Skip to content

Conch.app's commit graph: agents' branches, conflicts, CI and the autopilot - #250

Merged
Amitgb14 merged 38 commits into
masterfrom
conch/we-have-basic-git-view
Nov 16, 2026
Merged

Amitgb14 merged 38 commits into
masterfrom
conch/we-have-basic-git-view

Conversation

@Amitgb14

Copy link
Copy Markdown
Owner

Conch.app's commit graph, as in the git view mock (option A, Lanes: https://claude.ai/artifact/2C83Mjyp7fZkzR26iTHvqS), and what it needs from the server.

Commit graph

  • Branch list beside the graph: each agent's branch with its agent, state (waiting, CI failing, will conflict, stopped, resolving), ahead/behind and PR; whose commit is whose.
  • A stopped rebase or merge: a conflict banner, Ask X to resolve (the branch's agent, or a new agent in the worktree); the agent shown resolving while it works (branch.resolving.v1).
  • Hand to an Agent… from a commit: a task based on it.
  • A resolved rebase read back from the branch's reflog (branch.rebased.v1, range-diff old → new), reviewed, then pushed with a lease or undone.
  • CI panel: GitHub Actions history and statistics over 7/14/30 days, and re-run failed jobs (ci.stats, ci.rerun).
  • Autopilot activity beside the graph (autopilot.activity, kept in autopilot-activity.jsonl).
  • Git 3 readiness: reftable repositories' ref changes noticed, a moved repository made with its bundle's object format, project.log saying the object and ref format.

Merged with master

  • A checkout mid-rebase shows under the branch being rebased (#230) #245 already names the branch being rebased as a worktree's branch and adds operation; this branch's rebasing is kept and set alongside. The graph's mid-rebase check no longer needs branch to be nil (a test for the newer server form fails without that).
  • E2E rows renumbered 9.184–9.188 (master took 9.175–9.183).

Tested: make gate, swift test. Seen live in a scratch app earlier (conflict → ask agent, agent resolving, rebase review and push with lease, autopilot feed, CI panel at 14d); the e2e rows cover real agents and GitHub.

🤖 Generated with Claude Code

Amitgb14 and others added 30 commits November 12, 2026 23:22
A Compared page in the docs: one table across the four tools (shape,
platforms, persistence, agents, state, sessions, git and PRs, the
autopilot, MCP and scoping, remote machines, sandboxes, phone), then
when another tool is the better choice and when conch is. Dated, from
each project's own docs, since the others ship weekly. Linked from the
introduction's "Is conch for you?".

Tests: every sidebar entry has a page.mdx, and the compare page lists
all four tools, says when it was written, and has no row whose cells
would shift under the wrong column.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The comparison page named herdr, T3 Code and Orca row by row, which
dates quickly and reads as a claim about other projects. In its place,
Why conch: what conch does that other agent orchestrators leave to you
— the server-side autopilot, reloads that keep agents running, OSC 7501,
scoped agent-to-agent work, sandboxes as machines, cross-agent
sessions, a phone without a third party's server, the terminal first,
and the brain — each linking to the page that explains it.

Test: the sidebar and introduction point at /docs/why and not at the
removed /docs/compare, and every point on the page links to its docs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… task's flow, agent states

Built from the site's own type and colours rather than pictures, so they
stay sharp, follow the light and dark themes, and need no retaking when
the TUI changes. TuiShot puts a captured TUI screen in a docs page, and
the landing page's architecture drawing is now available to pages too.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…tures

The introduction opens with the TUI on a project and a What's new grid
(Conch.app's workbench, a branch's changes, browser tabs, files and Edit
with Agent, the server's autopilot, phones per device), draws how conch
fits together and the agent states, lists Devin with the other supported
agents, and ends on cards rather than a list. Why conch gets a summary
grid, the autopilot's loop, a section on the desktop app with its
screenshots, the phone, and the architecture; Tensorlake is left out of
it, since making a Tensorlake sandbox is hidden. Tasks draws a task's way
from prompt to merge, and Agents the states and how they are known.

Tests: every component a docs page uses is registered for MDX, every
screenshot and TUI screen it names exists, and these pages carry their
pictures.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The guides had grown to 88k words (the interface page alone 16k, the Mac
app 14k) and went unread; people look at the pictures and a line. Each
guide page is now a picture, one heading per feature, a sentence or two
on what it does and how to use it — key, command or menu — and a link
to the reference for the rest. The reference pages (keys, command line,
configuration, protocol) stay whole, since they are looked things up in.
Headings other pages link to were kept; three links that pointed at
headings no longer there (interface#workspaces, sessions#sharing-with-
another-agent, agents#two-tiers) now land. Tensorlake is one line, as
making one is hidden; Conch Cloud is not promoted.

Tests: each guide stays under 700 words of prose with no paragraph over
60, and every link to a docs page — from a page, the sidebar or the
site's components — reaches a page and heading that are there.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Conch Cloud and Cloud machine image pages stay but are no longer
linked from the sidebar, and open with how to turn Cloud on. The CLI
page's login, logout, account and cloud sections move to a Conch Cloud
section at its end under the same note; the interface and Mac app pages
say it is off and what a cloud machine somebody has keeps.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Conch Cloud is not enabled in 0.2.0. The test says the sidebar leaves
its two pages out, both pages say it is not enabled, and nothing else on
the site names it but the Command line page's own section about its
commands — so a guide page rewritten later cannot bring it back.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
# Conflicts:
#	web/src/app/docs/agents-together/page.mdx
#	web/src/app/docs/agents/page.mdx
#	web/src/app/docs/cli/page.mdx
#	web/src/app/docs/interface/page.mdx
#	web/src/app/docs/mac-app/page.mdx
#	web/src/app/docs/phone/page.mdx
#	web/src/app/docs/sandboxes/page.mdx
#	web/src/app/docs/sessions/page.mdx
#	web/src/app/docs/tasks/page.mdx
Git 3 makes new repositories with reftable refs, where branches and HEAD
live in reftable/ and refs/heads never moves. MetaStamp read only the
files backend, so the project ticker skipped every refresh there and the
branch list went stale. It now hashes reftable/tables.list for the
repository and each linked worktree.

The new git3_test.go runs conch's reading of history, branches,
worktrees and changes against a SHA-256 reftable repository.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
InitFromBundle ran a bare git init, which makes git's default hash: SHA-1
today, SHA-256 under git 3. A bundle from a repository of the other kind
was then refused (pack is corrupted / early EOF), so moving a project
between machines with different gits failed. It now reads the bundle's
@object-format header and inits to match.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The commit graph is to show whether a project is a git 3 repository
(SHA-256 objects, reftable refs) or an older one, since conch meets
both side by side. gitx.Format reads it with rev-parse; a git older
than 2.45 cannot name its ref backend and cannot open reftable, so its
refs are files. The field is optional: older servers leave it out.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The commit graph is to show which agent made each commit and what each
agent's branch needs from the person (option A of the git view mocks).
commitOwners gives a branch the commits its first parents reach before
the base, stopping at a branch stacked beneath it; graphBranches lists
the worktree branches with their agent and the most urgent state —
stopped on a conflict, agent waiting, CI red, a conflict coming (merge-
tree), working, done, idle. resolvePrompt is what an agent is asked to
fix a conflict; ciMarks are a pull request's checks for its commit; and
LogResult reads the repository format the server now sends.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Option A of the git view mocks, built on the graph the app already had.
Beside the graph, every agent's branch with its agent, what it needs
(stopped on a conflict, waiting, CI red, a conflict coming, working)
and how far it is from the base; picking one dims the rest. Each commit
says which agent made it, a branch's newest commit carries its state,
and a pull request's checks show as marks on the commit they ran on.

When a branch stops on a conflict, a banner above the graph asks its
agent to resolve it, or starts another agent in its worktree with the
same request; the commit panel offers the same for a conflict merge-tree
sees coming. A commit can be handed to an agent: the task sheet opens
with a new branch starting there (task.create's base). The header says
when a repository is stored the git 3 way, SHA-256 with reftable.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…format

The Mac app page says what the graph now shows and does (agents on
branches, Ask … to resolve, Hand to an Agent…, the git 3 format), the
protocol page names project.log's new format field, and end-to-end row
9.175 covers what fakes cannot: real agents resolving, real CI.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A rebase that stops on a conflict detaches the worktree's HEAD, so the
worktree named no branch and the commit graph could not show the case
that needs the person most: whose conflict it is. WorktreeInfo now
carries Rebasing (gitx.RebasingBranch, read only for a detached
checkout); the app takes it for the branch's name and finds that
branch's agent by the folder it works in, since its pane's branch is
blank too while the rebase is stopped.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A worktree stopped mid-rebase had no Conflicts, since merge-tree is only
tried for a branch git names, so asking an agent to resolve it said 'the
conflicting files' instead of which. A checkout with unmerged files now
lists them as its Conflicts (gitx.ConflictFiles), which is also what the
TUI's queue says of it: conflicts with the base in those files.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Run live against a scratch server, the graph tab with the branch list
and a commit's panel open was wider than a 1320-point window: the tree
and the panel were cut off, the header's title crushed to 'C…' and a
selected row's chips pushed its hash and subject out. The panels now
give way (the branch list can be hidden from the header), a row's chips
and subject share what is left and are clipped, and the compare hint
shows only while no commit is picked.

Commit rows and the branch list answered only a mouse tap, so VoiceOver
could not select a commit; both now have an accessibility action. A
server test pins that a branch can start at a commit, as Hand to an
Agent asks.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The commit graph's new section took the page to 880 words of prose,
past the limit TestDocsGuidesAreShort holds every guide to. It now says
what the graph shows in one line; the buttons explain themselves.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
For the commit graph's CI panel: Runs lists a repository's workflow
runs and RunJobs a run's jobs through gh; ComputeStats works out a
period of whole days — pass rate and the period before, median run and
queue time, how long a branch stayed red, runs per day, each job's
median and slowest tenth, flaky jobs (failed and passed on one commit,
or a run that passed on a later attempt) and the failures that can be
run again. Rerun asks gh to run a run's failed jobs again; FailedRunID
finds a branch's newest failed run, which FailedLog now shares.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ci.stats reads a project's GitHub Actions runs through gh and adds them
up with ghx.ComputeStats over the days asked (14 by default, 90 at most)
and as many before. Each finished run's jobs are read once and kept on
the project, a few dozen per call, four at a time; jobs_pending tells
the client to ask again. ci.rerun runs a run's failed jobs again, by
run or by a branch's newest failed run, then asks for the pull request's
checks sooner. Re-running spends the project's Actions minutes, so an
agent may only do it in its own project (scope). Each has a capability.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CIStats decodes ci.stats (optional fields, so an answer without the
lists still reads); passRate and its change in whole points, the median
run's change, durations as the panel says them, and the runs-per-day
chart's scale: its top is the busiest day rounded up to a number worth
labelling (11 runs to 12, not 20), so the tallest bar fills the chart.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CI beside the graph's search opens the project's GitHub Actions over 7,
14 or 30 days: pass rate and median run against the period before, how
long a branch stays red, the wait in the queue, runs per day, each job's
median and slowest tenth, flaky jobs, and the failures that can be run
again. While the server is still reading a busy repository's jobs the
panel says so and asks again. The branch list steps aside while it is
open, since graph, list and panel do not fit together.

Re-run failed jobs, in the panel and on a red branch's menus, asks
ci.rerun; the button says Re-running… until a later attempt shows up,
so a run that fails again can be re-run again. Tried live against a
fake gh serving a month of runs, at 14 days: the numbers, the chart and
the re-run (gh run rerun ID --failed) as expected. The 30-day chart and
the panel with the branch list stepped aside were not seen live.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The autopilot announced each thing it did (autopilot.note) to whoever
was connected and kept nothing, so what it did overnight was in the
server's log and nowhere a person looks. Every note now carries its
kind (ci, check, ship, merge, resolve, intake, triage) and its time,
and is kept: appended a line each to autopilot-activity.jsonl so a
restart or a reload keeps them, the newest 500, the file rewritten
through a rename once it doubles. A cut-short line loses itself and
nothing else. autopilot.activity lists them, newest first, for one
project or all (capability autopilot.activity.v1).

Intake taking an issue and triage reading one and its verdict only
went to the log; they are notes now too.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
AutopilotNote decodes autopilot.activity (an older server's note,
without kind or time, still reads) and says how the feed marks it:
merged, failed, handed to an agent, or said. activityDays groups the
notes by the person's day — Today, Yesterday, the date, then those an
older server sent without a time under Earlier. AutopilotPolicy and
AutopilotState read autopilot.get, and autopilotRules says the policy
rule by rule for one project: its own check, intake and triage labels,
not another project's.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Autopilot in the graph's header, a dot saying whether it is on, opens a
panel in the CI panel's place: the rules it follows in this project,
the branches it is minding and what each waits on, and what it did, a
day at a time — each entry marked by kind and how it ended, its branch
a link that brings that branch forward in the graph. It reads the
server's kept history (autopilot.activity), so what happened while the
app was closed is there, and reads it again whenever the server
announces a note. An older server's lack of history is said.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
LastRebase finds a rebase that is the newest thing done to a branch (its
tip before and after, and what it went onto); Rewritten pairs the old
commits with the new through git range-diff, so the one whose conflict
was resolved shows as changed; RewriteDiff is that diff of diffs; and
UndoRebase puts the branch back while it has not moved on.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…undone

An agent asked to resolve a branch's conflict leaves it rebased and not
pushed, and the remote still has the old commits. branch.rebased lists
such branches with each commit's fate and, asked for one, git range-diff's
account of what the resolution changed; a push with over replaces the
remote's branch only while it is still the commit rewritten, and
branch.rebase_undo puts the branch back (scoped like a push).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
branch.rebased's answer, what pushing each rewrite means (over the old
branch, a first push, or somebody pushed since), a rebased state in the
graph's branch list right after a stopped conflict, and git range-diff's
lines sorted into what the resolution added and took away.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…o it

Once an agent has resolved a branch's conflict, the commit graph says
so in a green banner and on the branch's card: each commit old -> new,
the one the conflict was resolved in marked. Review the resolution shows
git range-diff's account, the resolution's own lines marked; Push (with
lease) replaces only the old branch the remote still has, and is held
back when somebody pushed since; Undo the rebase puts it back after
asking. End-to-end row 9.178.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Amitgb14 and others added 8 commits November 14, 2026 22:35
The graph's header ran off its right edge with a commit's panel open, so
the repository format pill now waits for the room; the review sheet's
diff sat in the middle of the sheet rather than at its top; undo said a
conflict was back after a rebase that had none; and a branch brought into
view (after an undo, or opening the graph at a branch) was picked without
its panel reading it, which kept showing the commit before.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With either panel open the graph is narrow and the hint was cut to
"shi…"; it is left out then, as it already was with a commit picked.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…while it is there

An agent asked to resolve a conflict works for a while, and all git
shows meanwhile is a branch stuck mid-rebase, so a client could only
say it was stuck. The person's ask (branch.resolving) and the
autopilot's are kept on the branch's worktree (WorktreeInfo.Resolve)
until a fresh read of git shows the conflict gone, or the branch has
not moved for two minutes after it. Kept in memory: a reload forgets it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The conflict banner stayed red and said the branch was stopped for
as long as the agent asked to resolve it worked. Now it says the agent
is at it (blue, files left, who asked and when, Show), waiting on the
person (Answer), or stopped with the conflict still there, or its
pane ended (red, Ask again). The branch list, the graph's tag and the
commit's card say the same. Asking records it with branch.resolving;
from an older server an agent working in a worktree stopped
mid-rebase counts as resolving it.

Also: starting a new agent for a branch stopped mid-rebase said it had
no worktree, since git names no branch there; branchWorktree finds it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
rev-parse prints an option it does not know back as an argument and
exits 0, so with Apple's git 2.39 the graph's format pill read
"SHA-1 · --show-ref-format". An answer starting with - is no answer:
the refs are files, as for any git that cannot say.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…nt mid-way

- The banner turned red when an agent just asked had not started in
  time, but only once something else redrew it: it now looks again
  every few seconds while an agent has the conflict.
- "1 of 1 files left · you asked now" reads "1 file left · you asked
  just now"; "stopped resolving" was cut off in the branch list.
- While one agent is resolving, Another agent starts a different one
  in the worktree with the same request (the banner and the card).
- The ended headline fits on one line.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
#245 names the branch being rebased as a worktree's branch and adds
operation; this branch's rebasing field is kept and set alongside. The
graph's mid-rebase test no longer needs branch to be nil, which a newer
server never leaves it. E2E rows 9.175-9.179 become 9.184-9.188, since
master took those numbers.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@Amitgb14
Amitgb14 merged commit 45b55d9 into master Nov 16, 2026
6 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant