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 /// After a run's pull request merges, release the follow-up tasks the
458 /// merge approval's deputy filed held for that pull request
459 /// (`crate::followup::settle_deputy_tasks`): only a task the deputy filed
460 /// for this run whose hold reason names this pull request, exactly once,
461 /// and never one that duplicates an automatic follow-up (that one is left
462 /// held with the reason). The generation cap applies. Independent of
463 /// `file_followups`. On; off leaves them held until `magi task release`.
464 pub release_deputy_followups: bool,
465 /// How long to wait for an owner to answer a question before the run is
466 /// abandoned, seconds. A parked run costs nothing, so this is generous;
467 /// it exists so a forgotten question cannot pin a worktree forever.
468 pub answer_timeout: u64,
469 /// What a round does when one or more reviewer seats never answered
470 /// (timeout, crash, unparsable output).
471 pub incomplete_review: IncompleteReviewPolicy,
472 /// Run `verify.e2e` on every round, even one that already has blocking
473 /// findings and another round left to try.
474 ///
475 /// Off by default: a round with a blocking finding and rounds still left
476 /// is going back to the fixer regardless of what `verify.e2e` says, so
477 /// running it first only spends the round's slowest step (minutes, on a
478 /// Rust repo's `cargo test --all-targets`) on a head about to be
479 /// rewritten anyway. `verify.e2e` still runs once a round has no
480 /// blocking findings left (a round cannot go `clean` without it) and the
481 /// final `verify.gate` always runs on the actual tree that would land —
482 /// deferring is about *when* e2e runs mid-loop, never about skipping it.
483 ///
484 /// Set this to restore the old every-round diagnostic behaviour: e2e
485 /// output from a round that still has blocking findings is occasionally
486 /// useful on its own (a runtime failure a reviewer's panel would not
487 /// have caught by reading), and this is the way back to seeing it every
488 /// round instead of only once the panel has nothing left to flag.
489 pub e2e_every_round: bool,
490 /// Run the design-deliberation stage before `implement`: independent
491 /// advisor seats each sketch a design, and (when at least one produced a
492 /// usable proposal) a synthesis blends them into a brief carried in the
493 /// implementer's prompt. See `graph::Runner::advise`.
494 ///
495 /// On by default. A design sketch is a few paragraphs an agent can write
496 /// without touching the repository, where a full implementation is a
497 /// tool loop that re-reads the codebase on every turn - so three
498 /// sketches, gathered once before `implement` starts, cost a fraction of
499 /// a fourth candidate and buy back a form of the same disagreement
500 /// [`Self::implementers`]'s doc describes moving away from being the
501 /// default, on every run rather than only the ones an operator remembers
502 /// to ask for with `--implementers`.
503 pub advise: bool,
504 /// How many independent design proposals the deliberation stage gathers.
505 /// **Three by default** - the same number [`Self::implementers`]'s doc
506 /// names as the point where a fourth judge's first choice stopped
507 /// changing the tally.
508 pub advisors: usize,
509}
510
511impl Default for Graph {
512 fn default() -> Self {
513 Self {
514 github_text_guard: true,
515 implementers: 1,
516 judges: 3,
517 deliberate_rounds: 1,
518 reviewers: 3,
519 review_rounds: 6,
520 gate_fix_rounds: 1,
521 max_parallel: 4,
522 language: "en".to_owned(),
523 sessions: true,
524 timeout_implement: 3600,
525 timeout_judge: 1200,
526 timeout_review: 1200,
527 timeout_verify: None,
528 timeout_fix: 1800,
529 timeout_talk: 3600,
530 retries: 1,
531 worktree_root: None,
532 land: true,
533 land_rounds: 4,
534 land_approval: true,
535 hold_contested_merge: true,
536 file_followups: true,
537 release_deputy_followups: true,
538 answer_timeout: 86_400,
539 incomplete_review: IncompleteReviewPolicy::Block,
540 e2e_every_round: false,
541 advise: true,
542 advisors: 3,
543 }
544 }
545}
546
547impl Graph {
548 /// Effective machine-command budget. Older configuration files had only
549 /// `timeout_review`, which also governed verification, so absence is a
550 /// compatibility fallback rather than a new 1200-second default.
551 pub fn verify_timeout(&self) -> u64 {
552 self.timeout_verify.unwrap_or(self.timeout_review)
553 }
554}
555
556/// What a review round does when a reviewer seat never answered.
557///
558/// A round where half the panel timed out is not evidence of a clean patch —
559/// it is evidence of nothing. The default refuses to call that clean; `warn`
560/// exists for an operator who would rather keep a flaky seat from stalling
561/// every run, and accepts that the gap is on them to read in the report.
562#[derive(Debug, Clone, Copy, PartialEq, Eq, Deserialize, Serialize)]
563#[serde(rename_all = "lowercase")]
564pub enum IncompleteReviewPolicy {
565 /// A round with a missing seat is never `clean`: with nothing raised to
566 /// fix, the round is re-reviewed instead of gating; with the max rounds
567 /// exhausted, the run is left `Blocked` rather than declared ready.
568 Block,
569 /// A round with a missing seat can still gate as clean, once every seat
570 /// that *did* answer raised nothing blocking and verification is green.
571 /// The record keeps the gap visible (`magi show`, `magi stats`) even
572 /// though the run does not wait on it.
573 Warn,
574}
575
576/// What to do when vendor-identifying text is found in material shown to judges.
577#[derive(Debug, Clone, Copy, PartialEq, Eq, Deserialize, Serialize)]
578#[serde(rename_all = "lowercase")]
579pub enum LeakPolicy {
580 /// Record the leak, show the patch unmodified.
581 Warn,
582 /// Replace the token with `[REDACTED]` in the presented patch.
583 Redact,
584 /// Abort the run.
585 Fail,
586}
587
588/// Blindness policy.
589///
590/// Commit messages and candidate summaries are *always* stripped of
591/// attribution trailers and redacted — that is where signatures actually
592/// appear. [`Blind::on_leak`] governs the patch body only, where blanket
593/// redaction would corrupt the artifact under judgement.
594#[derive(Debug, Clone, Deserialize, Serialize)]
595#[serde(deny_unknown_fields, default)]
596pub struct Blind {
597 /// Install a per-worktree `commit-msg` hook that deletes attribution
598 /// trailers before they can land in a candidate's history.
599 pub commit_msg_hook: bool,
600 /// Literal, case-insensitive substrings. A line containing any of them is
601 /// dropped from commit messages and summaries; the `commit-msg` hook is
602 /// generated from the same list.
603 pub strip_lines: Vec<String>,
604 /// Case-insensitive substrings that identify a vendor or model.
605 pub vendor_tokens: Vec<String>,
606 /// Policy for vendor tokens found in the patch body.
607 pub on_leak: LeakPolicy,
608 /// Seed for label assignment and per-judge presentation order. Derived from
609 /// the run id when unset; set it to make a run reproducible.
610 pub seed: Option<u64>,
611}
612
613impl Default for Blind {
614 fn default() -> Self {
615 Self {
616 commit_msg_hook: true,
617 strip_lines: [
618 "Co-Authored-By:",
619 "Signed-off-by:",
620 "Assisted-by:",
621 "Generated-by:",
622 "Generated with",
623 "\u{1f916}",
624 ]
625 .iter()
626 .map(|s| (*s).to_owned())
627 .collect(),
628 vendor_tokens: [
629 "claude",
630 "anthropic",
631 "codex",
632 "openai",
633 "chatgpt",
634 "gemini",
635 "grok",
636 "xai",
637 "copilot",
638 "opencode",
639 "qoder",
640 "cursor",
641 "\u{1f916}",
642 ]
643 .iter()
644 .map(|s| (*s).to_owned())
645 .collect(),
646 on_leak: LeakPolicy::Warn,
647 seed: None,
648 }
649 }
650}
651
652/// Shell commands that gate the winner.
653#[derive(Debug, Clone, Default, Deserialize, Serialize)]
654#[serde(deny_unknown_fields, default)]
655pub struct Verify {
656 /// Run in the winner's worktree once per review round. Its output is fed
657 /// back to the fixer. This is the "real machine" leg of the review.
658 pub e2e: Vec<String>,
659 /// Final gate. Must all exit 0 before a merge is attempted.
660 pub gate: Vec<String>,
661 /// Mechanical fixers (formatters) run in the winner's worktree just before
662 /// `gate`, after review is clean. Anything they change is committed as one
663 /// neutral-identity commit so the gate and the PR see it. A command that
664 /// fails or times out is a warning, never a failed run: the gate stays the
665 /// single arbiter. Empty (the default) turns the feature off. Runs under
666 /// `graph.verify_timeout()` per command, like the gate. Meant for light
667 /// tools: it is not scanned for `CARGO_TARGET_DIR`, and it should be
668 /// idempotent since a resumed run may repeat it.
669 #[serde(default)]
670 pub pre_gate: Vec<String>,
671 /// Shell used to run the commands above. Defaults to `sh -c`, or
672 /// `cmd /C` when `sh` is not on `PATH`.
673 pub shell: Option<Vec<String>>,
674}
675
676impl Verify {
677 /// The `CARGO_TARGET_DIR=` value of the first rendered command that sets
678 /// one, if any. See [`crate::disk::extract_cargo_target_dir`] for the shape
679 /// this reads back. One rendering is enough - they all set the same
680 /// rendered `{{ vars.cache }}` path via the same shell - and the first e2e
681 /// command is checked before the gate because the e2e rebuilds the crate.
682 pub fn cache_dir(&self) -> Option<PathBuf> {
683 self.e2e
684 .iter()
685 .chain(self.gate.iter())
686 .find_map(|cmd| crate::disk::extract_cargo_target_dir(cmd))
687 }
688
689 /// Every `--manifest-path` the e2e and gate commands pass to cargo, for the
690 /// shared cache's freshness check in repositories with no root manifest.
691 pub fn cargo_manifests(&self) -> Vec<PathBuf> {
692 let all: Vec<String> = self.e2e.iter().chain(self.gate.iter()).cloned().collect();
693 crate::cache::extract_manifest_paths(&all)
694 }
695}
696
697/// Per-repository worktree preparation, run right after magi creates a worktree.
698#[derive(Debug, Clone, Default, PartialEq, Deserialize, Serialize)]
699#[serde(deny_unknown_fields, default)]
700pub struct Worktree {
701 /// Ordered steps. Declare in one layer only. Empty (the default) changes
702 /// nothing. See [`crate::worktree_setup`].
703 pub setup: Vec<SetupStep>,
704}
705
706/// One setup step: exactly one of `copy` / `run`.
707#[derive(Debug, Clone, Default, PartialEq, Deserialize, Serialize)]
708#[serde(deny_unknown_fields, default)]
709pub struct SetupStep {
710 /// `"src -> dst"` (or just `"path"`): copy from the primary checkout (an
711 /// absolute source is allowed) into the worktree. Never overwrites.
712 pub copy: Option<String>,
713 /// With `copy`: a missing source is skipped instead of an error.
714 pub optional: bool,
715 /// Shell command, run with the worktree as cwd.
716 pub run: Option<String>,
717 /// With `run`: seconds before the command is killed. Defaults to
718 /// `graph.verify_timeout()`.
719 pub timeout_secs: Option<u64>,
720}
721
722impl Worktree {
723 /// Refuse a malformed step at load time rather than in the middle of a run.
724 pub fn validate(&self) -> Result<()> {
725 for (i, step) in self.setup.iter().enumerate() {
726 let n = i + 1;
727 match (&step.copy, &step.run) {
728 (Some(c), None) => {
729 if step.timeout_secs.is_some() {
730 bail!("[worktree] setup step {n}: `timeout_secs` applies to `run` only");
731 }
732 crate::worktree_setup::validate_copy(c)
733 .with_context(|| format!("[worktree] setup step {n}"))?;
734 }
735 (None, Some(r)) => {
736 if r.trim().is_empty() {
737 bail!("[worktree] setup step {n}: `run` is empty");
738 }
739 if step.optional {
740 bail!("[worktree] setup step {n}: `optional` applies to `copy` only");
741 }
742 }
743 _ => bail!("[worktree] setup step {n} must set exactly one of `copy` and `run`"),
744 }
745 }
746 Ok(())
747 }
748}
749
750/// Disk hygiene: how hard magi is allowed to press on the machine's free space.
751///
752/// The numbers below come from one incident, not from theory: a machine with
753/// 951.8 GB free ran a few competitions and plans and best read 6.7 GB free.
754/// Three multi-gigabyte classes of junk accumulated side by side - per-run
755/// worktrees that end as `Merged`/`Ready`/`Failed`, a shared build cache whose
756/// each verify round and each implementation wave recompiles the derived
757/// section of the project into, and the outputs of runs that were removed but
758/// whose folders nobody deleted.
759#[derive(Debug, Clone, PartialEq, Deserialize, Serialize)]
760#[serde(deny_unknown_fields, default)]
761pub struct Disk {
762 /// Free space, in bytes, below which no new run may start: the daemon and
763 /// the `magi run` gate answer with "the disk is full" instead of letting
764 /// the graph fill it the rest of the way. `0` turns the gate off.
765 ///
766 /// Default 8 GiB. The incident ran down to 6.7 GB free of 951.8 GB total
767 /// before anybody noticed; 8 GiB is enough headroom for the compile a fresh
768 /// competition triggers and small enough that a 1 TB disk with 100 GB free
769 /// is nowhere near the threshold.
770 pub min_free_bytes: u64,
771 /// Fold finished runs without being asked. `Merged`, `Ready` and `Failed`
772 /// runs older than [`fold_grace_secs`](Self::fold_grace_secs) have their
773 /// worktrees removed. `0` turns the janitor off.
774 ///
775 /// Default true.
776 pub auto_fold: bool,
777 /// How old a finished run must be before the janitor folds it, seconds.
778 ///
779 /// Default 6 hours. A run that `Ready` at 8am is the operator's answer; a
780 /// run that `Ready` a week ago is worktrees holding a compile each. Six
781 /// hours is long enough that nobody loses an answer in the gap between
782 /// reading the report and starting from it, and short enough that a backlog
783 /// cannot pile up across two nights.
784 pub fold_grace_secs: u64,
785 /// Ceiling for the shared build cache (`CARGO_TARGET_DIR` in the rendered
786 /// verify commands), in bytes. When the janitor runs and the cache is over
787 /// it, files are dropped oldest-first until it is not. `0` turns pruning
788 /// off - the cache then only ever grows, which is the operator's call.
789 ///
790 /// Default 10 GiB. This is what the incident measured: 30.61 GB sat in the
791 /// shared cache on top of ~16 GB in the primary target directory and 6.7-
792 /// 11.15 GB in each of four per-worktree targets. 10 GiB holds a healthy
793 /// stack of prebuilt dependencies (cargo's per-file fingerprinting means
794 /// pruning only costs the rebuild of the dropped files, not of the world)
795 /// without letting one addled cache swallow the machine.
796 pub cache_limit_bytes: u64,
797}
798
799impl Default for Disk {
800 fn default() -> Self {
801 Self {
802 min_free_bytes: 8 * 1024 * 1024 * 1024,
803 auto_fold: true,
804 fold_grace_secs: 6 * 60 * 60,
805 cache_limit_bytes: 10 * 1024 * 1024 * 1024,
806 }
807 }
808}
809
810/// What to do with the winning branch.
811#[derive(Debug, Clone, Copy, PartialEq, Eq, Deserialize, Serialize)]
812#[serde(rename_all = "lowercase")]
813pub enum MergeMode {
814 /// Leave the branch alone and print the merge command matching
815 /// [`Merge::style`].
816 None,
817 /// Merge into the base branch in the primary worktree, using
818 /// [`Merge::style`].
819 Local,
820 /// Push the branch and open a PR with `gh pr create`. Landing this PR
821 /// (`[graph] land`) always squashes — see `land`'s module doc — so
822 /// [`Merge::style`] does not apply here.
823 Pr,
824}
825
826/// How the winning branch is attached to the base branch: what
827/// `mode = "local"` runs, and what `mode = "none"`'s printed guidance tells
828/// the operator to run by hand.
829///
830/// Read from configuration rather than asked of the repository at run time
831/// (e.g. `gh api repos/{owner}/{repo}/rulesets`) for two reasons: it keeps
832/// `mode = "none"`'s guidance a pure function of `RunState`, assertable in a
833/// unit test the same way `land::decide` is kept pure (see that module's
834/// doc), and it works for a base branch that is not hosted on GitHub, or not
835/// reachable at all, at the moment the report is rendered. An operator whose
836/// base branch enforces a ruleset already knows what it allows; declaring it
837/// once here is cheaper than magi re-discovering it, with a network call, on
838/// every render.
839#[derive(Debug, Clone, Copy, PartialEq, Eq, Default, Deserialize, Serialize)]
840#[serde(rename_all = "lowercase")]
841pub enum MergeStyle {
842 /// `git merge --no-ff`: every candidate commit lands, plus a merge
843 /// commit that records the merge as its own event in history. Rejected
844 /// by a base branch whose ruleset requires linear history or forbids
845 /// merge commits outright.
846 #[default]
847 Merge,
848 /// `git merge --squash` followed by a commit under an explicit message:
849 /// every candidate commit folds into one, and none of the candidate's
850 /// own placeholder subjects (`magi: candidate A (uncommitted work)`)
851 /// reach the base branch. No merge commit, so this satisfies a linear-
852 /// history ruleset.
853 Squash,
854 /// A fast-forward-only merge: every candidate commit lands verbatim, in
855 /// order, with no merge commit. Only succeeds because the winner was
856 /// already rebased onto the tracked base tip before this runs (see
857 /// `Runner::sync_to_base`) — equivalent to GitHub's "rebase and merge"
858 /// once that has happened.
859 Rebase,
860}
861
862/// Merge policy.
863#[derive(Debug, Clone, Deserialize, Serialize)]
864#[serde(deny_unknown_fields, default)]
865pub struct Merge {
866 /// Default is [`MergeMode::None`]: magi never touches your base branch
867 /// unless you ask it to.
868 pub mode: MergeMode,
869 /// Base branch. Defaults to the branch checked out when the run started.
870 pub base: Option<String>,
871 /// How the winner is attached to `base`; see [`MergeStyle`]. Ignored by
872 /// `mode = "pr"`.
873 pub style: MergeStyle,
874 /// Remote for `mode = "pr"`.
875 pub remote: String,
876 /// After a `mode = "pr"` run lands, open a `chore/release-vX.Y.Z` pull
877 /// request sized to the change by an agent's own judgement, so a version
878 /// bump does not depend on a human remembering to cut one.
879 ///
880 /// Only acts on Rust repositories (a `Cargo.toml` at the root of the base
881 /// branch); on any other repository it is a no-op that records one event.
882 ///
883 /// On by default. **Turning this off means a merge landed from the phone
884 /// never becomes a release**, so `POST /api/upgrade` keeps reporting
885 /// "already on the newest release" against a `main` that has moved past
886 /// it - the same gap this feature exists to close. Only for a repository
887 /// that wants to keep cutting releases by hand.
888 pub release_bump: bool,
889}
890
891impl Default for Merge {
892 fn default() -> Self {
893 Self {
894 mode: MergeMode::None,
895 base: None,
896 style: MergeStyle::default(),
897 remote: "origin".to_owned(),
898 release_bump: true,
899 }
900 }
901}
902
903/// How magi keeps itself current.
904#[derive(Debug, Clone, Copy, PartialEq, Eq, Deserialize, Serialize)]
905#[serde(rename_all = "lowercase")]
906pub enum UpdateMode {
907 /// Never check.
908 Off,
909 /// Check in the background and print a one-line banner when a newer
910 /// release exists.
911 Notify,
912 /// Check and install silently.
913 Install,
914}
915
916/// Self-update policy.
917#[derive(Debug, Clone, Deserialize, Serialize)]
918#[serde(deny_unknown_fields, default)]
919pub struct Update {
920 /// Default is [`UpdateMode::Notify`]: magi tells you, and lets you decide.
921 pub mode: UpdateMode,
922 /// Minimum time between checks, e.g. `24h`. kaishin's default when unset.
923 pub interval: Option<String>,
924}
925
926impl Default for Update {
927 fn default() -> Self {
928 Self {
929 mode: UpdateMode::Notify,
930 interval: None,
931 }
932 }
933}
934
935/// Who performs the release after a `chore/release-vX.Y.Z` pull request merges.
936#[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Deserialize, Serialize)]
937#[serde(rename_all = "lowercase")]
938pub enum ReleaseMode {
939 /// The repository's own workflows do (`auto-tag.yml` then `release.yml`),
940 /// and the release pull request merges on its checks.
941 #[default]
942 Actions,
943 /// magi does, from this machine: no CI is awaited, the owner approves the
944 /// merge, and magi then tags and runs [`Release::commands`]. See
945 /// [`crate::release_local`].
946 Local,
947}
948
949/// `[release]`: releasing without GitHub Actions. Opt-in per repository.
950#[derive(Debug, Clone, Deserialize, Serialize)]
951#[serde(deny_unknown_fields, default)]
952pub struct Release {
953 /// [`ReleaseMode::Actions`] (default: nothing changes) or
954 /// [`ReleaseMode::Local`].
955 pub mode: ReleaseMode,
956 /// Shell commands run in order, from a clean checkout of the merge commit,
957 /// once the tag is pushed. The first failure stops the list. They see
958 /// `MAGI_RELEASE_VERSION`, `MAGI_RELEASE_TAG`, `MAGI_RELEASE_COMMIT` and
959 /// `MAGI_RELEASE_PR` in the environment; `{{ ... }}` cannot carry the
960 /// version because teravars renders the whole file before any release
961 /// exists. Only read when `mode = "local"`.
962 pub commands: Vec<String>,
963 /// Minutes each command may run before it is killed and counted as failed.
964 pub timeout_minutes: u64,
965}
966
967impl Default for Release {
968 fn default() -> Self {
969 Self {
970 mode: ReleaseMode::Actions,
971 commands: Vec::new(),
972 timeout_minutes: 30,
973 }
974 }
975}
976
977impl Release {
978 /// Is this repository released by magi rather than by its workflows?
979 pub fn is_local(&self) -> bool {
980 self.mode == ReleaseMode::Local
981 }
982}
983
984/// Top-level configuration.
985#[derive(Debug, Clone, Default, Deserialize, Serialize)]
986#[serde(deny_unknown_fields, default)]
987pub struct Config {
988 /// Agent roster.
989 pub agents: Vec<AgentSpec>,
990 /// Role assignment.
991 pub roles: Roles,
992 /// Graph shape.
993 pub graph: Graph,
994 /// Blindness policy.
995 pub blind: Blind,
996 /// Verification commands.
997 pub verify: Verify,
998 /// Worktree preparation.
999 pub worktree: Worktree,
1000 /// Disk hygiene.
1001 pub disk: Disk,
1002 /// Merge policy.
1003 pub merge: Merge,
1004 /// Self-update policy.
1005 pub update: Update,
1006 /// Project-specific text appended to the node prompts.
1007 pub prompts: Prompts,
1008 /// How the operator is told a run is waiting on them.
1009 pub notify: Notify,
1010 /// Local repositories the plan surface can start or derive a conversation
1011 /// against.
1012 pub repos: Repos,
1013 /// Policy for the standing conversation ([`crate::talk`]).
1014 pub talk: Talk,
1015 /// How many runs `magi serve`'s own loop drives at once.
1016 pub daemon: Daemon,
1017 /// Who tags and publishes after a release pull request merges.
1018 pub release: Release,
1019 /// Context-window sizes in tokens, keyed by model name, merged over
1020 /// [`BUILTIN_CONTEXT_WINDOWS`]. See [`Config::context_window`].
1021 pub context_windows: BTreeMap<String, u64>,
1022}
1023
1024/// Context windows of models magi commonly drives, in tokens.
1025///
1026/// A convenience, not a source of truth: vendors resize windows and ship new
1027/// models faster than this table is edited, so an entry that is stale or
1028/// missing is fixed by `[context_windows]` in `magi.toml`, which wins. Matched
1029/// by [`Config::context_window`]: exact name first, then the longest key that
1030/// prefixes the model name (so a dated or suffixed id such as
1031/// `claude-sonnet-4-5-20250929` finds `claude-sonnet-4-5`). A model matching
1032/// nothing has no percentage - the UI shows the raw count.
1033const BUILTIN_CONTEXT_WINDOWS: &[(&str, u64)] = &[
1034 ("claude-opus-4", 200_000),
1035 ("claude-sonnet-4", 200_000),
1036 ("claude-haiku-4", 200_000),
1037 ("claude-3", 200_000),
1038 ("opus", 200_000),
1039 ("sonnet", 200_000),
1040 ("haiku", 200_000),
1041 ("gpt-5", 400_000),
1042 ("gpt-4.1", 1_000_000),
1043 ("gpt-4o", 128_000),
1044 ("o3", 200_000),
1045 ("o4-mini", 200_000),
1046 ("gemini-2.5", 1_000_000),
1047 ("deepseek", 128_000),
1048];
1049
1050impl Config {
1051 /// Context window in tokens for `model`: `[context_windows]` first, then
1052 /// the built-in table; within each, an exact name beats the longest
1053 /// prefix. `None` for an unknown model, which is a deliberate answer -
1054 /// guessing a denominator would print a confident wrong percentage.
1055 pub fn context_window(&self, model: &str) -> Option<u64> {
1056 fn lookup<'a>(model: &str, entries: impl Iterator<Item = (&'a str, u64)>) -> Option<u64> {
1057 let mut best: Option<(usize, u64)> = None;
1058 for (key, window) in entries {
1059 if key == model {
1060 return Some(window);
1061 }
1062 if model.starts_with(key) && best.is_none_or(|(len, _)| key.len() > len) {
1063 best = Some((key.len(), window));
1064 }
1065 }
1066 best.map(|(_, w)| w)
1067 }
1068 lookup(
1069 model,
1070 self.context_windows.iter().map(|(k, w)| (k.as_str(), *w)),
1071 )
1072 .or_else(|| lookup(model, BUILTIN_CONTEXT_WINDOWS.iter().copied()))
1073 }
1074}
1075
1076/// How the daemon loop itself behaves, as opposed to what one run does.
1077///
1078/// Machine-layer material in the same sense [`Repos::roots`] is: how many
1079/// competitions this machine's own loop is willing to babysit at once is a
1080/// fact about the machine running `magi serve`, not about any one
1081/// repository's task, so it belongs in `<config_dir>/magi/config.toml`
1082/// rather than a repository's own `magi.toml` - though, like `Repos::roots`,
1083/// nothing stops a repository from setting it too, since a scalar field
1084/// takes whichever layer has the highest precedence.
1085#[derive(Debug, Clone, Deserialize, Serialize)]
1086#[serde(deny_unknown_fields, default)]
1087pub struct Daemon {
1088 /// How many runs `magi serve` may have actively in flight at once.
1089 /// **One by default** - today's behaviour, one run at a time.
1090 ///
1091 /// This is a different knob from [`Graph::max_parallel`], and the two
1092 /// must not be confused: `max_parallel` bounds how many agent *processes*
1093 /// one run starts inside itself (implementers, judges, reviewers -
1094 /// candidates competing on a single task); this field bounds how many
1095 /// *runs* - whole competitions, each with its own `max_parallel` budget -
1096 /// the loop drives side by side, possibly across different tasks and
1097 /// different repositories. Raising `max_parallel` buys a bigger panel for
1098 /// one task; raising this buys more tasks worked at once. A config file
1099 /// that meant one and wrote the other would either starve a competition
1100 /// of judges or leave the rest of the backlog waiting for no reason.
1101 ///
1102 /// A run parked waiting on the operator's land-merge approval - see
1103 /// [`crate::land`] - does not hold one of these slots while it waits: the
1104 /// whole point of parking there is to let the loop spend the slot on
1105 /// something runnable instead of sitting on a decision only a human can
1106 /// make. So even at the default of `1`, an approval that comes back does
1107 /// not queue behind whatever else the loop happens to be running.
1108 pub max_concurrent_runs: usize,
1109 /// Let a task marked [`crate::queue::Task::interrupt`] (`magi task
1110 /// interrupt`) cut ahead of whatever `magi serve` already has in flight,
1111 /// instead of waiting for it to finish.
1112 ///
1113 /// **Off by default.** When enabled, a run that is in flight and is not
1114 /// itself the interrupt candidate may be paused at its next safe node
1115 /// boundary - see [`crate::graph::Runner::park_here`] - so the marked
1116 /// task can run alone; the paused run resumes automatically, through the
1117 /// same path any other parked run does, the moment the interrupting
1118 /// task's own run reaches a terminal status. This is opt-in because it
1119 /// bends the loop's own "one run at a time" principle (see this repository's
1120 /// `magi serve` help) for a specific, explicit operator request, and a
1121 /// repository that never files an interrupt task pays nothing for having
1122 /// it on - but an operator who does not want any run of theirs preempted,
1123 /// ever, should leave this `false`.
1124 pub pause_for_interrupts: bool,
1125 /// How many deputies `magi serve` runs at once. A deputy is the short-lived
1126 /// seat that waits on one open conductor question (see [`crate::deputy`]);
1127 /// each is a live agent conversation, so this is bounded - **two by
1128 /// default**, because the agent CLIs' quota is the scarce resource. A
1129 /// question over the limit keeps its place and starts when a deputy
1130 /// finishes; `0` runs none.
1131 pub max_deputies: usize,
1132 /// Minutes a release-bump pull request may make no progress before
1133 /// `magi serve` raises a notice and asks the owner (see
1134 /// [`crate::release_watch`]). Progress is a change of head commit or of any
1135 /// check's verdict. A required check that stays red after the one rerun the
1136 /// watcher is allowed escalates at once, whatever this says. **Sixty by
1137 /// default**; `0` turns the watcher off.
1138 pub release_stall_minutes: u64,
1139}
1140
1141impl Default for Daemon {
1142 fn default() -> Self {
1143 Self {
1144 max_concurrent_runs: 1,
1145 pause_for_interrupts: false,
1146 max_deputies: 2,
1147 release_stall_minutes: 60,
1148 }
1149 }
1150}
1151
1152/// Where `magi repos` and `GET /api/repos` look for local checkouts.
1153///
1154/// `roots` is one of the array keys [`array_merge_policy`] marks as
1155/// append-across-layers: which checkouts exist in general is a *machine*
1156/// fact in the same way the agent roster is - a repository's own `magi.toml`
1157/// cannot state where its siblings live before magi has resolved which
1158/// repository to read that file from in the first place - but a repository
1159/// that genuinely has an extra root worth scanning is not forced to choose
1160/// between an error and losing the machine's roots outright. Both layers'
1161/// roots are scanned; see [`Config::refuse_split_arrays`] for the keys that
1162/// are still refused.
1163#[derive(Debug, Clone, Deserialize, Serialize)]
1164#[serde(deny_unknown_fields, default)]
1165pub struct Repos {
1166 /// Roots to scan for a ghq-layout checkout: `<root>/<host>/<owner>/<repo>`
1167 /// with a `.git` directory. Empty by default - nothing is scanned unless
1168 /// asked to be.
1169 pub roots: Vec<PathBuf>,
1170 /// How long a scan is trusted before the next request re-scans it,
1171 /// seconds. `0` means never trust it: scan on every request. Defaults to
1172 /// a day, the same order of magnitude as [`Graph::answer_timeout`] for
1173 /// the same reason - a checkout does not usually appear or vanish inside
1174 /// a session, so there is little to gain from scanning more often than
1175 /// that, and an explicit refresh exists for the moment one does.
1176 pub scan_ttl: u64,
1177 /// How often `magi serve` runs `git fetch origin` in every checkout under
1178 /// `roots`, seconds, so `origin/main` stays fresh whatever state the
1179 /// working tree is in. Fetch only: no branch, HEAD or working tree is
1180 /// touched. `0` disables it.
1181 pub fetch_interval: u64,
1182}
1183
1184impl Default for Repos {
1185 fn default() -> Self {
1186 Self {
1187 roots: Vec::new(),
1188 scan_ttl: 86_400,
1189 fetch_interval: 600,
1190 }
1191 }
1192}
1193
1194/// Project-specific text appended to each node's prompt.
1195///
1196/// **Additive by construction.** These fields cannot replace magi's prompts,
1197/// only extend them, and that restriction is the whole design. The built-in
1198/// prompts carry the invariants the competition rests on: a judging prompt
1199/// names no authors, every structured answer must arrive as one fenced `json`
1200/// block, and a judge is told not to speculate about who wrote what. A config
1201/// that could overwrite them would let a typo silently un-blind the panel or
1202/// break the parser, and the symptom would be "the judges got worse" rather
1203/// than an error.
1204///
1205/// Repository-wide context belongs in `AGENTS.md`, which every agent already
1206/// reads from the checkout. Use these fields for the things a *magi node*
1207/// needs to know and a repository file cannot say - for instance that
1208/// reviewers here should ignore formatting because a hook owns it.
1209#[derive(Debug, Clone, Default, Deserialize, Serialize)]
1210#[serde(deny_unknown_fields, default)]
1211pub struct Prompts {
1212 /// Appended to every node's prompt.
1213 pub all: String,
1214 /// Appended for implementers.
1215 pub implement: String,
1216 /// Appended for judges, both ranking and voting.
1217 pub judge: String,
1218 /// Appended for reviewers.
1219 pub review: String,
1220 /// Appended for the fixer.
1221 pub fix: String,
1222}
1223
1224impl Prompts {
1225 /// The overlay for one node, or `None` when nothing is configured.
1226 ///
1227 /// `node` is the graph's own node name, so a new node gets no overlay
1228 /// rather than the wrong one.
1229 pub fn overlay(&self, node: &str) -> Option<String> {
1230 let specific = match node {
1231 "implement" => &self.implement,
1232 "judge" | "vote" | "deliberate" => &self.judge,
1233 "review" => &self.review,
1234 "fix" => &self.fix,
1235 _ => "",
1236 };
1237 let mut parts: Vec<&str> = Vec::new();
1238 for p in [self.all.trim(), specific.trim()] {
1239 if !p.is_empty() {
1240 parts.push(p);
1241 }
1242 }
1243 if parts.is_empty() {
1244 return None;
1245 }
1246 Some(parts.join("\n\n"))
1247 }
1248}
1249
1250/// How the operator is told that a run is waiting on them.
1251///
1252/// A command rather than a built-in integration: magi is one binary with no
1253/// network dependencies, and every operator's notification path is different -
1254/// ntfy, a Slack webhook, a Windows toast, an SSH to a machine that beeps.
1255/// Shelling out keeps all of them possible and none of them magi's problem.
1256#[derive(Debug, Clone, Default, Deserialize, Serialize)]
1257#[serde(deny_unknown_fields, default)]
1258pub struct Notify {
1259 /// Command and arguments. `{summary}`, `{run}` and `{url}` are replaced.
1260 /// Empty means no notification - the web UI is then the only surface.
1261 pub command: Vec<String>,
1262}
1263
1264/// Policy for [`crate::talk`], the standing conversation.
1265#[derive(Debug, Clone, Default, Deserialize, Serialize)]
1266#[serde(deny_unknown_fields, default)]
1267pub struct Talk {
1268 /// Let the conversation's agent edit files in the repository instead of
1269 /// filing a task for one. **Off by default** - see
1270 /// [`crate::talk`]'s module doc for why an edit made mid-conversation is
1271 /// an edit no run and no review can be attributed to, which is exactly
1272 /// the property a repository entered in a competition depends on. A
1273 /// repository that is never judged - dotfiles, a personal config
1274 /// checkout - has nothing to lose by turning this on in its own
1275 /// `magi.toml`, and a one-line fix stops costing a queued task to get.
1276 pub allow_write: bool,
1277 /// Extra chat personas, `[[talk.personas]]`. They add to the built-ins in
1278 /// [`crate::persona`]; an entry with a built-in's id replaces it. Like every
1279 /// array outside [`array_merge_policy`]'s append list, declaring it in two
1280 /// layers is refused.
1281 pub personas: Vec<PersonaSpec>,
1282 /// What the chat calls the operator: the persona addresses them by this
1283 /// name and the web Chat labels their turns with it instead of `YOU`.
1284 /// Blank is unset; at most [`OPERATOR_NAME_MAX`] characters, one line.
1285 pub operator_name: Option<String>,
1286}
1287
1288/// Longest accepted `[talk] operator_name`, in characters.
1289pub const OPERATOR_NAME_MAX: usize = 64;
1290
1291impl Talk {
1292 /// The trimmed operator name, `None` when unset or blank.
1293 pub fn operator_name(&self) -> Option<&str> {
1294 self.operator_name
1295 .as_deref()
1296 .map(str::trim)
1297 .filter(|n| !n.is_empty())
1298 }
1299}
1300
1301/// One `[[talk.personas]]` entry: a tone for the standing chat, nothing more.
1302#[derive(Debug, Clone, Default, PartialEq, Eq, Deserialize, Serialize)]
1303#[serde(deny_unknown_fields)]
1304pub struct PersonaSpec {
1305 /// Stable id the conversation stores. Not `default`.
1306 pub id: String,
1307 /// What the selector shows.
1308 pub name: String,
1309 /// The voice instruction appended to the chat's system prompt.
1310 pub prompt: String,
1311}
1312
1313/// Roles resolved to concrete agent specs for one run.
1314#[derive(Debug, Clone)]
1315pub struct ResolvedRoles {
1316 /// One per candidate.
1317 pub implementers: Vec<AgentSpec>,
1318 /// One per judge.
1319 pub judges: Vec<AgentSpec>,
1320 /// One per reviewer slot.
1321 pub reviewers: Vec<AgentSpec>,
1322 /// The configured fixer chain in order, if `[roles] fixer` is set
1323 /// (resolved like [`crate::agent::pick_chain`]: unknown or uninstalled ids
1324 /// skipped, duplicates keep their first place). `None` means the winner's
1325 /// own author.
1326 pub fixer: Option<Vec<AgentSpec>>,
1327 /// Queue conductor, explicitly selected or resolved by the standalone-seat fallback.
1328 pub conductor: AgentSpec,
1329 /// The full ordered implementer roster, in [`Roles::implementers`]'s own
1330 /// order — or `[[agents]]` in file order, when that list is empty. Unlike
1331 /// [`Self::implementers`], never truncated to `graph.implementers` and
1332 /// never `rotate`d/wrapped: a solo run (`implementers = 1`) resolves
1333 /// `implementers` down to a single slot, but `graph::Runner`'s per-seat
1334 /// quota fallback needs the *whole* list to walk forward through when
1335 /// that one slot's agent runs out of quota mid-run.
1336 pub implementer_roster: Vec<AgentSpec>,
1337 /// [`Self::implementer_roster`]'s counterpart for judge seats: the whole
1338 /// `[roles] judges` list (or every agent), untruncated and unrotated, for
1339 /// the per-seat handover when a judge's agent fails.
1340 pub judge_roster: Vec<AgentSpec>,
1341 /// Likewise for reviewer seats.
1342 pub reviewer_roster: Vec<AgentSpec>,
1343}
1344
1345/// Every array-valued key in a config table, as a dotted path.
1346///
1347/// Dotted so the error names `roles.implementers` rather than `implementers`:
1348/// an operator with three config files needs to know which key, not just that
1349/// there was one. `vars` is skipped because it is teravars' own input, merged
1350/// on purpose and never deserialised into `Config`.
1351fn array_keys(table: &toml::value::Table, prefix: &str) -> Vec<String> {
1352 let mut out = Vec::new();
1353 for (k, v) in table {
1354 if prefix.is_empty() && k == "vars" {
1355 continue;
1356 }
1357 let path = if prefix.is_empty() {
1358 k.clone()
1359 } else {
1360 format!("{prefix}.{k}")
1361 };
1362 match v {
1363 toml::Value::Array(_) => out.push(path),
1364 toml::Value::Table(t) => out.extend(array_keys(t, &path)),
1365 _ => {}
1366 }
1367 }
1368 out
1369}
1370
1371/// How an array key behaves when two config layers both declare it.
1372#[derive(Debug, Clone, Copy, PartialEq, Eq)]
1373enum ArrayMerge {
1374 /// Two layers may both declare it; the composed value is the
1375 /// low-to-high-priority concatenation teravars already produces (see
1376 /// [`Config::load_layers`]'s doc for why that order and no dedup).
1377 Append,
1378 /// Two layers declaring it is refused; see
1379 /// [`Config::refuse_split_arrays`].
1380 Replace,
1381}
1382
1383/// The single place that decides, for a dotted array key (as returned by
1384/// [`array_keys`]), whether declaring it in two config layers is a
1385/// concatenation the operator asked for or a silent accident.
1386///
1387/// Kept as one match so the whole policy is visible in one place - the same
1388/// reason `claude_quota` and `dropped_stream` close their own classification
1389/// in one spot elsewhere in this codebase. Anything not listed defaults to
1390/// [`ArrayMerge::Replace`]: refusing is the safe default for a key nobody has
1391/// reasoned about yet, and a new array key added later has to be added here
1392/// deliberately to become appendable.
1393///
1394/// - `verify.e2e` / `verify.gate` — a "run all of these, all must exit 0"
1395/// gate. Concatenating two of them is exactly the checks both layers
1396/// wanted, which is what lets a common gate (e.g. `editorconfig-checker`)
1397/// live in a shared layer while a repository's own layer adds its own
1398/// command, instead of every repository copying the shared command into
1399/// its own file.
1400/// - `repos.roots` — a set of directories to scan for checkouts. A
1401/// repository adding its own root on top of the machine's is additive by
1402/// nature, not a replacement of where the machine looks; see
1403/// [`Repos::roots`].
1404///
1405/// Left on the refuse side, and why:
1406/// - `roles.implementers` / `roles.judges` / `roles.reviewers` — an ordered
1407/// list of *seats*, not a set. A machine's two implementers plus a
1408/// repository's one is three seats nobody asked for and nobody is paying
1409/// for on purpose.
1410/// - `notify.command` — an argv. Concatenating two argvs does not produce a
1411/// program that runs; it produces `["ntfy", "publish", "curl", "-X"]`.
1412/// - `blind.strip_lines` — technically safe to concatenate (each entry is
1413/// matched as an independent substring, so a longer list only strips
1414/// *more*), but left on the refuse side anyway: the same list also drives
1415/// `commit_msg_hook`'s generated `sed` addresses, where position matters,
1416/// and a silent three-layer merge is exactly the kind of surprise
1417/// `refuse_split_arrays` exists to catch rather than to reason about
1418/// case-by-case. A repository that wants one more stripped phrase restates
1419/// the whole list; that restatement is visible in review, an accidental
1420/// concatenation would not be.
1421fn array_merge_policy(key: &str) -> ArrayMerge {
1422 match key {
1423 "verify.e2e" | "verify.gate" | "verify.pre_gate" | "repos.roots" => ArrayMerge::Append,
1424 _ => ArrayMerge::Replace,
1425 }
1426}
1427
1428impl Config {
1429 /// Load one file through teravars: Tera rendering, `[vars]` resolution,
1430 /// and the `include = [...]` directive.
1431 pub fn load(path: &Path) -> Result<Self> {
1432 Self::load_layers(&[path.to_path_buf()])
1433 }
1434
1435 /// The Tera render context shared by every layer: `system.*` (from
1436 /// teravars), `env` (magi's own addition - a config that names a shared
1437 /// build-cache directory or a machine-specific path needs
1438 /// `{{ env.NAME | default(value='...') }}`), and `repo` / `repo_name`
1439 /// derived from the last (highest-priority) path's parent directory.
1440 ///
1441 /// Factored out so [`Config::array_provenance`] can re-render a single
1442 /// layer under the exact same context [`Config::load_layers`] uses for
1443 /// the joint render, rather than drifting from it by accident.
1444 fn render_ctx(paths: &[PathBuf]) -> teravars::Context {
1445 let mut ctx = teravars::system_context();
1446 let env: std::collections::BTreeMap<String, String> = std::env::vars().collect();
1447 ctx.insert("env", &env);
1448 if let Some(last) = paths.last()
1449 && let Some(dir) = last.parent()
1450 {
1451 // A layer extracted from `<remote>/<base>` records the checkout it
1452 // stands for beside itself; its own directory is a cache path.
1453 let dir = std::fs::read_to_string(dir.join(REPO_MARKER))
1454 .map(|s| PathBuf::from(s.trim()))
1455 .unwrap_or_else(|_| dir.to_path_buf());
1456 ctx.insert("repo", &dir.to_string_lossy());
1457 ctx.insert(
1458 "repo_name",
1459 &dir.file_name().unwrap_or_default().to_string_lossy(),
1460 );
1461 }
1462 ctx
1463 }
1464
1465 /// Load and deep-merge a stack of config files, later files winning.
1466 ///
1467 /// This is why the config is TOML-through-teravars rather than plain serde:
1468 /// the roster is a *machine* fact (which CLIs and plans you pay for) while
1469 /// the gate is a *repository* fact (`cargo make check` here, `pnpm test`
1470 /// there). Picking one file and ignoring the other would force every repo
1471 /// to restate the roster.
1472 pub fn load_layers(paths: &[PathBuf]) -> Result<Self> {
1473 let mut engine = teravars::Engine::default();
1474 let ctx = Self::render_ctx(paths);
1475 if paths.len() > 1 {
1476 Self::refuse_split_arrays(paths, &mut engine, &ctx)?;
1477 }
1478 let merged = teravars::load_merged(paths, &mut engine, &ctx).with_context(|| {
1479 format!(
1480 "rendering config via teravars: {}",
1481 paths
1482 .iter()
1483 .map(|p| p.display().to_string())
1484 .collect::<Vec<_>>()
1485 .join(", ")
1486 )
1487 })?;
1488 let mut table = merged.config;
1489 Self::normalize_renamed_keys(paths, &mut engine, &ctx, &mut table)?;
1490 // `[vars]` is teravars' own input, already resolved into the render
1491 // context; `deny_unknown_fields` must not trip over it.
1492 table.remove("vars");
1493 // No layer says anything about `agents`: the roster is whatever is on
1494 // `PATH` now, exactly as when no layer exists at all. An explicit
1495 // `agents = []` is a statement and is kept.
1496 let detect_agents = !table.contains_key("agents");
1497 let mut cfg: Self = toml::Value::Table(table)
1498 .try_into()
1499 .context("deserializing magi config")?;
1500 crate::persona::validate(&cfg.talk.personas)?;
1501 cfg.worktree.validate()?;
1502 if let Some(name) = cfg.talk.operator_name() {
1503 if name.chars().count() > OPERATOR_NAME_MAX || name.chars().any(char::is_control) {
1504 bail!(
1505 "[talk] operator_name must be one line of at most {OPERATOR_NAME_MAX} characters"
1506 );
1507 }
1508 }
1509 if detect_agents {
1510 cfg.agents = Self::autodetected().agents;
1511 }
1512 Ok(cfg)
1513 }
1514
1515 /// `graph.candidates` was renamed `graph.implementers`. Both spellings are
1516 /// read, but the layers are merged into one table before serde sees it, so
1517 /// an old name in a lower layer and the new name in a higher one would
1518 /// arrive together and fail as a duplicate field instead of letting the
1519 /// higher layer win. Resolve it here: a single layer that writes both is
1520 /// refused (naming both keys and the file); otherwise the value from the
1521 /// last layer that wrote either key becomes `implementers`.
1522 fn normalize_renamed_keys(
1523 paths: &[PathBuf],
1524 engine: &mut teravars::Engine,
1525 ctx: &teravars::Context,
1526 table: &mut toml::map::Map<String, toml::Value>,
1527 ) -> Result<()> {
1528 let both = matches!(
1529 table.get("graph"),
1530 Some(toml::Value::Table(g)) if g.contains_key("candidates") && g.contains_key("implementers")
1531 );
1532 if !both {
1533 return Ok(());
1534 }
1535 let mut winner: Option<toml::Value> = None;
1536 for path in paths {
1537 let one = teravars::load_merged([path], engine, ctx)
1538 .with_context(|| format!("rendering {}", path.display()))?;
1539 let Some(toml::Value::Table(g)) = one.config.get("graph") else {
1540 continue;
1541 };
1542 if g.contains_key("candidates") && g.contains_key("implementers") {
1543 bail!(
1544 "`[graph]` sets both `implementers` and its deprecated alias \n `candidates` in {}; keep only `implementers`.",
1545 path.display()
1546 );
1547 }
1548 if let Some(v) = g.get("implementers").or_else(|| g.get("candidates")) {
1549 winner = Some(v.clone());
1550 }
1551 }
1552 if let Some(toml::Value::Table(g)) = table.get_mut("graph") {
1553 g.remove("candidates");
1554 if let Some(v) = winner {
1555 g.insert("implementers".to_owned(), v);
1556 }
1557 }
1558 Ok(())
1559 }
1560
1561 /// Refuse an array that two layers both declare, unless
1562 /// [`array_merge_policy`] says that key is meant to accumulate.
1563 ///
1564 /// teravars **appends** arrays when it merges layers, and that is wrong for
1565 /// most arrays magi has: `implementers` is an ordered list of seats,
1566 /// `notify.command` is an argv. Concatenating two of them yields something
1567 /// nobody wrote - three implementers out of a machine's two and a
1568 /// repository's one, or an argv of `["ntfy", "publish", "curl", "-X"]`.
1569 ///
1570 /// Replacing instead would be the right merge rule for those keys, but the
1571 /// rule lives in teravars, which several other projects depend on;
1572 /// changing it there is a decision for that crate, not something to fake
1573 /// here by re-reading the files with different semantics and hoping the
1574 /// two paths agree.
1575 ///
1576 /// So magi refuses the ambiguity rather than resolving it silently, for
1577 /// every array key except the short, deliberate list
1578 /// [`array_merge_policy`] marks [`ArrayMerge::Append`] - for those, the
1579 /// concatenation teravars already produces *is* what both files say, so
1580 /// there is nothing to refuse. The cost of guessing wrong on the refused
1581 /// keys is a roster the operator did not ask for and is paying for by the
1582 /// token; the append keys carry no such risk because every element runs
1583 /// (or every directory is scanned) regardless of order.
1584 fn refuse_split_arrays(
1585 paths: &[PathBuf],
1586 engine: &mut teravars::Engine,
1587 ctx: &teravars::Context,
1588 ) -> Result<()> {
1589 let mut seen: std::collections::BTreeMap<String, PathBuf> = Default::default();
1590 for path in paths {
1591 let one = teravars::load_merged([path], engine, ctx)
1592 .with_context(|| format!("rendering {}", path.display()))?;
1593 for key in array_keys(&one.config, "") {
1594 if array_merge_policy(&key) == ArrayMerge::Append {
1595 continue;
1596 }
1597 if let Some(first) = seen.get(&key) {
1598 bail!(
1599 "`{key}` is an array declared in two config layers:\n \
1600 {}\n {}\nteravars appends arrays when it merges, so \
1601 magi would run the concatenation of both - which is \
1602 not what either file says. Declare `{key}` in exactly \
1603 one of them.",
1604 first.display(),
1605 path.display()
1606 );
1607 }
1608 seen.insert(key, path.clone());
1609 }
1610 }
1611 Ok(())
1612 }
1613
1614 /// Which layers contributed to a composed, appendable array key (e.g.
1615 /// `"verify.gate"`), in the same low-to-high-priority order
1616 /// [`Config::load_layers`] concatenates them in. Layers that do not
1617 /// declare `key` at all are omitted.
1618 ///
1619 /// This is a **display aid for `magi doctor` only.** The command list
1620 /// that actually runs always comes from the one joint
1621 /// [`teravars::load_merged`] call in `load_layers`, never from this
1622 /// function - the exact hazard [`Config::refuse_split_arrays`] warns
1623 /// about is two merge paths that might disagree, so this function must
1624 /// never become a second source of the *composed* value, only of which
1625 /// file wrote which line in it.
1626 ///
1627 /// Re-rendering each layer alone can, in principle, resolve a
1628 /// `{{ vars.x }}` differently than the joint render would, if `x` is
1629 /// defined in one layer and referenced in another - the same caveat
1630 /// `refuse_split_arrays`'s structural, key-only check already lives with.
1631 /// None of magi's own gate commands cross that line, and a doctor listing
1632 /// is read by a human who can compare it against the joint one printed
1633 /// alongside it, so this is judged worth the simplicity of not
1634 /// threading provenance through the real load path.
1635 pub fn array_provenance(paths: &[PathBuf], key: &str) -> Vec<(PathBuf, Vec<String>)> {
1636 let mut engine = teravars::Engine::default();
1637 let ctx = Self::render_ctx(paths);
1638 let mut out = Vec::new();
1639 for path in paths {
1640 let Ok(one) = teravars::load_merged([path], &mut engine, &ctx) else {
1641 continue;
1642 };
1643 let mut cur = &one.config;
1644 let mut found = None;
1645 let parts: Vec<&str> = key.split('.').collect();
1646 for (i, part) in parts.iter().enumerate() {
1647 match cur.get(*part) {
1648 Some(toml::Value::Array(a)) if i == parts.len() - 1 => {
1649 found = Some(a);
1650 break;
1651 }
1652 Some(toml::Value::Table(t)) => cur = t,
1653 _ => break,
1654 }
1655 }
1656 let Some(values) = found else { continue };
1657 let strings: Vec<String> = values
1658 .iter()
1659 .filter_map(|v| v.as_str().map(str::to_owned))
1660 .collect();
1661 if !strings.is_empty() {
1662 out.push((path.clone(), strings));
1663 }
1664 }
1665 out
1666 }
1667
1668 /// Each layer rendered on its own, with the outcome per file: the
1669 /// settings screen's way of saying which layer a value came from and which
1670 /// layer is the one that does not parse. A **display aid only**, for the
1671 /// same reason as [`Config::array_provenance`]: the effective config
1672 /// always comes from the one joint [`Config::load_layers`].
1673 pub(crate) fn layer_tables(paths: &[PathBuf]) -> Vec<(PathBuf, Result<toml::Table>)> {
1674 let mut engine = teravars::Engine::default();
1675 let ctx = Self::render_ctx(paths);
1676 paths
1677 .iter()
1678 .map(|path| {
1679 let one = teravars::load_merged([path], &mut engine, &ctx)
1680 .map(|m| m.config)
1681 .with_context(|| format!("rendering {}", path.display()));
1682 (path.clone(), one)
1683 })
1684 .collect()
1685 }
1686
1687 /// Render a composed command list for `magi doctor`: the joined command
1688 /// line the run actually uses, plus - only when more than one layer
1689 /// contributed - which layer wrote which line.
1690 ///
1691 /// A single contributing layer (the common case today) stays the plain
1692 /// one-line summary magi has always printed, `empty` included: that
1693 /// honest "(none — ...)" is what caught a real gate-composition gap
1694 /// before this array could compose at all, and composition should not
1695 /// make the common case noisier.
1696 pub fn describe_composed(
1697 paths: &[PathBuf],
1698 commands: &[String],
1699 key: &str,
1700 empty: &str,
1701 ) -> String {
1702 if commands.is_empty() {
1703 return empty.to_owned();
1704 }
1705 let joined = commands.join(" && ");
1706 let provenance = Self::array_provenance(paths, key);
1707 if provenance.len() <= 1 {
1708 return joined;
1709 }
1710 let mut out = joined;
1711 for (path, cmds) in &provenance {
1712 out.push_str(&format!("\n [{}] {}", path.display(), cmds.join(" && ")));
1713 }
1714 out
1715 }
1716
1717 /// Resolve the config for `repo`, honouring an explicit `--config` path.
1718 ///
1719 /// Returns the config and the layers it came from, empty for built-in
1720 /// defaults. The repository layers come from the remote-tracking ref as it
1721 /// stands (see the module doc); use [`Config::discover_fetched`] where a
1722 /// stale ref would matter.
1723 pub fn discover(repo: &Path, explicit: Option<&Path>) -> Result<(Self, Vec<PathBuf>)> {
1724 if let Some(p) = explicit {
1725 let paths = vec![p.to_path_buf()];
1726 return Ok((Self::load_layers(&paths)?, paths));
1727 }
1728 let paths = Self::layers_checked(repo)?;
1729 if paths.is_empty() {
1730 return Ok((Self::autodetected(), paths));
1731 }
1732 Ok((Self::load_layers(&paths)?, paths))
1733 }
1734
1735 /// [`Config::discover`] after `git fetch <remote> <base>`: for every path
1736 /// that is about to start work. A failed fetch is an error.
1737 pub async fn discover_fetched(
1738 repo: &Path,
1739 explicit: Option<&Path>,
1740 ) -> Result<(Self, Vec<PathBuf>)> {
1741 if explicit.is_none() && is_git_worktree(repo) {
1742 // A checkout that was never cloned from the remote has no
1743 // `<remote>/HEAD`; learn it here (refs/remotes only) so an
1744 // ordinary start recovers instead of failing in `discover`.
1745 let (remote, cfg_base) = probe_remote_base(repo);
1746 let base = crate::git::merge_base_branch(repo, &remote, cfg_base.as_deref()).await?;
1747 let out = crate::git::fetch(repo, &remote, &base)
1748 .await
1749 .with_context(|| format!("fetching {remote}/{base} to read magi.toml"))?;
1750 if !out.ok() {
1751 bail!(
1752 "cannot read magi.toml: `git fetch {remote} {base}` failed ({}); refusing to use the local checkout's copy, which may be stale",
1753 out.stderr.lines().next().unwrap_or("").trim()
1754 );
1755 }
1756 if cfg_base.is_none() {
1757 crate::git::ensure_remote_head(repo, &remote, &base).await;
1758 }
1759 }
1760 Self::discover(repo, explicit)
1761 }
1762
1763 /// [`Config::layers`], but an unreadable remote ref is an error.
1764 fn layers_checked(repo: &Path) -> Result<Vec<PathBuf>> {
1765 if !is_git_worktree(repo) {
1766 return Ok(Self::layers(repo));
1767 }
1768 let mut paths: Vec<PathBuf> = Self::machine_layer()
1769 .into_iter()
1770 .filter(|p| p.is_file())
1771 .collect();
1772 paths.extend(Self::repo_layers(repo)?);
1773 Ok(paths)
1774 }
1775
1776 /// Only the repository layers (lowest first), from `<remote>/<base>` in a
1777 /// git working tree and from disk elsewhere.
1778 pub(crate) fn repo_layers(repo: &Path) -> Result<Vec<PathBuf>> {
1779 if !is_git_worktree(repo) {
1780 return Ok([
1781 repo.join(".magi").join("config.toml"),
1782 repo.join("magi.toml"),
1783 ]
1784 .into_iter()
1785 .filter(|p| p.is_file())
1786 .collect());
1787 }
1788 remote_layers(repo)
1789 }
1790
1791 /// Environment variable that relocates the machine-wide config layer.
1792 ///
1793 /// Set it to a directory and magi reads `<dir>/magi/config.toml` instead
1794 /// of the one under [`dirs::config_dir`]; set it to the empty string and
1795 /// magi reads no machine layer at all.
1796 ///
1797 /// This exists because the machine layer is otherwise unavoidable, and a
1798 /// test that builds a config fixture is not asking for the operator's
1799 /// preferences to be merged into it. Adding `[repos] roots` to the real
1800 /// machine config on a development box turned two passing tests red -
1801 /// `repos_list_returns_name_and_path_for_every_configured_root` and
1802 /// `repos_list_only_rescans_within_the_ttl_when_asked_to`, whose fixtures
1803 /// declare `[repos] roots` of their own, which [`Config::layers`] then
1804 /// found in two layers and [`Config::refuse_split_arrays`] correctly
1805 /// refused. CI never saw it: a runner has no machine config, so the suite
1806 /// was green there and red only where somebody actually uses magi.
1807 ///
1808 /// An operator gets the same escape hatch for free: a second machine
1809 /// config, or none, without moving files about.
1810 pub const CONFIG_DIR_ENV: &str = "MAGI_CONFIG_DIR";
1811
1812 /// Every config layer that applies to `repo`, in increasing precedence.
1813 ///
1814 /// The machine layer is whatever [`Config::machine_layer`] resolves to,
1815 /// which is nothing at all in a test build.
1816 pub fn layers(repo: &Path) -> Vec<PathBuf> {
1817 let mut paths = Vec::new();
1818 paths.extend(Self::machine_layer());
1819 paths.push(repo.join(".magi").join("config.toml"));
1820 paths.push(repo.join("magi.toml"));
1821 paths.retain(|p| p.is_file());
1822 paths
1823 }
1824
1825 /// The machine-wide layer's path, when there is one.
1826 ///
1827 /// **A test build has none unless it names one.** A fixture is a complete
1828 /// statement of the config under test, and the operator's own preferences
1829 /// have no business being merged into it - least of all silently, on one
1830 /// machine, in a suite that is green everywhere else.
1831 #[cfg(test)]
1832 pub(crate) fn machine_layer() -> Option<PathBuf> {
1833 std::env::var(Self::CONFIG_DIR_ENV)
1834 .ok()
1835 .filter(|dir| !dir.trim().is_empty())
1836 .map(|dir| PathBuf::from(dir).join("magi").join("config.toml"))
1837 }
1838
1839 /// The machine-wide layer's path, when there is one.
1840 #[cfg(not(test))]
1841 pub(crate) fn machine_layer() -> Option<PathBuf> {
1842 match std::env::var(Self::CONFIG_DIR_ENV) {
1843 // Named, and empty on purpose: no machine layer.
1844 Ok(dir) if dir.trim().is_empty() => None,
1845 Ok(dir) => Some(PathBuf::from(dir).join("magi").join("config.toml")),
1846 Err(_) => dirs::config_dir().map(|dir| dir.join("magi").join("config.toml")),
1847 }
1848 }
1849
1850 /// Built-in config whose roster is the agent CLIs found on `PATH`.
1851 pub fn autodetected() -> Self {
1852 let mut cfg = Self::default();
1853 for (kind, id, model) in [
1854 (AgentKind::Claude, "opus", Some("opus")),
1855 (AgentKind::Claude, "sonnet", Some("sonnet")),
1856 (AgentKind::Antigravity, "antigravity", None),
1857 (AgentKind::Opencode, "opencode", None),
1858 (AgentKind::Codex, "codex", None),
1859 (AgentKind::Omp, "omp", None),
1860 ] {
1861 if kind.program().is_some_and(which) && !cfg.agents.iter().any(|a| a.id == id) {
1862 cfg.agents.push(AgentSpec {
1863 id: id.to_owned(),
1864 kind,
1865 model: model.map(str::to_owned),
1866 command: Vec::new(),
1867 extra_args: Vec::new(),
1868 env: BTreeMap::new(),
1869 prompt_delivery: None,
1870 });
1871 }
1872 }
1873 cfg
1874 }
1875
1876 /// The shared build cache the verify commands and the agents both build
1877 /// into, when the config declares one. See [`Verify::cache_dir`].
1878 pub fn cache_dir(&self) -> Option<PathBuf> {
1879 self.verify.cache_dir()
1880 }
1881
1882 /// Look an agent up by id.
1883 pub fn agent(&self, id: &str) -> Result<&AgentSpec> {
1884 self.agents
1885 .iter()
1886 .find(|a| a.id == id)
1887 .with_context(|| format!("no agent with id `{id}` in the roster"))
1888 }
1889
1890 /// Rotate `count` seats out of `ids`, or out of the whole roster at
1891 /// `offset` when `ids` is empty.
1892 ///
1893 /// The one rotation rule - explicit ids cycle, an empty list rotates the
1894 /// roster - shared by every seat count [`Config::resolve_roles`] fills in,
1895 /// rather than each seat reimplementing it and drifting apart.
1896 fn rotate(&self, ids: &[String], count: usize, offset: usize) -> Result<Vec<AgentSpec>> {
1897 let mut out = Vec::with_capacity(count);
1898 for i in 0..count {
1899 let spec = if ids.is_empty() {
1900 self.agents[(i + offset) % self.agents.len()].clone()
1901 } else {
1902 self.agent(&ids[i % ids.len()])?.clone()
1903 };
1904 out.push(spec);
1905 }
1906 Ok(out)
1907 }
1908
1909 /// `ids`, resolved to specs in the order named, or the whole `[[agents]]`
1910 /// roster in file order when `ids` is empty — the same "explicit ids
1911 /// stand alone, an empty list means the whole roster" rule [`Self::rotate`]
1912 /// applies, but with no `count` to truncate to and no `offset` to wrap by.
1913 /// See [`ResolvedRoles::implementer_roster`] for why a truncated, rotated
1914 /// list cannot answer the question this exists for.
1915 fn full_roster(&self, ids: &[String]) -> Result<Vec<AgentSpec>> {
1916 if ids.is_empty() {
1917 Ok(self.agents.clone())
1918 } else {
1919 ids.iter().map(|id| self.agent(id).cloned()).collect()
1920 }
1921 }
1922
1923 /// Fill the roles out to the configured widths.
1924 ///
1925 /// An empty role list rotates through the whole roster, so a three-agent
1926 /// roster with `implementers = 3` gives one implementation per agent, and
1927 /// `judges = 3` rotates the judge seats by one so that judge *i* is not the
1928 /// author of candidate *i* whenever the roster has more than one agent.
1929 pub fn resolve_roles(&self) -> Result<ResolvedRoles> {
1930 if self.agents.is_empty() {
1931 bail!(
1932 "agent roster is empty: no agent CLI found on PATH and no \
1933 [[agents]] in the config. Run `magi init` to write a starter \
1934 magi.toml."
1935 );
1936 }
1937 Ok(ResolvedRoles {
1938 implementers: self.rotate(&self.roles.implementers, self.graph.implementers, 0)?,
1939 judges: self.rotate(&self.roles.judges, self.graph.judges, 1)?,
1940 reviewers: self.rotate(&self.roles.reviewers, self.graph.reviewers, 0)?,
1941 fixer: match self.roles.fixer.as_ref() {
1942 // Ids are validated against the roster here; whether the CLI
1943 // is installed is left to invocation, like every other seat.
1944 Some(choice) => Some(crate::agent::pick_chain(
1945 &self.agents,
1946 Some(choice),
1947 &|_| true,
1948 "fixer",
1949 )?),
1950 None => None,
1951 },
1952 // Role resolution validates roster shape, but deliberately does
1953 // not preflight a CLI. The other graph seats have always deferred
1954 // that failure to invocation; doing it only for the conductor
1955 // made otherwise usable graph commands and `doctor` fail as one.
1956 conductor: match self.roles.conductor.as_ref().map(AgentChoice::ids) {
1957 Some(ids) if ids.len() == 1 => self.agent(ids[0])?.clone(),
1958 // A fallback chain: the first id the roster knows. The rest
1959 // are `agent::pick_chain`'s business at invocation.
1960 Some(ids) if !ids.is_empty() => ids
1961 .iter()
1962 .find_map(|id| self.agent(id).ok())
1963 .cloned()
1964 .with_context(|| {
1965 format!("[roles] conductor names no agent in the roster: {ids:?}")
1966 })?,
1967 // Keep the normal standalone-seat preference when something
1968 // is installed, but retain a roster fallback when it is not.
1969 // Invocation then reports the unavailable CLI in the same
1970 // place it does for every other graph role.
1971 _ => crate::agent::pick(&self.agents, None, &crate::agent::installed)
1972 .unwrap_or_else(|_| self.agents[0].clone()),
1973 },
1974 implementer_roster: self.full_roster(&self.roles.implementers)?,
1975 judge_roster: self.full_roster(&self.roles.judges)?,
1976 reviewer_roster: self.full_roster(&self.roles.reviewers)?,
1977 })
1978 }
1979
1980 /// Advisor seats for the design-deliberation stage (see
1981 /// `graph::Runner::advise`): `[roles] advisors` when set, otherwise the
1982 /// judge roster - see [`Roles::advisors`] for why that fallback and not
1983 /// the whole roster.
1984 ///
1985 /// The fallback rotates with `offset = 1`, matching the judges line in
1986 /// [`Config::resolve_roles`] exactly, `ids` and offset both - not just
1987 /// `roles.judges`, which is empty whenever judges themselves are
1988 /// unconfigured and rotating the whole roster. Falling back with
1989 /// `offset = 0` there would silently hand the advisors a *different*
1990 /// agent set than the judges an unconfigured run would actually get,
1991 /// which is the one thing [`Roles::advisors`]'s doc promises will not
1992 /// happen.
1993 ///
1994 /// Called lazily from the graph node itself rather than folded into
1995 /// [`Config::resolve_roles`]: unlike the other roles, a failure here must
1996 /// not stop a run from starting at all - the deliberation stage is an
1997 /// enrichment `[graph] advise` can turn off, not a seat later nodes
1998 /// cannot proceed without - and at the point `resolve_roles` runs (before
1999 /// [`crate::run::RunState`] exists, on `Runner::start`) there would be no
2000 /// run yet for a resolution failure to be reported against.
2001 pub fn advisors(&self) -> Result<Vec<AgentSpec>> {
2002 if self.agents.is_empty() {
2003 bail!(
2004 "agent roster is empty: no agent CLI found on PATH and no \
2005 [[agents]] in the config. Run `magi init` to write a starter \
2006 magi.toml."
2007 );
2008 }
2009 if !self.roles.advisors.is_empty() {
2010 return self.rotate(&self.roles.advisors, self.graph.advisors, 0);
2011 }
2012 self.rotate(&self.roles.judges, self.graph.advisors, 1)
2013 }
2014
2015 /// The untruncated roster advisor seats hand over along: `[roles]
2016 /// advisors` when set, otherwise the judges' roster, exactly as
2017 /// [`Self::advisors`] chooses its own list.
2018 pub fn advisor_roster(&self) -> Result<Vec<AgentSpec>> {
2019 if self.roles.advisors.is_empty() {
2020 self.full_roster(&self.roles.judges)
2021 } else {
2022 self.full_roster(&self.roles.advisors)
2023 }
2024 }
2025
2026 /// Shell prefix for [`Verify`] commands.
2027 pub fn shell(&self) -> Vec<String> {
2028 if let Some(s) = &self.verify.shell {
2029 return s.clone();
2030 }
2031 if which("sh") {
2032 vec!["sh".to_owned(), "-c".to_owned()]
2033 } else {
2034 vec!["cmd".to_owned(), "/C".to_owned()]
2035 }
2036 }
2037
2038 /// Starter config, as written by `magi init`.
2039 pub fn starter_toml() -> String {
2040 let detected = Self::autodetected();
2041 let mut s = String::from(
2042 "# magi — blind multi-agent implementation competition.\n\
2043 # `magi run \"<task>\"` walks: implement (N parallel worktrees)\n\
2044 # -> blind judging -> deliberation -> private final vote\n\
2045 # -> fold losers -> review + E2E loop -> gate -> merge.\n\
2046 #\n\
2047 # Rendered by teravars: a `[vars]` table, env\n\
2048 # and system lookups, and `include = [...]` all work. Tera\n\
2049 # braces are live everywhere in this file, but comments are\n\
2050 # stripped before rendering (teravars >= 0.2.2), so a comment\n\
2051 # may quote `{{ ... }}` freely.\n\
2052 #\n\
2053 # Layers deep-merge in increasing\n\
2054 # precedence, so the roster can live once per machine in\n\
2055 # <config_dir>/magi/config.toml and each repo only states its own\n\
2056 # gate:\n\
2057 # <config_dir>/magi/config.toml < .magi/config.toml < magi.toml\n\n\
2058 [vars]\n\
2059 # Reference it as vars.cache inside Tera braces, anywhere below.\n\
2060 # Single quotes inside the braces: teravars renders the raw file\n\
2061 # text, so TOML's own \\\" escaping never reaches Tera.\n\
2062 cache = \"{{ env.MAGI_CACHE | default(value='/tmp') }}\"\n\n",
2063 );
2064 if detected.agents.is_empty() {
2065 s.push_str(
2066 "# No agent CLI was found on PATH. Fill this in by hand.\n\
2067 # kind = claude | opencode | antigravity | codex | command\n\
2068 [[agents]]\nid = \"opus\"\nkind = \"claude\"\nmodel = \"opus\"\n\n",
2069 );
2070 } else {
2071 for a in &detected.agents {
2072 s.push_str("[[agents]]\n");
2073 s.push_str(&format!("id = {:?}\n", a.id));
2074 s.push_str(&format!("kind = {:?}\n", a.kind.as_str()));
2075 if let Some(m) = &a.model {
2076 s.push_str(&format!("model = {m:?}\n"));
2077 }
2078 s.push('\n');
2079 }
2080 }
2081 s.push_str(
2082 "# Leave a role list empty to rotate through the roster.\n\
2083 # A non-empty list is also an ordered backup list: seats take entries from the\n\
2084 # front (see [graph] counts) and a failed or out-of-quota seat moves on to the\n\
2085 # next agent to its right, each tried once. Same for judges and reviewers.\n\
2086 # implementers = [\"opus\", \"codex\", \"sonnet\"] # with implementers = 1 below: opus sits, codex then sonnet are spares\n\
2087 [roles]\n\
2088 implementers = []\n\
2089 judges = []\n\
2090 reviewers = []\n\
2091 # conductor = \"opus\" # arranges the queue; unset picks a seat like chatter does\n\
2092 # synthesizer = \"opus\" # blends the advisors into one brief; unset picks a seat like chatter does\n\
2093 # synthesizer = [\"opus\", \"codex\"] # array form: fallback chain, each tried once on quota or failure\n\
2094 # language_judge = \"jev\" # kind = \"command\" wrapper deciding if PR text is English; unset keeps the built-in heuristics\n\
2095 # fixer = \"opus\" # applies review findings; unset keeps the winner's own author\n\
2096 # fixer = [\"opus\", \"codex\"] # array form: fallback chain, each tried once on quota or failure\n\n\
2097 [graph]\n\
2098 # Gate GitHub prose; secret redaction is always enabled.\n\
2099 github_text_guard = true\n\
2100 implementers = 3\n\
2101 judges = 3\n\
2102 deliberate_rounds = 1\n\
2103 reviewers = 3\n\
2104 review_rounds = 6\n\
2105 # Fix rounds a failing verify.gate gets before the run is blocked (0 = none).\n\
2106 gate_fix_rounds = 1\n\
2107 max_parallel = 4\n\
2108 language = \"en\"\n\
2109 # One CLI conversation per seat: judges keep their own argument\n\
2110 # across deliberation, the fixer keeps its implementation context.\n\
2111 sessions = true\n\
2112 # Reviewer-seat timeout. When timeout_verify is omitted, E2E and\n\
2113 # the final gate inherit this value for compatibility.\n\
2114 timeout_review = 1200\n\
2115 # Optional independent E2E/final-gate timeout; uncomment to keep\n\
2116 # verification independent if timeout_review changes later.\n\
2117 # timeout_verify = 1200\n\n\
2118 # What the chat calls you (persona address and the Chat's turn label).\n\
2119 # [talk]\n\
2120 # operator_name = \"Commander\"\n\n\
2121 # Chat personas (tone only; the web Chat picks one per conversation).\n\
2122 # Built in: magi, rei, misato, ritsuko, shinji, asuka, kaworu. Entries add to\n\
2123 # them, and one with a built-in's id replaces it. `default` is the plain\n\
2124 # voice and cannot be redefined. Declare this in one layer only.\n\
2125 # [[talk.personas]]\n\
2126 # id = \"gendo\"\n\
2127 # name = \"Gendo Ikari\"\n\
2128 # prompt = \"Speak coldly and tersely, hands folded.\"\n\n\
2129 [verify]\n\
2130 # Run once per review round in the winner's worktree; failures are\n\
2131 # fed back to the fixer.\n\
2132 e2e = []\n\
2133 # Final gate. Every command must exit 0 before a merge.\n\
2134 gate = []\n\
2135 # Mechanical fixers (formatters) run in the winner's worktree just\n\
2136 # before the gate; whatever they change becomes one commit. A failing\n\
2137 # command only warns - the gate stays the arbiter. Same timeout as the\n\
2138 # gate (timeout_verify). Keep it light and idempotent.\n\
2139 # pre_gate = []\n\n\
2140 # Prepare each worktree magi creates (candidates, judges, advisors,\n\
2141 # reviewers, the operator-fix and review-takeover checkouts), in order,\n\
2142 # before any agent or verify command touches it. `copy` reads from the\n\
2143 # primary checkout and never overwrites; `optional` skips a missing\n\
2144 # source. `run` goes through the shell with the worktree as cwd and\n\
2145 # fails the seat on a non-zero exit or timeout (timeout_secs, default\n\
2146 # the verify timeout). Setup must not edit tracked files, and what it\n\
2147 # leaves untracked is kept out of magi's commits. Declare this in one\n\
2148 # layer only.\n\
2149 # [worktree]\n\
2150 # setup = [\n\
2151 # { copy = \".env.example -> .env\" },\n\
2152 # { copy = \".env.local\", optional = true },\n\
2153 # { run = \"pnpm install --frozen-lockfile\", timeout_secs = 900 },\n\
2154 # ]\n\n\
2155 [merge]\n\
2156 # none | local | pr\n\
2157 mode = \"none\"\n\
2158 # release_bump = true # Rust (Cargo.toml) repositories only; a no-op elsewhere\n\n\
2159 [update]\n\
2160 # off | notify | install — checked in the background, throttled.\n\
2161 mode = \"notify\"\n\
2162 # interval = \"24h\"\n\n\
2163 [daemon]\n\
2164 # Minutes a release-bump pull request (chore/release-v*) may sit with\n\
2165 # no new commit and no check changing state before `magi serve`\n\
2166 # notices and asks you. A failed check is rerun once on its own;\n\
2167 # still red after that escalates at once. 0 turns the watcher off.\n\
2168 # release_stall_minutes = 60\n\n\
2169 [release]\n\
2170 # actions (default) | local. \"local\" is for repositories whose GitHub\n\
2171 # Actions never run (private, no minutes): the release pull request\n\
2172 # does not wait for CI, you approve the merge in magi, and then\n\
2173 # `magi serve` pushes the vX.Y.Z tag (never over an existing one) and\n\
2174 # runs `commands` from a clean checkout of the merge commit. A\n\
2175 # failure holds the release with a notice; it is never retried blind.\n\
2176 # Needs `magi serve` running and [daemon] release_stall_minutes > 0.\n\
2177 # mode = \"actions\"\n\
2178 # Run in order; the first failure stops. They see MAGI_RELEASE_VERSION,\n\
2179 # MAGI_RELEASE_TAG, MAGI_RELEASE_COMMIT and MAGI_RELEASE_PR. Make them\n\
2180 # safe to repeat: a crash mid-command is treated as unfinished.\n\
2181 # commands = [\"cargo build --release\", \"gh release create \\\"$MAGI_RELEASE_TAG\\\" --generate-notes\"]\n\
2182 # timeout_minutes = 30\n",
2183 );
2184 s
2185 }
2186}
2187
2188/// Is `program` on `PATH`?
2189pub fn which(program: &str) -> bool {
2190 find_program(program).is_some()
2191}
2192
2193/// The file a bare `program` name resolves to on `PATH`, the way the OS shell
2194/// would find it.
2195///
2196/// Spawn agents through this, not by bare name: on Windows Rust's std
2197/// only tries `name.exe`, so a CLI installed as an npm `.cmd` shim (codex
2198/// since its 2026-09-23 reinstall) is "not found" even though a shell runs it.
2199/// On Windows an extensionless file is never a match: npm writes an `sh`
2200/// script beside every `.cmd` shim, and spawning that fails with os error 193.
2201pub fn find_program(program: &str) -> Option<PathBuf> {
2202 find_program_on(program, &std::env::var_os("PATH")?)
2203}
2204
2205/// [`find_program`] against an explicit `PATH`, for a child whose environment
2206/// overrides it.
2207pub fn find_program_on(program: &str, paths: &std::ffi::OsStr) -> Option<PathBuf> {
2208 let pathext = std::env::var("PATHEXT").unwrap_or_else(|_| ".EXE;.CMD;.BAT".into());
2209 find_program_in(program, paths, cfg!(windows).then_some(pathext.as_str()))
2210}
2211
2212fn find_program_in(
2213 program: &str,
2214 paths: &std::ffi::OsStr,
2215 pathext: Option<&str>,
2216) -> Option<PathBuf> {
2217 let p = Path::new(program);
2218 if p.components().count() > 1 {
2219 return p.is_file().then(|| p.to_path_buf());
2220 }
2221 std::env::split_paths(paths).find_map(|dir| match pathext {
2222 Some(exts) => {
2223 let bare = dir.join(program);
2224 if p.extension().is_some() && bare.is_file() {
2225 return Some(bare);
2226 }
2227 exts.split(';')
2228 .filter(|e| !e.is_empty())
2229 .map(|e| dir.join(format!("{program}{}", e.to_lowercase())))
2230 .find(|c| c.is_file())
2231 }
2232 None => Some(dir.join(program)).filter(|c| c.is_file()),
2233 })
2234}
2235
2236/// Sidecar naming the checkout an extracted layer stands for.
2237const REPO_MARKER: &str = ".magi-repo";
2238
2239fn git_out(repo: &Path, args: &[&str]) -> Option<String> {
2240 use crate::proc::Quiet as _;
2241 let out = std::process::Command::new("git")
2242 .args(args)
2243 .quiet()
2244 .current_dir(repo)
2245 .stdin(std::process::Stdio::null())
2246 .output()
2247 .ok()?;
2248 out.status
2249 .success()
2250 .then(|| String::from_utf8_lossy(&out.stdout).into_owned())
2251}
2252
2253/// Is `repo` inside a git working tree? A directory that is not has no branch
2254/// to be stale against, so its files are read as they are.
2255fn is_git_worktree(repo: &Path) -> bool {
2256 git_out(repo, &["rev-parse", "--git-dir"]).is_some()
2257}
2258
2259#[derive(Deserialize, Default)]
2260struct ProbeMerge {
2261 base: Option<String>,
2262 remote: Option<String>,
2263}
2264
2265#[derive(Deserialize, Default)]
2266struct Probe {
2267 merge: Option<ProbeMerge>,
2268}
2269
2270/// `[merge]` base / remote from one file, nothing else. The file is rendered
2271/// (env templates, includes) exactly as a real load would, so a machine layer
2272/// saying `base = "{{ env.X }}"` locates the same ref; when rendering fails
2273/// (a checkout file that needs its own repo context) the plain TOML is read
2274/// instead. Unreadable is `None`: the real load reports it.
2275fn probe_merge(path: &Path, repo: &Path) -> Option<ProbeMerge> {
2276 if !path.is_file() {
2277 return None;
2278 }
2279 let paths = [path.to_path_buf()];
2280 let mut engine = teravars::Engine::default();
2281 let mut ctx = Config::render_ctx(&paths);
2282 // The layered load renders every layer for the checkout being resolved;
2283 // so does this, or `{{ repo_name }}` would name the file's own directory.
2284 ctx.insert("repo", &repo.to_string_lossy());
2285 ctx.insert(
2286 "repo_name",
2287 &repo.file_name().unwrap_or_default().to_string_lossy(),
2288 );
2289 if let Ok(merged) = teravars::load_merged(&paths, &mut engine, &ctx)
2290 && let Ok(probe) = toml::Value::Table(merged.config).try_into::<Probe>()
2291 {
2292 return probe.merge;
2293 }
2294 let text = std::fs::read_to_string(path).ok()?;
2295 toml::from_str::<Probe>(&text).ok()?.merge
2296}
2297
2298/// The `[merge] remote` / `base` the layers name, before any git lookup.
2299fn probe_remote_base(repo: &Path) -> (String, Option<String>) {
2300 let mut remote = None;
2301 let mut base = None;
2302 let mut sources: Vec<PathBuf> = Config::machine_layer().into_iter().collect();
2303 sources.push(repo.join("magi.toml"));
2304 sources.push(repo.join(".magi").join("config.toml"));
2305 for path in sources {
2306 if let Some(m) = probe_merge(&path, repo) {
2307 remote = remote.or(m.remote.filter(|s| !s.trim().is_empty()));
2308 base = base.or(m.base.filter(|s| !s.trim().is_empty()));
2309 }
2310 }
2311 (remote.unwrap_or_else(|| "origin".to_owned()), base)
2312}
2313
2314/// Where the remote ref is: see the module doc for the order.
2315fn resolve_remote_base(repo: &Path) -> Result<(String, String)> {
2316 let (remote, base) = probe_remote_base(repo);
2317 let base = match base {
2318 Some(b) => b,
2319 None => {
2320 let head = format!("refs/remotes/{remote}/HEAD");
2321 git_out(repo, &["symbolic-ref", "--quiet", &head])
2322 .and_then(|s| crate::git::remote_head_branch(&remote, &s))
2323 .with_context(|| {
2324 format!(
2325 "cannot tell which branch of `{remote}` holds magi.toml: \
2326 set [merge] base, or run `git remote set-head {remote} -a`"
2327 )
2328 })?
2329 }
2330 };
2331 Ok((remote, base))
2332}
2333
2334fn fnv_hex(text: &str) -> String {
2335 let mut hash: u64 = 0xcbf2_9ce4_8422_2325;
2336 for b in text.bytes() {
2337 hash ^= u64::from(b);
2338 hash = hash.wrapping_mul(0x0100_0000_01b3);
2339 }
2340 format!("{hash:016x}")
2341}
2342
2343/// Does the text declare an `include` directive in any form teravars accepts
2344/// (root `include`, bare or quoted, or `[teravars] include`, as a table or an
2345/// inline table)? Parsed as TOML when it parses; a templated file that does
2346/// not parse is judged by a plain search for the word, because over-detecting
2347/// only costs an extra extraction while missing one fails the load.
2348fn declares_include(text: &str) -> bool {
2349 match toml::from_str::<toml::Table>(text) {
2350 Ok(table) => {
2351 table.contains_key("include")
2352 || table
2353 .get("teravars")
2354 .and_then(toml::Value::as_table)
2355 .is_some_and(|t| t.contains_key("include"))
2356 }
2357 Err(_) => text.contains("include"),
2358 }
2359}
2360
2361/// Unpack the whole tree of `sha` into `dest`, so relative `include`s of a
2362/// repository layer resolve against the same commit.
2363fn extract_tree(repo: &Path, sha: &str, dest: &Path) -> Result<()> {
2364 use crate::proc::Quiet as _;
2365 let mut archive = std::process::Command::new("git")
2366 .args(["archive", "--format=tar", sha])
2367 .quiet()
2368 .current_dir(repo)
2369 .stdin(std::process::Stdio::null())
2370 .stdout(std::process::Stdio::piped())
2371 .stderr(std::process::Stdio::null())
2372 .spawn()
2373 .context("running git archive")?;
2374 let tar = std::process::Command::new("tar")
2375 .arg("-x")
2376 .arg("-C")
2377 .arg(dest)
2378 .quiet()
2379 .stdin(archive.stdout.take().context("git archive has no stdout")?)
2380 .output()
2381 .context("running tar")?;
2382 let status = archive.wait().context("waiting for git archive")?;
2383 if !status.success() || !tar.status.success() {
2384 bail!(
2385 "cannot unpack {sha} for the config's includes: {}",
2386 String::from_utf8_lossy(&tar.stderr).trim()
2387 );
2388 }
2389 Ok(())
2390}
2391
2392/// The repository layers (`.magi/config.toml`, `magi.toml`; lowest first) of
2393/// `<remote>/<base>`, extracted from one commit so both files agree.
2394fn remote_layers(repo: &Path) -> Result<Vec<PathBuf>> {
2395 let (remote, base) = resolve_remote_base(repo)?;
2396 let tracking = format!("refs/remotes/{remote}/{base}");
2397 let spec = format!("{tracking}^{{commit}}");
2398 let sha = git_out(repo, &["rev-parse", "--verify", "--quiet", &spec])
2399 .map(|s| s.trim().to_owned())
2400 .filter(|s| !s.is_empty())
2401 .with_context(|| {
2402 format!(
2403 "`{remote}/{base}` does not exist in this checkout; fetch it \
2404 (`git fetch {remote} {base}`) or set [merge] base/remote. magi will not read \
2405 a possibly stale local magi.toml instead"
2406 )
2407 })?;
2408 let key = format!("{sha}-{}", fnv_hex(&repo.to_string_lossy()));
2409 let rels = [".magi/config.toml", "magi.toml"];
2410 let mut texts: Vec<(&str, String)> = Vec::new();
2411 for rel in rels {
2412 let object = format!("{sha}:{rel}");
2413 if git_out(repo, &["cat-file", "-e", &object]).is_none() {
2414 continue;
2415 }
2416 let Some(text) = git_out(repo, &["show", &object]) else {
2417 bail!("cannot read {rel} from {remote}/{base} ({sha})");
2418 };
2419 texts.push((rel, text));
2420 }
2421 let wants_tree = texts.iter().any(|(_, t)| declares_include(t));
2422 // Content-addressed and built beside the final name, then renamed in: a
2423 // concurrent `discover` sees the whole directory or none of it, never a
2424 // half-written file.
2425 let build = |root: &Path| -> Result<PathBuf> {
2426 let final_dir = root.join(&key);
2427 if !final_dir.is_dir() {
2428 std::fs::create_dir_all(root)?;
2429 let tmp = root.join(format!(
2430 "{key}.tmp-{}-{}",
2431 std::process::id(),
2432 std::time::SystemTime::now()
2433 .duration_since(std::time::UNIX_EPOCH)
2434 .map(|d| d.as_nanos())
2435 .unwrap_or_default()
2436 ));
2437 let filled = (|| -> Result<()> {
2438 std::fs::create_dir_all(&tmp)?;
2439 if wants_tree {
2440 extract_tree(repo, &sha, &tmp)?;
2441 }
2442 for (rel, text) in &texts {
2443 let file = tmp.join(rel);
2444 let parent = file.parent().unwrap_or(&tmp);
2445 std::fs::create_dir_all(parent)?;
2446 std::fs::write(&file, text)?;
2447 std::fs::write(parent.join(REPO_MARKER), repo.to_string_lossy().as_bytes())?;
2448 }
2449 Ok(())
2450 })();
2451 if let Err(e) = filled {
2452 let _ = std::fs::remove_dir_all(&tmp);
2453 return Err(e);
2454 }
2455 if std::fs::rename(&tmp, &final_dir).is_err() {
2456 // Lost the race to another process building the same commit.
2457 let _ = std::fs::remove_dir_all(&tmp);
2458 if !final_dir.is_dir() {
2459 bail!("cannot place {}", final_dir.display());
2460 }
2461 }
2462 }
2463 Ok(final_dir)
2464 };
2465 let common = git_out(
2466 repo,
2467 &["rev-parse", "--path-format=absolute", "--git-common-dir"],
2468 )
2469 .map(|s| PathBuf::from(s.trim()).join("magi").join("config"));
2470 let dir = match common.map(|c| build(&c)) {
2471 Some(Ok(d)) => d,
2472 _ => build(&std::env::temp_dir().join("magi-config"))
2473 .with_context(|| format!("extracting config from {remote}/{base}"))?,
2474 };
2475 Ok(texts.iter().map(|(rel, _)| dir.join(rel)).collect())
2476}
2477
2478#[cfg(test)]
2479mod tests {
2480 #[test]
2481 fn context_window_prefers_exact_then_longest_prefix_and_user_over_builtin() {
2482 let mut cfg = Config::default();
2483 assert_eq!(
2484 cfg.context_window("claude-sonnet-4-5-20250929"),
2485 Some(200_000)
2486 );
2487 assert_eq!(cfg.context_window("gpt-4.1-mini"), Some(1_000_000));
2488 assert_eq!(cfg.context_window("mystery-model"), None);
2489 cfg.context_windows
2490 .insert("claude-sonnet-4-5".into(), 1_000_000);
2491 cfg.context_windows.insert("claude-sonnet-4".into(), 5);
2492 cfg.context_windows.insert("mystery".into(), 32_768);
2493 // Longest user prefix beats a shorter one and the built-in table.
2494 assert_eq!(
2495 cfg.context_window("claude-sonnet-4-5-20250929"),
2496 Some(1_000_000)
2497 );
2498 // An exact name beats any prefix.
2499 cfg.context_windows
2500 .insert("claude-sonnet-4-5-20250929".into(), 7);
2501 assert_eq!(cfg.context_window("claude-sonnet-4-5-20250929"), Some(7));
2502 assert_eq!(cfg.context_window("mystery-model"), Some(32_768));
2503 }
2504
2505 use super::*;
2506
2507 /// npm writes an extensionless `sh` script beside `codex.cmd`; on Windows
2508 /// the `.cmd` is the spawnable one (os error 193 otherwise, 2026-09-23).
2509 #[test]
2510 fn windows_lookup_picks_the_cmd_shim_over_the_extensionless_script() {
2511 let dir = tempfile::tempdir().unwrap();
2512 std::fs::write(dir.path().join("codex"), "#!/bin/sh\n").unwrap();
2513 std::fs::write(dir.path().join("codex.cmd"), "@echo off\n").unwrap();
2514 let paths = dir.path().as_os_str();
2515 assert_eq!(
2516 find_program_in("codex", paths, Some(".EXE;.CMD")),
2517 Some(dir.path().join("codex.cmd"))
2518 );
2519 assert_eq!(
2520 find_program_in("codex", paths, None),
2521 Some(dir.path().join("codex"))
2522 );
2523 assert_eq!(find_program_in("claude", paths, Some(".EXE;.CMD")), None);
2524 }
2525
2526 fn spec(id: &str) -> AgentSpec {
2527 AgentSpec {
2528 id: id.to_owned(),
2529 kind: AgentKind::Command,
2530 model: None,
2531 command: vec!["true".to_owned()],
2532 extra_args: Vec::new(),
2533 env: BTreeMap::new(),
2534 prompt_delivery: None,
2535 }
2536 }
2537
2538 #[test]
2539 fn github_text_guard_defaults_on_and_can_be_disabled() {
2540 let default: Graph = toml::from_str("").unwrap();
2541 assert!(default.github_text_guard);
2542 let disabled: Graph = toml::from_str("github_text_guard = false").unwrap();
2543 assert!(!disabled.github_text_guard);
2544 let config = layered("", "[graph]\ngithub_text_guard = false\n").unwrap();
2545 assert!(!config.graph.github_text_guard);
2546 assert!(Config::starter_toml().contains("github_text_guard = true"));
2547 }
2548
2549 #[test]
2550 fn timeout_verify_omitted_from_old_toml_inherits_timeout_review() {
2551 // `timeout_verify` used to not exist: `verify.e2e`/`verify.gate` ran
2552 // under `timeout_review`. A config written before this field existed
2553 // must run exactly as before, which means its default has to be the
2554 // same 1200s `timeout_review` has always defaulted to.
2555 let g: Graph = toml::from_str("timeout_review = 3600").expect("parse");
2556 assert_eq!(g.timeout_verify, None);
2557 assert_eq!(g.verify_timeout(), 3600);
2558 }
2559
2560 #[test]
2561 fn shrinking_timeout_review_does_not_shrink_timeout_verify() {
2562 // The bug this field exists to close: `[graph] timeout_review = 45`
2563 // used to shrink the real-machine `verify.e2e`/`verify.gate` budget
2564 // along with the reviewer seats' own timeout, because both read the
2565 // same field.
2566 let g: Graph = toml::from_str("timeout_review = 45").expect("parse");
2567 assert_eq!(g.timeout_review, 45);
2568 assert_eq!(
2569 g.verify_timeout(),
2570 45,
2571 "an omitted legacy value follows review"
2572 );
2573 let explicit: Graph = toml::from_str("timeout_review = 45\ntimeout_verify = 1200")
2574 .expect("parse explicit override");
2575 assert_eq!(explicit.verify_timeout(), 1200);
2576 }
2577
2578 #[test]
2579 fn graph_implementers_new_key_loads_and_old_key_is_an_alias() {
2580 let g: Graph = toml::from_str("implementers = 4\n").unwrap();
2581 assert_eq!(g.implementers, 4);
2582 let g: Graph = toml::from_str("candidates = 5\n").unwrap();
2583 assert_eq!(g.implementers, 5);
2584 }
2585
2586 fn layered(lower: &str, upper: &str) -> Result<Config> {
2587 let dir = tempfile::tempdir().unwrap();
2588 let a = dir.path().join("a.toml");
2589 let b = dir.path().join("b.toml");
2590 std::fs::write(&a, lower).unwrap();
2591 std::fs::write(&b, upper).unwrap();
2592 Config::load_layers(&[a, b])
2593 }
2594
2595 #[test]
2596 fn graph_implementers_higher_layer_wins_across_spellings() {
2597 let c = layered("[graph]\ncandidates = 2\n", "[graph]\nimplementers = 5\n").unwrap();
2598 assert_eq!(c.graph.implementers, 5);
2599 let c = layered("[graph]\nimplementers = 2\n", "[graph]\ncandidates = 5\n").unwrap();
2600 assert_eq!(c.graph.implementers, 5);
2601 let c = layered("[graph]\ncandidates = 2\n", "[graph]\njudges = 3\n").unwrap();
2602 assert_eq!(c.graph.implementers, 2);
2603 }
2604
2605 #[test]
2606 fn graph_implementers_both_keys_in_one_layer_is_refused() {
2607 let err = layered("", "[graph]\ncandidates = 2\nimplementers = 3\n")
2608 .unwrap_err()
2609 .to_string();
2610 assert!(
2611 err.contains("candidates") && err.contains("implementers"),
2612 "{err}"
2613 );
2614 assert!(err.contains("b.toml"), "{err}");
2615 }
2616
2617 #[test]
2618 fn a_toml_layer_written_before_these_fields_existed_still_parses() {
2619 // `deny_unknown_fields` cuts both ways: a config from before
2620 // `timeout_verify`/`e2e_every_round` existed must still parse, with
2621 // both defaulted rather than refused as unknown-in-reverse.
2622 let g: Graph =
2623 toml::from_str("candidates = 1\nreviewers = 3\nreview_rounds = 6\nmax_parallel = 4\n")
2624 .expect("an old-shaped [graph] table must still parse");
2625 assert_eq!(g.verify_timeout(), Graph::default().timeout_review);
2626 assert!(
2627 !g.e2e_every_round,
2628 "off by default, same as before this field existed"
2629 );
2630 }
2631
2632 #[test]
2633 fn empty_roles_rotate_judges_off_their_own_candidate() {
2634 // Three seats, said out loud: this is a test about *rotation*, and it
2635 // has nothing to say about how many candidates a task buys by default.
2636 let cfg = Config {
2637 agents: vec![spec("a"), spec("b"), spec("c")],
2638 graph: Graph {
2639 implementers: 3,
2640 ..Graph::default()
2641 },
2642 ..Config::default()
2643 };
2644 let roles = cfg.resolve_roles().unwrap();
2645 let impls: Vec<&str> = roles.implementers.iter().map(|a| a.id.as_str()).collect();
2646 let judges: Vec<&str> = roles.judges.iter().map(|a| a.id.as_str()).collect();
2647 assert_eq!(impls, ["a", "b", "c"]);
2648 assert_eq!(judges, ["b", "c", "a"]);
2649 for (i, j) in judges.iter().enumerate() {
2650 assert_ne!(*j, impls[i], "judge {i} must not sit on its own candidate");
2651 }
2652 }
2653
2654 #[test]
2655 fn single_agent_roster_fills_every_seat() {
2656 let cfg = Config {
2657 agents: vec![spec("solo")],
2658 graph: Graph {
2659 implementers: 3,
2660 ..Graph::default()
2661 },
2662 ..Config::default()
2663 };
2664 let roles = cfg.resolve_roles().unwrap();
2665 assert_eq!(roles.implementers.len(), 3);
2666 assert!(roles.judges.iter().all(|a| a.id == "solo"));
2667 }
2668
2669 #[test]
2670 fn implementer_roster_is_the_whole_agent_list_untruncated_when_unset() {
2671 // `implementers = 1` (a solo run) still resolves `implementers` down to
2672 // one slot, but `implementer_roster` must carry every agent in the
2673 // roster, in file order, for `graph::Runner`'s quota fallback to walk
2674 // forward through once that one slot's agent runs out of quota.
2675 let cfg = Config {
2676 agents: vec![spec("a"), spec("b"), spec("c")],
2677 graph: Graph {
2678 implementers: 1,
2679 ..Graph::default()
2680 },
2681 ..Config::default()
2682 };
2683 let roles = cfg.resolve_roles().unwrap();
2684 assert_eq!(roles.implementers.len(), 1);
2685 let roster: Vec<&str> = roles
2686 .implementer_roster
2687 .iter()
2688 .map(|a| a.id.as_str())
2689 .collect();
2690 assert_eq!(roster, ["a", "b", "c"]);
2691 }
2692
2693 #[test]
2694 fn implementer_roster_follows_explicit_ids_in_the_order_named() {
2695 let cfg = Config {
2696 agents: vec![spec("a"), spec("b"), spec("c")],
2697 roles: Roles {
2698 implementers: vec!["c".to_owned(), "a".to_owned()],
2699 ..Roles::default()
2700 },
2701 graph: Graph {
2702 implementers: 1,
2703 ..Graph::default()
2704 },
2705 ..Config::default()
2706 };
2707 let roles = cfg.resolve_roles().unwrap();
2708 // Unaffected by the new field: still just the first named id.
2709 assert_eq!(roles.implementers.len(), 1);
2710 assert_eq!(roles.implementers[0].id, "c");
2711 let roster: Vec<&str> = roles
2712 .implementer_roster
2713 .iter()
2714 .map(|a| a.id.as_str())
2715 .collect();
2716 assert_eq!(roster, ["c", "a"]);
2717 }
2718
2719 #[test]
2720 fn explicit_roles_win() {
2721 let cfg = Config {
2722 agents: vec![spec("a"), spec("b")],
2723 roles: Roles {
2724 implementers: vec!["b".to_owned()],
2725 judges: vec!["a".to_owned()],
2726 reviewers: Vec::new(),
2727 fixer: Some("a".into()),
2728 ..Roles::default()
2729 },
2730 ..Config::default()
2731 };
2732 let roles = cfg.resolve_roles().unwrap();
2733 assert!(roles.implementers.iter().all(|a| a.id == "b"));
2734 assert!(roles.judges.iter().all(|a| a.id == "a"));
2735 let fixer = roles.fixer.unwrap();
2736 assert_eq!(fixer.len(), 1);
2737 assert_eq!(fixer[0].id, "a");
2738 assert_eq!(roles.conductor.id, "a");
2739 }
2740
2741 #[test]
2742 fn fixer_takes_a_string_or_an_ordered_chain() {
2743 let one: Roles = toml::from_str("fixer = \"a\"").unwrap();
2744 assert_eq!(one.fixer.as_ref().unwrap().ids(), ["a"]);
2745 let chain: Roles = toml::from_str("fixer = [\"b\", \"a\"]").unwrap();
2746 assert_eq!(chain.fixer.as_ref().unwrap().ids(), ["b", "a"]);
2747
2748 let resolve = |fixer: Option<AgentChoice>| {
2749 Config {
2750 agents: vec![spec("a"), spec("b"), spec("c")],
2751 roles: Roles {
2752 fixer,
2753 ..Roles::default()
2754 },
2755 ..Config::default()
2756 }
2757 .resolve_roles()
2758 };
2759 let ids = |r: &ResolvedRoles| -> Vec<String> {
2760 r.fixer
2761 .as_ref()
2762 .unwrap()
2763 .iter()
2764 .map(|s| s.id.clone())
2765 .collect()
2766 };
2767 // The string form is a chain of one; unset stays unset (the winner's author).
2768 assert_eq!(ids(&resolve(Some("c".into())).unwrap()), ["c"]);
2769 assert!(resolve(None).unwrap().fixer.is_none());
2770 // Order kept, an unknown id skipped, a duplicate keeps its first place.
2771 let chain = AgentChoice::Chain(vec!["b".into(), "zzz".into(), "a".into(), "b".into()]);
2772 assert_eq!(ids(&resolve(Some(chain)).unwrap()), ["b", "a"]);
2773 // Nothing resolving names the role.
2774 let err = resolve(Some(AgentChoice::Chain(vec!["zzz".into()]))).unwrap_err();
2775 assert!(format!("{err:#}").contains("fixer"), "{err:#}");
2776 }
2777
2778 #[test]
2779 fn unknown_agent_id_is_an_error() {
2780 let cfg = Config {
2781 agents: vec![spec("a")],
2782 roles: Roles {
2783 judges: vec!["nope".to_owned()],
2784 ..Roles::default()
2785 },
2786 ..Config::default()
2787 };
2788 assert!(cfg.resolve_roles().is_err());
2789 }
2790
2791 #[test]
2792 fn conductor_role_is_resolved_validated_and_has_a_fallback() {
2793 let mut cfg = Config {
2794 agents: vec![spec("a"), spec("b")],
2795 ..Config::default()
2796 };
2797 assert_eq!(cfg.resolve_roles().unwrap().conductor.id, "a");
2798
2799 cfg.roles.conductor = Some("b".into());
2800 assert_eq!(cfg.resolve_roles().unwrap().conductor.id, "b");
2801
2802 cfg.roles.conductor = Some("missing".into());
2803 assert!(cfg.resolve_roles().is_err());
2804 }
2805
2806 #[test]
2807 fn a_standalone_role_takes_a_string_or_an_array_and_round_trips() {
2808 let one: Roles = toml::from_str("synthesizer = \"opus\"").unwrap();
2809 assert_eq!(one.synthesizer, Some(AgentChoice::One("opus".to_owned())));
2810 assert_eq!(primary(one.synthesizer.as_ref()), Some("opus"));
2811
2812 let chain: Roles = toml::from_str("chatter = [\"opus\", \"codex\"]").unwrap();
2813 assert_eq!(chain.chatter.as_ref().unwrap().ids(), ["opus", "codex"]);
2814 assert_eq!(primary(chain.chatter.as_ref()), Some("opus"));
2815
2816 for roles in [one, chain] {
2817 let text = toml::to_string(&roles).unwrap();
2818 let back: Roles = toml::from_str(&text).unwrap();
2819 assert_eq!(back.synthesizer, roles.synthesizer);
2820 assert_eq!(back.chatter, roles.chatter);
2821 }
2822 let text = toml::to_string(&Roles {
2823 conductor: Some("x".into()),
2824 ..Roles::default()
2825 })
2826 .unwrap();
2827 assert!(
2828 text.contains("conductor = \"x\""),
2829 "string stays a string: {text}"
2830 );
2831 assert!(toml::from_str::<Roles>("synthesizer = 3").is_err());
2832 }
2833
2834 #[test]
2835 fn an_empty_choice_counts_as_unset() {
2836 assert!(AgentChoice::Chain(Vec::new()).ids().is_empty());
2837 assert!(AgentChoice::One(" ".to_owned()).ids().is_empty());
2838 assert_eq!(primary(Some(&AgentChoice::Chain(vec![]))), None);
2839 }
2840
2841 #[test]
2842 fn synthesizer_role_is_a_lazy_lookup_with_the_same_fallback_as_chatter() {
2843 // Unlike `conductor`, `synthesizer` is never validated by
2844 // `resolve_roles` — it is looked up lazily by
2845 // `graph::Runner::synthesize_brief` the same way `[roles] chatter`
2846 // is looked up by `talk::begin`, so this exercises `agent::pick`
2847 // directly instead of going through `resolve_roles`.
2848 let mut cfg = Config {
2849 agents: vec![spec("a"), spec("b")],
2850 ..Config::default()
2851 };
2852 let want = primary(cfg.roles.synthesizer.as_ref());
2853 assert_eq!(
2854 crate::agent::pick(&cfg.agents, want, &crate::agent::installed)
2855 .unwrap()
2856 .id,
2857 "a",
2858 "unset falls back to agent::pick's own default order"
2859 );
2860
2861 cfg.roles.synthesizer = Some("b".into());
2862 let want = primary(cfg.roles.synthesizer.as_ref());
2863 assert_eq!(
2864 crate::agent::pick(&cfg.agents, want, &crate::agent::installed)
2865 .unwrap()
2866 .id,
2867 "b",
2868 "the named agent wins over the default order"
2869 );
2870 }
2871
2872 #[test]
2873 fn empty_roster_is_an_error() {
2874 assert!(Config::default().resolve_roles().is_err());
2875 }
2876
2877 #[test]
2878 fn advise_defaults_to_on_with_three_proposals() {
2879 let g = Graph::default();
2880 assert!(g.advise);
2881 assert_eq!(g.advisors, 3);
2882 }
2883
2884 #[test]
2885 fn unset_advisors_falls_back_to_the_judge_roster() {
2886 let cfg = Config {
2887 agents: vec![spec("a"), spec("b")],
2888 roles: Roles {
2889 judges: vec!["b".to_owned()],
2890 ..Roles::default()
2891 },
2892 graph: Graph {
2893 advisors: 2,
2894 ..Graph::default()
2895 },
2896 ..Config::default()
2897 };
2898 let advisors = cfg.advisors().expect("advisors resolve");
2899 assert_eq!(advisors.len(), 2);
2900 assert!(
2901 advisors.iter().all(|a| a.id == "b"),
2902 "an unset [roles] advisors must fall back to [roles] judges: {advisors:?}"
2903 );
2904 }
2905
2906 #[test]
2907 fn an_explicit_advisor_roster_wins_over_the_judge_fallback() {
2908 let cfg = Config {
2909 agents: vec![spec("a"), spec("b")],
2910 roles: Roles {
2911 judges: vec!["b".to_owned()],
2912 advisors: vec!["a".to_owned()],
2913 ..Roles::default()
2914 },
2915 graph: Graph {
2916 advisors: 2,
2917 ..Graph::default()
2918 },
2919 ..Config::default()
2920 };
2921 let advisors = cfg.advisors().expect("advisors resolve");
2922 assert!(advisors.iter().all(|a| a.id == "a"));
2923 }
2924
2925 /// Neither `[roles] advisors` nor `[roles] judges` set: an unconfigured
2926 /// advisor roster must resolve to the exact same agents an unconfigured
2927 /// judge panel would get - same ids, same rotation offset - or the
2928 /// promise in [`Roles::advisors`]'s doc ("advisor diversity for free")
2929 /// does not actually hold.
2930 #[test]
2931 fn an_unconfigured_advisor_and_judge_roster_resolve_to_the_same_agents() {
2932 let cfg = Config {
2933 agents: vec![spec("a"), spec("b"), spec("c")],
2934 graph: Graph {
2935 advisors: 3,
2936 judges: 3,
2937 ..Graph::default()
2938 },
2939 ..Config::default()
2940 };
2941 let advisors = cfg.advisors().expect("advisors resolve");
2942 let judges = cfg.resolve_roles().expect("roles resolve").judges;
2943 let advisor_ids: Vec<&str> = advisors.iter().map(|a| a.id.as_str()).collect();
2944 let judge_ids: Vec<&str> = judges.iter().map(|a| a.id.as_str()).collect();
2945 assert_eq!(
2946 advisor_ids, judge_ids,
2947 "an unconfigured advisor roster must be the same seats an unconfigured judge panel gets"
2948 );
2949 }
2950
2951 #[test]
2952 fn an_unresolvable_advisor_seat_is_an_error_naming_the_id() {
2953 let cfg = Config {
2954 agents: vec![spec("a")],
2955 roles: Roles {
2956 advisors: vec!["nope".to_owned()],
2957 ..Roles::default()
2958 },
2959 graph: Graph {
2960 advisors: 1,
2961 ..Graph::default()
2962 },
2963 ..Config::default()
2964 };
2965 let err = cfg.advisors().expect_err("`nope` is not in the roster");
2966 assert!(format!("{err:#}").contains("nope"));
2967 }
2968
2969 #[test]
2970 fn repos_default_to_no_roots_and_a_day_of_trust() {
2971 assert_eq!(Config::default().repos.roots, Vec::<PathBuf>::new());
2972 assert_eq!(Config::default().repos.scan_ttl, 86_400);
2973 assert_eq!(Config::default().repos.fetch_interval, 600);
2974 }
2975
2976 #[test]
2977 fn a_config_file_with_no_repos_table_still_loads() {
2978 let dir = tempfile::tempdir().unwrap();
2979 let path = dir.path().join("magi.toml");
2980 std::fs::write(&path, "[graph]\ncandidates = 2\n").unwrap();
2981 let cfg = Config::load(&path).expect("must load without [repos]");
2982 assert_eq!(cfg.repos.roots, Vec::<PathBuf>::new());
2983 assert_eq!(cfg.repos.scan_ttl, 86_400);
2984 }
2985
2986 /// A fixture is the whole config under test.
2987 ///
2988 /// `layers` used to reach for `dirs::config_dir()` unconditionally, so on
2989 /// a machine where somebody had written `<config_dir>/magi/config.toml`
2990 /// the suite silently loaded it as the lowest layer. Adding `[repos]
2991 /// roots` there turned two web tests red - their fixtures declare
2992 /// `[repos] roots` too, and `refuse_split_arrays` rightly refuses one
2993 /// array key spread across two layers. CI stayed green throughout,
2994 /// because a runner has no such file: the suite failed only where magi is
2995 /// actually used.
2996 ///
2997 /// So a test build has no machine layer unless it asks for one, and this
2998 /// is that promise. Written against a real file at the real location so
2999 /// it fails if `machine_layer` starts reading it again.
3000 #[test]
3001 fn a_test_build_does_not_read_the_operators_machine_config() {
3002 let repo = tempfile::tempdir().unwrap();
3003 std::fs::write(repo.path().join("magi.toml"), "[graph]\ncandidates = 2\n").unwrap();
3004
3005 let layers = Config::layers(repo.path());
3006 assert_eq!(
3007 layers,
3008 vec![repo.path().join("magi.toml")],
3009 "only the fixture's own file may be a layer"
3010 );
3011 if let Some(real) = dirs::config_dir() {
3012 let machine = real.join("magi").join("config.toml");
3013 assert!(
3014 !layers.contains(&machine),
3015 "the operator's {} must not be a layer in a test build",
3016 machine.display()
3017 );
3018 }
3019 }
3020
3021 #[test]
3022 fn starter_toml_loads_through_teravars() {
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 parsed = Config::load(&path).expect("starter config must load");
3027 assert_eq!(parsed.graph.implementers, 3);
3028 assert_eq!(parsed.merge.mode, MergeMode::None);
3029 assert_eq!(parsed.merge.style, MergeStyle::Merge);
3030 assert!(parsed.graph.sessions);
3031 assert!(parsed.graph.github_text_guard);
3032 assert_eq!(parsed.graph.timeout_review, 1200);
3033 assert_eq!(parsed.graph.verify_timeout(), 1200);
3034 assert_eq!(parsed.update.mode, UpdateMode::Notify);
3035 }
3036
3037 #[test]
3038 fn release_defaults_to_actions_and_the_starter_documents_it() {
3039 let dir = tempfile::tempdir().unwrap();
3040 let path = dir.path().join("magi.toml");
3041 std::fs::write(&path, Config::starter_toml()).unwrap();
3042 let r = Config::load(&path).unwrap().release;
3043 assert_eq!(r.mode, ReleaseMode::Actions);
3044 assert!(!r.is_local());
3045 assert!(r.commands.is_empty());
3046 assert_eq!(r.timeout_minutes, 30);
3047 assert!(Config::starter_toml().contains("# mode = \"actions\""));
3048 }
3049
3050 #[test]
3051 fn release_local_parses_and_unknown_keys_are_refused() {
3052 let dir = tempfile::tempdir().unwrap();
3053 let path = dir.path().join("magi.toml");
3054 std::fs::write(
3055 &path,
3056 "[release]\nmode = \"local\"\ncommands = [\"true\", \"echo $MAGI_RELEASE_TAG\"]\ntimeout_minutes = 5\n",
3057 )
3058 .unwrap();
3059 let r = Config::load(&path).unwrap().release;
3060 assert!(r.is_local());
3061 assert_eq!(r.commands.len(), 2);
3062 assert_eq!(r.timeout_minutes, 5);
3063 std::fs::write(&path, "[release]\nmode = \"local\"\nbogus = 1\n").unwrap();
3064 assert!(Config::load(&path).is_err());
3065 std::fs::write(&path, "[release]\nmode = \"gha\"\n").unwrap();
3066 assert!(Config::load(&path).is_err());
3067 }
3068
3069 #[test]
3070 fn pre_gate_defaults_to_empty_and_the_starter_documents_it() {
3071 let dir = tempfile::tempdir().unwrap();
3072 let path = dir.path().join("magi.toml");
3073 std::fs::write(&path, Config::starter_toml()).unwrap();
3074 assert!(Config::load(&path).unwrap().verify.pre_gate.is_empty());
3075 assert!(Config::starter_toml().contains("# pre_gate = []"));
3076 }
3077
3078 #[test]
3079 fn starter_toml_explains_inherited_and_explicit_verify_timeouts() {
3080 let starter = Config::starter_toml();
3081 assert!(starter.contains("When timeout_verify is omitted, E2E and"));
3082 assert!(starter.contains("verification independent if timeout_review changes later"));
3083 assert!(starter.contains("# timeout_verify = 1200"));
3084 }
3085
3086 #[test]
3087 fn starter_toml_documents_chat_personas() {
3088 let starter = Config::starter_toml();
3089 assert!(starter.contains("# [[talk.personas]]"));
3090 assert!(starter.contains("tone only"));
3091 }
3092
3093 #[test]
3094 fn talk_personas_parse_and_are_validated() {
3095 let dir = tempfile::tempdir().unwrap();
3096 let path = dir.path().join("magi.toml");
3097 std::fs::write(
3098 &path,
3099 "[[talk.personas]]\nid = \"rei\"\nname = \"Quiet\"\nprompt = \"Be brief.\"\n\n\
3100 [[talk.personas]]\nid = \"gendo\"\nname = \"Gendo\"\nprompt = \"Be cold.\"\n",
3101 )
3102 .unwrap();
3103 let cfg = Config::load(&path).expect("loads");
3104 assert_eq!(cfg.talk.personas.len(), 2);
3105 let catalog = crate::persona::catalog(&cfg.talk.personas);
3106 let rei = catalog.iter().find(|p| p.id == "rei").unwrap();
3107 assert_eq!(
3108 (rei.name.as_str(), rei.prompt.as_str()),
3109 ("Quiet", "Be brief.")
3110 );
3111 assert_eq!(catalog.last().unwrap().id, "gendo");
3112
3113 for (body, needle) in [
3114 (
3115 "[[talk.personas]]\nid = \"\"\nname = \"n\"\nprompt = \"p\"\n",
3116 "empty `id`",
3117 ),
3118 (
3119 "[[talk.personas]]\nid = \"a\"\nname = \"n\"\nprompt = \" \"\n",
3120 "empty `prompt`",
3121 ),
3122 (
3123 "[[talk.personas]]\nid = \"a\"\nname = \"n\"\nprompt = \"p\"\n\
3124 [[talk.personas]]\nid = \"a\"\nname = \"m\"\nprompt = \"q\"\n",
3125 "more than once",
3126 ),
3127 (
3128 "[[talk.personas]]\nid = \"default\"\nname = \"n\"\nprompt = \"p\"\n",
3129 "cannot be redefined",
3130 ),
3131 ] {
3132 std::fs::write(&path, body).unwrap();
3133 let err = format!("{:#}", Config::load(&path).unwrap_err());
3134 assert!(err.contains(needle), "{needle}: {err}");
3135 }
3136 }
3137
3138 #[test]
3139 fn talk_operator_name_is_trimmed_and_validated() {
3140 let dir = tempfile::tempdir().unwrap();
3141 let path = dir.path().join("magi.toml");
3142 let load = |body: &str| {
3143 std::fs::write(&path, body).unwrap();
3144 Config::load(&path)
3145 };
3146 let ok = load("[talk]\noperator_name = \" Commander \"\n").expect("ok");
3147 assert_eq!(ok.talk.operator_name(), Some("Commander"));
3148 let blank = load("[talk]\noperator_name = \" \"\n").expect("blank");
3149 assert_eq!(blank.talk.operator_name(), None);
3150 assert!(load("[talk]\noperator_name = \"a\\nb\"\n").is_err());
3151 let long = "x".repeat(65);
3152 assert!(load(&format!("[talk]\noperator_name = \"{long}\"\n")).is_err());
3153 }
3154
3155 #[test]
3156 fn talk_personas_in_two_layers_are_refused() {
3157 let dir = tempfile::tempdir().unwrap();
3158 let machine = dir.path().join("machine.toml");
3159 let repo = dir.path().join("repo.toml");
3160 let one = "[[talk.personas]]\nid = \"a\"\nname = \"n\"\nprompt = \"p\"\n";
3161 std::fs::write(&machine, one).unwrap();
3162 std::fs::write(&repo, one.replace("\"a\"", "\"b\"")).unwrap();
3163 let err = format!("{:#}", Config::load_layers(&[machine, repo]).unwrap_err());
3164 assert!(err.contains("talk.personas"), "{err}");
3165 }
3166
3167 #[test]
3168 fn starter_toml_documents_role_lists_as_backup_order() {
3169 let starter = Config::starter_toml();
3170 assert!(starter.contains("also an ordered backup list"));
3171 assert!(starter.contains("codex then sonnet are spares"));
3172 }
3173
3174 /// A repository whose ruleset forbids merge commits declares that once,
3175 /// here, rather than magi asking GitHub about it on every render (see
3176 /// [`MergeStyle`]'s own doc for why).
3177 #[test]
3178 fn a_repository_can_declare_a_linear_history_merge_style() {
3179 let dir = tempfile::tempdir().unwrap();
3180 let path = dir.path().join("magi.toml");
3181 std::fs::write(&path, "[merge]\nmode = \"none\"\nstyle = \"squash\"\n").unwrap();
3182 let parsed = Config::load(&path).expect("config must load");
3183 assert_eq!(parsed.merge.style, MergeStyle::Squash);
3184 }
3185
3186 #[test]
3187 fn later_layers_win_and_vars_render() {
3188 let dir = tempfile::tempdir().unwrap();
3189 let machine = dir.path().join("machine.toml");
3190 let project = dir.path().join("magi.toml");
3191 // The machine layer owns the roster...
3192 std::fs::write(
3193 &machine,
3194 "[[agents]]\nid = \"opus\"\nkind = \"claude\"\nmodel = \"opus\"\n\n\
3195 [graph]\ncandidates = 3\nmax_parallel = 8\n",
3196 )
3197 .unwrap();
3198 // ...and the project layer only states what is repo-specific, plus a
3199 // `[vars]` value interpolated into a command.
3200 std::fs::write(
3201 &project,
3202 "[vars]\ncache = \"/shared\"\n\n\
3203 [graph]\ncandidates = 2\n\n\
3204 [verify]\ngate = [\"CARGO_TARGET_DIR={{ vars.cache }}/t cargo test\"]\n",
3205 )
3206 .unwrap();
3207
3208 let cfg = Config::load_layers(&[machine, project]).expect("layered load");
3209 assert_eq!(cfg.agents.len(), 1, "roster comes from the machine layer");
3210 assert_eq!(cfg.graph.implementers, 2, "project layer wins");
3211 assert_eq!(cfg.graph.max_parallel, 8, "machine layer survives");
3212 assert_eq!(
3213 cfg.verify.gate,
3214 ["CARGO_TARGET_DIR=/shared/t cargo test".to_owned()]
3215 );
3216 // The rendered command is where the cache path is read back from.
3217 assert_eq!(cfg.cache_dir(), Some(PathBuf::from("/shared/t")));
3218 }
3219
3220 #[test]
3221 fn talk_defaults_to_an_hour_and_an_unwritten_config_still_gets_it() {
3222 // An operator who writes no `[graph]` timeout keys at all must still
3223 // land on the hour, not on the five/fifteen minutes this turn used
3224 // to hardcode before it read from config.
3225 let g = Graph::default();
3226 assert_eq!(g.timeout_talk, 3600);
3227
3228 let dir = tempfile::tempdir().unwrap();
3229 let path = dir.path().join("magi.toml");
3230 std::fs::write(&path, "[graph]\ncandidates = 2\n").unwrap();
3231 let cfg = Config::load(&path).expect("must load without timeout_talk set");
3232 assert_eq!(cfg.graph.timeout_talk, 3600);
3233 }
3234
3235 #[test]
3236 fn an_overridden_talk_timeout_reaches_the_loaded_config() {
3237 let dir = tempfile::tempdir().unwrap();
3238 let path = dir.path().join("magi.toml");
3239 std::fs::write(&path, "[graph]\ntimeout_talk = 120\n").unwrap();
3240 let cfg = Config::load(&path).expect("must load");
3241 assert_eq!(cfg.graph.timeout_talk, 120);
3242 }
3243
3244 #[test]
3245 fn the_disk_defaults_are_the_measurements_made_up_front() {
3246 let cfg = Config::default();
3247 assert_eq!(cfg.disk.min_free_bytes, 8 * 1024 * 1024 * 1024);
3248 assert!(cfg.disk.auto_fold);
3249 assert_eq!(cfg.disk.fold_grace_secs, 6 * 60 * 60);
3250 assert_eq!(cfg.disk.cache_limit_bytes, 10 * 1024 * 1024 * 1024);
3251 }
3252
3253 #[test]
3254 fn an_unset_disk_section_is_the_safe_default() {
3255 let dir = tempfile::tempdir().unwrap();
3256 std::fs::write(dir.path().join("magi.toml"), "[graph]\ncandidates = 1\n").unwrap();
3257 let cfg = Config::load(&dir.path().join("magi.toml")).expect("load");
3258 assert_eq!(cfg.disk, Disk::default());
3259 }
3260
3261 #[test]
3262 fn env_is_available_to_templates_with_a_default() {
3263 let dir = tempfile::tempdir().unwrap();
3264 let path = dir.path().join("magi.toml");
3265 // teravars ships no `env`; magi adds it, and the `default` filter has
3266 // to cover the unset case or every machine would need the variable.
3267 //
3268 // Deliberately no named variable: `env` is keyed by the exact spelling
3269 // the OS reports, and Windows says `Path` where POSIX says `PATH`, so a
3270 // test asserting `env.PATH` passes on one runner and fails on another.
3271 // The map's non-emptiness is the platform-neutral claim.
3272 std::fs::write(
3273 &path,
3274 "[verify]\n\
3275 gate = [\"cache={{ env.MAGI_TEST_UNSET_XYZ | default(value='fallback') }}\", \
3276 \"populated={{ env | length > 0 }}\"]\n",
3277 )
3278 .unwrap();
3279 let cfg = Config::load(&path).expect("env lookup must render");
3280 assert_eq!(cfg.verify.gate[0], "cache=fallback");
3281 assert_eq!(cfg.verify.gate[1], "populated=true");
3282 }
3283
3284 #[test]
3285 fn a_broken_template_names_the_file() {
3286 let dir = tempfile::tempdir().unwrap();
3287 let path = dir.path().join("magi.toml");
3288 std::fs::write(&path, "[graph]\nlanguage = \"{{ nope.\"\n").unwrap();
3289 let err = Config::load(&path).expect_err("must not silently ignore");
3290 assert!(err.to_string().contains("teravars"), "{err}");
3291 }
3292
3293 #[test]
3294 fn tera_syntax_in_comments_is_inert() {
3295 // teravars >= 0.2.2 strips `#` comments before Tera sees the file, so a
3296 // comment may quote template syntax without rendering. Before 0.2.2 this
3297 // load failed: the commented-out braces reached the template parser.
3298 let dir = tempfile::tempdir().unwrap();
3299 let path = dir.path().join("magi.toml");
3300 std::fs::write(
3301 &path,
3302 "# a comment may quote templates: `{{ env.NOPE | default(value='x') }}` and `{% if %}`\n\
3303 [graph]\ncandidates = 2\n",
3304 )
3305 .unwrap();
3306 let cfg = Config::load(&path).expect("comments must be inert, not rendered");
3307 assert_eq!(cfg.graph.implementers, 2);
3308 }
3309
3310 #[test]
3311 fn opencode_defaults_to_file_delivery() {
3312 let mut s = spec("oc");
3313 s.kind = AgentKind::Opencode;
3314 assert_eq!(s.delivery(), Delivery::File);
3315 s.prompt_delivery = Some(Delivery::Argv);
3316 assert_eq!(s.delivery(), Delivery::Argv);
3317 }
3318 #[test]
3319 fn the_land_loop_is_on_but_it_cannot_merge_without_being_asked() {
3320 // Both default on, and that pair is the safety property: `land` takes
3321 // over the watching an operator was doing by hand, `land_approval`
3322 // keeps the irreversible step a human decision. An unattended merge
3323 // needs BOTH flipped, which has to be chosen deliberately twice.
3324 let g = Graph::default();
3325 assert!(
3326 g.land,
3327 "stopping at an open PR left the watching to a human"
3328 );
3329 assert!(
3330 g.land_approval,
3331 "on-by-default land is only defensible while this is also on"
3332 );
3333 assert!(g.land_rounds > 0, "a loop with no budget never terminates");
3334 assert!(
3335 g.hold_contested_merge,
3336 "a contested hand-off must still reach a human when approvals are off"
3337 );
3338 }
3339 #[test]
3340 fn an_array_declared_in_two_layers_is_refused_instead_of_concatenated() {
3341 // teravars appends arrays. For an ordered list of seats, or an argv,
3342 // the concatenation is something neither file says - and the operator
3343 // pays for the extra seats by the token.
3344 let dir = tempfile::tempdir().unwrap();
3345 let machine = dir.path().join("machine.toml");
3346 let repo = dir.path().join("magi.toml");
3347 std::fs::write(&machine, "[roles]\nimplementers = [\"a\", \"b\"]\n").unwrap();
3348 std::fs::write(&repo, "[roles]\nimplementers = [\"oc\"]\n").unwrap();
3349
3350 let err = Config::load_layers(&[machine.clone(), repo.clone()])
3351 .expect_err("two layers naming one array must not merge silently")
3352 .to_string();
3353 assert!(err.contains("roles.implementers"), "{err}");
3354 // Both files are named: the fix is to delete one of them, and the
3355 // operator has to know which two to choose between.
3356 assert!(err.contains("machine.toml"), "{err}");
3357 assert!(err.contains("magi.toml"), "{err}");
3358 }
3359
3360 #[test]
3361 fn a_scalar_in_one_layer_and_an_array_in_another_still_merges() {
3362 // The split the layering exists for: state a preference machine-wide,
3363 // let the repository own its own lists.
3364 let dir = tempfile::tempdir().unwrap();
3365 let machine = dir.path().join("machine.toml");
3366 let repo = dir.path().join("magi.toml");
3367 std::fs::write(&machine, "[roles]\nchatter = \"opus\"\n").unwrap();
3368 std::fs::write(
3369 &repo,
3370 "[[agents]]\nid = \"oc\"\nkind = \"opencode\"\n\n\
3371 [roles]\nimplementers = [\"oc\"]\n",
3372 )
3373 .unwrap();
3374
3375 let cfg = Config::load_layers(&[machine, repo]).expect("layers merge");
3376 assert_eq!(primary(cfg.roles.chatter.as_ref()), Some("opus"));
3377 assert_eq!(cfg.roles.implementers, ["oc"]);
3378 assert_eq!(cfg.agents.len(), 1, "the roster is not doubled");
3379 }
3380
3381 #[test]
3382 fn two_layers_declaring_verify_gate_run_both_in_priority_order() {
3383 // The `editorconfig-checker` distribution problem: a shared layer
3384 // wants to add a gate command without erasing the repository's own.
3385 let dir = tempfile::tempdir().unwrap();
3386 let machine = dir.path().join("machine.toml");
3387 let repo = dir.path().join("magi.toml");
3388 std::fs::write(&machine, "[verify]\ngate = [\"editorconfig-checker\"]\n").unwrap();
3389 std::fs::write(&repo, "[verify]\ngate = [\"cargo make check\"]\n").unwrap();
3390
3391 let cfg = Config::load_layers(&[machine, repo]).expect("appendable arrays must merge");
3392 assert_eq!(
3393 cfg.verify.gate,
3394 [
3395 "editorconfig-checker".to_owned(),
3396 "cargo make check".to_owned()
3397 ],
3398 "low-priority (machine) command first, high-priority (repo) command after"
3399 );
3400 }
3401
3402 fn load_one(body: &str) -> Result<Config> {
3403 let dir = tempfile::tempdir().unwrap();
3404 let path = dir.path().join("magi.toml");
3405 std::fs::write(&path, body).unwrap();
3406 Config::load_layers(&[path])
3407 }
3408
3409 #[test]
3410 fn worktree_setup_defaults_to_empty_and_parses_steps() {
3411 assert!(Config::default().worktree.setup.is_empty());
3412 let cfg = load_one(
3413 "[worktree]\nsetup = [\n { copy = \".env.example -> .env\" },\n \
3414 { copy = \"x\", optional = true },\n { run = \"true\", timeout_secs = 5 },\n]\n",
3415 )
3416 .unwrap();
3417 assert_eq!(cfg.worktree.setup.len(), 3);
3418 assert!(cfg.worktree.setup[1].optional);
3419 assert_eq!(cfg.worktree.setup[2].timeout_secs, Some(5));
3420 }
3421
3422 #[test]
3423 fn worktree_setup_rejects_malformed_steps() {
3424 for bad in [
3425 "{ copy = \"a\", run = \"b\" }",
3426 "{ optional = true }",
3427 "{ copy = \"a -> ../b\" }",
3428 "{ copy = \"a -> /abs\" }",
3429 "{ copy = \"a -> .git/config\" }",
3430 "{ run = \"b\", optional = true }",
3431 "{ copy = \"a\", timeout_secs = 3 }",
3432 "{ run = \"b\", bogus = 1 }",
3433 ] {
3434 let body = format!("[worktree]\nsetup = [{bad}]\n");
3435 assert!(load_one(&body).is_err(), "{bad}");
3436 }
3437 }
3438
3439 #[test]
3440 fn worktree_setup_in_two_layers_is_refused() {
3441 let dir = tempfile::tempdir().unwrap();
3442 let machine = dir.path().join("machine.toml");
3443 let repo = dir.path().join("magi.toml");
3444 std::fs::write(&machine, "[worktree]\nsetup = [{ run = \"a\" }]\n").unwrap();
3445 std::fs::write(&repo, "[worktree]\nsetup = [{ run = \"b\" }]\n").unwrap();
3446 assert!(Config::load_layers(&[machine, repo]).is_err());
3447 }
3448
3449 #[test]
3450 fn starter_toml_documents_worktree_setup() {
3451 let starter = Config::starter_toml();
3452 assert!(starter.contains("# [worktree]") && starter.contains("setup = ["));
3453 // The commented example must be valid once uncommented.
3454 let body = starter.replace("# [worktree]", "[worktree]");
3455 let mut out = String::new();
3456 let mut live = false;
3457 for line in body.lines() {
3458 if line == "[worktree]" {
3459 live = true;
3460 } else if live && !line.starts_with("# ") {
3461 live = false;
3462 }
3463 out.push_str(&if live {
3464 line.replacen("# ", "", 1)
3465 } else {
3466 line.to_owned()
3467 });
3468 out.push('\n');
3469 }
3470 let cfg = load_one(&out).expect("uncommented example loads");
3471 assert_eq!(cfg.worktree.setup.len(), 3);
3472 }
3473
3474 #[test]
3475 fn two_layers_declaring_verify_e2e_run_both_in_priority_order() {
3476 let dir = tempfile::tempdir().unwrap();
3477 let machine = dir.path().join("machine.toml");
3478 let repo = dir.path().join("magi.toml");
3479 std::fs::write(&machine, "[verify]\ne2e = [\"shared-smoke-test\"]\n").unwrap();
3480 std::fs::write(&repo, "[verify]\ne2e = [\"cargo test\"]\n").unwrap();
3481
3482 let cfg = Config::load_layers(&[machine, repo]).expect("appendable arrays must merge");
3483 assert_eq!(
3484 cfg.verify.e2e,
3485 ["shared-smoke-test".to_owned(), "cargo test".to_owned()]
3486 );
3487 }
3488
3489 #[test]
3490 fn two_layers_declaring_repos_roots_are_both_scanned() {
3491 let dir = tempfile::tempdir().unwrap();
3492 let machine = dir.path().join("machine.toml");
3493 let repo = dir.path().join("magi.toml");
3494 std::fs::write(&machine, "[repos]\nroots = [\"/machine/root\"]\n").unwrap();
3495 std::fs::write(&repo, "[repos]\nroots = [\"/repo/root\"]\n").unwrap();
3496
3497 let cfg = Config::load_layers(&[machine, repo]).expect("appendable arrays must merge");
3498 assert_eq!(
3499 cfg.repos.roots,
3500 [PathBuf::from("/machine/root"), PathBuf::from("/repo/root")]
3501 );
3502 }
3503
3504 #[test]
3505 fn duplicate_gate_commands_across_layers_both_run() {
3506 // Dropping the duplicate would be a silent surprise; the operator
3507 // sees a slower gate, never a missing one.
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, "[verify]\ngate = [\"same-command\"]\n").unwrap();
3512 std::fs::write(&repo, "[verify]\ngate = [\"same-command\"]\n").unwrap();
3513
3514 let cfg = Config::load_layers(&[machine, repo]).expect("appendable arrays must merge");
3515 assert_eq!(
3516 cfg.verify.gate,
3517 ["same-command".to_owned(), "same-command".to_owned()]
3518 );
3519 }
3520
3521 #[test]
3522 fn notify_command_is_still_refused_across_two_layers() {
3523 // An argv, not a set: concatenating two of them is not a program.
3524 let dir = tempfile::tempdir().unwrap();
3525 let machine = dir.path().join("machine.toml");
3526 let repo = dir.path().join("magi.toml");
3527 std::fs::write(&machine, "[notify]\ncommand = [\"ntfy\", \"publish\"]\n").unwrap();
3528 std::fs::write(&repo, "[notify]\ncommand = [\"curl\", \"-X\"]\n").unwrap();
3529
3530 let err = Config::load_layers(&[machine.clone(), repo.clone()])
3531 .expect_err("an argv split across layers must not concatenate")
3532 .to_string();
3533 assert!(err.contains("notify.command"), "{err}");
3534 assert!(err.contains("machine.toml"), "{err}");
3535 assert!(err.contains("magi.toml"), "{err}");
3536 }
3537
3538 #[test]
3539 fn one_layer_declaring_verify_gate_runs_unchanged() {
3540 // The classification must not change behaviour for the configuration
3541 // this very repository has today: exactly one layer names the gate.
3542 let dir = tempfile::tempdir().unwrap();
3543 let path = dir.path().join("magi.toml");
3544 std::fs::write(&path, "[verify]\ngate = [\"cargo make check\"]\n").unwrap();
3545
3546 let cfg = Config::load(&path).expect("single layer must still load");
3547 assert_eq!(cfg.verify.gate, ["cargo make check".to_owned()]);
3548 }
3549
3550 #[test]
3551 fn describe_composed_names_the_contributing_layers_only_when_there_are_two() {
3552 let dir = tempfile::tempdir().unwrap();
3553 let machine = dir.path().join("machine.toml");
3554 let repo = dir.path().join("magi.toml");
3555 std::fs::write(&machine, "[verify]\ngate = [\"editorconfig-checker\"]\n").unwrap();
3556 std::fs::write(&repo, "[verify]\ngate = [\"cargo make check\"]\n").unwrap();
3557 let paths = vec![machine.clone(), repo.clone()];
3558
3559 let cfg = Config::load_layers(&paths).expect("appendable arrays must merge");
3560 let described =
3561 Config::describe_composed(&paths, &cfg.verify.gate, "verify.gate", "(none)");
3562 assert!(described.contains("editorconfig-checker && cargo make check"));
3563 assert!(
3564 described.contains(&machine.display().to_string()),
3565 "{described}"
3566 );
3567 assert!(
3568 described.contains(&repo.display().to_string()),
3569 "{described}"
3570 );
3571
3572 // A single contributing layer stays the plain one-line summary.
3573 let single = vec![repo.clone()];
3574 let solo_cfg = Config::load_layers(&single).expect("single layer loads");
3575 let solo_described =
3576 Config::describe_composed(&single, &solo_cfg.verify.gate, "verify.gate", "(none)");
3577 assert_eq!(solo_described, "cargo make check");
3578 }
3579
3580 mod remote_layers {
3581 use super::*;
3582
3583 fn git(dir: &Path, args: &[&str]) {
3584 use crate::proc::Quiet as _;
3585 let out = std::process::Command::new("git")
3586 .args(args)
3587 .quiet()
3588 .current_dir(dir)
3589 .output()
3590 .unwrap();
3591 assert!(out.status.success(), "git {args:?}: {out:?}");
3592 }
3593
3594 /// A checkout whose origin/main carries `remote` as magi.toml and whose
3595 /// working tree carries `disk`.
3596 fn fixture(remote: &str, disk: &str) -> (tempfile::TempDir, PathBuf) {
3597 let tmp = tempfile::tempdir().unwrap();
3598 let bare = tmp.path().join("o.git");
3599 git(tmp.path(), &["init", "--bare", "-b", "trunk", "o.git"]);
3600 let repo = tmp.path().join("repo");
3601 std::fs::create_dir_all(&repo).unwrap();
3602 git(&repo, &["init", "-b", "trunk"]);
3603 git(&repo, &["config", "user.name", "t"]);
3604 git(&repo, &["config", "user.email", "t@example.com"]);
3605 std::fs::write(repo.join("magi.toml"), remote).unwrap();
3606 git(&repo, &["add", "-A"]);
3607 git(&repo, &["commit", "-m", "x"]);
3608 git(&repo, &["remote", "add", "origin", &bare.to_string_lossy()]);
3609 git(&repo, &["push", "origin", "trunk"]);
3610 git(&repo, &["fetch", "origin"]);
3611 git(&repo, &["remote", "set-head", "origin", "trunk"]);
3612 std::fs::write(repo.join("magi.toml"), disk).unwrap();
3613 (tmp, repo)
3614 }
3615
3616 #[test]
3617 fn reads_the_remote_blob_not_the_working_tree() {
3618 let (_t, repo) = fixture("[graph]\ncandidates = 2\n", "[graph]\ncandidates = 5\n");
3619 let (cfg, _) = Config::discover(&repo, None).unwrap();
3620 assert_eq!(cfg.graph.implementers, 2);
3621 }
3622
3623 #[test]
3624 fn detached_head_still_reads_the_remote() {
3625 let (_t, repo) = fixture("[graph]\ncandidates = 2\n", "[graph]\ncandidates = 5\n");
3626 git(&repo, &["checkout", "--detach"]);
3627 let (cfg, _) = Config::discover(&repo, None).unwrap();
3628 assert_eq!(cfg.graph.implementers, 2);
3629 }
3630
3631 #[test]
3632 fn a_missing_ref_is_an_error_not_a_fallback() {
3633 let (_t, repo) = fixture("[graph]\ncandidates = 2\n", "[graph]\ncandidates = 5\n");
3634 git(&repo, &["update-ref", "-d", "refs/remotes/origin/trunk"]);
3635 let err = format!("{:#}", Config::discover(&repo, None).unwrap_err());
3636 assert!(err.contains("origin"), "{err}");
3637 }
3638
3639 #[test]
3640 fn merge_base_in_the_checkout_locates_the_ref() {
3641 let (_t, repo) = fixture("[graph]\ncandidates = 2\n", "[merge]\nbase = \"nope\"\n");
3642 assert!(Config::discover(&repo, None).is_err());
3643 }
3644
3645 #[test]
3646 fn the_magi_dir_layer_is_read_too() {
3647 let (_t, repo) = fixture("[graph]\ncandidates = 2\n", "");
3648 std::fs::write(repo.join("magi.toml"), "[graph]\ncandidates = 2\n").unwrap();
3649 std::fs::create_dir_all(repo.join(".magi")).unwrap();
3650 std::fs::write(repo.join(".magi/config.toml"), "[graph]\njudges = 3\n").unwrap();
3651 git(&repo, &["add", "-A"]);
3652 git(&repo, &["commit", "-m", "y"]);
3653 git(&repo, &["push", "origin", "trunk"]);
3654 git(&repo, &["fetch", "origin"]);
3655 let (cfg, layers) = Config::discover(&repo, None).unwrap();
3656 assert_eq!((cfg.graph.implementers, cfg.graph.judges), (2, 3));
3657 assert_eq!(layers.len(), 2);
3658 }
3659
3660 #[test]
3661 fn an_include_in_a_repository_layer_resolves_against_the_same_commit() {
3662 let (_t, repo) = fixture("include = [\"gate.toml\"]\n", "");
3663 std::fs::write(repo.join("magi.toml"), "include = [\"gate.toml\"]\n").unwrap();
3664 std::fs::write(repo.join("gate.toml"), "[graph]\ncandidates = 2\n").unwrap();
3665 git(&repo, &["add", "-A"]);
3666 git(&repo, &["commit", "-m", "inc"]);
3667 git(&repo, &["push", "origin", "trunk"]);
3668 git(&repo, &["fetch", "origin"]);
3669 // The working tree's copy differs and must not be used.
3670 std::fs::write(repo.join("gate.toml"), "[graph]\ncandidates = 9\n").unwrap();
3671 let (cfg, _) = Config::discover(&repo, None).unwrap();
3672 assert_eq!(cfg.graph.implementers, 2);
3673 }
3674
3675 #[test]
3676 fn a_machine_layer_base_is_rendered_before_locating_the_ref() {
3677 let (t, repo) = fixture("[graph]\ncandidates = 2\n", "[merge]\nbase = \"nope\"\n");
3678 let m = t.path().join("machine.toml");
3679 std::fs::write(
3680 &m,
3681 "[merge]\nbase = \"{{ env.MAGI_TEST_NO_SUCH | default(value='trunk') }}\"\n",
3682 )
3683 .unwrap();
3684 let (remote, base) = {
3685 let probe = probe_merge(&m, &repo).unwrap();
3686 (probe.remote, probe.base)
3687 };
3688 assert_eq!((remote, base), (None, Some("trunk".to_owned())));
3689 }
3690
3691 #[test]
3692 fn a_machine_layer_base_sees_the_checkouts_repo_name() {
3693 let (t, repo) = fixture("[graph]\ncandidates = 2\n", "");
3694 let m = t.path().join("machine.toml");
3695 std::fs::write(&m, "[merge]\nbase = \"{{ repo_name }}\"\n").unwrap();
3696 let probe = probe_merge(&m, &repo).unwrap();
3697 assert_eq!(probe.base.as_deref(), Some("repo"));
3698 }
3699
3700 #[test]
3701 fn every_include_form_is_recognised() {
3702 assert!(declares_include("include = [\"a.toml\"]\n"));
3703 assert!(declares_include("\"include\" = [\"a.toml\"]\n"));
3704 assert!(declares_include("teravars = { include = [\"a.toml\"] }\n"));
3705 assert!(declares_include("[teravars]\ninclude = [\"a.toml\"]\n"));
3706 assert!(!declares_include("[graph]\ncandidates = 2\n"));
3707 }
3708
3709 #[test]
3710 fn an_explicit_config_is_read_as_named() {
3711 let (t, repo) = fixture("[graph]\ncandidates = 2\n", "");
3712 let own = t.path().join("own.toml");
3713 std::fs::write(&own, "[graph]\ncandidates = 4\n").unwrap();
3714 let (cfg, _) = Config::discover(&repo, Some(&own)).unwrap();
3715 assert_eq!(cfg.graph.implementers, 4);
3716 }
3717
3718 #[test]
3719 fn discover_fetched_fails_when_the_fetch_does() {
3720 let (t, repo) = fixture("[graph]\ncandidates = 2\n", "");
3721 std::fs::remove_dir_all(t.path().join("o.git")).unwrap();
3722 let rt = tokio::runtime::Builder::new_current_thread()
3723 .enable_all()
3724 .build()
3725 .unwrap();
3726 let err = rt
3727 .block_on(Config::discover_fetched(&repo, None))
3728 .unwrap_err();
3729 assert!(format!("{err:#}").contains("fetch"), "{err:#}");
3730 }
3731
3732 #[test]
3733 fn discover_fetched_survives_a_fetch_that_leaves_no_origin_head() {
3734 // Never-fetched remote + followRemoteHEAD=never: fetch creates
3735 // origin/trunk but not origin/HEAD, which discovery then re-reads.
3736 let (_t, repo) = fixture("[graph]\ncandidates = 2\n", "");
3737 git(&repo, &["remote", "set-head", "origin", "-d"]);
3738 git(&repo, &["update-ref", "-d", "refs/remotes/origin/trunk"]);
3739 git(
3740 &repo,
3741 &["config", "remote.origin.followRemoteHEAD", "never"],
3742 );
3743 let rt = tokio::runtime::Builder::new_current_thread()
3744 .enable_all()
3745 .build()
3746 .unwrap();
3747 let (cfg, _) = rt.block_on(Config::discover_fetched(&repo, None)).unwrap();
3748 assert_eq!(cfg.graph.implementers, 2);
3749 }
3750
3751 #[test]
3752 fn a_non_git_directory_reads_its_files() {
3753 let dir = tempfile::tempdir().unwrap();
3754 std::fs::write(dir.path().join("magi.toml"), "[graph]\ncandidates = 4\n").unwrap();
3755 let (cfg, _) = Config::discover(dir.path(), None).unwrap();
3756 assert_eq!(cfg.graph.implementers, 4);
3757 }
3758 }
3759}