senior-dev: a coding agent built into codeaf takes a whole task, in the folder itself - #1488

Merged

AbirAbbas merged 205 commits into

devfrom

zeropoint95/senior-dev-harness

Sep 26, 2026

Commits on Sep 24, 2026

design: delegates — handing a task to an outside harness (draft)

ZeroPoint95
and
claude
committed
1329d4e
design: delegates — one word per delegate starts a task; harness commands drawn beside it

ZeroPoint95
and
claude
committed
d215975
design: delegates — swe-pro's control plane is optional and it streams spend

ZeroPoint95
and
claude
committed
8a70b38
design: delegates — generated command rows, the manual law moves to load time

ZeroPoint95
and
claude
committed
7ae4151
design: delegates — swe-pro's eager commits are squashed at landing

ZeroPoint95
and
claude
committed
9eaf4ab
design: delegates — rewritten for readability

ZeroPoint95
and
claude
committed
21e6d77
design: delegates — one generic reader over a four-record protocol; text-landing delegates

ZeroPoint95
and
claude
committed
83829f4
design: delegates — swe-pro speaks the protocol whole: top-level spend and a step record

ZeroPoint95
and
claude
committed
dc7ff2f
docs: the delegate protocol, version 1, with swe-pro's conformance table

ZeroPoint95
and
claude
committed
9621d37
delegate: the manifest, the loader, the one protocol reader, the launch, and the run's delegate worker

ZeroPoint95
and
claude
committed
d3b990e
delegate: /<name> <brief> starts a task the program does on its own; /delegate lists them; via on propose_task

ZeroPoint95
and
claude
committed
7a6ad9e
delegate: the chat's manual layers each installed delegate's own page over the packed corpus

ZeroPoint95
and
claude
committed
b2fa854
session: the stop-law ledger names the door that now publishes a run's row

ZeroPoint95
and
claude
committed
9a5c16d
delegate: the door crosses the wire, every delegate door is asked off the loop, and the laws hold

ZeroPoint95
and
claude
committed
fcde8b9
delegate: the delegates folder is made at launch, so nobody has to mkdir it before installing one

ZeroPoint95
and
claude
committed
3df3eb0
delegate: the first delegate is senior-dev, the name swe-pro took in its own repository

ZeroPoint95
and
claude
committed
7909bb6
delegate: a brief reaches the program as written, even when it spells a placeholder

ZeroPoint95
and
claude
committed
4237e01
delegate: the programs are built in, and the manifest road is gone

ZeroPoint95
and
claude
committed
e49bdd1
delegate: the model route and the program record are fixed before the lanes

ZeroPoint95
and
claude
committed
4f647c1
session: a program's task page carries its conversation with codeaf

ZeroPoint95
and
claude
committed
3cce8db
tui3: a program's task page is its conversation with codeaf

ZeroPoint95
and
claude
committed
d16758a
tui3: the manual says the program's record keeps more of each call, not all of it

ZeroPoint95
and
claude
committed
aede274
delegate: the program record carries the run's ceiling, so the page can show it

ZeroPoint95
and
claude
committed
0187ca4
seniordev: senior-dev's engine is copied in as it was, before it is adapted

ZeroPoint95
and
claude
committed
63e7259
seniordev: senior-dev's commits survive a symlinked folder and its own folder stays out of a worktree's git

ZeroPoint95
and
claude
committed
790ee07
seniordev: senior-dev runs on the model API codeaf serves it and reports only through the host

ZeroPoint95
and
claude
committed
4a5b872
seniordev: codeaf carries senior-dev, and the chat's manual knows it

ZeroPoint95
and
claude
committed
0bae907
modelapi: every run of a carried program gets a model API of its own

ZeroPoint95
and
claude
committed
6a8daf4
run: a delegated run serves its program the model API and meters every call

ZeroPoint95
and
claude
committed
66983f5
cli: codeaf <name> runs a carried program from a shell, and --help lists it

ZeroPoint95
and
claude
committed
27b57ab
modelapi: a program's reasoning depth rides codeaf's own ladder, xhigh and max included

ZeroPoint95
and
claude
committed
c2e5bd9
run: the program record carries the run's ceiling

ZeroPoint95
and
claude
committed
fd77921
modelapi: a model's working goes out under the router's name and comes back under its own

ZeroPoint95
and
claude
committed
bf2d98b
cli: senior-dev itself works a task through a shell run's model API, pinned by a test

ZeroPoint95
and
claude
committed
e1ee5bd
delegate: the three lanes meet — help fits eighty columns, and the pages say what the API does

ZeroPoint95
and
claude
committed
8702e3e
cli: senior-dev itself works a task the chat's way, through the run's worker, pinned by a test

ZeroPoint95
and
claude
committed
8abbd11
perf: SIZE-BUDGET rises by what senior-dev weighs, and the bill already owed is named

ZeroPoint95
and
claude
committed
be2a083
delegate: a bare --help is the program's own page

ZeroPoint95
and
claude
committed
81b726c
delegate: the program's page says a bare brief is the same as its default command

ZeroPoint95
and
claude
committed
045a6f0
delegate: a program describes itself to the model, and codeaf says which folder to hand it

ZeroPoint95
and
claude
committed
9c4076a
perf: the prompt caps rise by what the programs paragraph costs, on the owner's call

ZeroPoint95
and
claude
committed
2ad0a12
tui3: every door into a program's task opens its conversation, not a blank room

ZeroPoint95
and
claude
committed
d4ad04c
session: a program's task page is read with the belt switch off

ZeroPoint95
and
claude
committed
2f72e15
tui3: an approval card stands, and the work above it stays folded

ZeroPoint95
and
claude
committed
2e9630c
tui3: a program run's tab shows its conversation, for the run still working

ZeroPoint95
and
claude
committed
301982e
tui3: a room opened on a program's task becomes its page, from every door

ZeroPoint95
and
claude
committed
ca679d1
seniordev: writes stay inside the workspace; reads may leave it

ZeroPoint95
and
claude
committed
990345f
delegate: a plain folder is worked in place, and the program is told so

ZeroPoint95
and
claude
committed
4f40fe7
delegate: work lands as its branch, the brief names the copy, the row says why

ZeroPoint95
and
claude
committed
c45d045
delegate: a program works on the conversation's crew by default

ZeroPoint95
and
claude
committed
4576593
scripts: clean-run opens bin/codeaf on a fresh state root with only settings and keys

ZeroPoint95
and
claude
committed
9f283de
enginehost: a socket path may weigh 103 bytes, not 104

ZeroPoint95
and
claude
committed
cb16b0c
config: the talk lane's borrow row is read, and the model API pins are plumbing

ZeroPoint95
and
claude
committed
14e1c34
run: a run whose own task failed is ended in its store

ZeroPoint95
and
claude
committed
545ff02
delegate: the program record carries the program's own clock

ZeroPoint95
and
claude
committed
d747342
run: the worker stamps the program's own clock on its record and ending line

ZeroPoint95
and
claude
committed
ebf014c
session: a run's row, checkpoint and page read one pair, the hand-off to the program's exit

ZeroPoint95
and
claude
committed
a3a717f
session: the tasks tool reads a senior-dev run, and it and the landing note say how long it took

ZeroPoint95
and
claude
committed
8db3d53
session: a hand-off's run reaches the project's index and its conversation's presence

ZeroPoint95
and
claude
committed
d10ffcb
provider: a queued receipt is owed until it is banked, and gets a fifth ask

ZeroPoint95
and
claude
committed
70a92ac
modelapi: a run's books close with its receipts in, and name who answered

ZeroPoint95
and
claude
committed
2512ba6
run: a program's calls are filed under their conversation and folded whole

ZeroPoint95
and
claude
committed
44e99d3
codeaf senior-dev: a shell run waits for its last price, files it, and says how long

ZeroPoint95
and
claude
committed
ff4848d
provider: a run's answer with no usage block is priced by its receipt

ZeroPoint95
and
claude
committed
8e4e4be
manual: the dollar ceiling cannot hold on a service that reports no prices

ZeroPoint95
and
claude
committed
1cc7452
senior-dev: a step costs what the service charged, not the catalog's guess

ZeroPoint95
and
claude
committed
963e38b
modelapi: the owed-receipt count unlocks from a defer

ZeroPoint95
and
claude
committed
06ebe4c
run: a program's live step goes while its last price is owed

ZeroPoint95
and
claude
committed
9b5dd81
modelapi: the call settings comment names every door it arms

ZeroPoint95
and
claude
committed
bd27116
run: the delegate worker's header says where each call's price goes now

ZeroPoint95
and
claude
committed
8ca394a
plandb: a run can be ended at the instant it was last seen

ZeroPoint95
and
claude
committed
51481d0
session: a run no process holds is a record, never more work

ZeroPoint95
and
claude
committed
af8547e
run: a run the caller cut is left open, whichever way the select fell

ZeroPoint95
and
claude
committed
e5e8d42
session: a program's work lands on codeaf's task branch wherever it left HEAD

ZeroPoint95
and
claude
committed
b820b74
session: a program run that changed nothing leaves no branch behind

ZeroPoint95
and
claude
committed
0ee788f
session: the chat is told a program's work was not merged, where it is, and how to bring it in

ZeroPoint95
and
claude
committed
f565eee
delegate: a brief names the same place in the copy, once

ZeroPoint95
and
claude
committed
905dcb0
session: a program on a plain folder leaves its records in the task, not the folder

ZeroPoint95
and
claude
committed
6a8c853
seniordev: the crew's planning seat is not passed, because nothing would use it

ZeroPoint95
and
claude
committed
5a062fe
plandb: the comments stop saying the next hand-off adopts an open run

ZeroPoint95
and
claude
committed
122987a
seniordev: the frontier flag's help fits eighty columns

ZeroPoint95
and
claude
committed
f83bcf9
session: a branch-only landing is a road of its own

ZeroPoint95
and
claude
committed
eac21fb
session: a moved HEAD's note says only what is true of the landing

ZeroPoint95
and
claude
committed
5544b35
session: pin what a moved HEAD's note claims over nothing to land

ZeroPoint95
and
claude
committed
be29f76
tui3: a run's clock counts from the record's start and its landed span is the record's

ZeroPoint95
and
claude
committed
63070a7
tui3: the task pages the belt switch draws over the conversation obey the tab bar

ZeroPoint95
and
claude
committed
8d5e75f
tui3: a program's task is a room in its conversation's tab

ZeroPoint95
and
claude
committed
c994ffb
session: a run's reading and a question asked of it are in the books, and a program's run buys none

ZeroPoint95
and
claude
committed
af77735
delegate: a program's stage is shown in the word it gives a person

ZeroPoint95
and
claude
committed
3164fee
session: a run codeaf closed under reads ended on its page and its row, with its time stopped

ZeroPoint95
and
claude
committed
4d9f6b1
session: a run row codeaf closed under is kept, not the reason the whole task list is thrown away

ZeroPoint95
and
claude
committed
4e5c372
session: a program row's clock is read in three steps, and a run row's state has a rule of its own

ZeroPoint95
and
claude
committed
21af51b
session, tui3: a finished run's time is rounded to the second wherever it is read

ZeroPoint95
and
claude
committed
06564ad
session: a program's commits on its task branch survive the landing, and a stopped program's work goes to that branch

ZeroPoint95
and
claude
committed
2438f4e
tui3: a run's clocks stop where the run did, and a program room reads its landing

ZeroPoint95
and
claude
committed
a2c450e
seniordev: the landing stage reads `checking its work`

ZeroPoint95
and
claude
committed
ad4ca20
session: a run settled after its process went away says what ended it and when

ZeroPoint95
and
claude
committed
623912f
session, tui3: home counts a conversation's dollars once, and its books reach meta.json as work folds

ZeroPoint95
and
claude
committed
120a73b
modelapi: a sibling model that answered is named, and only a dated build or variant counts as the ask

ZeroPoint95
and
claude
committed
9151a5d
provider: a paid answer cut for its text is priced by its receipt on work that asked for that

ZeroPoint95
and
claude
committed
a155d49
provider: a receipt's 70-second bound is counted from when it was owed, not when a worker reached it

ZeroPoint95
and
claude
committed
aa28935
codeaf, delegate: a shell run's time is the program's, its summary is its last line, and a second ctrl-c leaves

ZeroPoint95
and
claude
committed
c066b0a
delegate: the ignores-TERM test waits for its program's trap before it stops it

ZeroPoint95
and
claude
committed
2dd087a
session, seniordev: a program works in the folder it was named, never the home folder

ZeroPoint95
and
claude
committed
ed62886
session: a program's card names the project it will work in

ZeroPoint95
and
claude
committed
2091b56
session, seniordev, delegate: senior-dev works with the models the person asked for

ZeroPoint95
and
claude
committed
61a349f
session: a model no connected service serves is refused for a program, by name

ZeroPoint95
and
claude
committed
5f57c2d
session: a model spelled with a connected service's prefix is taken as written for a program

ZeroPoint95
and
claude
committed
134d850
delegate, seniordev: senior-dev says which step of its process each action served

ZeroPoint95
and
claude
committed
c167939
delegate, run, session, seniordev: a program's task keeps the actions it took, read in its own words

ZeroPoint95
and
claude
committed
d979f47
tui3, seniordev, manual: senior-dev's page shows the actions it took, step by step, with its raw calls one key away

ZeroPoint95
and
claude
committed
e574ecc
seniordev: the scripted run proves a model's shell command reports its exit code

ZeroPoint95
and
claude
committed
f94355f
run: a program's live step is written when its step changes, not on every action

ZeroPoint95
and
claude
committed
362dd41
session, remote: a program's work says which program has it, from the card to the index

ZeroPoint95
and
claude
committed
b033c9c
tui3, manual: a program's tasks wear its badge wherever a task is named

ZeroPoint95
and
claude
committed
b4d860c
session, seniordev: the chat hands complex coding work to senior-dev by itself, and uses it when asked

ZeroPoint95
and
claude
committed
2249cc7
tui3: a task's row leaves its clock out while its room is open, rather than show the second it was clicked

ZeroPoint95
and
claude
committed
fba1baa
session, run, delegate, codeaf: a program works in the folder it is given, on a branch of its own in a repository

ZeroPoint95
and
claude
committed
3ad6f4c
manual: a program's refusals name both roads, and a moved HEAD reads as one sentence

ZeroPoint95
and
claude
committed
8b6f34a
codeaf, manual: the merged lanes build together, and a landing question still reaches its page

ZeroPoint95
and
claude
committed
d3189ea
tui3, manual: a program's card no longer says its work starts from unsaved edits

ZeroPoint95
and
claude
committed
d451c57
seniordev: a refused submit leaves the steps after it where they were

ZeroPoint95
and
claude
committed
5569e01
tui3, manual: a long run from before the action log says how many calls its page leaves out

ZeroPoint95
and
claude
committed
5002e5e
tui3: a page onto another conversation's task wears that task's badge, never a local task's of the same number

ZeroPoint95
and
claude
committed
a1536bb
manual: a page onto another conversation's task wears that task's own badge

ZeroPoint95
and
claude
committed
33320cb
tui3, manual: home's needs and since-you-left rows and the @ list pay for a program's badge out of the title

ZeroPoint95
and
claude
committed
f90024f
session, manual: a proposal leaving out the program the person named passes only after the model has read the bounce

ZeroPoint95
and
claude
committed
c63b4bf
session, manual: a correction typed into the turn does not unsay the program the person named

ZeroPoint95
and
claude
committed
cecfb5f
session, manual: a commit, undo or revert that names a program stays in the conversation, and only an ask for the program lifts the floor

ZeroPoint95
and
claude
committed
f605e81
session, manual: a program's folder is read again before codeaf says the person's branch is safe

ZeroPoint95
and
claude
committed
a12f607
session, manual: a program's run whose process went away is settled without a commit

ZeroPoint95
and
claude
committed
bfe3df2
session, manual: a program's run is refused a folder inside or around one another run holds

ZeroPoint95
and
claude
committed
6e38731
session, manual: nothing else of codeaf's writes in a folder a program's run holds

ZeroPoint95
and
claude
committed
2dec537
delegate, codeaf, manual: a shell run finishes its folder when its program ends, and a hangup stops it

ZeroPoint95
and
claude
committed
0661692
session, manual: a reopened program's row says where its work went, and the receipt names a repository at home

ZeroPoint95
and
claude
committed
a9cf23b
session: a run's folder is readied in a function of its own, and the hold waits by a deadline

ZeroPoint95
and
claude
committed
c8d3a1f
delegate: a child whose host has gone is ended when its grace has passed

ZeroPoint95
and
claude
committed
76d034d
session: the folder's comments say a run that went away is settled, and where its ending is kept

ZeroPoint95
and
claude
committed
7db7483
tui3, manual: senior-dev's pages read the home tab as dev now spells it, and an earlier-messages question still reaches its page

ZeroPoint95
and
claude
committed
2b538eb

Commits on Sep 25, 2026

changes: senior-dev, a coding agent built into codeaf, takes a whole task in the folder itself

ZeroPoint95
and
claude
committed
148ebf5
session: opening a conversation reads its task graph and never builds one for a program's ending

ZeroPoint95
and
claude
committed
af8517d
cmd/codeaf: the ssh policy test gives the control socket a home short enough to fit

ZeroPoint95
and
claude
committed
2f2dc74
session: a program's run is never carried on, and its row and ContinueRun say so in one sentence

ZeroPoint95
and
claude
committed
b6e3cb0
tui3: a program's task offers no steer in its @ block and no retry on its card

ZeroPoint95
and
claude
committed
7c568b7
session, tui3, manual: a program's run says what it cost on its row, its landed card and the tasks tool

ZeroPoint95
and
claude
committed
eec8237
session, manual: a program's proposal names only finished work in depends_on, and any task may name a program's finished run

ZeroPoint95
and
claude
committed
0b53366
tui3: a senior-dev task opened from another window shows its actions, read-only

ZeroPoint95
and
claude
committed
378a0fd
Merge branch 'claude/sd-guest' into zeropoint95/senior-dev-harness
ZeroPoint95
committed
aa31490
changes: a senior-dev run cannot wait and is never carried on, and its price is on its card and in the tasks tool

ZeroPoint95
and
claude
committed
89784d1
tui3: a window on the engine reads the other conversations' presence, so their rows open

ZeroPoint95
and
claude
committed
a80a0b5
Merge branch 'claude/sd-guest' into zeropoint95/senior-dev-harness
ZeroPoint95
committed
0437179
Merge origin/dev (#1410: the worker harness as default, runs as work) into senior-dev

ZeroPoint95
and
claude
committed
fff9033
session, tui3, manual: the merge's loose ends — the page's trail wears the program's badge, the ratchet holds

ZeroPoint95
and
claude
committed
5f8c691
seniordev: git if it is there — a folder with no git history runs on the snapshot recorder, flag or no flag

ZeroPoint95
and
claude
committed
06e129a
session, run, tui3: senior-dev's ending is codeaf's to act on — checked, sent back at most twice, one branch

ZeroPoint95
and
claude
committed
7f6f6d6
tui3, session, seniordev: senior-dev's page names its task once, keeps its brief behind a dropdown, and opens every step

ZeroPoint95
and
claude
committed
92742cf
seniordev, delegate, tui3: every change to the work wears git's +N,-M on senior-dev's page

ZeroPoint95
and
claude
committed
86a8c50
Merge dev into senior-dev-harness
AbirAbbas
committed
9abc110
Merge dev into zeropoint95/senior-dev-harness

ZeroPoint95
and
claude
committed
45a7671
session: fence the chat's remaining file writers while a program holds a folder

AbirAbbas
and
claude
committed
ecc2793
seniordev, session: the person's folder keeps their edits, their branch and their ignored files

AbirAbbas
and
claude
committed
c31bad3
seniordev: the verifier runs unittest when a Python project has no pytest

AbirAbbas
and
claude
committed
ca28d6b
delegate, seniordev: the model's commands get the chat's scrubbed shell, and a wrong protocol lands nothing

AbirAbbas
and
claude
committed
19c5c1a
seniordev: nothing a run starts outlives it, a models.dev outage refuses nothing, and question leaves the hosted belt

AbirAbbas
and
claude
committed
41684fb
session: propose_task offers via only where a program can run

AbirAbbas
and
claude
committed
b96d331
tui3: /senior-dev on a dirty checkout refuses without promising a copy

AbirAbbas
and
claude
committed
a55f8ab
manual: quote the stale guest footer as the page draws it, and say where via is offered

AbirAbbas
and
claude
committed
339dcff
Merge dev into zeropoint95/senior-dev-harness

ZeroPoint95
and
claude
committed
2efde83
README: DeepSWE benchmark in place of "coming soon"

agentfield-bot
and
claude
committed
d8f8d3d
senior-dev: finite ceilings, a ceiling that holds, a limit the person is told about, and a hand-off cap that lasts

AbirAbbas
and
claude
committed
a952053
README: the benchmark chart points to its numbers

agentfield-bot
and
claude
committed
9bdf25e
README: lead the benchmark line with #1 on DeepSWE

agentfield-bot
and
claude
committed
7c92173
README: fold version pinning and building from source into a collapse

agentfield-bot
and
claude
committed
b1ac028
installer: drop the telemetry notice

agentfield-bot
and
claude
committed
53e9fd9
installer: a failed install marker is silent

agentfield-bot
and
claude
committed
df57499
tui3: keep senior-dev's ending outside the wake fold

AbirAbbas
and
claude
committed
d2415a8
session: weigh the prompt prefix with a declared program, and bring it under its cap

AbirAbbas
and
claude
committed
91f6940
session: a run in a gitignored folder says at its start why it has no branch

AbirAbbas
and
claude
committed
e54fb9b
run: a call refused at the estimated ceiling ends the run on its cost limit

AbirAbbas
and
claude
committed
475a76e
seniordev, delegate: the restore accounts for every change the person made, and the engine reaps its whole process tree

AbirAbbas
and
claude
committed
7fb4a47
manual: the general task pages stop denying senior-dev's own ceiling

AbirAbbas
and
claude
committed
cee995f
changes: write the review round's fixes as what was true and what is true now

AbirAbbas
and
claude
committed
7f9f566
senior-dev: a commit the model makes itself is the run's, and carries its credit

AbirAbbas
and
claude
committed
be9ee9f
tui3: pin a senior-dev landing that arrives after the chat's wake reply

AbirAbbas
and
claude
committed
91e6d17
session, tui3, remote: /budget binds the open conversation before it says so

AbirAbbas
and
claude
committed
44444f7
changes: the live /budget bind and the model's own commits

AbirAbbas
and
claude
committed
df7f7ae
seniordev: name rescued edits in every submitted ending

AbirAbbas
and
claude
committed
d866bd9
seniordev/util: the run facts codeaf reads build on every platform

AbirAbbas
and
claude
committed
19eb19f
session: credit the engine's own checkpoint at the tip, and never amend a pushed one

AbirAbbas
and
claude
committed
9515279
seniordev: a restore that could not run ends unchecked, and says why

AbirAbbas
and
claude
committed
8a937a0
changes: a failed restore, the engine's checkpoint credit, and the Windows build

AbirAbbas
and
claude
committed
b9edfa1
Merge the PR branch's newer head into the review fixes
AbirAbbas
committed
7a690c8
Merge dev into senior-dev-harness: team delegation and Pareto crew beside senior-dev

AbirAbbas
and
claude
committed
93e9801
gitidentity: the run facts codeaf reads live outside the engine

AbirAbbas
and
claude
committed
64564dc
seniordev: no file of the engine reaches a Windows build, again without exception

AbirAbbas
and
claude
committed
1e2b177
delegate, config: register the run marker, and let the key suffix rule name the rest

AbirAbbas
and
claude
committed
58e8b78
changes: where the run facts codeaf reads now live

AbirAbbas
and
claude
committed
898c3d1
seniordev: the recovery test prepares its recorder the way a real run does

AbirAbbas
and
claude
committed
5086e5c
Merge dev into senior-dev-harness
AbirAbbas
committed
22867a9
manual: the stop heading says "if it will not stop" in the asker's words

AbirAbbas
and
claude
committed
3b97085
session: codeaf never rewrites a commit; the credit rides only on a commit it makes

AbirAbbas
and
claude
committed
f4a9362
codeaf: an explicit shell model starts senior-dev without a crew seat

AbirAbbas
and
claude
committed
dbb5e95
seniordev: nothing unchecked is called passed, and the rescue is bounded

AbirAbbas
and
claude
committed
e236d44
Merge dev into senior-dev-harness: the Pareto crew's follow-ups and the task room

AbirAbbas
and
claude
committed
5713b91
manual: the crew's $5 task limit is pinned where the pages say "ordinary"

AbirAbbas
and
claude
committed
49ad1ea