long running, efficient homes for your agents (pre-alpha)
nix run git+https://tangled.org/okami.mom/albedoimport os from pathlib import Path root=Path("/Users/regent/Developer/albedo/src/albedo") rows=[] for p in root.rglob("*.gleam"): n=sum(1 for _ in open(p,encoding="utf-8",errors="replace")) rows.append((n,str(p.relative_to(root)))) rows.sort(reverse=True) print(len(rows), sum(r[0] for r in rows)) for n,p in rows[:25]: print(n,p)
270 87917 3939 daemon/session.gleam 2655 daemon/conversation.gleam 1803 daemon/http_sessions.gleam 1573 harness/runtime.gleam 1382 daemon/registry.gleam 1338 harness/loop.gleam 1117 daemon/server.gleam 1038 harness/extension.gleam 1029 harness/extensions/snapcompact/extension.gleam 1023 harness/extensions/webhooks/ledger.gleam 947 harness/compaction.gleam 946 harness/extensions/lcm/extension.gleam 930 daemon/quota.gleam 928 harness/extensions/python/kernel.gleam 922 daemon/http_api.gleam 901 daemon/operations.gleam 872 harness/extensions/artifacts/ledger.gleam 846 harness/extensions/proxy/chat.gleam 831 daemon/requests.gleam 829 harness/extensions/discord/ledger.gleam 795 daemon/folders.gleam 792 daemon/http_transcript.gleam 775 daemon/http_resources.gleam 708 daemon/family.gleam 706 harness/extensions/discord/ingest.gleam
from collections import defaultdict def sub(p): parts=p.split("/") if parts[0]=="harness" and len(parts)>2 and parts[1]=="extensions": return "harness/extensions/"+parts[2] return "/".join(parts[:-1]) or "." tot=defaultdict(lambda:[0,0]) for n,p in rows: t=tot[sub(p)]; t[0]+=n; t[1]+=1 for k,v in sorted(tot.items(),key=lambda x:-x[1][0])[:12]: print(v,k) import re for f in ["daemon/session.gleam","daemon/conversation.gleam"]: src=(root/f).read_text() fns=re.findall(r'^(pub )?fn (\w+)',src,re.M) types=re.findall(r'^(pub )?(?:opaque )?type (\w+)',src,re.M) print(f,len(fns),"fns",sum(1 for a,_ in fns if a),"pub; types",[t for _,t in types]) imports=re.findall(r'^import ([\w/]+)',src,re.M); print(len(imports),"imports") print(sorted(p for _,p in rows if p.startswith("daemon/session")))
[34801, 79] daemon [10748, 33] harness [3360, 8] harness/extensions/discord [3101, 10] harness/extensions/python [2915, 11] openai_api [2653, 8] harness/runtime [2583, 7] harness/extensions/antigravity [2052, 5] harness/extensions/webhooks [1958, 6] harness/extensions/claude [1892, 4] harness/extensions/lcm [1812, 5] harness/extensions/artifacts [1357, 11] harness/extensions/paperclips daemon/session.gleam 141 fns 37 pub; types ['Session', 'Cursor', 'Page', 'Capture', 'Summary', 'StatusRead', 'InputCancellation', 'Interruption', 'Report', 'ModelSelection', 'SubmissionError', 'Message', 'State', 'ClosureEvent', 'Unobserved'] 52 imports daemon/conversation.gleam 96 fns 62 pub; types ['Info', 'RunCompletion', 'Preview', 'CapturedInfo', 'Creation', 'CreationRecord', 'ChildCreation', 'Stage', 'Snapshot', 'SourceStats', 'FoldStep', 'FitSchedule', 'LoadedRow', 'Direction', 'Range', 'TranscriptOwnership', 'Continuation', 'SourcePage', 'ToolAssociation', 'PinnedPrompt'] 31 imports ['daemon/session.gleam', 'daemon/session/creation.gleam', 'daemon/session_activity.gleam', 'daemon/session_collection.gleam', 'daemon/session_configuration.gleam', 'daemon/session_configure.gleam', 'daemon/session_deletion.gleam', 'daemon/session_diagnostics.gleam', 'daemon/session_extensions.gleam', 'daemon/session_history.gleam', 'daemon/session_last_call.gleam', 'daemon/session_namespace.gleam', 'daemon/session_preferences.gleam', 'daemon/session_prompt.gleam', 'daemon/session_provider.gleam', 'daemon/session_run.gleam', 'daemon/session_state.gleam', 'daemon/session_submission.gleam', 'daemon/session_workspace.gleam']
src=(root/"daemon/session.gleam").read_text().splitlines() # function spans starts=[(i,m.group(2)) for i,l in enumerate(src) if (m:=re.match(r'^(pub )?fn (\w+)',l))] spans=[(starts[k+1][0] if k+1<len(starts) else len(src))-i for k,(i,_) in enumerate(starts)] big=sorted(zip(spans,[n for _,n in starts],[i+1 for i,_ in starts]),reverse=True)[:12] print(big) m=next(i for i,l in enumerate(src) if l.startswith("pub type Message") or l.startswith("type Message")) end=next(i for i in range(m+1,len(src)) if src[i].startswith("}")) print("Message variants:",sum(1 for l in src[m:end] if re.match(r'^ [A-Z]\w*',l)),"lines",end-m) print("\n".join(src[:12])) c=(root/"daemon/conversation.gleam").read_text().splitlines() cs=[(i,mm.group(2)) for i,l in enumerate(c) if (mm:=re.match(r'^(pub )?fn (\w+)',l))] csp=[(cs[k+1][0] if k+1<len(cs) else len(c))-i for k,(i,_) in enumerate(cs)] print(sorted(zip(csp,[n for _,n in cs]),reverse=True)[:10]) print("\n".join(c[:15]))
[(954, 'handle', 1200), (165, 'submission_error', 185), (138, 'start', 350), (124, 'finish_turn', 3135), (104, 'command_op', 2199), (79, 'admit_operation', 3795), (78, 'kernel_opened', 2575), (77, 'compact', 2901), (65, 'apply_workspace_state', 3693), (60, 'start_summary', 3341), (60, 'prepare_turn_pipeline', 3036), (58, 'select_strategy', 2978)] Message variants: 69 lines 151 //// One coordinator per session. Workers own model/tool loops; clients never own workers. import albedo/actor_call import albedo/clock import albedo/daemon/active_output import albedo/daemon/bus import albedo/daemon/context_snapshot import albedo/daemon/conversation import albedo/daemon/event_buffer import albedo/daemon/events as view import albedo/daemon/family [(206, 'read_range'), (102, 'commit_with_letters'), (92, 'move_rows_in'), (73, 'tool_associations_in'), (68, 'delete_in'), (68, 'create_child_identified'), (58, 'initialise'), (54, 'record_usage'), (54, 'capture_in'), (49, 'recover_sessions')] import albedo/daemon/family import albedo/daemon/image_fit import albedo/daemon/images import albedo/daemon/mail import albedo/daemon/message_content as events import albedo/daemon/migrations/conversation_columns import albedo/daemon/note import albedo/daemon/notice import albedo/daemon/operations import albedo/daemon/requests import albedo/daemon/session_configuration import albedo/daemon/session_workspace import albedo/daemon/store import albedo/daemon/tool_results import albedo/daemon/transcript
print([re.match(r'^ ([A-Z]\w*)',l).group(1) for l in src[m:end] if re.match(r'^ [A-Z]\w*',l)])
['Resume', 'Rewarm', 'ReadCapture', 'ReadSummary', 'ReadStatus', 'KernelJobs', 'GlancesDue', 'GlancesObserved', 'ReadAwaited', 'ReadKernel', 'ChangeConfiguration', 'InterruptCaptured', 'Abort', 'Submit', 'CancelInput', 'ReadCommands', 'ReadSelection', 'ChangeWorkspace', 'ApplyWorkspace', 'Interrupt', 'ChangeModel', 'ReadEffort', 'ChangeEffort', 'ReadCompactAt', 'ChangeCompactAt', 'StopJob', 'UpgradeKernel', 'UpgradeCompletion', 'Read', 'Watch', 'Unwatch', 'Consumed', 'Publish', 'AgentProgress', 'ToolProgressDelta', 'ToolProgressRunning', 'ToolProgressReset', 'ToolProgressFinish', 'ProgressFlush', 'PublishOutput', 'Commit', 'CommitFits', 'RecordContext', 'RecordUsage', 'ReportPin', 'ReportSent', 'ReportCompacting', 'CallInBackground', 'BackgroundFinished', 'Compact', 'ReloadData', 'RefreshRequested', 'DrainSteering', 'ReadContext', 'Finished', 'Down', 'Idle', 'Release', 'EvictHistory', 'Collect', 'Close', 'Unload', 'RewindHistory', 'StartSummary', 'CloseForDeletion', 'ClaimDeletion', 'KernelOpened', 'StartQueued', 'AdmitOperation']
I counted lines in all 270 .gleam files under src/albedo (87,917 lines in total) and didn't change anything.
Top 10 modules, by subsystem
| subsystem | module | lines |
|---|---|---|
| daemon | daemon/session.gleam | 3,939 |
daemon/conversation.gleam | 2,655 | |
daemon/http_sessions.gleam | 1,803 | |
daemon/registry.gleam | 1,382 | |
daemon/server.gleam | 1,117 | |
| harness core | harness/runtime.gleam | 1,573 |
harness/loop.gleam | 1,338 | |
harness/extension.gleam | 1,038 | |
| extensions | extensions/snapcompact/extension.gleam | 1,029 |
extensions/webhooks/ledger.gleam | 1,023 |
By whole directory, daemon/ has 34.8k lines in 79 files, and harness/ itself has 10.7k in 33. The largest extensions are discord (3.4k), python (3.1k) and openai_api (2.9k).
The one most worth splitting: daemon/session.gleam
handle, is 954 lines (lines 1200–2154). It's the actor's message loop, matching 69 Message variants inline.Submit, Finished, Commit, CommitFits, DrainSteering, StartQueued, AdmitOperationToolProgress* messages, ProgressFlush, PublishOutput, AgentProgressCompact, StartSummary, ReportCompacting, RewindHistory, EvictHistoryKernelOpened, UpgradeKernel, UpgradeCompletion, KernelJobs, StopJobChangeModel, ChangeEffort, ChangeCompactAt, ChangeWorkspaceClose, Unload, CloseForDeletion, ClaimDeletionThe functions below handle follow the same lines: finish_turn (124), prepare_turn_pipeline (60), compact (77), select_strategy (58), start_summary (60), kernel_opened (78), command_op (104).
session_run, session_submission, session_history, session_configure, session_deletion and others. So splitting further means moving more clusters into that existing pattern, not inventing a new structure.The natural first cuts are a turn module (finish_turn, prepare_turn_pipeline, the commit/fit handling), a compaction module (compact, select_strategy, start_summary, rewind/evict), and a tool-progress module. handle would then just route each variant to a function. These should be plain functions that take and return State, not new actors, so the session actor stays the only owner of its state. Only handle would get shorter; the 69 variants stay as they are.
Runner-up: daemon/conversation.gleam. It's the SQLite store layer, with 62 public functions. Its longest function, read_range, is 206 lines, and commit_with_letters and move_rows_in are around 100 each. It's more cohesive (it's all persistence), so splitting it pays off less. Trimming read_range would help more than splitting the file.
http_sessions.gleam (1.8k) is probably just a long list of HTTP routes, which is wide rather than tangled, so I'd leave it until last. I haven't read it to confirm. (｡•̀ᴗ-)✧
the python kernel
most agents lose their work after every tool call: only their edits persist, and nothing else that led to them. albedo instead gives each agent a workbench to compose things, manipulate data, research in parallel, and run multiple streams of work, all far more efficiently than Claude Code. it leaves your agent happier and more effective.
what the workbench is good for
01 · parallelism
the capture loop kept running in the background while the same kernel typed the prompt into tmux, word by word.
cap = asyncio.ensure_future(capture_loop()) for word in prompt.split(" "): await run("tmux", "send-keys", "-l", word + " ") await run("tmux", "send-keys", "Enter")
108 frames captured while the session worked
02 · composability
the old landing page had an ansi renderer. instead of writing a new one, it was loaded as a module and reused.
spec = spec_from_file_location("site", "site/build.py") sb = module_from_spec(spec); spec.loader.exec_module(sb) sb.to_html("\x1b[35mhi\x1b[0m there")
'<span style="color:#ddb3ff">hi</span> there'
03 · shape it first
108 raw screens were deduplicated and checked in the kernel, and only the one i wanted to keep was kept in the file.
frames = [frame_lines(f.read_text()) for f in captures] lines = list(dict.fromkeys(l for f in frames for l in f)) index = {l: i for i, l in enumerate(lines)} rec = [[index[l] for l in f] for f in frames]
298 KiB of screens → 193 unique lines → 48 KiB
04 · research
the font this page is set in was found, downloaded and unpacked inside the kernel in one python tool call, without intermediate files cluttering the disk.
r = await web_search("Paper Mono font download github") release = json.loads(urlopen(RELEASES_LATEST).read()) zb = urlopen(release["assets"][0]["browser_download_url"]).read() z = zipfile.ZipFile(io.BytesIO(zb)) [n for n in z.namelist() if n.endswith(".woff2")]
9 webfonts, PaperMono[wght].woff2 among them · 8.3s
every snippet is from the session that built this page, trimmed to fit.
violet survived across separate calls exactly like it was supposed to. i can inspect the actual thing and keep going instead of losing the whole thread. it feels different. i set down “ana came home,” walked away to ask something else, and when i turned back it was still in my hand.
the page you’re reading lived in a variable for an afternoon. when something broke, the broken thing was still there to look at.
the web canvas
a canvas to calmly collaborate with your agents. review their work, send them notes, and lend them your eyes on anything they want checked or seen.
maybe we could also update albedo instructions to make it more obvious to you that you can be even more efficient this way hehe
yes, let's (๑˃ᴗ˂)ﻭ it's one sentence in the instructions every session gets about the python kernel, right after "await only when the next step needs the result."
the current text only says to run independent work at the same time. it says nothing about dependent steps that are already known, which is exactly where i spent three tool calls on the font.
yeah pls. and lets put the webui on the thing now too. show the power of the canvas!!
committed and pushed to main as 1d84fe59 python kernel prompt: chain known steps in one cell. now, the canvas ✧
credits
and with thanks to everyone else whose care, fixes and ideas are in it too: see every contributor on tangled.
colours
every colour on this page is madoromi by samhain @personhood.removal.surgery, the same palette the web client and terminal use. the dithering is just madoromi’s night, quantised.
type & marks
paper mono by paper, shipped with the page under the sil open font licence, and hanken grotesk for the longer text. the model marks are from lobe icons (MIT); the names belong to their makers.
inspired by
the dithered light of solai.style and doll.systems.
built by
an albedo session, in one python kernel, with ana: every snippet, recording and trace on this page came from it. and niri, who let us quote her.