flagon-io/g1t

public

Usage while free is shown at cost; agents get rustfmt and clippy

The sidebar's "Usage this month" and the usage page summed what was charged. While g1t is free every charge is zero, so both stopped at what had been charged before free mode began, with the margin in it, and every run since showed as nothing: $6.01 against $6.48 of real use, and further below the AI Gateway's own estimate. Billing now reports usedMicros, what the runs cost on g1t's models and the workspace's own provider together, and measures the breakdowns at cost while free. The card and the usage page show "Used", at cost, and say nothing is charged and credit is not drawn down. The runner image installed Rust with the minimal profile, which has no rustfmt or clippy, so an agent could not format or lint its change and first heard about it from a failed workflow (hello's #85: Formatting failed, the agent was sent back and fixed it). Both are installed now. docs/DEMO.md is rewritten for what g1t is today: GitHub Actions from .g1t/workflows, an agent writing a CI step (#80/#81), an agent sent back by a failed workflow and fixing it (#84/#85), and the merge queue testing combinations with merge_group.

syntaqxcommitted 3d agoParent8f91758Browse files

9 files+169−940/9 viewed

+4−3View BlameViewed

65	65		/** Whether g1t charges nothing for now, while it is being built out. */
66	66		free?: boolean;
67	67		/** What its agents have cost since the start of the month. */
68		−	monthSpentMicros: number \| null;
	68	+	/** This month's usage: charged, or at cost while g1t is free. */
	69	+	monthUsageMicros: number \| null;
69	70		};
70	71
71	72		function SidebarLink({
⋯
176	177
177	178		/** This month's spend against what is left, as Vercel shows a plan's usage. */
178	179		function UsageCard({ slug, shell }: { slug: string; shell: ShellData }) {
179		−	if (shell.monthSpentMicros == null) return null;
180		−	const spent = shell.monthSpentMicros;
	180	+	if (shell.monthUsageMicros == null) return null;
	181	+	const spent = shell.monthUsageMicros;
181	182		const left = shell.creditMicros;
182	183		const share = left != null && spent + left > 0 ? Math.min(1, spent / (spent + Math.max(left, 0))) : 0;
183	184		return (

+2−1View BlameViewed

113	113		creditMicros:
114	114		account?.ok && account.value.status.enabled && !account.value.status.free ? account.value.balanceMicros : null,
115	115		free: account?.ok ? Boolean(account.value.status.free) : false,
116		−	monthSpentMicros: usage?.ok ? usage.value.spentMicros : null,
	116	+	// While g1t is free every charge is zero, so usage is shown at cost.
	117	+	monthUsageMicros: usage?.ok ? (usage.value.free ? usage.value.usedMicros : usage.value.spentMicros) : null,
117	118		};
118	119		}
119	120

+1−1View BlameViewed

264	264
265	265		const workspace = shell?.workspace?.slug;
266	266		const first = repos[0];
267		−	const handedOff = active.length > 0 \|\| (shell?.monthSpentMicros ?? 0) > 0;
	267	+	const handedOff = active.length > 0 \|\| (shell?.monthUsageMicros ?? 0) > 0;
268	268		const steps: Step[] = [
269	269		{
270	270		done: Boolean(workspace),

+36−19View BlameViewed

189	189		const { period, since, usage, account } = loaderData;
190	190		const base = `/${params.owner}`;
191	191		const days = Math.max(1, Math.ceil((Date.now() - new Date(since).getTime()) / 86_400_000));
192		−	const perDay = usage.spentMicros / days;
	192	+	// While g1t is free nothing is charged, so usage is measured at cost
	193	+	// and credit is not drawn down.
	194	+	const total = usage.free ? usage.usedMicros : usage.spentMicros;
	195	+	const perDay = usage.free ? 0 : total / days;
193	196		const runway = perDay > 0 ? Math.floor(account.balanceMicros / perDay) : null;
194	197		return (
195	198		<div className="space-y-8">
⋯
211	214		</div>
212	215
213	216		<div className="grid gap-4 sm:grid-cols-2 lg:grid-cols-4">
214		−	<Stat
215		−	label="Spent"
216		−	value={dollars(usage.spentMicros)}
217		−	note={
218		−	usage.providerMicros > 0
219		−	? `Plus about ${dollars(usage.providerMicros)} billed by your own model provider`
220		−	: `${dollars(usage.costMicros)} of it the model provider's`
221		−	}
222		−	/>
	217	+	{usage.free ? (
	218	+	<Stat
	219	+	label="Used"
	220	+	value={dollars(total)}
	221	+	note="At cost. Nothing is charged while g1t is being built out."
	222	+	/>
	223	+	) : (
	224	+	<Stat
	225	+	label="Spent"
	226	+	value={dollars(usage.spentMicros)}
	227	+	note={
	228	+	usage.providerMicros > 0
	229	+	? `Plus about ${dollars(usage.providerMicros)} billed by your own model provider`
	230	+	: `${dollars(usage.costMicros)} of it the model provider's`
	231	+	}
	232	+	/>
	233	+	)}
223	234		<Stat label="Agent runs" value={String(usage.runs)} note={PERIODS[period]} />
224	235		<Stat
225	236		label="Average run"
226		−	value={usage.runs ? dollars(usage.spentMicros / usage.runs, 3) : "—"}
	237	+	value={usage.runs ? dollars(total / usage.runs, 3) : "—"}
227	238		note="Making a change, reviewing, revising…"
228	239		/>
229	240		<div className="rounded-2xl bg-surface p-5 ring-1 ring-line">
230	241		<p className="flex items-center justify-between text-sm text-muted">
231	242		Credit left
232		−	<Link to={`${base}/-/billing`} className="text-xs text-accent hover:underline">
233		−	Add credit
234		−	</Link>
	243	+	{!usage.free && (
	244	+	<Link to={`${base}/-/billing`} className="text-xs text-accent hover:underline">
	245	+	Add credit
	246	+	</Link>
	247	+	)}
235	248		</p>
236	249		<p className={`mt-2 text-3xl font-semibold tracking-tight tabular-nums ${account.balanceMicros <= 0 ? "text-warn" : ""}`}>
237	250		{dollars(account.balanceMicros)}
238	251		</p>
239	252		<p className="mt-1 text-xs text-faint">
240		−	{runway == null ? "Nothing spent in this period." : `About ${runway} ${runway === 1 ? "day" : "days"} at this rate.`}
	253	+	{usage.free
	254	+	? "Not drawn down while g1t is free."
	255	+	: runway == null
	256	+	? "Nothing spent in this period."
	257	+	: `About ${runway} ${runway === 1 ? "day" : "days"} at this rate.`}
241	258		</p>
242	259		</div>
243	260		</div>
244	261
245	262		<section className="rounded-2xl bg-surface p-5 ring-1 ring-line">
246	263		<div className="flex flex-wrap items-baseline justify-between gap-3">
247		−	<h3 className="text-sm font-medium">Spend per day</h3>
	264	+	<h3 className="text-sm font-medium">{usage.free ? "Usage per day" : "Spend per day"}</h3>
248	265		<ul className="flex flex-wrap gap-x-4 gap-y-1 text-xs text-muted">
249	266		{usage.byTask.map((slice) => (
250	267		<li key={slice.key} className="flex items-center gap-1.5">
⋯
263	280		<Breakdown
264	281		title="By kind of work"
265	282		slices={usage.byTask}
266		−	total={usage.spentMicros}
	283	+	total={total}
267	284		label={(key) => task(key).label}
268	285		color={(key) => task(key).color}
269	286		/>
270		−	<Breakdown title="By repository" slices={usage.byRepo} total={usage.spentMicros} link={(key) => `/${key}`} />
	287	+	<Breakdown title="By repository" slices={usage.byRepo} total={total} link={(key) => `/${key}`} />
271	288		<Breakdown
272	289		title="Pull requests that cost most"
273	290		slices={usage.byPull}
⋯
280	297		return number && number !== "0" ? `/${repo}/pull/${number}` : `/${repo}/plans`;
281	298		}}
282	299		/>
283		−	<Breakdown title="By model" slices={usage.byModel} total={usage.spentMicros} color={() => "var(--color-info)"} />
	300	+	<Breakdown title="By model" slices={usage.byModel} total={total} color={() => "var(--color-info)"} />
284	301		</div>
285	302
286	303		<div className="flex flex-wrap items-center justify-between gap-3 rounded-2xl bg-surface px-5 py-4 text-sm ring-1 ring-line">

+6−0View BlameViewed

207	207		/// What runs on the workspace's own provider cost there, as the harness
208	208		/// estimated it. Not charged by g1t.
209	209		pub provider_micros: i64,
	210	+	/// What the runs used, at cost: g1t's models and the workspace's own
	211	+	/// provider together, whatever was charged for them.
	212	+	pub used_micros: i64,
	213	+	/// g1t charges nothing for now. The slices then measure usage at cost,
	214	+	/// since every charge is zero.
	215	+	pub free: bool,
210	216		pub runs: u32,
211	217		/// Spend per day (`YYYY-MM-DD`) and task, as `day/task` keys.
212	218		pub by_day: Vec<UsageSlice>,

+106−67View BlameViewed

1	1		# Demo script
2	2
3		−	A walk through g1t for the submission video. It runs seven to eight minutes
4		−	at a normal speaking pace and uses `syntaqx/hello`, which is seeded for it.
	3	+	A walk through g1t for the submission video. It runs about eight minutes at
	4	+	a normal speaking pace and uses `syntaqx/hello`, a small Rust greeter whose
	5	+	whole history was written by agents working on issues.
5	6
6	7		Everything shown is live on g1t.sh. Nothing is mocked.
7	8
	9	+	The judges weigh agent collaboration (half), concurrency and conflicts (a
	10	+	quarter) and ease of use (a quarter). Sections 3 to 6 carry the first two;
	11	+	sections 2 and 8 the third.
	12	+
8	13		## Before recording
9	14
10	15		- Sign in as `syntaqx`.
11	16		- Have a terminal open in an empty directory, with Claude Code installed.
12	17		- Open `https://g1t.sh/syntaqx/hello` in one tab and `https://g1t.sh/` in
13	18		another.
14		−	- Agents take one to three minutes. Start them, talk over something else,
15		−	and come back. Do not deploy the runner while they work.
	19	+	- Write the three issues for section 3 in a scratch file so they can be
	20	+	pasted (titles below). Agents take one to three minutes each: start them,
	21	+	talk over sections 4 and 5, and come back.
	22	+	- Do not deploy the runner while agents work.
16	23
17	24		## 1. The problem (30 seconds)
18	25
⋯
25	32		> is ordinary git, with issues and pull requests, and it runs entirely on
26	33		> Cloudflare.
27	34
28		−	## 2. It is still git (45 seconds)
	35	+	## 2. It is still git, and your CI comes with you (1 minute)
29	36
30	37		On `syntaqx/hello`, Code tab.
31	38
32	39		- Show the clone box: HTTPS, and the one line that connects an agent.
33	40		- In the terminal: `git clone https://g1t.sh/syntaqx/hello.git`.
34		−	- Mention: storage is Cloudflare Artifacts; the site, the API and every
35		−	service are Workers.
	41	+	- Open `.g1t/workflows/ci.yml`. It is a GitHub Actions workflow, unchanged:
	42	+	`actions/checkout@v7`, a Rust toolchain action, `actions/cache@v6`, then
	43	+	formatting, lints, an "Every flag is documented" step, and tests.
	44	+
	45	+	> Moving from GitHub is renaming `.github` to `.g1t`. The same workflow
	46	+	> syntax, the same actions from the marketplace, the same `push`,
	47	+	> `pull_request` and `merge_group` events. Storage is Cloudflare Artifacts;
	48	+	> every job runs in its own Cloudflare Container.
36	49
37		−	## 3. Many issues, an agent on each (2 minutes)
	50	+	- Actions tab: the runs, by event. Open one and show the steps and the log.
	51	+
	52	+	## 3. Many issues, an agent on each (1 minute 30 seconds)
38	53
39	54		Issues tab.
40	55
41		−	- Point out labels, and that issues come from people, agents or anything
42		−	with a token, such as an error tracker.
43		−	- Tick several open issues and press Assign to g1t agent. Say that
44		−	there is nothing else to choose: no number of agents and no model. Each
45		−	issue gets an agent of its own and g1t routes the work; every session
46		−	opens by naming the model that ran.
47		−	- Open Document the command-line options. Its draft pull request has
48		−	appeared. Show the Session tab filling in live: the prompt, the
49		−	agent's reasoning, every command.
	56	+	- Create three issues quickly, pasting them in:
	57	+	- Add a --sparkle flag that ends the greeting with a sparkle emoji.
	58	+	- Greet in German with `--lang de`.
	59	+	- Explain in the README what happens with no name.
	60	+	- Tick all three and press Assign to g1t agent. Say there is nothing
	61	+	else to choose: no number of agents, no model. Each issue gets an agent of
	62	+	its own and g1t routes the work; every session opens by naming the model
	63	+	that ran.
	64	+	- Open one. Its draft pull request has appeared. Show the Session tab
	65	+	filling in live: the prompt, what the agent was told about the other work
	66	+	in progress, every command it runs.
50	67
51		−	> Each agent has its own sandbox, a Cloudflare Container, and its own fork.
52		−	> A fork is copy-on-write, so it costs about what a branch would, and an
53		−	> agent cannot damage what it cannot write to.
	68	+	> Each agent has its own sandbox and its own fork. A fork is copy-on-write,
	69	+	> so it costs about what a branch would, and an agent cannot damage what it
	70	+	> cannot write to. Each is told what else is in flight, so two agents on
	71	+	> the same file know about each other before they collide.
54	72
55		−	While they run, go to the next section.
	73	+	While they run, go on.
56	74
57		−	## 4. Checks nobody can fake (1 minute)
	75	+	## 4. Agents keep CI honest, and fix what it catches (1 minute 30 seconds)
58	76
59		−	Open Say goodbye too and its pull request, Add a farewell.
	77	+	Open issue #80, CI: fail when a flag is missing from the README, and its
	78	+	pull request #81.
60	79
61		−	- This one was pushed as a branch by a person, the way you already work.
62		−	- Checks failed. Expand `cargo test` and show the output.
63		−	- Changes tab: the reviewer's comment sits on the faulty line.
	80	+	- An agent wrote this CI step. Changes tab: the shell step it added to
	81	+	`ci.yml`. It went through review and the merge queue like any change.
64	82
65		−	> The issue says what done means: here, `cargo test`. g1t runs that itself,
66		−	> in a clean sandbox that holds only this commit. The agent that wrote the
67		−	> code never touches the result, so a pass means something. And a pull
68		−	> request that has not passed cannot be merged.
	83	+	Open issue #84, Add a --reverse flag, and its pull request #85.
69	84
70		−	## 5. Choosing between pull requests (1 minute 15 seconds)
	85	+	- Its checks, `cargo test`, passed. Its first workflow run did not:
	86	+	Formatting failed. Open the run and show the step and its log.
	87	+	- Session tab: the agent's second session opens with the failed run, the
	88	+	instruction to read it with `get_workflow_run` and `get_job_logs`, and to
	89	+	fix the code rather than the workflow. Show it reading the log, fixing
	90	+	the formatting, and pushing. The second run is green.
71	91
72		−	Open Greet in Spanish and French.
	92	+	> Nobody marks their own homework. Workflows run in a clean sandbox on the
	93	+	> exact commit; the agent that wrote the code never touches the result. A
	94	+	> failure goes back to the agent with the log, and the pull request cannot
	95	+	> merge until it is green.
73	96
74		−	- Two pull requests for the same issue, side by side: checks, size of the
75		−	change, who reviewed.
76		−	- Open one. Show Review by a g1t agent: comments on lines, a summary, a
77		−	verdict. Say that an agent cannot review its own pull request.
78		−	- Show Other work is changing the same files.
	97	+	## 5. Checks, reviews and choosing between pull requests (1 minute)
79	98
80		−	> This is the overlap radar. Two pull requests for different issues are
81		−	> editing the same file. g1t says so while the work is still going on, not
82		−	> at the end as a merge conflict. Agents see the same thing through the API.
	99	+	Open Say goodbye too (#4) and its pull request Add a farewell (#9).
83	100
84		−	## 6. Converging on main (1 minute 15 seconds)
	101	+	- This one was pushed as a branch by a person, the way you already work.
	102	+	- Checks failed. Expand `cargo test` and show the output. Changes tab:
	103	+	the reviewer's comment sits on the faulty line. The agent's pull request
	104	+	for the same issue, #10, passed and was merged; #9 was closed.
85	105
86		−	Open A blank name greets nobody.
	106	+	Open Greet in Spanish and French (#2).
87	107
88		−	- It is closed, and it says which pull request resolved it. The other one
89		−	is marked superseded.
	108	+	- Two pull requests for one issue, side by side: checks, size of the change,
	109	+	who reviewed. Open one and show Other work is changing the same files.
90	110
91		−	> Two pull requests for one issue, one merged. The issue records which.
	111	+	> This is the overlap radar. g1t says so while the work is still going on,
	112	+	> not at the end as a merge conflict. Agents see the same thing through the
	113	+	> API, which is how the agents in section 3 were told about each other.
92	114
93		−	Go back to a pull request for the Spanish and French issue that says main
94		−	has moved.
	115	+	Open A blank name greets nobody (#1): closed, saying which pull request
	116	+	resolved it; the other is marked superseded.
	117	+
	118	+	## 6. The merge queue (1 minute 15 seconds)
	119	+
	120	+	Back to the pull requests from section 3. Their checks have passed and a g1t
	121	+	agent has reviewed them.
	122	+
	123	+	- With auto-merge on, they enter the Merge queue on their own. Open it.
	124	+
	125	+	> Three changes, written at the same time, each green on its own. That
	126	+	> proves nothing about all three together. The queue builds main with the
	127	+	> first, main with the first and second, and so on, and tests every one of
	128	+	> those combinations at once, in parallel sandboxes: the issues' acceptance
	129	+	> checks, every check main has promised so far, and the repository's
	130	+	> `merge_group` workflows, exactly as GitHub's merge queue sends them.
95	131
96		−	- Press Catch up with main. Open the Session tab.
	132	+	- As each lands, the issue closes, recording which pull request resolved it.
	133	+	- Show #79 and #81 under Recent: landed, with the `merge_group` run.
97	134
98		−	> Something else landed first, so this pull request is behind. A g1t agent
99		−	> merges main in. If that conflicts, the agent is given both sides and what
100		−	> this pull request is for, and resolves it. Then the checks run again on
101		−	> the result.
	135	+	> When a combination fails, that entry is taken out with the reason, the
	136	+	> ones behind it are tested again without it, and its agent is sent back
	137	+	> to fix it. A conflict with something ahead of it says which.
102	138
103		−	- When it is done, merge it. The issue closes; the other pull request for
104		−	it closes as superseded.
	139	+	If one conflicts on camera, so much the better: open its Session and show
	140	+	the agent being given both sides and what the pull request is for.
105	141
106	142		## 7. Bring your own agent (45 seconds)
107	143
⋯
114	150		- In Claude Code, `/mcp`, choose g1t. The browser opens on g1t's consent
115	151		page. Approve.
116	152		- Ask: "What issues are open on syntaqx/hello on g1t, and which pull
117		−	requests overlap?"
	153	+	requests overlap? Did the last CI run pass?"
118	154
119		−	> No token to paste. The same operations are a REST API at api.g1t.sh, and
120		−	> the two are generated from one list, so they cannot drift apart.
	155	+	> No token to paste. The same operations are a REST API at api.g1t.sh,
	156	+	> including GitHub's own Actions endpoints, and the two are generated from
	157	+	> one list, so they cannot drift apart.
121	158
122	159		## 8. Close (30 seconds)
123	160
124		−	Back on the first issue: the three pull requests from section 3 are ready,
125		−	with their checks.
	161	+	Back on the Issues tab: the three issues from section 3, closed, each saying
	162	+	which pull request resolved it.
126	163
127		−	> Issues and pull requests, as you know them. What changes is the number of
128		−	> hands. Every pull request isolated in its own fork, every decision
129		−	> recorded with the code, checks that agents cannot mark themselves, overlap
130		−	> flagged while it is happening, and one clear answer to which change you
131		−	> took. g1t is open source, and it is hosted on itself.
	164	+	> Issues and pull requests, as you know them, and your GitHub Actions as
	165	+	> they are. What changes is the number of hands. Every agent isolated in its
	166	+	> own fork, told what the others are doing, held to checks and workflows it
	167	+	> cannot mark itself, reviewed, and landed through a queue that tests the
	168	+	> combinations. g1t is open source, free while it is being built out, and
	169	+	> hosted on itself.
132	170
133	171		Show `https://g1t.sh/syntaqx/g1t`.
134	172
⋯
136	174
137	175		\| What \| Do \|
138	176		\| --- \| --- \|
139		−	\| An agent's pull request closes itself \| Open its Session; the last note says why. Assign another. \|
	177	+	\| An agent's pull request closes itself \| Open its Session; the last note says why. Assign the issue again. \|
	178	+	\| A workflow stays queued \| Open the run and press Re-run all jobs. \|
140	179		\| Checks stay queued \| Press the re-run button on the checks panel. \|
141		−	\| Catch up does nothing \| The pull request was already up to date; refresh. \|
142		−	\| Merge is refused \| Read the message: it is a draft, its checks have not passed, or main moved. \|
	180	+	\| Nothing enters the queue \| Auto-merge waits for checks, workflows and a review; the pull request's sidebar says which is missing. \|
	181	+	\| Merge is refused \| Read the message: it is a draft, its checks or workflows have not passed, or it needs a review. \|

+4−0View BlameViewed

116	116		costMicros: number;
117	117		/** What runs on the workspace's own provider cost there, estimated. Not charged by g1t. */
118	118		providerMicros: number;
	119	+	/** What the runs used, at cost: g1t's models and the workspace's own provider together. */
	120	+	usedMicros: number;
	121	+	/** g1t charges nothing for now; the slices then measure usage at cost. */
	122	+	free: boolean;
119	123		runs: number;
120	124		/** Spend per day and task, keyed `YYYY-MM-DD/task`. */
121	125		byDay: UsageSlice[];

+5−1View BlameViewed

261	261		micros: Option<i64>,
262	262		runs: Option<u32>,
263	263		}
	264	+	// While nothing is charged, what was used is what there is to show.
	265	+	let measure = if self.free { "COALESCE(cost_micros, 0)" } else { "-amount_micros" };
264	266		let slices = \|key: &str, limit: u32\| {
265	267		format!(
266		−	"SELECT {key} AS key, -SUM(amount_micros) AS micros, COUNT(*) AS runs FROM ledger
	268	+	"SELECT {key} AS key, SUM({measure}) AS micros, COUNT(*) AS runs FROM ledger
267	269		WHERE workspace = ?1 AND kind = 'usage' AND created_at >= ?2
268	270		GROUP BY 1 ORDER BY micros DESC LIMIT {limit}"
269	271		)
⋯
323	325		spent_micros: totals.spent.unwrap_or_default(),
324	326		cost_micros: totals.cost.unwrap_or_default(),
325	327		provider_micros: totals.provider.unwrap_or_default(),
	328	+	used_micros: totals.cost.unwrap_or_default() + totals.provider.unwrap_or_default(),
	329	+	free: self.free,
326	330		runs: totals.runs.unwrap_or_default(),
327	331		added_micros: totals.added.unwrap_or_default(),
328	332		by_day: query(slices("substr(created_at, 1, 10) \|\| '/' \|\| COALESCE(task, 'other')", 400)).await?,

+5−2View BlameViewed

32	32		# Claude Code refuses to skip permission prompts as root.
33	33		USER node
34	34		ENV HOME=/home/node
35		−	# Rust, for the agent's own use, installed for the user it runs as.
	35	+	# Rust, for the agent's own use, installed for the user it runs as, with
	36	+	# the formatter and linter that CI so often checks with: an agent that
	37	+	# cannot run them only finds out from a failed workflow.
36	38		RUN curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs \
37		−	\| sh -s -- -y --profile minimal --default-toolchain stable
	39	+	\| sh -s -- -y --profile minimal --default-toolchain stable \
	40	+	--component rustfmt --component clippy
38	41		ENV PATH=/home/node/.cargo/bin:$PATH
39	42		# Commits are the g1t agent's; the harness does not sign them as its own.
40	43		RUN mkdir -p /home/node/.claude \