not. `0` turns pruning
772 /// off - the cache then only ever grows, which is the operator's call.
773 ///
774 /// Default 10 GiB. This is what the incident measured: 30.61 GB sat in the
775 /// shared cache on top of ~16 GB in the primary target directory and 6.7-
776 /// 11.15 GB in each of four per-worktree targets. 10 GiB holds a healthy
777 /// stack of prebuilt dependencies (cargo's per-file fingerprinting means
778 /// pruning only costs the rebuild of the dropped files, not of the world)
779 /// without letting one addled cache swallow the machine.
780 pub cache_limit_bytes: u64,
781}
782
783impl Default for Disk {
784 fn default() -> Self {
785 Self {
786 min_free_bytes: 8 * 1024 * 1024 * 1024,
787 auto_fold: true,
788 fold_grace_secs: 6 * 60 * 60,
789 cache_limit_bytes: 10 * 1024 * 1024 * 1024,
790 }
791 }
792}
793
794/// What to do with the winning branch.
795#[derive(Debug, Clone, Copy, PartialEq, Eq, Deserialize, Serialize)]
796#[serde(rename_all = "lowercase")]
797pub enum MergeMode {
798 /// Leave the branch alone and print the merge command matching
799 /// [`Merge::style`].
800 None,
801 /// Merge into the base branch in the primary worktree, using
802 /// [`Merge::style`].
803 Local,
804 /// Push the branch and open a PR with `gh pr create`. Landing this PR
805 /// (`[graph] land`) always squashes — see `land`'s module doc — so
806 /// [`Merge::style`] does not apply here.
807 Pr,
808}
809
810/// How the winning branch is attached to the base branch: what
811/// `mode = "local"` runs, and what `mode = "none"`'s printed guidance tells
812/// the operator to run by hand.
813///
814/// Read from configuration rather than asked of the repository at run time
815/// (e.g. `gh api repos/{owner}/{repo}/rulesets`) for two reasons: it keeps
816/// `mode = "none"`'s guidance a pure function of `RunState`, assertable in a
817/// unit test the same way `land::decide` is kept pure (see that module's
818/// doc), and it works for a base branch that is not hosted on GitHub, or not
819/// reachable at all, at the moment the report is rendered. An operator whose
820/// base branch enforces a ruleset already knows what it allows; declaring it
821/// once here is cheaper than magi re-discovering it, with a network call, on
822/// every render.
823#[derive(Debug, Clone, Copy, PartialEq, Eq, Default, Deserialize, Serialize)]
824#[serde(rename_all = "lowercase")]
825pub enum MergeStyle {
826 /// `git merge --no-ff`: every candidate commit lands, plus a merge
827 /// commit that records the merge as its own event in history. Rejected
828 /// by a base branch whose ruleset requires linear history or forbids
829 /// merge commits outright.
830 #[default]
831 Merge,
832 /// `git merge --squash` followed by a commit under an explicit message:
833 /// every candidate commit folds into one, and none of the candidate's
834 /// own placeholder subjects (`magi: candidate A (uncommitted work)`)
835 /// reach the base branch. No merge commit, so this satisfies a linear-
836 /// history ruleset.
837 Squash,
838 /// A fast-forward-only merge: every candidate commit lands verbatim, in
839 /// order, with no merge commit. Only succeeds because the winner was
840 /// already rebased onto the tracked base tip before this runs (see
841 /// `Runner::sync_to_base`) — equivalent to GitHub's "rebase and merge"
842 /// once that has happened.
843 Rebase,
844}
845
846/// Merge policy.
847#[derive(Debug, Clone, Deserialize, Serialize)]
848#[serde(deny_unknown_fields, default)]
849pub struct Merge {
850 /// Default is [`MergeMode::None`]: magi never touches your base branch
851 /// unless you ask it to.
852 pub mode: MergeMode,
853 /// Base branch. Defaults to the branch checked out when the run started.
854 pub base: Option<String>,
855 /// How the winner is attached to `base`; see [`MergeStyle`]. Ignored by
856 /// `mode = "pr"`.
857 pub style: MergeStyle,
858 /// Remote for `mode = "pr"`.
859 pub remote: String,
860 /// After a `mode = "pr"` run lands, open a `chore/release-vX.Y.Z` pull
861 /// request sized to the change by an agent's own judgement, so a version
862 /// bump does not depend on a human remembering to cut one.
863 ///
864 /// Only acts on Rust repositories (a `Cargo.toml` at the root of the base
865 /// branch); on any other repository it is a no-op that records one event.
866 ///
867 /// On by default. **Turning this off means a merge landed from the phone
868 /// never becomes a release**, so `POST /api/upgrade` keeps reporting
869 /// "already on the newest release" against a `main` that has moved past
870 /// it - the same gap this feature exists to close. Only for a repository
871 /// that wants to keep cutting releases by hand.
872 pub release_bump: bool,
873}
874
875impl Default for Merge {
876 fn default() -> Self {
877 Self {
878 mode: MergeMode::None,
879 base: None,
880 style: MergeStyle::default(),
881 remote: "origin".to_owned(),
882 release_bump: true,
883 }
884 }
885}
886
887/// How magi keeps itself current.
888#[derive(Debug, Clone, Copy, PartialEq, Eq, Deserialize, Serialize)]
889#[serde(rename_all = "lowercase")]
890pub enum UpdateMode {
891 /// Never check.
892 Off,
893 /// Check in the background and print a one-line banner when a newer
894 /// release exists.
895 Notify,
896 /// Check and install silently.
897 Install,
898}
899
900/// Self-update policy.
901#[derive(Debug, Clone, Deserialize, Serialize)]
902#[serde(deny_unknown_fields, default)]
903pub struct Update {
904 /// Default is [`UpdateMode::Notify`]: magi tells you, and lets you decide.
905 pub mode: UpdateMode,
906 /// Minimum time between checks, e.g. `24h`. kaishin's default when unset.
907 pub interval: Option<String>,
908}
909
910impl Default for Update {
911 fn default() -> Self {
912 Self {
913 mode: UpdateMode::Notify,
914 interval: None,
915 }
916 }
917}
918
919/// Who performs the release after a `chore/release-vX.Y.Z` pull request merges.
920#[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Deserialize, Serialize)]
921#[serde(rename_all = "lowercase")]
922pub enum ReleaseMode {
923 /// The repository's own workflows do (`auto-tag.yml` then `release.yml`),
924 /// and the release pull request merges on its checks.
925 #[default]
926 Actions,
927 /// magi does, from this machine: no CI is awaited, the owner approves the
928 /// merge, and magi then tags and runs [`Release::commands`]. See
929 /// [`crate::release_local`].
930 Local,
931}
932
933/// `[release]`: releasing without GitHub Actions. Opt-in per repository.
934#[derive(Debug, Clone, Deserialize, Serialize)]
935#[serde(deny_unknown_fields, default)]
936pub struct Release {
937 /// [`ReleaseMode::Actions`] (default: nothing changes) or
938 /// [`ReleaseMode::Local`].
939 pub mode: ReleaseMode,
940 /// Shell commands run in order, from a clean checkout of the merge commit,
941 /// once the tag is pushed. The first failure stops the list. They see
942 /// `MAGI_RELEASE_VERSION`, `MAGI_RELEASE_TAG`, `MAGI_RELEASE_COMMIT` and
943 /// `MAGI_RELEASE_PR` in the environment; `{{ ... }}` cannot carry the
944 /// version because teravars renders the whole file before any release
945 /// exists. Only read when `mode = "local"`.
946 pub commands: Vec<String>,
947 /// Minutes each command may run before it is killed and counted as failed.
948 pub timeout_minutes: u64,
949}
950
951impl Default for Release {
952 fn default() -> Self {
953 Self {
954 mode: ReleaseMode::Actions,
955 commands: Vec::new(),
956 timeout_minutes: 30,
957 }
958 }
959}
960
961impl Release {
962 /// Is this repository released by magi rather than by its workflows?
963 pub fn is_local(&self) -> bool {
964 self.mode == ReleaseMode::Local
965 }
966}
967
968/// Top-level configuration.
969#[derive(Debug, Clone, Default, Deserialize, Serialize)]
970#[serde(deny_unknown_fields, default)]
971pub struct Config {
972 /// Agent roster.
973 pub agents: Vec<AgentSpec>,
974 /// Role assignment.
975 pub roles: Roles,
976 /// Graph shape.
977 pub graph: Graph,
978 /// Blindness policy.
979 pub blind: Blind,
980 /// Verification commands.
981 pub verify: Verify,
982 /// Worktree preparation.
983 pub worktree: Worktree,
984 /// Disk hygiene.
985 pub disk: Disk,
986 /// Merge policy.
987 pub merge: Merge,
988 /// Self-update policy.
989 pub update: Update,
990 /// Project-specific text appended to the node prompts.
991 pub prompts: Prompts,
992 /// How the operator is told a run is waiting on them.
993 pub notify: Notify,
994 /// Local repositories the plan surface can start or derive a conversation
995 /// against.
996 pub repos: Repos,
997 /// Policy for the standing conversation ([`crate::talk`]).
998 pub talk: Talk,
999 /// How many runs `magi serve`'s own loop drives at once.
1000 pub daemon: Daemon,
1001 /// Who tags and publishes after a release pull request merges.
1002 pub release: Release,
1003 /// Context-window sizes in tokens, keyed by model name, merged over
1004 /// [`BUILTIN_CONTEXT_WINDOWS`]. See [`Config::context_window`].
1005 pub context_windows: BTreeMap<String, u64>,
1006}
1007
1008/// Context windows of models magi commonly drives, in tokens.
1009///
1010/// A convenience, not a source of truth: vendors resize windows and ship new
1011/// models faster than this table is edited, so an entry that is stale or
1012/// missing is fixed by `[context_windows]` in `magi.toml`, which wins. Matched
1013/// by [`Config::context_window`]: exact name first, then the longest key that
1014/// prefixes the model name (so a dated or suffixed id such as
1015/// `claude-sonnet-4-5-20250929` finds `claude-sonnet-4-5`). A model matching
1016/// nothing has no percentage - the UI shows the raw count.
1017const BUILTIN_CONTEXT_WINDOWS: &[(&str, u64)] = &[
1018 ("claude-opus-4", 200_000),
1019 ("claude-sonnet-4", 200_000),
1020 ("claude-haiku-4", 200_000),
1021 ("claude-3", 200_000),
1022 ("opus", 200_000),
1023 ("sonnet", 200_000),
1024 ("haiku", 200_000),
1025 ("gpt-5", 400_000),
1026 ("gpt-4.1", 1_000_000),
1027 ("gpt-4o", 128_000),
1028 ("o3", 200_000),
1029 ("o4-mini", 200_000),
1030 ("gemini-2.5", 1_000_000),
1031 ("deepseek", 128_000),
1032];
1033
1034impl Config {
1035 /// Context window in tokens for `model`: `[context_windows]` first, then
1036 /// the built-in table; within each, an exact name beats the longest
1037 /// prefix. `None` for an unknown model, which is a deliberate answer -
1038 /// guessing a denominator would print a confident wrong percentage.
1039 pub fn context_window(&self, model: &str) -> Option<u64> {
1040 fn lookup<'a>(model: &str, entries: impl Iterator<Item = (&'a str, u64)>) -> Option<u64> {
1041 let mut best: Option<(usize, u64)> = None;
1042 for (key, window) in entries {
1043 if key == model {
1044 return Some(window);
1045 }
1046 if model.starts_with(key) && best.is_none_or(|(len, _)| key.len() > len) {
1047 best = Some((key.len(), window));
1048 }
1049 }
1050 best.map(|(_, w)| w)
1051 }
1052 lookup(
1053 model,
1054 self.context_windows.iter().map(|(k, w)| (k.as_str(), *w)),
1055 )
1056 .or_else(|| lookup(model, BUILTIN_CONTEXT_WINDOWS.iter().copied()))
1057 }
1058}
1059
1060/// How the daemon loop itself behaves, as opposed to what one run does.
1061///
1062/// Machine-layer material in the same sense [`Repos::roots`] is: how many
1063/// competitions this machine's own loop is willing to babysit at once is a
1064/// fact about the machine running `magi serve`, not about any one
1065/// repository's task, so it belongs in `<config_dir>/magi/config.toml`
1066/// rather than a repository's own `magi.toml` - though, like `Repos::roots`,
1067/// nothing stops a repository from setting it too, since a scalar field
1068/// takes whichever layer has the highest precedence.
1069#[derive(Debug, Clone, Deserialize, Serialize)]
1070#[serde(deny_unknown_fields, default)]
1071pub struct Daemon {
1072 /// How many runs `magi serve` may have actively in flight at once.
1073 /// **One by default** - today's behaviour, one run at a time.
1074 ///
1075 /// This is a different knob from [`Graph::max_parallel`], and the two
1076 /// must not be confused: `max_parallel` bounds how many agent *processes*
1077 /// one run starts inside itself (implementers, judges, reviewers -
1078 /// candidates competing on a single task); this field bounds how many
1079 /// *runs* - whole competitions, each with its own `max_parallel` budget -
1080 /// the loop drives side by side, possibly across different tasks and
1081 /// different repositories. Raising `max_parallel` buys a bigger panel for
1082 /// one task; raising this buys more tasks worked at once. A config file
1083 /// that meant one and wrote the other would either starve a competition
1084 /// of judges or leave the rest of the backlog waiting for no reason.
1085 ///
1086 /// A run parked waiting on the operator's land-merge approval - see
1087 /// [`crate::land`] - does not hold one of these slots while it waits: the
1088 /// whole point of parking there is to let the loop spend the slot on
1089 /// something runnable instead of sitting on a decision only a human can
1090 /// make. So even at the default of `1`, an approval that comes back does
1091 /// not queue behind whatever else the loop happens to be running.
1092 pub max_concurrent_runs: usize,
1093 /// Let a task marked [`crate::queue::Task::interrupt`] (`magi task
1094 /// interrupt`) cut ahead of whatever `magi serve` already has in flight,
1095 /// instead of waiting for it to finish.
1096 ///
1097 /// **Off by default.** When enabled, a run that is in flight and is not
1098 /// itself the interrupt candidate may be paused at its next safe node
1099 /// boundary - see [`crate::graph::Runner::park_here`] - so the marked
1100 /// task can run alone; the paused run resumes automatically, through the
1101 /// same path any other parked run does, the moment the interrupting
1102 /// task's own run reaches a terminal status. This is opt-in because it
1103 /// bends the loop's own "one run at a time" principle (see this repository's
1104 /// `magi serve` help) for a specific, explicit operator request, and a
1105 /// repository that never files an interrupt task pays nothing for having
1106 /// it on - but an operator who does not want any run of theirs preempted,
1107 /// ever, should leave this `false`.
1108 pub pause_for_interrupts: bool,
1109 /// How many deputies `magi serve` runs at once. A deputy is the short-lived
1110 /// seat that waits on one open conductor question (see [`crate::deputy`]);
1111 /// each is a live agent conversation, so this is bounded - **two by
1112 /// default**, because the agent CLIs' quota is the scarce resource. A
1113 /// question over the limit keeps its place and starts when a deputy
1114 /// finishes; `0` runs none.
1115 pub max_deputies: usize,
1116 /// Minutes a release-bump pull request may make no progress before
1117 /// `magi serve` raises a notice and asks the owner (see
1118 /// [`crate::release_watch`]). Progress is a change of head commit or of any
1119 /// check's verdict. A required check that stays red after the one rerun the
1120 /// watcher is allowed escalates at once, whatever this says. **Sixty by
1121 /// default**; `0` turns the watcher off.
1122 pub release_stall_minutes: u64,
1123}
1124
1125impl Default for Daemon {
1126 fn default() -> Self {
1127 Self {
1128 max_concurrent_runs: 1,
1129 pause_for_interrupts: false,
1130 max_deputies: 2,
1131 release_stall_minutes: 60,
1132 }
1133 }
1134}
1135
1136/// Where `magi repos` and `GET /api/repos` look for local checkouts.
1137///
1138/// `roots` is one of the array keys [`array_merge_policy`] marks as
1139/// append-across-layers: which checkouts exist in general is a *machine*
1140/// fact in the same way the agent roster is - a repository's own `magi.toml`
1141/// cannot state where its siblings live before magi has resolved which
1142/// repository to read that file from in the first place - but a repository
1143/// that genuinely has an extra root worth scanning is not forced to choose
1144/// between an error and losing the machine's roots outright. Both layers'
1145/// roots are scanned; see [`Config::refuse_split_arrays`] for the keys that
1146/// are still refused.
1147#[derive(Debug, Clone, Deserialize, Serialize)]
1148#[serde(deny_unknown_fields, default)]
1149pub struct Repos {
1150 /// Roots to scan for a ghq-layout checkout: `<root>/<host>/<owner>/<repo>`
1151 /// with a `.git` directory. Empty by default - nothing is scanned unless
1152 /// asked to be.
1153 pub roots: Vec<PathBuf>,
1154 /// How long a scan is trusted before the next request re-scans it,
1155 /// seconds. `0` means never trust it: scan on every request. Defaults to
1156 /// a day, the same order of magnitude as [`Graph::answer_timeout`] for
1157 /// the same reason - a checkout does not usually appear or vanish inside
1158 /// a session, so there is little to gain from scanning more often than
1159 /// that, and an explicit refresh exists for the moment one does.
1160 pub scan_ttl: u64,
1161 /// How often `magi serve` runs `git fetch origin` in every checkout under
1162 /// `roots`, seconds, so `origin/main` stays fresh whatever state the
1163 /// working tree is in. Fetch only: no branch, HEAD or working tree is
1164 /// touched. `0` disables it.
1165 pub fetch_interval: u64,
1166}
1167
1168impl Default for Repos {
1169 fn default() -> Self {
1170 Self {
1171 roots: Vec::new(),
1172 scan_ttl: 86_400,
1173 fetch_interval: 600,
1174 }
1175 }
1176}
1177
1178/// Project-specific text appended to each node's prompt.
1179///
1180/// **Additive by construction.** These fields cannot replace magi's prompts,
1181/// only extend them, and that restriction is the whole design. The built-in
1182/// prompts carry the invariants the competition rests on: a judging prompt
1183/// names no authors, every structured answer must arrive as one fenced `json`
1184/// block, and a judge is told not to speculate about who wrote what. A config
1185/// that could overwrite them would let a typo silently un-blind the panel or
1186/// break the parser, and the symptom would be "the judges got worse" rather
1187/// than an error.
1188///
1189/// Repository-wide context belongs in `AGENTS.md`, which every agent already
1190/// reads from the checkout. Use these fields for the things a *magi node*
1191/// needs to know and a repository file cannot say - for instance that
1192/// reviewers here should ignore formatting because a hook owns it.
1193#[derive(Debug, Clone, Default, Deserialize, Serialize)]
1194#[serde(deny_unknown_fields, default)]
1195pub struct Prompts {
1196 /// Appended to every node's prompt.
1197 pub all: String,
1198 /// Appended for implementers.
1199 pub implement: String,
1200 /// Appended for judges, both ranking and voting.
1201 pub judge: String,
1202 /// Appended for reviewers.
1203 pub review: String,
1204 /// Appended for the fixer.
1205 pub fix: String,
1206}
1207
1208impl Prompts {
1209 /// The overlay for one node, or `None` when nothing is configured.
1210 ///
1211 /// `node` is the graph's own node name, so a new node gets no overlay
1212 /// rather than the wrong one.
1213 pub fn overlay(&self, node: &str) -> Option<String> {
1214 let specific = match node {
1215 "implement" => &self.implement,
1216 "judge" | "vote" | "deliberate" => &self.judge,
1217 "review" => &self.review,
1218 "fix" => &self.fix,
1219 _ => "",
1220 };
1221 let mut parts: Vec<&str> = Vec::new();
1222 for p in [self.all.trim(), specific.trim()] {
1223 if !p.is_empty() {
1224 parts.push(p);
1225 }
1226 }
1227 if parts.is_empty() {
1228 return None;
1229 }
1230 Some(parts.join("\n\n"))
1231 }
1232}
1233
1234/// How the operator is told that a run is waiting on them.
1235///
1236/// A command rather than a built-in integration: magi is one binary with no
1237/// network dependencies, and every operator's notification path is different -
1238/// ntfy, a Slack webhook, a Windows toast, an SSH to a machine that beeps.
1239/// Shelling out keeps all of them possible and none of them magi's problem.
1240#[derive(Debug, Clone, Default, Deserialize, Serialize)]
1241#[serde(deny_unknown_fields, default)]
1242pub struct Notify {
1243 /// Command and arguments. `{summary}`, `{run}` and `{url}` are replaced.
1244 /// Empty means no notification - the web UI is then the only surface.
1245 pub command: Vec<String>,
1246}
1247
1248/// Policy for [`crate::talk`], the standing conversation.
1249#[derive(Debug, Clone, Default, Deserialize, Serialize)]
1250#[serde(deny_unknown_fields, default)]
1251pub struct Talk {
1252 /// Let the conversation's agent edit files in the repository instead of
1253 /// filing a task for one. **Off by default** - see
1254 /// [`crate::talk`]'s module doc for why an edit made mid-conversation is
1255 /// an edit no run and no review can be attributed to, which is exactly
1256 /// the property a repository entered in a competition depends on. A
1257 /// repository that is never judged - dotfiles, a personal config
1258 /// checkout - has nothing to lose by turning this on in its own
1259 /// `magi.toml`, and a one-line fix stops costing a queued task to get.
1260 pub allow_write: bool,
1261 /// Extra chat personas, `[[talk.personas]]`. They add to the built-ins in
1262 /// [`crate::persona`]; an entry with a built-in's id replaces it. Like every
1263 /// array outside [`array_merge_policy`]'s append list, declaring it in two
1264 /// layers is refused.
1265 pub personas: Vec<PersonaSpec>,
1266 /// What the chat calls the operator: the persona addresses them by this
1267 /// name and the web Chat labels their turns with it instead of `YOU`.
1268 /// Blank is unset; at most [`OPERATOR_NAME_MAX`] characters, one line.
1269 pub operator_name: Option<String>,
1270}
1271
1272/// Longest accepted `[talk] operator_name`, in characters.
1273pub const OPERATOR_NAME_MAX: usize = 64;
1274
1275impl Talk {
1276 /// The trimmed operator name, `None` when unset or blank.
1277 pub fn operator_name(&self) -> Option<&str> {
1278 self.operator_name
1279 .as_deref()
1280 .map(str::trim)
1281 .filter(|n| !n.is_empty())
1282 }
1283}
1284
1285/// One `[[talk.personas]]` entry: a tone for the standing chat, nothing more.
1286#[derive(Debug, Clone, Default, PartialEq, Eq, Deserialize, Serialize)]
1287#[serde(deny_unknown_fields)]
1288pub struct PersonaSpec {
1289 /// Stable id the conversation stores. Not `default`.
1290 pub id: String,
1291 /// What the selector shows.
1292 pub name: String,
1293 /// The voice instruction appended to the chat's system prompt.
1294 pub prompt: String,
1295}
1296
1297/// Roles resolved to concrete agent specs for one run.
1298#[derive(Debug, Clone)]
1299pub struct ResolvedRoles {
1300 /// One per candidate.
1301 pub implementers: Vec<AgentSpec>,
1302 /// One per judge.
1303 pub judges: Vec<AgentSpec>,
1304 /// One per reviewer slot.
1305 pub reviewers: Vec<AgentSpec>,
1306 /// The configured fixer chain in order, if `[roles] fixer` is set
1307 /// (resolved like [`crate::agent::pick_chain`]: unknown or uninstalled ids
1308 /// skipped, duplicates keep their first place). `None` means the winner's
1309 /// own author.
1310 pub fixer: Option<Vec<AgentSpec>>,
1311 /// Queue conductor, explicitly selected or resolved by the standalone-seat fallback.
1312 pub conductor: AgentSpec,
1313 /// The full ordered implementer roster, in [`Roles::implementers`]'s own
1314 /// order — or `[[agents]]` in file order, when that list is empty. Unlike
1315 /// [`Self::implementers`], never truncated to `graph.implementers` and
1316 /// never `rotate`d/wrapped: a solo run (`implementers = 1`) resolves
1317 /// `implementers` down to a single slot, but `graph::Runner`'s per-seat
1318 /// quota fallback needs the *whole* list to walk forward through when
1319 /// that one slot's agent runs out of quota mid-run.
1320 pub implementer_roster: Vec<AgentSpec>,
1321 /// [`Self::implementer_roster`]'s counterpart for judge seats: the whole
1322 /// `[roles] judges` list (or every agent), untruncated and unrotated, for
1323 /// the per-seat handover when a judge's agent fails.
1324 pub judge_roster: Vec<AgentSpec>,
1325 /// Likewise for reviewer seats.
1326 pub reviewer_roster: Vec<AgentSpec>,
1327}
1328
1329/// Every array-valued key in a config table, as a dotted path.
1330///
1331/// Dotted so the error names `roles.implementers` rather than `implementers`:
1332/// an operator with three config files needs to know which key, not just that
1333/// there was one. `vars` is skipped because it is teravars' own input, merged
1334/// on purpose and never deserialised into `Config`.
1335fn array_keys(table: &toml::value::Table, prefix: &str) -> Vec<String> {
1336 let mut out = Vec::new();
1337 for (k, v) in table {
1338 if prefix.is_empty() && k == "vars" {
1339 continue;
1340 }
1341 let path = if prefix.is_empty() {
1342 k.clone()
1343 } else {
1344 format!("{prefix}.{k}")
1345 };
1346 match v {
1347 toml::Value::Array(_) => out.push(path),
1348 toml::Value::Table(t) => out.extend(array_keys(t, &path)),
1349 _ => {}
1350 }
1351 }
1352 out
1353}
1354
1355/// How an array key behaves when two config layers both declare it.
1356#[derive(Debug, Clone, Copy, PartialEq, Eq)]
1357enum ArrayMerge {
1358 /// Two layers may both declare it; the composed value is the
1359 /// low-to-high-priority concatenation teravars already produces (see
1360 /// [`Config::load_layers`]'s doc for why that order and no dedup).
1361 Append,
1362 /// Two layers declaring it is refused; see
1363 /// [`Config::refuse_split_arrays`].
1364 Replace,
1365}
1366
1367/// The single place that decides, for a dotted array key (as returned by
1368/// [`array_keys`]), whether declaring it in two config layers is a
1369/// concatenation the operator asked for or a silent accident.
1370///
1371/// Kept as one match so the whole policy is visible in one place - the same
1372/// reason `claude_quota` and `dropped_stream` close their own classification
1373/// in one spot elsewhere in this codebase. Anything not listed defaults to
1374/// [`ArrayMerge::Replace`]: refusing is the safe default for a key nobody has
1375/// reasoned about yet, and a new array key added later has to be added here
1376/// deliberately to become appendable.
1377///
1378/// - `verify.e2e` / `verify.gate` — a "run all of these, all must exit 0"
1379/// gate. Concatenating two of them is exactly the checks both layers
1380/// wanted, which is what lets a common gate (e.g. `editorconfig-checker`)
1381/// live in a shared layer while a repository's own layer adds its own
1382/// command, instead of every repository copying the shared command into
1383/// its own file.
1384/// - `repos.roots` — a set of directories to scan for checkouts. A
1385/// repository adding its own root on top of the machine's is additive by
1386/// nature, not a replacement of where the machine looks; see
1387/// [`Repos::roots`].
1388///
1389/// Left on the refuse side, and why:
1390/// - `roles.implementers` / `roles.judges` / `roles.reviewers` — an ordered
1391/// list of *seats*, not a set. A machine's two implementers plus a
1392/// repository's one is three seats nobody asked for and nobody is paying
1393/// for on purpose.
1394/// - `notify.command` — an argv. Concatenating two argvs does not produce a
1395/// program that runs; it produces `["ntfy", "publish", "curl", "-X"]`.
1396/// - `blind.strip_lines` — technically safe to concatenate (each entry is
1397/// matched as an independent substring, so a longer list only strips
1398/// *more*), but left on the refuse side anyway: the same list also drives
1399/// `commit_msg_hook`'s generated `sed` addresses, where position matters,
1400/// and a silent three-layer merge is exactly the kind of surprise
1401/// `refuse_split_arrays` exists to catch rather than to reason about
1402/// case-by-case. A repository that wants one more stripped phrase restates
1403/// the whole list; that restatement is visible in review, an accidental
1404/// concatenation would not be.
1405fn array_merge_policy(key: &str) -> ArrayMerge {
1406 match key {
1407 "verify.e2e" | "verify.gate" | "verify.pre_gate" | "repos.roots" => ArrayMerge::Append,
1408 _ => ArrayMerge::Replace,
1409 }
1410}
1411
1412impl Config {
1413 /// Load one file through teravars: Tera rendering, `[vars]` resolution,
1414 /// and the `include = [...]` directive.
1415 pub fn load(path: &Path) -> Result<Self> {
1416 Self::load_layers(&[path.to_path_buf()])
1417 }
1418
1419 /// The Tera render context shared by every layer: `system.*` (from
1420 /// teravars), `env` (magi's own addition - a config that names a shared
1421 /// build-cache directory or a machine-specific path needs
1422 /// `{{ env.NAME | default(value='...') }}`), and `repo` / `repo_name`
1423 /// derived from the last (highest-priority) path's parent directory.
1424 ///
1425 /// Factored out so [`Config::array_provenance`] can re-render a single
1426 /// layer under the exact same context [`Config::load_layers`] uses for
1427 /// the joint render, rather than drifting from it by accident.
1428 fn render_ctx(paths: &[PathBuf]) -> teravars::Context {
1429 let mut ctx = teravars::system_context();
1430 let env: std::collections::BTreeMap<String, String> = std::env::vars().collect();
1431 ctx.insert("env", &env);
1432 if let Some(last) = paths.last()
1433 && let Some(dir) = last.parent()
1434 {
1435 // A layer extracted from `<remote>/<base>` records the checkout it
1436 // stands for beside itself; its own directory is a cache path.
1437 let dir = std::fs::read_to_string(dir.join(REPO_MARKER))
1438 .map(|s| PathBuf::from(s.trim()))
1439 .unwrap_or_else(|_| dir.to_path_buf());
1440 ctx.insert("repo", &dir.to_string_lossy());
1441 ctx.insert(
1442 "repo_name",
1443 &dir.file_name().unwrap_or_default().to_string_lossy(),
1444 );
1445 }
1446 ctx
1447 }
1448
1449 /// Load and deep-merge a stack of config files, later files winning.
1450 ///
1451 /// This is why the config is TOML-through-teravars rather than plain serde:
1452 /// the roster is a *machine* fact (which CLIs and plans you pay for) while
1453 /// the gate is a *repository* fact (`cargo make check` here, `pnpm test`
1454 /// there). Picking one file and ignoring the other would force every repo
1455 /// to restate the roster.
1456 pub fn load_layers(paths: &[PathBuf]) -> Result<Self> {
1457 let mut engine = teravars::Engine::default();
1458 let ctx = Self::render_ctx(paths);
1459 if paths.len() > 1 {
1460 Self::refuse_split_arrays(paths, &mut engine, &ctx)?;
1461 }
1462 let merged = teravars::load_merged(paths, &mut engine, &ctx).with_context(|| {
1463 format!(
1464 "rendering config via teravars: {}",
1465 paths
1466 .iter()
1467 .map(|p| p.display().to_string())
1468 .collect::<Vec<_>>()
1469 .join(", ")
1470 )
1471 })?;
1472 let mut table = merged.config;
1473 Self::normalize_renamed_keys(paths, &mut engine, &ctx, &mut table)?;
1474 // `[vars]` is teravars' own input, already resolved into the render
1475 // context; `deny_unknown_fields` must not trip over it.
1476 table.remove("vars");
1477 // No layer says anything about `agents`: the roster is whatever is on
1478 // `PATH` now, exactly as when no layer exists at all. An explicit
1479 // `agents = []` is a statement and is kept.
1480 let detect_agents = !table.contains_key("agents");
1481 let mut cfg: Self = toml::Value::Table(table)
1482 .try_into()
1483 .context("deserializing magi config")?;
1484 crate::persona::validate(&cfg.talk.personas)?;
1485 cfg.worktree.validate()?;
1486 if let Some(name) = cfg.talk.operator_name() {
1487 if name.chars().count() > OPERATOR_NAME_MAX || name.chars().any(char::is_control) {
1488 bail!(
1489 "[talk] operator_name must be one line of at most {OPERATOR_NAME_MAX} characters"
1490 );
1491 }
1492 }
1493 if detect_agents {
1494 cfg.agents = Self::autodetected().agents;
1495 }
1496 Ok(cfg)
1497 }
1498
1499 /// `graph.candidates` was renamed `graph.implementers`. Both spellings are
1500 /// read, but the layers are merged into one table before serde sees it, so
1501 /// an old name in a lower layer and the new name in a higher one would
1502 /// arrive together and fail as a duplicate field instead of letting the
1503 /// higher layer win. Resolve it here: a single layer that writes both is
1504 /// refused (naming both keys and the file); otherwise the value from the
1505 /// last layer that wrote either key becomes `implementers`.
1506 fn normalize_renamed_keys(
1507 paths: &[PathBuf],
1508 engine: &mut teravars::Engine,
1509 ctx: &teravars::Context,
1510 table: &mut toml::map::Map<String, toml::Value>,
1511 ) -> Result<()> {
1512 let both = matches!(
1513 table.get("graph"),
1514 Some(toml::Value::Table(g)) if g.contains_key("candidates") && g.contains_key("implementers")
1515 );
1516 if !both {
1517 return Ok(());
1518 }
1519 let mut winner: Option<toml::Value> = None;
1520 for path in paths {
1521 let one = teravars::load_merged([path], engine, ctx)
1522 .with_context(|| format!("rendering {}", path.display()))?;
1523 let Some(toml::Value::Table(g)) = one.config.get("graph") else {
1524 continue;
1525 };
1526 if g.contains_key("candidates") && g.contains_key("implementers") {
1527 bail!(
1528 "`[graph]` sets both `implementers` and its deprecated alias \n `candidates` in {}; keep only `implementers`.",
1529 path.display()
1530 );
1531 }
1532 if let Some(v) = g.get("implementers").or_else(|| g.get("candidates")) {
1533 winner = Some(v.clone());
1534 }
1535 }
1536 if let Some(toml::Value::Table(g)) = table.get_mut("graph") {
1537 g.remove("candidates");
1538 if let Some(v) = winner {
1539 g.insert("implementers".to_owned(), v);
1540 }
1541 }
1542 Ok(())
1543 }
1544
1545 /// Refuse an array that two layers both declare, unless
1546 /// [`array_merge_policy`] says that key is meant to accumulate.
1547 ///
1548 /// teravars **appends** arrays when it merges layers, and that is wrong for
1549 /// most arrays magi has: `implementers` is an ordered list of seats,
1550 /// `notify.command` is an argv. Concatenating two of them yields something
1551 /// nobody wrote - three implementers out of a machine's two and a
1552 /// repository's one, or an argv of `["ntfy", "publish", "curl", "-X"]`.
1553 ///
1554 /// Replacing instead would be the right merge rule for those keys, but the
1555 /// rule lives in teravars, which several other projects depend on;
1556 /// changing it there is a decision for that crate, not something to fake
1557 /// here by re-reading the files with different semantics and hoping the
1558 /// two paths agree.
1559 ///
1560 /// So magi refuses the ambiguity rather than resolving it silently, for
1561 /// every array key except the short, deliberate list
1562 /// [`array_merge_policy`] marks [`ArrayMerge::Append`] - for those, the
1563 /// concatenation teravars already produces *is* what both files say, so
1564 /// there is nothing to refuse. The cost of guessing wrong on the refused
1565 /// keys is a roster the operator did not ask for and is paying for by the
1566 /// token; the append keys carry no such risk because every element runs
1567 /// (or every directory is scanned) regardless of order.
1568 fn refuse_split_arrays(
1569 paths: &[PathBuf],
1570 engine: &mut teravars::Engine,
1571 ctx: &teravars::Context,
1572 ) -> Result<()> {
1573 let mut seen: std::collections::BTreeMap<String, PathBuf> = Default::default();
1574 for path in paths {
1575 let one = teravars::load_merged([path], engine, ctx)
1576 .with_context(|| format!("rendering {}", path.display()))?;
1577 for key in array_keys(&one.config, "") {
1578 if array_merge_policy(&key) == ArrayMerge::Append {
1579 continue;
1580 }
1581 if let Some(first) = seen.get(&key) {
1582 bail!(
1583 "`{key}` is an array declared in two config layers:\n \
1584 {}\n {}\nteravars appends arrays when it merges, so \
1585 magi would run the concatenation of both - which is \
1586 not what either file says. Declare `{key}` in exactly \
1587 one of them.",
1588 first.display(),
1589 path.display()
1590 );
1591 }
1592 seen.insert(key, path.clone());
1593 }
1594 }
1595 Ok(())
1596 }
1597
1598 /// Which layers contributed to a composed, appendable array key (e.g.
1599 /// `"verify.gate"`), in the same low-to-high-priority order
1600 /// [`Config::load_layers`] concatenates them in. Layers that do not
1601 /// declare `key` at all are omitted.
1602 ///
1603 /// This is a **display aid for `magi doctor` only.** The command list
1604 /// that actually runs always comes from the one joint
1605 /// [`teravars::load_merged`] call in `load_layers`, never from this
1606 /// function - the exact hazard [`Config::refuse_split_arrays`] warns
1607 /// about is two merge paths that might disagree, so this function must
1608 /// never become a second source of the *composed* value, only of which
1609 /// file wrote which line in it.
1610 ///
1611 /// Re-rendering each layer alone can, in principle, resolve a
1612 /// `{{ vars.x }}` differently than the joint render would, if `x` is
1613 /// defined in one layer and referenced in another - the same caveat
1614 /// `refuse_split_arrays`'s structural, key-only check already lives with.
1615 /// None of magi's own gate commands cross that line, and a doctor listing
1616 /// is read by a human who can compare it against the joint one printed
1617 /// alongside it, so this is judged worth the simplicity of not
1618 /// threading provenance through the real load path.
1619 pub fn array_provenance(paths: &[PathBuf], key: &str) -> Vec<(PathBuf, Vec<String>)> {
1620 let mut engine = teravars::Engine::default();
1621 let ctx = Self::render_ctx(paths);
1622 let mut out = Vec::new();
1623 for path in paths {
1624 let Ok(one) = teravars::load_merged([path], &mut engine, &ctx) else {
1625 continue;
1626 };
1627 let mut cur = &one.config;
1628 let mut found = None;
1629 let parts: Vec<&str> = key.split('.').collect();
1630 for (i, part) in parts.iter().enumerate() {
1631 match cur.get(*part) {
1632 Some(toml::Value::Array(a)) if i == parts.len() - 1 => {
1633 found = Some(a);
1634 break;
1635 }
1636 Some(toml::Value::Table(t)) => cur = t,
1637 _ => break,
1638 }
1639 }
1640 let Some(values) = found else { continue };
1641 let strings: Vec<String> = values
1642 .iter()
1643 .filter_map(|v| v.as_str().map(str::to_owned))
1644 .collect();
1645 if !strings.is_empty() {
1646 out.push((path.clone(), strings));
1647 }
1648 }
1649 out
1650 }
1651
1652 /// Each layer rendered on its own, with the outcome per file: the
1653 /// settings screen's way of saying which layer a value came from and which
1654 /// layer is the one that does not parse. A **display aid only**, for the
1655 /// same reason as [`Config::array_provenance`]: the effective config
1656 /// always comes from the one joint [`Config::load_layers`].
1657 pub(crate) fn layer_tables(paths: &[PathBuf]) -> Vec<(PathBuf, Result<toml::Table>)> {
1658 let mut engine = teravars::Engine::default();
1659 let ctx = Self::render_ctx(paths);
1660 paths
1661 .iter()
1662 .map(|path| {
1663 let one = teravars::load_merged([path], &mut engine, &ctx)
1664 .map(|m| m.config)
1665 .with_context(|| format!("rendering {}", path.display()));
1666 (path.clone(), one)
1667 })
1668 .collect()
1669 }
1670
1671 /// Render a composed command list for `magi doctor`: the joined command
1672 /// line the run actually uses, plus - only when more than one layer
1673 /// contributed - which layer wrote which line.
1674 ///
1675 /// A single contributing layer (the common case today) stays the plain
1676 /// one-line summary magi has always printed, `empty` included: that
1677 /// honest "(none — ...)" is what caught a real gate-composition gap
1678 /// before this array could compose at all, and composition should not
1679 /// make the common case noisier.
1680 pub fn describe_composed(
1681 paths: &[PathBuf],
1682 commands: &[String],
1683 key: &str,
1684 empty: &str,
1685 ) -> String {
1686 if commands.is_empty() {
1687 return empty.to_owned();
1688 }
1689 let joined = commands.join(" && ");
1690 let provenance = Self::array_provenance(paths, key);
1691 if provenance.len() <= 1 {
1692 return joined;
1693 }
1694 let mut out = joined;
1695 for (path, cmds) in &provenance {
1696 out.push_str(&format!("\n [{}] {}", path.display(), cmds.join(" && ")));
1697 }
1698 out
1699 }
1700
1701 /// Resolve the config for `repo`, honouring an explicit `--config` path.
1702 ///
1703 /// Returns the config and the layers it came from, empty for built-in
1704 /// defaults. The repository layers come from the remote-tracking ref as it
1705 /// stands (see the module doc); use [`Config::discover_fetched`] where a
1706 /// stale ref would matter.
1707 pub fn discover(repo: &Path, explicit: Option<&Path>) -> Result<(Self, Vec<PathBuf>)> {
1708 if let Some(p) = explicit {
1709 let paths = vec![p.to_path_buf()];
1710 return Ok((Self::load_layers(&paths)?, paths));
1711 }
1712 let paths = Self::layers_checked(repo)?;
1713 if paths.is_empty() {
1714 return Ok((Self::autodetected(), paths));
1715 }
1716 Ok((Self::load_layers(&paths)?, paths))
1717 }
1718
1719 /// [`Config::discover`] after `git fetch <remote> <base>`: for every path
1720 /// that is about to start work. A failed fetch is an error.
1721 pub async fn discover_fetched(
1722 repo: &Path,
1723 explicit: Option<&Path>,
1724 ) -> Result<(Self, Vec<PathBuf>)> {
1725 if explicit.is_none() && is_git_worktree(repo) {
1726 // A checkout that was never cloned from the remote has no
1727 // `<remote>/HEAD`; learn it here (refs/remotes only) so an
1728 // ordinary start recovers instead of failing in `discover`.
1729 let (remote, cfg_base) = probe_remote_base(repo);
1730 let base = crate::git::merge_base_branch(repo, &remote, cfg_base.as_deref()).await?;
1731 let out = crate::git::fetch(repo, &remote, &base)
1732 .await
1733 .with_context(|| format!("fetching {remote}/{base} to read magi.toml"))?;
1734 if !out.ok() {
1735 bail!(
1736 "cannot read magi.toml: `git fetch {remote} {base}` failed ({}); refusing to use the local checkout's copy, which may be stale",
1737 out.stderr.lines().next().unwrap_or("").trim()
1738 );
1739 }
1740 if cfg_base.is_none() {
1741 crate::git::ensure_remote_head(repo, &remote, &base).await;
1742 }
1743 }
1744 Self::discover(repo, explicit)
1745 }
1746
1747 /// [`Config::layers`], but an unreadable remote ref is an error.
1748 fn layers_checked(repo: &Path) -> Result<Vec<PathBuf>> {
1749 if !is_git_worktree(repo) {
1750 return Ok(Self::layers(repo));
1751 }
1752 let mut paths: Vec<PathBuf> = Self::machine_layer()
1753 .into_iter()
1754 .filter(|p| p.is_file())
1755 .collect();
1756 paths.extend(Self::repo_layers(repo)?);
1757 Ok(paths)
1758 }
1759
1760 /// Only the repository layers (lowest first), from `<remote>/<base>` in a
1761 /// git working tree and from disk elsewhere.
1762 pub(crate) fn repo_layers(repo: &Path) -> Result<Vec<PathBuf>> {
1763 if !is_git_worktree(repo) {
1764 return Ok([
1765 repo.join(".magi").join("config.toml"),
1766 repo.join("magi.toml"),
1767 ]
1768 .into_iter()
1769 .filter(|p| p.is_file())
1770 .collect());
1771 }
1772 remote_layers(repo)
1773 }
1774
1775 /// Environment variable that relocates the machine-wide config layer.
1776 ///
1777 /// Set it to a directory and magi reads `<dir>/magi/config.toml` instead
1778 /// of the one under [`dirs::config_dir`]; set it to the empty string and
1779 /// magi reads no machine layer at all.
1780 ///
1781 /// This exists because the machine layer is otherwise unavoidable, and a
1782 /// test that builds a config fixture is not asking for the operator's
1783 /// preferences to be merged into it. Adding `[repos] roots` to the real
1784 /// machine config on a development box turned two passing tests red -
1785 /// `repos_list_returns_name_and_path_for_every_configured_root` and
1786 /// `repos_list_only_rescans_within_the_ttl_when_asked_to`, whose fixtures
1787 /// declare `[repos] roots` of their own, which [`Config::layers`] then
1788 /// found in two layers and [`Config::refuse_split_arrays`] correctly
1789 /// refused. CI never saw it: a runner has no machine config, so the suite
1790 /// was green there and red only where somebody actually uses magi.
1791 ///
1792 /// An operator gets the same escape hatch for free: a second machine
1793 /// config, or none, without moving files about.
1794 pub const CONFIG_DIR_ENV: &str = "MAGI_CONFIG_DIR";
1795
1796 /// Every config layer that applies to `repo`, in increasing precedence.
1797 ///
1798 /// The machine layer is whatever [`Config::machine_layer`] resolves to,
1799 /// which is nothing at all in a test build.
1800 pub fn layers(repo: &Path) -> Vec<PathBuf> {
1801 let mut paths = Vec::new();
1802 paths.extend(Self::machine_layer());
1803 paths.push(repo.join(".magi").join("config.toml"));
1804 paths.push(repo.join("magi.toml"));
1805 paths.retain(|p| p.is_file());
1806 paths
1807 }
1808
1809 /// The machine-wide layer's path, when there is one.
1810 ///
1811 /// **A test build has none unless it names one.** A fixture is a complete
1812 /// statement of the config under test, and the operator's own preferences
1813 /// have no business being merged into it - least of all silently, on one
1814 /// machine, in a suite that is green everywhere else.
1815 #[cfg(test)]
1816 pub(crate) fn machine_layer() -> Option<PathBuf> {
1817 std::env::var(Self::CONFIG_DIR_ENV)
1818 .ok()
1819 .filter(|dir| !dir.trim().is_empty())
1820 .map(|dir| PathBuf::from(dir).join("magi").join("config.toml"))
1821 }
1822
1823 /// The machine-wide layer's path, when there is one.
1824 #[cfg(not(test))]
1825 pub(crate) fn machine_layer() -> Option<PathBuf> {
1826 match std::env::var(Self::CONFIG_DIR_ENV) {
1827 // Named, and empty on purpose: no machine layer.
1828 Ok(dir) if dir.trim().is_empty() => None,
1829 Ok(dir) => Some(PathBuf::from(dir).join("magi").join("config.toml")),
1830 Err(_) => dirs::config_dir().map(|dir| dir.join("magi").join("config.toml")),
1831 }
1832 }
1833
1834 /// Built-in config whose roster is the agent CLIs found on `PATH`.
1835 pub fn autodetected() -> Self {
1836 let mut cfg = Self::default();
1837 for (kind, id, model) in [
1838 (AgentKind::Claude, "opus", Some("opus")),
1839 (AgentKind::Claude, "sonnet", Some("sonnet")),
1840 (AgentKind::Antigravity, "antigravity", None),
1841 (AgentKind::Opencode, "opencode", None),
1842 (AgentKind::Codex, "codex", None),
1843 (AgentKind::Omp, "omp", None),
1844 ] {
1845 if kind.program().is_some_and(which) && !cfg.agents.iter().any(|a| a.id == id) {
1846 cfg.agents.push(AgentSpec {
1847 id: id.to_owned(),
1848 kind,
1849 model: model.map(str::to_owned),
1850 command: Vec::new(),
1851 extra_args: Vec::new(),
1852 env: BTreeMap::new(),
1853 prompt_delivery: None,
1854 });
1855 }
1856 }
1857 cfg
1858 }
1859
1860 /// The shared build cache the verify commands and the agents both build
1861 /// into, when the config declares one. See [`Verify::cache_dir`].
1862 pub fn cache_dir(&self) -> Option<PathBuf> {
1863 self.verify.cache_dir()
1864 }
1865
1866 /// Look an agent up by id.
1867 pub fn agent(&self, id: &str) -> Result<&AgentSpec> {
1868 self.agents
1869 .iter()
1870 .find(|a| a.id == id)
1871 .with_context(|| format!("no agent with id `{id}` in the roster"))
1872 }
1873
1874 /// Rotate `count` seats out of `ids`, or out of the whole roster at
1875 /// `offset` when `ids` is empty.
1876 ///
1877 /// The one rotation rule - explicit ids cycle, an empty list rotates the
1878 /// roster - shared by every seat count [`Config::resolve_roles`] fills in,
1879 /// rather than each seat reimplementing it and drifting apart.
1880 fn rotate(&self, ids: &[String], count: usize, offset: usize) -> Result<Vec<AgentSpec>> {
1881 let mut out = Vec::with_capacity(count);
1882 for i in 0..count {
1883 let spec = if ids.is_empty() {
1884 self.agents[(i + offset) % self.agents.len()].clone()
1885 } else {
1886 self.agent(&ids[i % ids.len()])?.clone()
1887 };
1888 out.push(spec);
1889 }
1890 Ok(out)
1891 }
1892
1893 /// `ids`, resolved to specs in the order named, or the whole `[[agents]]`
1894 /// roster in file order when `ids` is empty — the same "explicit ids
1895 /// stand alone, an empty list means the whole roster" rule [`Self::rotate`]
1896 /// applies, but with no `count` to truncate to and no `offset` to wrap by.
1897 /// See [`ResolvedRoles::implementer_roster`] for why a truncated, rotated
1898 /// list cannot answer the question this exists for.
1899 fn full_roster(&self, ids: &[String]) -> Result<Vec<AgentSpec>> {
1900 if ids.is_empty() {
1901 Ok(self.agents.clone())
1902 } else {
1903 ids.iter().map(|id| self.agent(id).cloned()).collect()
1904 }
1905 }
1906
1907 /// Fill the roles out to the configured widths.
1908 ///
1909 /// An empty role list rotates through the whole roster, so a three-agent
1910 /// roster with `implementers = 3` gives one implementation per agent, and
1911 /// `judges = 3` rotates the judge seats by one so that judge *i* is not the
1912 /// author of candidate *i* whenever the roster has more than one agent.
1913 pub fn resolve_roles(&self) -> Result<ResolvedRoles> {
1914 if self.agents.is_empty() {
1915 bail!(
1916 "agent roster is empty: no agent CLI found on PATH and no \
1917 [[agents]] in the config. Run `magi init` to write a starter \
1918 magi.toml."
1919 );
1920 }
1921 Ok(ResolvedRoles {
1922 implementers: self.rotate(&self.roles.implementers, self.graph.implementers, 0)?,
1923 judges: self.rotate(&self.roles.judges, self.graph.judges, 1)?,
1924 reviewers: self.rotate(&self.roles.reviewers, self.graph.reviewers, 0)?,
1925 fixer: match self.roles.fixer.as_ref() {
1926 // Ids are validated against the roster here; whether the CLI
1927 // is installed is left to invocation, like every other seat.
1928 Some(choice) => Some(crate::agent::pick_chain(
1929 &self.agents,
1930 Some(choice),
1931 &|_| true,
1932 "fixer",
1933 )?),
1934 None => None,
1935 },
1936 // Role resolution validates roster shape, but deliberately does
1937 // not preflight a CLI. The other graph seats have always deferred
1938 // that failure to invocation; doing it only for the conductor
1939 // made otherwise usable graph commands and `doctor` fail as one.
1940 conductor: match self.roles.conductor.as_ref().map(AgentChoice::ids) {
1941 Some(ids) if ids.len() == 1 => self.agent(ids[0])?.clone(),
1942 // A fallback chain: the first id the roster knows. The rest
1943 // are `agent::pick_chain`'s business at invocation.
1944 Some(ids) if !ids.is_empty() => ids
1945 .iter()
1946 .find_map(|id| self.agent(id).ok())
1947 .cloned()
1948 .with_context(|| {
1949 format!("[roles] conductor names no agent in the roster: {ids:?}")
1950 })?,
1951 // Keep the normal standalone-seat preference when something
1952 // is installed, but retain a roster fallback when it is not.
1953 // Invocation then reports the unavailable CLI in the same
1954 // place it does for every other graph role.
1955 _ => crate::agent::pick(&self.agents, None, &crate::agent::installed)
1956 .unwrap_or_else(|_| self.agents[0].clone()),
1957 },
1958 implementer_roster: self.full_roster(&self.roles.implementers)?,
1959 judge_roster: self.full_roster(&self.roles.judges)?,
1960 reviewer_roster: self.full_roster(&self.roles.reviewers)?,
1961 })
1962 }
1963
1964 /// Advisor seats for the design-deliberation stage (see
1965 /// `graph::Runner::advise`): `[roles] advisors` when set, otherwise the
1966 /// judge roster - see [`Roles::advisors`] for why that fallback and not
1967 /// the whole roster.
1968 ///
1969 /// The fallback rotates with `offset = 1`, matching the judges line in
1970 /// [`Config::resolve_roles`] exactly, `ids` and offset both - not just
1971 /// `roles.judges`, which is empty whenever judges themselves are
1972 /// unconfigured and rotating the whole roster. Falling back with
1973 /// `offset = 0` there would silently hand the advisors a *different*
1974 /// agent set than the judges an unconfigured run would actually get,
1975 /// which is the one thing [`Roles::advisors`]'s doc promises will not
1976 /// happen.
1977 ///
1978 /// Called lazily from the graph node itself rather than folded into
1979 /// [`Config::resolve_roles`]: unlike the other roles, a failure here must
1980 /// not stop a run from starting at all - the deliberation stage is an
1981 /// enrichment `[graph] advise` can turn off, not a seat later nodes
1982 /// cannot proceed without - and at the point `resolve_roles` runs (before
1983 /// [`crate::run::RunState`] exists, on `Runner::start`) there would be no
1984 /// run yet for a resolution failure to be reported against.
1985 pub fn advisors(&self) -> Result<Vec<AgentSpec>> {
1986 if self.agents.is_empty() {
1987 bail!(
1988 "agent roster is empty: no agent CLI found on PATH and no \
1989 [[agents]] in the config. Run `magi init` to write a starter \
1990 magi.toml."
1991 );
1992 }
1993 if !self.roles.advisors.is_empty() {
1994 return self.rotate(&self.roles.advisors, self.graph.advisors, 0);
1995 }
1996 self.rotate(&self.roles.judges, self.graph.advisors, 1)
1997 }
1998
1999 /// The untruncated roster advisor seats hand over along: `[roles]
2000 /// advisors` when set, otherwise the judges' roster, exactly as
2001 /// [`Self::advisors`] chooses its own list.
2002 pub fn advisor_roster(&self) -> Result<Vec<AgentSpec>> {
2003 if self.roles.advisors.is_empty() {
2004 self.full_roster(&self.roles.judges)
2005 } else {
2006 self.full_roster(&self.roles.advisors)
2007 }
2008 }
2009
2010 /// Shell prefix for [`Verify`] commands.
2011 pub fn shell(&self) -> Vec<String> {
2012 if let Some(s) = &self.verify.shell {
2013 return s.clone();
2014 }
2015 if which("sh") {
2016 vec!["sh".to_owned(), "-c".to_owned()]
2017 } else {
2018 vec!["cmd".to_owned(), "/C".to_owned()]
2019 }
2020 }
2021
2022 /// Starter config, as written by `magi init`.
2023 pub fn starter_toml() -> String {
2024 let detected = Self::autodetected();
2025 let mut s = String::from(
2026 "# magi — blind multi-agent implementation competition.\n\
2027 # `magi run \"<task>\"` walks: implement (N parallel worktrees)\n\
2028 # -> blind judging -> deliberation -> private final vote\n\
2029 # -> fold losers -> review + E2E loop -> gate -> merge.\n\
2030 #\n\
2031 # Rendered by teravars: a `[vars]` table, env\n\
2032 # and system lookups, and `include = [...]` all work. Tera\n\
2033 # braces are live everywhere in this file, but comments are\n\
2034 # stripped before rendering (teravars >= 0.2.2), so a comment\n\
2035 # may quote `{{ ... }}` freely.\n\
2036 #\n\
2037 # Layers deep-merge in increasing\n\
2038 # precedence, so the roster can live once per machine in\n\
2039 # <config_dir>/magi/config.toml and each repo only states its own\n\
2040 # gate:\n\
2041 # <config_dir>/magi/config.toml < .magi/config.toml < magi.toml\n\n\
2042 [vars]\n\
2043 # Reference it as vars.cache inside Tera braces, anywhere below.\n\
2044 # Single quotes inside the braces: teravars renders the raw file\n\
2045 # text, so TOML's own \\\" escaping never reaches Tera.\n\
2046 cache = \"{{ env.MAGI_CACHE | default(value='/tmp') }}\"\n\n",
2047 );
2048 if detected.agents.is_empty() {
2049 s.push_str(
2050 "# No agent CLI was found on PATH. Fill this in by hand.\n\
2051 # kind = claude | opencode | antigravity | codex | command\n\
2052 [[agents]]\nid = \"opus\"\nkind = \"claude\"\nmodel = \"opus\"\n\n",
2053 );
2054 } else {
2055 for a in &detected.agents {
2056 s.push_str("[[agents]]\n");
2057 s.push_str(&format!("id = {:?}\n", a.id));
2058 s.push_str(&format!("kind = {:?}\n", a.kind.as_str()));
2059 if let Some(m) = &a.model {
2060 s.push_str(&format!("model = {m:?}\n"));
2061 }
2062 s.push('\n');
2063 }
2064 }
2065 s.push_str(
2066 "# Leave a role list empty to rotate through the roster.\n\
2067 # A non-empty list is also an ordered backup list: seats take entries from the\n\
2068 # front (see [graph] counts) and a failed or out-of-quota seat moves on to the\n\
2069 # next agent to its right, each tried once. Same for judges and reviewers.\n\
2070 # implementers = [\"opus\", \"codex\", \"sonnet\"] # with implementers = 1 below: opus sits, codex then sonnet are spares\n\
2071 [roles]\n\
2072 implementers = []\n\
2073 judges = []\n\
2074 reviewers = []\n\
2075 # conductor = \"opus\" # arranges the queue; unset picks a seat like chatter does\n\
2076 # synthesizer = \"opus\" # blends the advisors into one brief; unset picks a seat like chatter does\n\
2077 # synthesizer = [\"opus\", \"codex\"] # array form: fallback chain, each tried once on quota or failure\n\
2078 # language_judge = \"jev\" # kind = \"command\" wrapper deciding if PR text is English; unset keeps the built-in heuristics\n\
2079 # fixer = \"opus\" # applies review findings; unset keeps the winner's own author\n\
2080 # fixer = [\"opus\", \"codex\"] # array form: fallback chain, each tried once on quota or failure\n\n\
2081 [graph]\n\
2082 # Gate GitHub prose; secret redaction is always enabled.\n\
2083 github_text_guard = true\n\
2084 implementers = 3\n\
2085 judges = 3\n\
2086 deliberate_rounds = 1\n\
2087 reviewers = 3\n\
2088 review_rounds = 6\n\
2089 # Fix rounds a failing verify.gate gets before the run is blocked (0 = none).\n\
2090 gate_fix_rounds = 1\n\
2091 max_parallel = 4\n\
2092 language = \"en\"\n\
2093 # One CLI conversation per seat: judges keep their own argument\n\
2094 # across deliberation, the fixer keeps its implementation context.\n\
2095 sessions = true\n\
2096 # Reviewer-seat timeout. When timeout_verify is omitted, E2E and\n\
2097 # the final gate inherit this value for compatibility.\n\
2098 timeout_review = 1200\n\
2099 # Optional independent E2E/final-gate timeout; uncomment to keep\n\
2100 # verification independent if timeout_review changes later.\n\
2101 # timeout_verify = 1200\n\n\
2102 # What the chat calls you (persona address and the Chat's turn label).\n\
2103 # [talk]\n\
2104 # operator_name = \"Commander\"\n\n\
2105 # Chat personas (tone only; the web Chat picks one per conversation).\n\
2106 # Built in: magi, rei, misato, ritsuko, shinji, asuka, kaworu. Entries add to\n\
2107 # them, and one with a built-in's id replaces it. `default` is the plain\n\
2108 # voice and cannot be redefined. Declare this in one layer only.\n\
2109 # [[talk.personas]]\n\
2110 # id = \"gendo\"\n\
2111 # name = \"Gendo Ikari\"\n\
2112 # prompt = \"Speak coldly and tersely, hands folded.\"\n\n\
2113 [verify]\n\
2114 # Run once per review round in the winner's worktree; failures are\n\
2115 # fed back to the fixer.\n\
2116 e2e = []\n\
2117 # Final gate. Every command must exit 0 before a merge.\n\
2118 gate = []\n\
2119 # Mechanical fixers (formatters) run in the winner's worktree just\n\
2120 # before the gate; whatever they change becomes one commit. A failing\n\
2121 # command only warns - the gate stays the arbiter. Same timeout as the\n\
2122 # gate (timeout_verify). Keep it light and idempotent.\n\
2123 # pre_gate = []\n\n\
2124 # Prepare each worktree magi creates (candidates, judges, advisors,\n\
2125 # reviewers, the operator-fix and review-takeover checkouts), in order,\n\
2126 # before any agent or verify command touches it. `copy` reads from the\n\
2127 # primary checkout and never overwrites; `optional` skips a missing\n\
2128 # source. `run` goes through the shell with the worktree as cwd and\n\
2129 # fails the seat on a non-zero exit or timeout (timeout_secs, default\n\
2130 # the verify timeout). Setup must not edit tracked files, and what it\n\
2131 # leaves untracked is kept out of magi's commits. Declare this in one\n\
2132 # layer only.\n\
2133 # [worktree]\n\
2134 # setup = [\n\
2135 # { copy = \".env.example -> .env\" },\n\
2136 # { copy = \".env.local\", optional = true },\n\
2137 # { run = \"pnpm install --frozen-lockfile\", timeout_secs = 900 },\n\
2138 # ]\n\n\
2139 [merge]\n\
2140 # none | local | pr\n\
2141 mode = \"none\"\n\
2142 # release_bump = true # Rust (Cargo.toml) repositories only; a no-op elsewhere\n\n\
2143 [update]\n\
2144 # off | notify | install — checked in the background, throttled.\n\
2145 mode = \"notify\"\n\
2146 # interval = \"24h\"\n\n\
2147 [daemon]\n\
2148 # Minutes a release-bump pull request (chore/release-v*) may sit with\n\
2149 # no new commit and no check changing state before `magi serve`\n\
2150 # notices and asks you. A failed check is rerun once on its own;\n\
2151 # still red after that escalates at once. 0 turns the watcher off.\n\
2152 # release_stall_minutes = 60\n\n\
2153 [release]\n\
2154 # actions (default) | local. \"local\" is for repositories whose GitHub\n\
2155 # Actions never run (private, no minutes): the release pull request\n\
2156 # does not wait for CI, you approve the merge in magi, and then\n\
2157 # `magi serve` pushes the vX.Y.Z tag (never over an existing one) and\n\
2158 # runs `commands` from a clean checkout of the merge commit. A\n\
2159 # failure holds the release with a notice; it is never retried blind.\n\
2160 # Needs `magi serve` running and [daemon] release_stall_minutes > 0.\n\
2161 # mode = \"actions\"\n\
2162 # Run in order; the first failure stops. They see MAGI_RELEASE_VERSION,\n\
2163 # MAGI_RELEASE_TAG, MAGI_RELEASE_COMMIT and MAGI_RELEASE_PR. Make them\n\
2164 # safe to repeat: a crash mid-command is treated as unfinished.\n\
2165 # commands = [\"cargo build --release\", \"gh release create \\\"$MAGI_RELEASE_TAG\\\" --generate-notes\"]\n\
2166 # timeout_minutes = 30\n",
2167 );
2168 s
2169 }
2170}
2171
2172/// Is `program` on `PATH`?
2173pub fn which(program: &str) -> bool {
2174 find_program(program).is_some()
2175}
2176
2177/// The file a bare `program` name resolves to on `PATH`, the way the OS shell
2178/// would find it.
2179///
2180/// Spawn agents through this, not by bare name: on Windows Rust's std
2181/// only tries `name.exe`, so a CLI installed as an npm `.cmd` shim (codex
2182/// since its 2026-09-23 reinstall) is "not found" even though a shell runs it.
2183/// On Windows an extensionless file is never a match: npm writes an `sh`
2184/// script beside every `.cmd` shim, and spawning that fails with os error 193.
2185pub fn find_program(program: &str) -> Option<PathBuf> {
2186 find_program_on(program, &std::env::var_os("PATH")?)
2187}
2188
2189/// [`find_program`] against an explicit `PATH`, for a child whose environment
2190/// overrides it.
2191pub fn find_program_on(program: &str, paths: &std::ffi::OsStr) -> Option<PathBuf> {
2192 let pathext = std::env::var("PATHEXT").unwrap_or_else(|_| ".EXE;.CMD;.BAT".into());
2193 find_program_in(program, paths, cfg!(windows).then_some(pathext.as_str()))
2194}
2195
2196fn find_program_in(
2197 program: &str,
2198 paths: &std::ffi::OsStr,
2199 pathext: Option<&str>,
2200) -> Option<PathBuf> {
2201 let p = Path::new(program);
2202 if p.components().count() > 1 {
2203 return p.is_file().then(|| p.to_path_buf());
2204 }
2205 std::env::split_paths(paths).find_map(|dir| match pathext {
2206 Some(exts) => {
2207 let bare = dir.join(program);
2208 if p.extension().is_some() && bare.is_file() {
2209 return Some(bare);
2210 }
2211 exts.split(';')
2212 .filter(|e| !e.is_empty())
2213 .map(|e| dir.join(format!("{program}{}", e.to_lowercase())))
2214 .find(|c| c.is_file())
2215 }
2216 None => Some(dir.join(program)).filter(|c| c.is_file()),
2217 })
2218}
2219
2220/// Sidecar naming the checkout an extracted layer stands for.
2221const REPO_MARKER: &str = ".magi-repo";
2222
2223fn git_out(repo: &Path, args: &[&str]) -> Option<String> {
2224 use crate::proc::Quiet as _;
2225 let out = std::process::Command::new("git")
2226 .args(args)
2227 .quiet()
2228 .current_dir(repo)
2229 .stdin(std::process::Stdio::null())
2230 .output()
2231 .ok()?;
2232 out.status
2233 .success()
2234 .then(|| String::from_utf8_lossy(&out.stdout).into_owned())
2235}
2236
2237/// Is `repo` inside a git working tree? A directory that is not has no branch
2238/// to be stale against, so its files are read as they are.
2239fn is_git_worktree(repo: &Path) -> bool {
2240 git_out(repo, &["rev-parse", "--git-dir"]).is_some()
2241}
2242
2243#[derive(Deserialize, Default)]
2244struct ProbeMerge {
2245 base: Option<String>,
2246 remote: Option<String>,
2247}
2248
2249#[derive(Deserialize, Default)]
2250struct Probe {
2251 merge: Option<ProbeMerge>,
2252}
2253
2254/// `[merge]` base / remote from one file, nothing else. The file is rendered
2255/// (env templates, includes) exactly as a real load would, so a machine layer
2256/// saying `base = "{{ env.X }}"` locates the same ref; when rendering fails
2257/// (a checkout file that needs its own repo context) the plain TOML is read
2258/// instead. Unreadable is `None`: the real load reports it.
2259fn probe_merge(path: &Path, repo: &Path) -> Option<ProbeMerge> {
2260 if !path.is_file() {
2261 return None;
2262 }
2263 let paths = [path.to_path_buf()];
2264 let mut engine = teravars::Engine::default();
2265 let mut ctx = Config::render_ctx(&paths);
2266 // The layered load renders every layer for the checkout being resolved;
2267 // so does this, or `{{ repo_name }}` would name the file's own directory.
2268 ctx.insert("repo", &repo.to_string_lossy());
2269 ctx.insert(
2270 "repo_name",
2271 &repo.file_name().unwrap_or_default().to_string_lossy(),
2272 );
2273 if let Ok(merged) = teravars::load_merged(&paths, &mut engine, &ctx)
2274 && let Ok(probe) = toml::Value::Table(merged.config).try_into::<Probe>()
2275 {
2276 return probe.merge;
2277 }
2278 let text = std::fs::read_to_string(path).ok()?;
2279 toml::from_str::<Probe>(&text).ok()?.merge
2280}
2281
2282/// The `[merge] remote` / `base` the layers name, before any git lookup.
2283fn probe_remote_base(repo: &Path) -> (String, Option<String>) {
2284 let mut remote = None;
2285 let mut base = None;
2286 let mut sources: Vec<PathBuf> = Config::machine_layer().into_iter().collect();
2287 sources.push(repo.join("magi.toml"));
2288 sources.push(repo.join(".magi").join("config.toml"));
2289 for path in sources {
2290 if let Some(m) = probe_merge(&path, repo) {
2291 remote = remote.or(m.remote.filter(|s| !s.trim().is_empty()));
2292 base = base.or(m.base.filter(|s| !s.trim().is_empty()));
2293 }
2294 }
2295 (remote.unwrap_or_else(|| "origin".to_owned()), base)
2296}
2297
2298/// Where the remote ref is: see the module doc for the order.
2299fn resolve_remote_base(repo: &Path) -> Result<(String, String)> {
2300 let (remote, base) = probe_remote_base(repo);
2301 let base = match base {
2302 Some(b) => b,
2303 None => {
2304 let head = format!("refs/remotes/{remote}/HEAD");
2305 git_out(repo, &["symbolic-ref", "--quiet", &head])
2306 .and_then(|s| crate::git::remote_head_branch(&remote, &s))
2307 .with_context(|| {
2308 format!(
2309 "cannot tell which branch of `{remote}` holds magi.toml: \
2310 set [merge] base, or run `git remote set-head {remote} -a`"
2311 )
2312 })?
2313 }
2314 };
2315 Ok((remote, base))
2316}
2317
2318fn fnv_hex(text: &str) -> String {
2319 let mut hash: u64 = 0xcbf2_9ce4_8422_2325;
2320 for b in text.bytes() {
2321 hash ^= u64::from(b);
2322 hash = hash.wrapping_mul(0x0100_0000_01b3);
2323 }
2324 format!("{hash:016x}")
2325}
2326
2327/// Does the text declare an `include` directive in any form teravars accepts
2328/// (root `include`, bare or quoted, or `[teravars] include`, as a table or an
2329/// inline table)? Parsed as TOML when it parses; a templated file that does
2330/// not parse is judged by a plain search for the word, because over-detecting
2331/// only costs an extra extraction while missing one fails the load.
2332fn declares_include(text: &str) -> bool {
2333 match toml::from_str::<toml::Table>(text) {
2334 Ok(table) => {
2335 table.contains_key("include")
2336 || table
2337 .get("teravars")
2338 .and_then(toml::Value::as_table)
2339 .is_some_and(|t| t.contains_key("include"))
2340 }
2341 Err(_) => text.contains("include"),
2342 }
2343}
2344
2345/// Unpack the whole tree of `sha` into `dest`, so relative `include`s of a
2346/// repository layer resolve against the same commit.
2347fn extract_tree(repo: &Path, sha: &str, dest: &Path) -> Result<()> {
2348 use crate::proc::Quiet as _;
2349 let mut archive = std::process::Command::new("git")
2350 .args(["archive", "--format=tar", sha])
2351 .quiet()
2352 .current_dir(repo)
2353 .stdin(std::process::Stdio::null())
2354 .stdout(std::process::Stdio::piped())
2355 .stderr(std::process::Stdio::null())
2356 .spawn()
2357 .context("running git archive")?;
2358 let tar = std::process::Command::new("tar")
2359 .arg("-x")
2360 .arg("-C")
2361 .arg(dest)
2362 .quiet()
2363 .stdin(archive.stdout.take().context("git archive has no stdout")?)
2364 .output()
2365 .context("running tar")?;
2366 let status = archive.wait().context("waiting for git archive")?;
2367 if !status.success() || !tar.status.success() {
2368 bail!(
2369 "cannot unpack {sha} for the config's includes: {}",
2370 String::from_utf8_lossy(&tar.stderr).trim()
2371 );
2372 }
2373 Ok(())
2374}
2375
2376/// The repository layers (`.magi/config.toml`, `magi.toml`; lowest first) of
2377/// `<remote>/<base>`, extracted from one commit so both files agree.
2378fn remote_layers(repo: &Path) -> Result<Vec<PathBuf>> {
2379 let (remote, base) = resolve_remote_base(repo)?;
2380 let tracking = format!("refs/remotes/{remote}/{base}");
2381 let spec = format!("{tracking}^{{commit}}");
2382 let sha = git_out(repo, &["rev-parse", "--verify", "--quiet", &spec])
2383 .map(|s| s.trim().to_owned())
2384 .filter(|s| !s.is_empty())
2385 .with_context(|| {
2386 format!(
2387 "`{remote}/{base}` does not exist in this checkout; fetch it \
2388 (`git fetch {remote} {base}`) or set [merge] base/remote. magi will not read \
2389 a possibly stale local magi.toml instead"
2390 )
2391 })?;
2392 let key = format!("{sha}-{}", fnv_hex(&repo.to_string_lossy()));
2393 let rels = [".magi/config.toml", "magi.toml"];
2394 let mut texts: Vec<(&str, String)> = Vec::new();
2395 for rel in rels {
2396 let object = format!("{sha}:{rel}");
2397 if git_out(repo, &["cat-file", "-e", &object]).is_none() {
2398 continue;
2399 }
2400 let Some(text) = git_out(repo, &["show", &object]) else {
2401 bail!("cannot read {rel} from {remote}/{base} ({sha})");
2402 };
2403 texts.push((rel, text));
2404 }
2405 let wants_tree = texts.iter().any(|(_, t)| declares_include(t));
2406 // Content-addressed and built beside the final name, then renamed in: a
2407 // concurrent `discover` sees the whole directory or none of it, never a
2408 // half-written file.
2409 let build = |root: &Path| -> Result<PathBuf> {
2410 let final_dir = root.join(&key);
2411 if !final_dir.is_dir() {
2412 std::fs::create_dir_all(root)?;
2413 let tmp = root.join(format!(
2414 "{key}.tmp-{}-{}",
2415 std::process::id(),
2416 std::time::SystemTime::now()
2417 .duration_since(std::time::UNIX_EPOCH)
2418 .map(|d| d.as_nanos())
2419 .unwrap_or_default()
2420 ));
2421 let filled = (|| -> Result<()> {
2422 std::fs::create_dir_all(&tmp)?;
2423 if wants_tree {
2424 extract_tree(repo, &sha, &tmp)?;
2425 }
2426 for (rel, text) in &texts {
2427 let file = tmp.join(rel);
2428 let parent = file.parent().unwrap_or(&tmp);
2429 std::fs::create_dir_all(parent)?;
2430 std::fs::write(&file, text)?;
2431 std::fs::write(parent.join(REPO_MARKER), repo.to_string_lossy().as_bytes())?;
2432 }
2433 Ok(())
2434 })();
2435 if let Err(e) = filled {
2436 let _ = std::fs::remove_dir_all(&tmp);
2437 return Err(e);
2438 }
2439 if std::fs::rename(&tmp, &final_dir).is_err() {
2440 // Lost the race to another process building the same commit.
2441 let _ = std::fs::remove_dir_all(&tmp);
2442 if !final_dir.is_dir() {
2443 bail!("cannot place {}", final_dir.display());
2444 }
2445 }
2446 }
2447 Ok(final_dir)
2448 };
2449 let common = git_out(
2450 repo,
2451 &["rev-parse", "--path-format=absolute", "--git-common-dir"],
2452 )
2453 .map(|s| PathBuf::from(s.trim()).join("magi").join("config"));
2454 let dir = match common.map(|c| build(&c)) {
2455 Some(Ok(d)) => d,
2456 _ => build(&std::env::temp_dir().join("magi-config"))
2457 .with_context(|| format!("extracting config from {remote}/{base}"))?,
2458 };
2459 Ok(texts.iter().map(|(rel, _)| dir.join(rel)).collect())
2460}
2461
2462#[cfg(test)]
2463mod tests {
2464 #[test]
2465 fn context_window_prefers_exact_then_longest_prefix_and_user_over_builtin() {
2466 let mut cfg = Config::default();
2467 assert_eq!(
2468 cfg.context_window("claude-sonnet-4-5-20250929"),
2469 Some(200_000)
2470 );
2471 assert_eq!(cfg.context_window("gpt-4.1-mini"), Some(1_000_000));
2472 assert_eq!(cfg.context_window("mystery-model"), None);
2473 cfg.context_windows
2474 .insert("claude-sonnet-4-5".into(), 1_000_000);
2475 cfg.context_windows.insert("claude-sonnet-4".into(), 5);
2476 cfg.context_windows.insert("mystery".into(), 32_768);
2477 // Longest user prefix beats a shorter one and the built-in table.
2478 assert_eq!(
2479 cfg.context_window("claude-sonnet-4-5-20250929"),
2480 Some(1_000_000)
2481 );
2482 // An exact name beats any prefix.
2483 cfg.context_windows
2484 .insert("claude-sonnet-4-5-20250929".into(), 7);
2485 assert_eq!(cfg.context_window("claude-sonnet-4-5-20250929"), Some(7));
2486 assert_eq!(cfg.context_window("mystery-model"), Some(32_768));
2487 }
2488
2489 use super::*;
2490
2491 /// npm writes an extensionless `sh` script beside `codex.cmd`; on Windows
2492 /// the `.cmd` is the spawnable one (os error 193 otherwise, 2026-09-23).
2493 #[test]
2494 fn windows_lookup_picks_the_cmd_shim_over_the_extensionless_script() {
2495 let dir = tempfile::tempdir().unwrap();
2496 std::fs::write(dir.path().join("codex"), "#!/bin/sh\n").unwrap();
2497 std::fs::write(dir.path().join("codex.cmd"), "@echo off\n").unwrap();
2498 let paths = dir.path().as_os_str();
2499 assert_eq!(
2500 find_program_in("codex", paths, Some(".EXE;.CMD")),
2501 Some(dir.path().join("codex.cmd"))
2502 );
2503 assert_eq!(
2504 find_program_in("codex", paths, None),
2505 Some(dir.path().join("codex"))
2506 );
2507 assert_eq!(find_program_in("claude", paths, Some(".EXE;.CMD")), None);
2508 }
2509
2510 fn spec(id: &str) -> AgentSpec {
2511 AgentSpec {
2512 id: id.to_owned(),
2513 kind: AgentKind::Command,
2514 model: None,
2515 command: vec!["true".to_owned()],
2516 extra_args: Vec::new(),
2517 env: BTreeMap::new(),
2518 prompt_delivery: None,
2519 }
2520 }
2521
2522 #[test]
2523 fn github_text_guard_defaults_on_and_can_be_disabled() {
2524 let default: Graph = toml::from_str("").unwrap();
2525 assert!(default.github_text_guard);
2526 let disabled: Graph = toml::from_str("github_text_guard = false").unwrap();
2527 assert!(!disabled.github_text_guard);
2528 let config = layered("", "[graph]\ngithub_text_guard = false\n").unwrap();
2529 assert!(!config.graph.github_text_guard);
2530 assert!(Config::starter_toml().contains("github_text_guard = true"));
2531 }
2532
2533 #[test]
2534 fn timeout_verify_omitted_from_old_toml_inherits_timeout_review() {
2535 // `timeout_verify` used to not exist: `verify.e2e`/`verify.gate` ran
2536 // under `timeout_review`. A config written before this field existed
2537 // must run exactly as before, which means its default has to be the
2538 // same 1200s `timeout_review` has always defaulted to.
2539 let g: Graph = toml::from_str("timeout_review = 3600").expect("parse");
2540 assert_eq!(g.timeout_verify, None);
2541 assert_eq!(g.verify_timeout(), 3600);
2542 }
2543
2544 #[test]
2545 fn shrinking_timeout_review_does_not_shrink_timeout_verify() {
2546 // The bug this field exists to close: `[graph] timeout_review = 45`
2547 // used to shrink the real-machine `verify.e2e`/`verify.gate` budget
2548 // along with the reviewer seats' own timeout, because both read the
2549 // same field.
2550 let g: Graph = toml::from_str("timeout_review = 45").expect("parse");
2551 assert_eq!(g.timeout_review, 45);
2552 assert_eq!(
2553 g.verify_timeout(),
2554 45,
2555 "an omitted legacy value follows review"
2556 );
2557 let explicit: Graph = toml::from_str("timeout_review = 45\ntimeout_verify = 1200")
2558 .expect("parse explicit override");
2559 assert_eq!(explicit.verify_timeout(), 1200);
2560 }
2561
2562 #[test]
2563 fn graph_implementers_new_key_loads_and_old_key_is_an_alias() {
2564 let g: Graph = toml::from_str("implementers = 4\n").unwrap();
2565 assert_eq!(g.implementers, 4);
2566 let g: Graph = toml::from_str("candidates = 5\n").unwrap();
2567 assert_eq!(g.implementers, 5);
2568 }
2569
2570 fn layered(lower: &str, upper: &str) -> Result<Config> {
2571 let dir = tempfile::tempdir().unwrap();
2572 let a = dir.path().join("a.toml");
2573 let b = dir.path().join("b.toml");
2574 std::fs::write(&a, lower).unwrap();
2575 std::fs::write(&b, upper).unwrap();
2576 Config::load_layers(&[a, b])
2577 }
2578
2579 #[test]
2580 fn graph_implementers_higher_layer_wins_across_spellings() {
2581 let c = layered("[graph]\ncandidates = 2\n", "[graph]\nimplementers = 5\n").unwrap();
2582 assert_eq!(c.graph.implementers, 5);
2583 let c = layered("[graph]\nimplementers = 2\n", "[graph]\ncandidates = 5\n").unwrap();
2584 assert_eq!(c.graph.implementers, 5);
2585 let c = layered("[graph]\ncandidates = 2\n", "[graph]\njudges = 3\n").unwrap();
2586 assert_eq!(c.graph.implementers, 2);
2587 }
2588
2589 #[test]
2590 fn graph_implementers_both_keys_in_one_layer_is_refused() {
2591 let err = layered("", "[graph]\ncandidates = 2\nimplementers = 3\n")
2592 .unwrap_err()
2593 .to_string();
2594 assert!(
2595 err.contains("candidates") && err.contains("implementers"),
2596 "{err}"
2597 );
2598 assert!(err.contains("b.toml"), "{err}");
2599 }
2600
2601 #[test]
2602 fn a_toml_layer_written_before_these_fields_existed_still_parses() {
2603 // `deny_unknown_fields` cuts both ways: a config from before
2604 // `timeout_verify`/`e2e_every_round` existed must still parse, with
2605 // both defaulted rather than refused as unknown-in-reverse.
2606 let g: Graph =
2607 toml::from_str("candidates = 1\nreviewers = 3\nreview_rounds = 6\nmax_parallel = 4\n")
2608 .expect("an old-shaped [graph] table must still parse");
2609 assert_eq!(g.verify_timeout(), Graph::default().timeout_review);
2610 assert!(
2611 !g.e2e_every_round,
2612 "off by default, same as before this field existed"
2613 );
2614 }
2615
2616 #[test]
2617 fn empty_roles_rotate_judges_off_their_own_candidate() {
2618 // Three seats, said out loud: this is a test about *rotation*, and it
2619 // has nothing to say about how many candidates a task buys by default.
2620 let cfg = Config {
2621 agents: vec![spec("a"), spec("b"), spec("c")],
2622 graph: Graph {
2623 implementers: 3,
2624 ..Graph::default()
2625 },
2626 ..Config::default()
2627 };
2628 let roles = cfg.resolve_roles().unwrap();
2629 let impls: Vec<&str> = roles.implementers.iter().map(|a| a.id.as_str()).collect();
2630 let judges: Vec<&str> = roles.judges.iter().map(|a| a.id.as_str()).collect();
2631 assert_eq!(impls, ["a", "b", "c"]);
2632 assert_eq!(judges, ["b", "c", "a"]);
2633 for (i, j) in judges.iter().enumerate() {
2634 assert_ne!(*j, impls[i], "judge {i} must not sit on its own candidate");
2635 }
2636 }
2637
2638 #[test]
2639 fn single_agent_roster_fills_every_seat() {
2640 let cfg = Config {
2641 agents: vec![spec("solo")],
2642 graph: Graph {
2643 implementers: 3,
2644 ..Graph::default()
2645 },
2646 ..Config::default()
2647 };
2648 let roles = cfg.resolve_roles().unwrap();
2649 assert_eq!(roles.implementers.len(), 3);
2650 assert!(roles.judges.iter().all(|a| a.id == "solo"));
2651 }
2652
2653 #[test]
2654 fn implementer_roster_is_the_whole_agent_list_untruncated_when_unset() {
2655 // `implementers = 1` (a solo run) still resolves `implementers` down to
2656 // one slot, but `implementer_roster` must carry every agent in the
2657 // roster, in file order, for `graph::Runner`'s quota fallback to walk
2658 // forward through once that one slot's agent runs out of quota.
2659 let cfg = Config {
2660 agents: vec![spec("a"), spec("b"), spec("c")],
2661 graph: Graph {
2662 implementers: 1,
2663 ..Graph::default()
2664 },
2665 ..Config::default()
2666 };
2667 let roles = cfg.resolve_roles().unwrap();
2668 assert_eq!(roles.implementers.len(), 1);
2669 let roster: Vec<&str> = roles
2670 .implementer_roster
2671 .iter()
2672 .map(|a| a.id.as_str())
2673 .collect();
2674 assert_eq!(roster, ["a", "b", "c"]);
2675 }
2676
2677 #[test]
2678 fn implementer_roster_follows_explicit_ids_in_the_order_named() {
2679 let cfg = Config {
2680 agents: vec![spec("a"), spec("b"), spec("c")],
2681 roles: Roles {
2682 implementers: vec!["c".to_owned(), "a".to_owned()],
2683 ..Roles::default()
2684 },
2685 graph: Graph {
2686 implementers: 1,
2687 ..Graph::default()
2688 },
2689 ..Config::default()
2690 };
2691 let roles = cfg.resolve_roles().unwrap();
2692 // Unaffected by the new field: still just the first named id.
2693 assert_eq!(roles.implementers.len(), 1);
2694 assert_eq!(roles.implementers[0].id, "c");
2695 let roster: Vec<&str> = roles
2696 .implementer_roster
2697 .iter()
2698 .map(|a| a.id.as_str())
2699 .collect();
2700 assert_eq!(roster, ["c", "a"]);
2701 }
2702
2703 #[test]
2704 fn explicit_roles_win() {
2705 let cfg = Config {
2706 agents: vec![spec("a"), spec("b")],
2707 roles: Roles {
2708 implementers: vec!["b".to_owned()],
2709 judges: vec!["a".to_owned()],
2710 reviewers: Vec::new(),
2711 fixer: Some("a".into()),
2712 ..Roles::default()
2713 },
2714 ..Config::default()
2715 };
2716 let roles = cfg.resolve_roles().unwrap();
2717 assert!(roles.implementers.iter().all(|a| a.id == "b"));
2718 assert!(roles.judges.iter().all(|a| a.id == "a"));
2719 let fixer = roles.fixer.unwrap();
2720 assert_eq!(fixer.len(), 1);
2721 assert_eq!(fixer[0].id, "a");
2722 assert_eq!(roles.conductor.id, "a");
2723 }
2724
2725 #[test]
2726 fn fixer_takes_a_string_or_an_ordered_chain() {
2727 let one: Roles = toml::from_str("fixer = \"a\"").unwrap();
2728 assert_eq!(one.fixer.as_ref().unwrap().ids(), ["a"]);
2729 let chain: Roles = toml::from_str("fixer = [\"b\", \"a\"]").unwrap();
2730 assert_eq!(chain.fixer.as_ref().unwrap().ids(), ["b", "a"]);
2731
2732 let resolve = |fixer: Option<AgentChoice>| {
2733 Config {
2734 agents: vec![spec("a"), spec("b"), spec("c")],
2735 roles: Roles {
2736 fixer,
2737 ..Roles::default()
2738 },
2739 ..Config::default()
2740 }
2741 .resolve_roles()
2742 };
2743 let ids = |r: &ResolvedRoles| -> Vec<String> {
2744 r.fixer
2745 .as_ref()
2746 .unwrap()
2747 .iter()
2748 .map(|s| s.id.clone())
2749 .collect()
2750 };
2751 // The string form is a chain of one; unset stays unset (the winner's author).
2752 assert_eq!(ids(&resolve(Some("c".into())).unwrap()), ["c"]);
2753 assert!(resolve(None).unwrap().fixer.is_none());
2754 // Order kept, an unknown id skipped, a duplicate keeps its first place.
2755 let chain = AgentChoice::Chain(vec!["b".into(), "zzz".into(), "a".into(), "b".into()]);
2756 assert_eq!(ids(&resolve(Some(chain)).unwrap()), ["b", "a"]);
2757 // Nothing resolving names the role.
2758 let err = resolve(Some(AgentChoice::Chain(vec!["zzz".into()]))).unwrap_err();
2759 assert!(format!("{err:#}").contains("fixer"), "{err:#}");
2760 }
2761
2762 #[test]
2763 fn unknown_agent_id_is_an_error() {
2764 let cfg = Config {
2765 agents: vec![spec("a")],
2766 roles: Roles {
2767 judges: vec!["nope".to_owned()],
2768 ..Roles::default()
2769 },
2770 ..Config::default()
2771 };
2772 assert!(cfg.resolve_roles().is_err());
2773 }
2774
2775 #[test]
2776 fn conductor_role_is_resolved_validated_and_has_a_fallback() {
2777 let mut cfg = Config {
2778 agents: vec![spec("a"), spec("b")],
2779 ..Config::default()
2780 };
2781 assert_eq!(cfg.resolve_roles().unwrap().conductor.id, "a");
2782
2783 cfg.roles.conductor = Some("b".into());
2784 assert_eq!(cfg.resolve_roles().unwrap().conductor.id, "b");
2785
2786 cfg.roles.conductor = Some("missing".into());
2787 assert!(cfg.resolve_roles().is_err());
2788 }
2789
2790 #[test]
2791 fn a_standalone_role_takes_a_string_or_an_array_and_round_trips() {
2792 let one: Roles = toml::from_str("synthesizer = \"opus\"").unwrap();
2793 assert_eq!(one.synthesizer, Some(AgentChoice::One("opus".to_owned())));
2794 assert_eq!(primary(one.synthesizer.as_ref()), Some("opus"));
2795
2796 let chain: Roles = toml::from_str("chatter = [\"opus\", \"codex\"]").unwrap();
2797 assert_eq!(chain.chatter.as_ref().unwrap().ids(), ["opus", "codex"]);
2798 assert_eq!(primary(chain.chatter.as_ref()), Some("opus"));
2799
2800 for roles in [one, chain] {
2801 let text = toml::to_string(&roles).unwrap();
2802 let back: Roles = toml::from_str(&text).unwrap();
2803 assert_eq!(back.synthesizer, roles.synthesizer);
2804 assert_eq!(back.chatter, roles.chatter);
2805 }
2806 let text = toml::to_string(&Roles {
2807 conductor: Some("x".into()),
2808 ..Roles::default()
2809 })
2810 .unwrap();
2811 assert!(
2812 text.contains("conductor = \"x\""),
2813 "string stays a string: {text}"
2814 );
2815 assert!(toml::from_str::<Roles>("synthesizer = 3").is_err());
2816 }
2817
2818 #[test]
2819 fn an_empty_choice_counts_as_unset() {
2820 assert!(AgentChoice::Chain(Vec::new()).ids().is_empty());
2821 assert!(AgentChoice::One(" ".to_owned()).ids().is_empty());
2822 assert_eq!(primary(Some(&AgentChoice::Chain(vec![]))), None);
2823 }
2824
2825 #[test]
2826 fn synthesizer_role_is_a_lazy_lookup_with_the_same_fallback_as_chatter() {
2827 // Unlike `conductor`, `synthesizer` is never validated by
2828 // `resolve_roles` — it is looked up lazily by
2829 // `graph::Runner::synthesize_brief` the same way `[roles] chatter`
2830 // is looked up by `talk::begin`, so this exercises `agent::pick`
2831 // directly instead of going through `resolve_roles`.
2832 let mut cfg = Config {
2833 agents: vec![spec("a"), spec("b")],
2834 ..Config::default()
2835 };
2836 let want = primary(cfg.roles.synthesizer.as_ref());
2837 assert_eq!(
2838 crate::agent::pick(&cfg.agents, want, &crate::agent::installed)
2839 .unwrap()
2840 .id,
2841 "a",
2842 "unset falls back to agent::pick's own default order"
2843 );
2844
2845 cfg.roles.synthesizer = Some("b".into());
2846 let want = primary(cfg.roles.synthesizer.as_ref());
2847 assert_eq!(
2848 crate::agent::pick(&cfg.agents, want, &crate::agent::installed)
2849 .unwrap()
2850 .id,
2851 "b",
2852 "the named agent wins over the default order"
2853 );
2854 }
2855
2856 #[test]
2857 fn empty_roster_is_an_error() {
2858 assert!(Config::default().resolve_roles().is_err());
2859 }
2860
2861 #[test]
2862 fn advise_defaults_to_on_with_three_proposals() {
2863 let g = Graph::default();
2864 assert!(g.advise);
2865 assert_eq!(g.advisors, 3);
2866 }
2867
2868 #[test]
2869 fn unset_advisors_falls_back_to_the_judge_roster() {
2870 let cfg = Config {
2871 agents: vec![spec("a"), spec("b")],
2872 roles: Roles {
2873 judges: vec!["b".to_owned()],
2874 ..Roles::default()
2875 },
2876 graph: Graph {
2877 advisors: 2,
2878 ..Graph::default()
2879 },
2880 ..Config::default()
2881 };
2882 let advisors = cfg.advisors().expect("advisors resolve");
2883 assert_eq!(advisors.len(), 2);
2884 assert!(
2885 advisors.iter().all(|a| a.id == "b"),
2886 "an unset [roles] advisors must fall back to [roles] judges: {advisors:?}"
2887 );
2888 }
2889
2890 #[test]
2891 fn an_explicit_advisor_roster_wins_over_the_judge_fallback() {
2892 let cfg = Config {
2893 agents: vec![spec("a"), spec("b")],
2894 roles: Roles {
2895 judges: vec!["b".to_owned()],
2896 advisors: vec!["a".to_owned()],
2897 ..Roles::default()
2898 },
2899 graph: Graph {
2900 advisors: 2,
2901 ..Graph::default()
2902 },
2903 ..Config::default()
2904 };
2905 let advisors = cfg.advisors().expect("advisors resolve");
2906 assert!(advisors.iter().all(|a| a.id == "a"));
2907 }
2908
2909 /// Neither `[roles] advisors` nor `[roles] judges` set: an unconfigured
2910 /// advisor roster must resolve to the exact same agents an unconfigured
2911 /// judge panel would get - same ids, same rotation offset - or the
2912 /// promise in [`Roles::advisors`]'s doc ("advisor diversity for free")
2913 /// does not actually hold.
2914 #[test]
2915 fn an_unconfigured_advisor_and_judge_roster_resolve_to_the_same_agents() {
2916 let cfg = Config {
2917 agents: vec![spec("a"), spec("b"), spec("c")],
2918 graph: Graph {
2919 advisors: 3,
2920 judges: 3,
2921 ..Graph::default()
2922 },
2923 ..Config::default()
2924 };
2925 let advisors = cfg.advisors().expect("advisors resolve");
2926 let judges = cfg.resolve_roles().expect("roles resolve").judges;
2927 let advisor_ids: Vec<&str> = advisors.iter().map(|a| a.id.as_str()).collect();
2928 let judge_ids: Vec<&str> = judges.iter().map(|a| a.id.as_str()).collect();
2929 assert_eq!(
2930 advisor_ids, judge_ids,
2931 "an unconfigured advisor roster must be the same seats an unconfigured judge panel gets"
2932 );
2933 }
2934
2935 #[test]
2936 fn an_unresolvable_advisor_seat_is_an_error_naming_the_id() {
2937 let cfg = Config {
2938 agents: vec![spec("a")],
2939 roles: Roles {
2940 advisors: vec!["nope".to_owned()],
2941 ..Roles::default()
2942 },
2943 graph: Graph {
2944 advisors: 1,
2945 ..Graph::default()
2946 },
2947 ..Config::default()
2948 };
2949 let err = cfg.advisors().expect_err("`nope` is not in the roster");
2950 assert!(format!("{err:#}").contains("nope"));
2951 }
2952
2953 #[test]
2954 fn repos_default_to_no_roots_and_a_day_of_trust() {
2955 assert_eq!(Config::default().repos.roots, Vec::<PathBuf>::new());
2956 assert_eq!(Config::default().repos.scan_ttl, 86_400);
2957 assert_eq!(Config::default().repos.fetch_interval, 600);
2958 }
2959
2960 #[test]
2961 fn a_config_file_with_no_repos_table_still_loads() {
2962 let dir = tempfile::tempdir().unwrap();
2963 let path = dir.path().join("magi.toml");
2964 std::fs::write(&path, "[graph]\ncandidates = 2\n").unwrap();
2965 let cfg = Config::load(&path).expect("must load without [repos]");
2966 assert_eq!(cfg.repos.roots, Vec::<PathBuf>::new());
2967 assert_eq!(cfg.repos.scan_ttl, 86_400);
2968 }
2969
2970 /// A fixture is the whole config under test.
2971 ///
2972 /// `layers` used to reach for `dirs::config_dir()` unconditionally, so on
2973 /// a machine where somebody had written `<config_dir>/magi/config.toml`
2974 /// the suite silently loaded it as the lowest layer. Adding `[repos]
2975 /// roots` there turned two web tests red - their fixtures declare
2976 /// `[repos] roots` too, and `refuse_split_arrays` rightly refuses one
2977 /// array key spread across two layers. CI stayed green throughout,
2978 /// because a runner has no such file: the suite failed only where magi is
2979 /// actually used.
2980 ///
2981 /// So a test build has no machine layer unless it asks for one, and this
2982 /// is that promise. Written against a real file at the real location so
2983 /// it fails if `machine_layer` starts reading it again.
2984 #[test]
2985 fn a_test_build_does_not_read_the_operators_machine_config() {
2986 let repo = tempfile::tempdir().unwrap();
2987 std::fs::write(repo.path().join("magi.toml"), "[graph]\ncandidates = 2\n").unwrap();
2988
2989 let layers = Config::layers(repo.path());
2990 assert_eq!(
2991 layers,
2992 vec![repo.path().join("magi.toml")],
2993 "only the fixture's own file may be a layer"
2994 );
2995 if let Some(real) = dirs::config_dir() {
2996 let machine = real.join("magi").join("config.toml");
2997 assert!(
2998 !layers.contains(&machine),
2999 "the operator's {} must not be a layer in a test build",
3000 machine.display()
3001 );
3002 }
3003 }
3004
3005 #[test]
3006 fn starter_toml_loads_through_teravars() {
3007 let dir = tempfile::tempdir().unwrap();
3008 let path = dir.path().join("magi.toml");
3009 std::fs::write(&path, Config::starter_toml()).unwrap();
3010 let parsed = Config::load(&path).expect("starter config must load");
3011 assert_eq!(parsed.graph.implementers, 3);
3012 assert_eq!(parsed.merge.mode, MergeMode::None);
3013 assert_eq!(parsed.merge.style, MergeStyle::Merge);
3014 assert!(parsed.graph.sessions);
3015 assert!(parsed.graph.github_text_guard);
3016 assert_eq!(parsed.graph.timeout_review, 1200);
3017 assert_eq!(parsed.graph.verify_timeout(), 1200);
3018 assert_eq!(parsed.update.mode, UpdateMode::Notify);
3019 }
3020
3021 #[test]
3022 fn release_defaults_to_actions_and_the_starter_documents_it() {
3023 let dir = tempfile::tempdir().unwrap();
3024 let path = dir.path().join("magi.toml");
3025 std::fs::write(&path, Config::starter_toml()).unwrap();
3026 let r = Config::load(&path).unwrap().release;
3027 assert_eq!(r.mode, ReleaseMode::Actions);
3028 assert!(!r.is_local());
3029 assert!(r.commands.is_empty());
3030 assert_eq!(r.timeout_minutes, 30);
3031 assert!(Config::starter_toml().contains("# mode = \"actions\""));
3032 }
3033
3034 #[test]
3035 fn release_local_parses_and_unknown_keys_are_refused() {
3036 let dir = tempfile::tempdir().unwrap();
3037 let path = dir.path().join("magi.toml");
3038 std::fs::write(
3039 &path,
3040 "[release]\nmode = \"local\"\ncommands = [\"true\", \"echo $MAGI_RELEASE_TAG\"]\ntimeout_minutes = 5\n",
3041 )
3042 .unwrap();
3043 let r = Config::load(&path).unwrap().release;
3044 assert!(r.is_local());
3045 assert_eq!(r.commands.len(), 2);
3046 assert_eq!(r.timeout_minutes, 5);
3047 std::fs::write(&path, "[release]\nmode = \"local\"\nbogus = 1\n").unwrap();
3048 assert!(Config::load(&path).is_err());
3049 std::fs::write(&path, "[release]\nmode = \"gha\"\n").unwrap();
3050 assert!(Config::load(&path).is_err());
3051 }
3052
3053 #[test]
3054 fn pre_gate_defaults_to_empty_and_the_starter_documents_it() {
3055 let dir = tempfile::tempdir().unwrap();
3056 let path = dir.path().join("magi.toml");
3057 std::fs::write(&path, Config::starter_toml()).unwrap();
3058 assert!(Config::load(&path).unwrap().verify.pre_gate.is_empty());
3059 assert!(Config::starter_toml().contains("# pre_gate = []"));
3060 }
3061
3062 #[test]
3063 fn starter_toml_explains_inherited_and_explicit_verify_timeouts() {
3064 let starter = Config::starter_toml();
3065 assert!(starter.contains("When timeout_verify is omitted, E2E and"));
3066 assert!(starter.contains("verification independent if timeout_review changes later"));
3067 assert!(starter.contains("# timeout_verify = 1200"));
3068 }
3069
3070 #[test]
3071 fn starter_toml_documents_chat_personas() {
3072 let starter = Config::starter_toml();
3073 assert!(starter.contains("# [[talk.personas]]"));
3074 assert!(starter.contains("tone only"));
3075 }
3076
3077 #[test]
3078 fn talk_personas_parse_and_are_validated() {
3079 let dir = tempfile::tempdir().unwrap();
3080 let path = dir.path().join("magi.toml");
3081 std::fs::write(
3082 &path,
3083 "[[talk.personas]]\nid = \"rei\"\nname = \"Quiet\"\nprompt = \"Be brief.\"\n\n\
3084 [[talk.personas]]\nid = \"gendo\"\nname = \"Gendo\"\nprompt = \"Be cold.\"\n",
3085 )
3086 .unwrap();
3087 let cfg = Config::load(&path).expect("loads");
3088 assert_eq!(cfg.talk.personas.len(), 2);
3089 let catalog = crate::persona::catalog(&cfg.talk.personas);
3090 let rei = catalog.iter().find(|p| p.id == "rei").unwrap();
3091 assert_eq!(
3092 (rei.name.as_str(), rei.prompt.as_str()),
3093 ("Quiet", "Be brief.")
3094 );
3095 assert_eq!(catalog.last().unwrap().id, "gendo");
3096
3097 for (body, needle) in [
3098 (
3099 "[[talk.personas]]\nid = \"\"\nname = \"n\"\nprompt = \"p\"\n",
3100 "empty `id`",
3101 ),
3102 (
3103 "[[talk.personas]]\nid = \"a\"\nname = \"n\"\nprompt = \" \"\n",
3104 "empty `prompt`",
3105 ),
3106 (
3107 "[[talk.personas]]\nid = \"a\"\nname = \"n\"\nprompt = \"p\"\n\
3108 [[talk.personas]]\nid = \"a\"\nname = \"m\"\nprompt = \"q\"\n",
3109 "more than once",
3110 ),
3111 (
3112 "[[talk.personas]]\nid = \"default\"\nname = \"n\"\nprompt = \"p\"\n",
3113 "cannot be redefined",
3114 ),
3115 ] {
3116 std::fs::write(&path, body).unwrap();
3117 let err = format!("{:#}", Config::load(&path).unwrap_err());
3118 assert!(err.contains(needle), "{needle}: {err}");
3119 }
3120 }
3121
3122 #[test]
3123 fn talk_operator_name_is_trimmed_and_validated() {
3124 let dir = tempfile::tempdir().unwrap();
3125 let path = dir.path().join("magi.toml");
3126 let load = |body: &str| {
3127 std::fs::write(&path, body).unwrap();
3128 Config::load(&path)
3129 };
3130 let ok = load("[talk]\noperator_name = \" Commander \"\n").expect("ok");
3131 assert_eq!(ok.talk.operator_name(), Some("Commander"));
3132 let blank = load("[talk]\noperator_name = \" \"\n").expect("blank");
3133 assert_eq!(blank.talk.operator_name(), None);
3134 assert!(load("[talk]\noperator_name = \"a\\nb\"\n").is_err());
3135 let long = "x".repeat(65);
3136 assert!(load(&format!("[talk]\noperator_name = \"{long}\"\n")).is_err());
3137 }
3138
3139 #[test]
3140 fn talk_personas_in_two_layers_are_refused() {
3141 let dir = tempfile::tempdir().unwrap();
3142 let machine = dir.path().join("machine.toml");
3143 let repo = dir.path().join("repo.toml");
3144 let one = "[[talk.personas]]\nid = \"a\"\nname = \"n\"\nprompt = \"p\"\n";
3145 std::fs::write(&machine, one).unwrap();
3146 std::fs::write(&repo, one.replace("\"a\"", "\"b\"")).unwrap();
3147 let err = format!("{:#}", Config::load_layers(&[machine, repo]).unwrap_err());
3148 assert!(err.contains("talk.personas"), "{err}");
3149 }
3150
3151 #[test]
3152 fn starter_toml_documents_role_lists_as_backup_order() {
3153 let starter = Config::starter_toml();
3154 assert!(starter.contains("also an ordered backup list"));
3155 assert!(starter.contains("codex then sonnet are spares"));
3156 }
3157
3158 /// A repository whose ruleset forbids merge commits declares that once,
3159 /// here, rather than magi asking GitHub about it on every render (see
3160 /// [`MergeStyle`]'s own doc for why).
3161 #[test]
3162 fn a_repository_can_declare_a_linear_history_merge_style() {
3163 let dir = tempfile::tempdir().unwrap();
3164 let path = dir.path().join("magi.toml");
3165 std::fs::write(&path, "[merge]\nmode = \"none\"\nstyle = \"squash\"\n").unwrap();
3166 let parsed = Config::load(&path).expect("config must load");
3167 assert_eq!(parsed.merge.style, MergeStyle::Squash);
3168 }
3169
3170 #[test]
3171 fn later_layers_win_and_vars_render() {
3172 let dir = tempfile::tempdir().unwrap();
3173 let machine = dir.path().join("machine.toml");
3174 let project = dir.path().join("magi.toml");
3175 // The machine layer owns the roster...
3176 std::fs::write(
3177 &machine,
3178 "[[agents]]\nid = \"opus\"\nkind = \"claude\"\nmodel = \"opus\"\n\n\
3179 [graph]\ncandidates = 3\nmax_parallel = 8\n",
3180 )
3181 .unwrap();
3182 // ...and the project layer only states what is repo-specific, plus a
3183 // `[vars]` value interpolated into a command.
3184 std::fs::write(
3185 &project,
3186 "[vars]\ncache = \"/shared\"\n\n\
3187 [graph]\ncandidates = 2\n\n\
3188 [verify]\ngate = [\"CARGO_TARGET_DIR={{ vars.cache }}/t cargo test\"]\n",
3189 )
3190 .unwrap();
3191
3192 let cfg = Config::load_layers(&[machine, project]).expect("layered load");
3193 assert_eq!(cfg.agents.len(), 1, "roster comes from the machine layer");
3194 assert_eq!(cfg.graph.implementers, 2, "project layer wins");
3195 assert_eq!(cfg.graph.max_parallel, 8, "machine layer survives");
3196 assert_eq!(
3197 cfg.verify.gate,
3198 ["CARGO_TARGET_DIR=/shared/t cargo test".to_owned()]
3199 );
3200 // The rendered command is where the cache path is read back from.
3201 assert_eq!(cfg.cache_dir(), Some(PathBuf::from("/shared/t")));
3202 }
3203
3204 #[test]
3205 fn talk_defaults_to_an_hour_and_an_unwritten_config_still_gets_it() {
3206 // An operator who writes no `[graph]` timeout keys at all must still
3207 // land on the hour, not on the five/fifteen minutes this turn used
3208 // to hardcode before it read from config.
3209 let g = Graph::default();
3210 assert_eq!(g.timeout_talk, 3600);
3211
3212 let dir = tempfile::tempdir().unwrap();
3213 let path = dir.path().join("magi.toml");
3214 std::fs::write(&path, "[graph]\ncandidates = 2\n").unwrap();
3215 let cfg = Config::load(&path).expect("must load without timeout_talk set");
3216 assert_eq!(cfg.graph.timeout_talk, 3600);
3217 }
3218
3219 #[test]
3220 fn an_overridden_talk_timeout_reaches_the_loaded_config() {
3221 let dir = tempfile::tempdir().unwrap();
3222 let path = dir.path().join("magi.toml");
3223 std::fs::write(&path, "[graph]\ntimeout_talk = 120\n").unwrap();
3224 let cfg = Config::load(&path).expect("must load");
3225 assert_eq!(cfg.graph.timeout_talk, 120);
3226 }
3227
3228 #[test]
3229 fn the_disk_defaults_are_the_measurements_made_up_front() {
3230 let cfg = Config::default();
3231 assert_eq!(cfg.disk.min_free_bytes, 8 * 1024 * 1024 * 1024);
3232 assert!(cfg.disk.auto_fold);
3233 assert_eq!(cfg.disk.fold_grace_secs, 6 * 60 * 60);
3234 assert_eq!(cfg.disk.cache_limit_bytes, 10 * 1024 * 1024 * 1024);
3235 }
3236
3237 #[test]
3238 fn an_unset_disk_section_is_the_safe_default() {
3239 let dir = tempfile::tempdir().unwrap();
3240 std::fs::write(dir.path().join("magi.toml"), "[graph]\ncandidates = 1\n").unwrap();
3241 let cfg = Config::load(&dir.path().join("magi.toml")).expect("load");
3242 assert_eq!(cfg.disk, Disk::default());
3243 }
3244
3245 #[test]
3246 fn env_is_available_to_templates_with_a_default() {
3247 let dir = tempfile::tempdir().unwrap();
3248 let path = dir.path().join("magi.toml");
3249 // teravars ships no `env`; magi adds it, and the `default` filter has
3250 // to cover the unset case or every machine would need the variable.
3251 //
3252 // Deliberately no named variable: `env` is keyed by the exact spelling
3253 // the OS reports, and Windows says `Path` where POSIX says `PATH`, so a
3254 // test asserting `env.PATH` passes on one runner and fails on another.
3255 // The map's non-emptiness is the platform-neutral claim.
3256 std::fs::write(
3257 &path,
3258 "[verify]\n\
3259 gate = [\"cache={{ env.MAGI_TEST_UNSET_XYZ | default(value='fallback') }}\", \
3260 \"populated={{ env | length > 0 }}\"]\n",
3261 )
3262 .unwrap();
3263 let cfg = Config::load(&path).expect("env lookup must render");
3264 assert_eq!(cfg.verify.gate[0], "cache=fallback");
3265 assert_eq!(cfg.verify.gate[1], "populated=true");
3266 }
3267
3268 #[test]
3269 fn a_broken_template_names_the_file() {
3270 let dir = tempfile::tempdir().unwrap();
3271 let path = dir.path().join("magi.toml");
3272 std::fs::write(&path, "[graph]\nlanguage = \"{{ nope.\"\n").unwrap();
3273 let err = Config::load(&path).expect_err("must not silently ignore");
3274 assert!(err.to_string().contains("teravars"), "{err}");
3275 }
3276
3277 #[test]
3278 fn tera_syntax_in_comments_is_inert() {
3279 // teravars >= 0.2.2 strips `#` comments before Tera sees the file, so a
3280 // comment may quote template syntax without rendering. Before 0.2.2 this
3281 // load failed: the commented-out braces reached the template parser.
3282 let dir = tempfile::tempdir().unwrap();
3283 let path = dir.path().join("magi.toml");
3284 std::fs::write(
3285 &path,
3286 "# a comment may quote templates: `{{ env.NOPE | default(value='x') }}` and `{% if %}`\n\
3287 [graph]\ncandidates = 2\n",
3288 )
3289 .unwrap();
3290 let cfg = Config::load(&path).expect("comments must be inert, not rendered");
3291 assert_eq!(cfg.graph.implementers, 2);
3292 }
3293
3294 #[test]
3295 fn opencode_defaults_to_file_delivery() {
3296 let mut s = spec("oc");
3297 s.kind = AgentKind::Opencode;
3298 assert_eq!(s.delivery(), Delivery::File);
3299 s.prompt_delivery = Some(Delivery::Argv);
3300 assert_eq!(s.delivery(), Delivery::Argv);
3301 }
3302 #[test]
3303 fn the_land_loop_is_on_but_it_cannot_merge_without_being_asked() {
3304 // Both default on, and that pair is the safety property: `land` takes
3305 // over the watching an operator was doing by hand, `land_approval`
3306 // keeps the irreversible step a human decision. An unattended merge
3307 // needs BOTH flipped, which has to be chosen deliberately twice.
3308 let g = Graph::default();
3309 assert!(
3310 g.land,
3311 "stopping at an open PR left the watching to a human"
3312 );
3313 assert!(
3314 g.land_approval,
3315 "on-by-default land is only defensible while this is also on"
3316 );
3317 assert!(g.land_rounds > 0, "a loop with no budget never terminates");
3318 assert!(
3319 g.hold_contested_merge,
3320 "a contested hand-off must still reach a human when approvals are off"
3321 );
3322 }
3323 #[test]
3324 fn an_array_declared_in_two_layers_is_refused_instead_of_concatenated() {
3325 // teravars appends arrays. For an ordered list of seats, or an argv,
3326 // the concatenation is something neither file says - and the operator
3327 // pays for the extra seats by the token.
3328 let dir = tempfile::tempdir().unwrap();
3329 let machine = dir.path().join("machine.toml");
3330 let repo = dir.path().join("magi.toml");
3331 std::fs::write(&machine, "[roles]\nimplementers = [\"a\", \"b\"]\n").unwrap();
3332 std::fs::write(&repo, "[roles]\nimplementers = [\"oc\"]\n").unwrap();
3333
3334 let err = Config::load_layers(&[machine.clone(), repo.clone()])
3335 .expect_err("two layers naming one array must not merge silently")
3336 .to_string();
3337 assert!(err.contains("roles.implementers"), "{err}");
3338 // Both files are named: the fix is to delete one of them, and the
3339 // operator has to know which two to choose between.
3340 assert!(err.contains("machine.toml"), "{err}");
3341 assert!(err.contains("magi.toml"), "{err}");
3342 }
3343
3344 #[test]
3345 fn a_scalar_in_one_layer_and_an_array_in_another_still_merges() {
3346 // The split the layering exists for: state a preference machine-wide,
3347 // let the repository own its own lists.
3348 let dir = tempfile::tempdir().unwrap();
3349 let machine = dir.path().join("machine.toml");
3350 let repo = dir.path().join("magi.toml");
3351 std::fs::write(&machine, "[roles]\nchatter = \"opus\"\n").unwrap();
3352 std::fs::write(
3353 &repo,
3354 "[[agents]]\nid = \"oc\"\nkind = \"opencode\"\n\n\
3355 [roles]\nimplementers = [\"oc\"]\n",
3356 )
3357 .unwrap();
3358
3359 let cfg = Config::load_layers(&[machine, repo]).expect("layers merge");
3360 assert_eq!(primary(cfg.roles.chatter.as_ref()), Some("opus"));
3361 assert_eq!(cfg.roles.implementers, ["oc"]);
3362 assert_eq!(cfg.agents.len(), 1, "the roster is not doubled");
3363 }
3364
3365 #[test]
3366 fn two_layers_declaring_verify_gate_run_both_in_priority_order() {
3367 // The `editorconfig-checker` distribution problem: a shared layer
3368 // wants to add a gate command without erasing the repository's own.
3369 let dir = tempfile::tempdir().unwrap();
3370 let machine = dir.path().join("machine.toml");
3371 let repo = dir.path().join("magi.toml");
3372 std::fs::write(&machine, "[verify]\ngate = [\"editorconfig-checker\"]\n").unwrap();
3373 std::fs::write(&repo, "[verify]\ngate = [\"cargo make check\"]\n").unwrap();
3374
3375 let cfg = Config::load_layers(&[machine, repo]).expect("appendable arrays must merge");
3376 assert_eq!(
3377 cfg.verify.gate,
3378 [
3379 "editorconfig-checker".to_owned(),
3380 "cargo make check".to_owned()
3381 ],
3382 "low-priority (machine) command first, high-priority (repo) command after"
3383 );
3384 }
3385
3386 fn load_one(body: &str) -> Result<Config> {
3387 let dir = tempfile::tempdir().unwrap();
3388 let path = dir.path().join("magi.toml");
3389 std::fs::write(&path, body).unwrap();
3390 Config::load_layers(&[path])
3391 }
3392
3393 #[test]
3394 fn worktree_setup_defaults_to_empty_and_parses_steps() {
3395 assert!(Config::default().worktree.setup.is_empty());
3396 let cfg = load_one(
3397 "[worktree]\nsetup = [\n { copy = \".env.example -> .env\" },\n \
3398 { copy = \"x\", optional = true },\n { run = \"true\", timeout_secs = 5 },\n]\n",
3399 )
3400 .unwrap();
3401 assert_eq!(cfg.worktree.setup.len(), 3);
3402 assert!(cfg.worktree.setup[1].optional);
3403 assert_eq!(cfg.worktree.setup[2].timeout_secs, Some(5));
3404 }
3405
3406 #[test]
3407 fn worktree_setup_rejects_malformed_steps() {
3408 for bad in [
3409 "{ copy = \"a\", run = \"b\" }",
3410 "{ optional = true }",
3411 "{ copy = \"a -> ../b\" }",
3412 "{ copy = \"a -> /abs\" }",
3413 "{ copy = \"a -> .git/config\" }",
3414 "{ run = \"b\", optional = true }",
3415 "{ copy = \"a\", timeout_secs = 3 }",
3416 "{ run = \"b\", bogus = 1 }",
3417 ] {
3418 let body = format!("[worktree]\nsetup = [{bad}]\n");
3419 assert!(load_one(&body).is_err(), "{bad}");
3420 }
3421 }
3422
3423 #[test]
3424 fn worktree_setup_in_two_layers_is_refused() {
3425 let dir = tempfile::tempdir().unwrap();
3426 let machine = dir.path().join("machine.toml");
3427 let repo = dir.path().join("magi.toml");
3428 std::fs::write(&machine, "[worktree]\nsetup = [{ run = \"a\" }]\n").unwrap();
3429 std::fs::write(&repo, "[worktree]\nsetup = [{ run = \"b\" }]\n").unwrap();
3430 assert!(Config::load_layers(&[machine, repo]).is_err());
3431 }
3432
3433 #[test]
3434 fn starter_toml_documents_worktree_setup() {
3435 let starter = Config::starter_toml();
3436 assert!(starter.contains("# [worktree]") && starter.contains("setup = ["));
3437 // The commented example must be valid once uncommented.
3438 let body = starter.replace("# [worktree]", "[worktree]");
3439 let mut out = String::new();
3440 let mut live = false;
3441 for line in body.lines() {
3442 if line == "[worktree]" {
3443 live = true;
3444 } else if live && !line.starts_with("# ") {
3445 live = false;
3446 }
3447 out.push_str(&if live {
3448 line.replacen("# ", "", 1)
3449 } else {
3450 line.to_owned()
3451 });
3452 out.push('\n');
3453 }
3454 let cfg = load_one(&out).expect("uncommented example loads");
3455 assert_eq!(cfg.worktree.setup.len(), 3);
3456 }
3457
3458 #[test]
3459 fn two_layers_declaring_verify_e2e_run_both_in_priority_order() {
3460 let dir = tempfile::tempdir().unwrap();
3461 let machine = dir.path().join("machine.toml");
3462 let repo = dir.path().join("magi.toml");
3463 std::fs::write(&machine, "[verify]\ne2e = [\"shared-smoke-test\"]\n").unwrap();
3464 std::fs::write(&repo, "[verify]\ne2e = [\"cargo test\"]\n").unwrap();
3465
3466 let cfg = Config::load_layers(&[machine, repo]).expect("appendable arrays must merge");
3467 assert_eq!(
3468 cfg.verify.e2e,
3469 ["shared-smoke-test".to_owned(), "cargo test".to_owned()]
3470 );
3471 }
3472
3473 #[test]
3474 fn two_layers_declaring_repos_roots_are_both_scanned() {
3475 let dir = tempfile::tempdir().unwrap();
3476 let machine = dir.path().join("machine.toml");
3477 let repo = dir.path().join("magi.toml");
3478 std::fs::write(&machine, "[repos]\nroots = [\"/machine/root\"]\n").unwrap();
3479 std::fs::write(&repo, "[repos]\nroots = [\"/repo/root\"]\n").unwrap();
3480
3481 let cfg = Config::load_layers(&[machine, repo]).expect("appendable arrays must merge");
3482 assert_eq!(
3483 cfg.repos.roots,
3484 [PathBuf::from("/machine/root"), PathBuf::from("/repo/root")]
3485 );
3486 }
3487
3488 #[test]
3489 fn duplicate_gate_commands_across_layers_both_run() {
3490 // Dropping the duplicate would be a silent surprise; the operator
3491 // sees a slower gate, never a missing one.
3492 let dir = tempfile::tempdir().unwrap();
3493 let machine = dir.path().join("machine.toml");
3494 let repo = dir.path().join("magi.toml");
3495 std::fs::write(&machine, "[verify]\ngate = [\"same-command\"]\n").unwrap();
3496 std::fs::write(&repo, "[verify]\ngate = [\"same-command\"]\n").unwrap();
3497
3498 let cfg = Config::load_layers(&[machine, repo]).expect("appendable arrays must merge");
3499 assert_eq!(
3500 cfg.verify.gate,
3501 ["same-command".to_owned(), "same-command".to_owned()]
3502 );
3503 }
3504
3505 #[test]
3506 fn notify_command_is_still_refused_across_two_layers() {
3507 // An argv, not a set: concatenating two of them is not a program.
3508 let dir = tempfile::tempdir().unwrap();
3509 let machine = dir.path().join("machine.toml");
3510 let repo = dir.path().join("magi.toml");
3511 std::fs::write(&machine, "[notify]\ncommand = [\"ntfy\", \"publish\"]\n").unwrap();
3512 std::fs::write(&repo, "[notify]\ncommand = [\"curl\", \"-X\"]\n").unwrap();
3513
3514 let err = Config::load_layers(&[machine.clone(), repo.clone()])
3515 .expect_err("an argv split across layers must not concatenate")
3516 .to_string();
3517 assert!(err.contains("notify.command"), "{err}");
3518 assert!(err.contains("machine.toml"), "{err}");
3519 assert!(err.contains("magi.toml"), "{err}");
3520 }
3521
3522 #[test]
3523 fn one_layer_declaring_verify_gate_runs_unchanged() {
3524 // The classification must not change behaviour for the configuration
3525 // this very repository has today: exactly one layer names the gate.
3526 let dir = tempfile::tempdir().unwrap();
3527 let path = dir.path().join("magi.toml");
3528 std::fs::write(&path, "[verify]\ngate = [\"cargo make check\"]\n").unwrap();
3529
3530 let cfg = Config::load(&path).expect("single layer must still load");
3531 assert_eq!(cfg.verify.gate, ["cargo make check".to_owned()]);
3532 }
3533
3534 #[test]
3535 fn describe_composed_names_the_contributing_layers_only_when_there_are_two() {
3536 let dir = tempfile::tempdir().unwrap();
3537 let machine = dir.path().join("machine.toml");
3538 let repo = dir.path().join("magi.toml");
3539 std::fs::write(&machine, "[verify]\ngate = [\"editorconfig-checker\"]\n").unwrap();
3540 std::fs::write(&repo, "[verify]\ngate = [\"cargo make check\"]\n").unwrap();
3541 let paths = vec![machine.clone(), repo.clone()];
3542
3543 let cfg = Config::load_layers(&paths).expect("appendable arrays must merge");
3544 let described =
3545 Config::describe_composed(&paths, &cfg.verify.gate, "verify.gate", "(none)");
3546 assert!(described.contains("editorconfig-checker && cargo make check"));
3547 assert!(
3548 described.contains(&machine.display().to_string()),
3549 "{described}"
3550 );
3551 assert!(
3552 described.contains(&repo.display().to_string()),
3553 "{described}"
3554 );
3555
3556 // A single contributing layer stays the plain one-line summary.
3557 let single = vec![repo.clone()];
3558 let solo_cfg = Config::load_layers(&single).expect("single layer loads");
3559 let solo_described =
3560 Config::describe_composed(&single, &solo_cfg.verify.gate, "verify.gate", "(none)");
3561 assert_eq!(solo_described, "cargo make check");
3562 }
3563
3564 mod remote_layers {
3565 use super::*;
3566
3567 fn git(dir: &Path, args: &[&str]) {
3568 use crate::proc::Quiet as _;
3569 let out = std::process::Command::new("git")
3570 .args(args)
3571 .quiet()
3572 .current_dir(dir)
3573 .output()
3574 .unwrap();
3575 assert!(out.status.success(), "git {args:?}: {out:?}");
3576 }
3577
3578 /// A checkout whose origin/main carries `remote` as magi.toml and whose
3579 /// working tree carries `disk`.
3580 fn fixture(remote: &str, disk: &str) -> (tempfile::TempDir, PathBuf) {
3581 let tmp = tempfile::tempdir().unwrap();
3582 let bare = tmp.path().join("o.git");
3583 git(tmp.path(), &["init", "--bare", "-b", "trunk", "o.git"]);
3584 let repo = tmp.path().join("repo");
3585 std::fs::create_dir_all(&repo).unwrap();
3586 git(&repo, &["init", "-b", "trunk"]);
3587 git(&repo, &["config", "user.name", "t"]);
3588 git(&repo, &["config", "user.email", "t@example.com"]);
3589 std::fs::write(repo.join("magi.toml"), remote).unwrap();
3590 git(&repo, &["add", "-A"]);
3591 git(&repo, &["commit", "-m", "x"]);
3592 git(&repo, &["remote", "add", "origin", &bare.to_string_lossy()]);
3593 git(&repo, &["push", "origin", "trunk"]);
3594 git(&repo, &["fetch", "origin"]);
3595 git(&repo, &["remote", "set-head", "origin", "trunk"]);
3596 std::fs::write(repo.join("magi.toml"), disk).unwrap();
3597 (tmp, repo)
3598 }
3599
3600 #[test]
3601 fn reads_the_remote_blob_not_the_working_tree() {
3602 let (_t, repo) = fixture("[graph]\ncandidates = 2\n", "[graph]\ncandidates = 5\n");
3603 let (cfg, _) = Config::discover(&repo, None).unwrap();
3604 assert_eq!(cfg.graph.implementers, 2);
3605 }
3606
3607 #[test]
3608 fn detached_head_still_reads_the_remote() {
3609 let (_t, repo) = fixture("[graph]\ncandidates = 2\n", "[graph]\ncandidates = 5\n");
3610 git(&repo, &["checkout", "--detach"]);
3611 let (cfg, _) = Config::discover(&repo, None).unwrap();
3612 assert_eq!(cfg.graph.implementers, 2);
3613 }
3614
3615 #[test]
3616 fn a_missing_ref_is_an_error_not_a_fallback() {
3617 let (_t, repo) = fixture("[graph]\ncandidates = 2\n", "[graph]\ncandidates = 5\n");
3618 git(&repo, &["update-ref", "-d", "refs/remotes/origin/trunk"]);
3619 let err = format!("{:#}", Config::discover(&repo, None).unwrap_err());
3620 assert!(err.contains("origin"), "{err}");
3621 }
3622
3623 #[test]
3624 fn merge_base_in_the_checkout_locates_the_ref() {
3625 let (_t, repo) = fixture("[graph]\ncandidates = 2\n", "[merge]\nbase = \"nope\"\n");
3626 assert!(Config::discover(&repo, None).is_err());
3627 }
3628
3629 #[test]
3630 fn the_magi_dir_layer_is_read_too() {
3631 let (_t, repo) = fixture("[graph]\ncandidates = 2\n", "");
3632 std::fs::write(repo.join("magi.toml"), "[graph]\ncandidates = 2\n").unwrap();
3633 std::fs::create_dir_all(repo.join(".magi")).unwrap();
3634 std::fs::write(repo.join(".magi/config.toml"), "[graph]\njudges = 3\n").unwrap();
3635 git(&repo, &["add", "-A"]);
3636 git(&repo, &["commit", "-m", "y"]);
3637 git(&repo, &["push", "origin", "trunk"]);
3638 git(&repo, &["fetch", "origin"]);
3639 let (cfg, layers) = Config::discover(&repo, None).unwrap();
3640 assert_eq!((cfg.graph.implementers, cfg.graph.judges), (2, 3));
3641 assert_eq!(layers.len(), 2);
3642 }
3643
3644 #[test]
3645 fn an_include_in_a_repository_layer_resolves_against_the_same_commit() {
3646 let (_t, repo) = fixture("include = [\"gate.toml\"]\n", "");
3647 std::fs::write(repo.join("magi.toml"), "include = [\"gate.toml\"]\n").unwrap();
3648 std::fs::write(repo.join("gate.toml"), "[graph]\ncandidates = 2\n").unwrap();
3649 git(&repo, &["add", "-A"]);
3650 git(&repo, &["commit", "-m", "inc"]);
3651 git(&repo, &["push", "origin", "trunk"]);
3652 git(&repo, &["fetch", "origin"]);
3653 // The working tree's copy differs and must not be used.
3654 std::fs::write(repo.join("gate.toml"), "[graph]\ncandidates = 9\n").unwrap();
3655 let (cfg, _) = Config::discover(&repo, None).unwrap();
3656 assert_eq!(cfg.graph.implementers, 2);
3657 }
3658
3659 #[test]
3660 fn a_machine_layer_base_is_rendered_before_locating_the_ref() {
3661 let (t, repo) = fixture("[graph]\ncandidates = 2\n", "[merge]\nbase = \"nope\"\n");
3662 let m = t.path().join("machine.toml");
3663 std::fs::write(
3664 &m,
3665 "[merge]\nbase = \"{{ env.MAGI_TEST_NO_SUCH | default(value='trunk') }}\"\n",
3666 )
3667 .unwrap();
3668 let (remote, base) = {
3669 let probe = probe_merge(&m, &repo).unwrap();
3670 (probe.remote, probe.base)
3671 };
3672 assert_eq!((remote, base), (None, Some("trunk".to_owned())));
3673 }
3674
3675 #[test]
3676 fn a_machine_layer_base_sees_the_checkouts_repo_name() {
3677 let (t, repo) = fixture("[graph]\ncandidates = 2\n", "");
3678 let m = t.path().join("machine.toml");
3679 std::fs::write(&m, "[merge]\nbase = \"{{ repo_name }}\"\n").unwrap();
3680 let probe = probe_merge(&m, &repo).unwrap();
3681 assert_eq!(probe.base.as_deref(), Some("repo"));
3682 }
3683
3684 #[test]
3685 fn every_include_form_is_recognised() {
3686 assert!(declares_include("include = [\"a.toml\"]\n"));
3687 assert!(declares_include("\"include\" = [\"a.toml\"]\n"));
3688 assert!(declares_include("teravars = { include = [\"a.toml\"] }\n"));
3689 assert!(declares_include("[teravars]\ninclude = [\"a.toml\"]\n"));
3690 assert!(!declares_include("[graph]\ncandidates = 2\n"));
3691 }
3692
3693 #[test]
3694 fn an_explicit_config_is_read_as_named() {
3695 let (t, repo) = fixture("[graph]\ncandidates = 2\n", "");
3696 let own = t.path().join("own.toml");
3697 std::fs::write(&own, "[graph]\ncandidates = 4\n").unwrap();
3698 let (cfg, _) = Config::discover(&repo, Some(&own)).unwrap();
3699 assert_eq!(cfg.graph.implementers, 4);
3700 }
3701
3702 #[test]
3703 fn discover_fetched_fails_when_the_fetch_does() {
3704 let (t, repo) = fixture("[graph]\ncandidates = 2\n", "");
3705 std::fs::remove_dir_all(t.path().join("o.git")).unwrap();
3706 let rt = tokio::runtime::Builder::new_current_thread()
3707 .enable_all()
3708 .build()
3709 .unwrap();
3710 let err = rt
3711 .block_on(Config::discover_fetched(&repo, None))
3712 .unwrap_err();
3713 assert!(format!("{err:#}").contains("fetch"), "{err:#}");
3714 }
3715
3716 #[test]
3717 fn discover_fetched_survives_a_fetch_that_leaves_no_origin_head() {
3718 // Never-fetched remote + followRemoteHEAD=never: fetch creates
3719 // origin/trunk but not origin/HEAD, which discovery then re-reads.
3720 let (_t, repo) = fixture("[graph]\ncandidates = 2\n", "");
3721 git(&repo, &["remote", "set-head", "origin", "-d"]);
3722 git(&repo, &["update-ref", "-d", "refs/remotes/origin/trunk"]);
3723 git(
3724 &repo,
3725 &["config", "remote.origin.followRemoteHEAD", "never"],
3726 );
3727 let rt = tokio::runtime::Builder::new_current_thread()
3728 .enable_all()
3729 .build()
3730 .unwrap();
3731 let (cfg, _) = rt.block_on(Config::discover_fetched(&repo, None)).unwrap();
3732 assert_eq!(cfg.graph.implementers, 2);
3733 }
3734
3735 #[test]
3736 fn a_non_git_directory_reads_its_files() {
3737 let dir = tempfile::tempdir().unwrap();
3738 std::fs::write(dir.path().join("magi.toml"), "[graph]\ncandidates = 4\n").unwrap();
3739 let (cfg, _) = Config::discover(dir.path(), None).unwrap();
3740 assert_eq!(cfg.graph.implementers, 4);
3741 }
3742 }
3743}