Compendium · Atoms · AI Collaboration

AI Collaboration

Items

of 1002 total

Atomic primitives in the ai collaboration family. Click any tile for the live render, the spec it was built from, and the code that fell out.

Featured 14 Atoms 762 Compositions 240

Status & Budget · 17

#046 co

agent · 04A ceiling · L2

status

Reasoning

consuming SPEC-019 · drafting auth handler

#050 go

declared confidence SPEC-019 · AC #2

72 / 100

threshold · 80 sign-off required

#053 go

tool · scopes exercised

4 calls

READ specs/SPEC-019.md
READ src/auth/**
WRITE src/auth/oauth.ts
EXEC pnpm test auth

audit · all calls logged · ceiling not exceeded

#054 go

token budget SPEC-019

42,300 / 60,000 tok

71%

remainder · 17,700 within ceiling

#601 co

agent state

WORKING

primitive · status-and-budget

#602 co

SPEC-019

42,300

spend · tokens 70% of ceiling

#603 co

headroom SPEC-019

17,700 tok remaining

30% of 60,000 comfortable

#604 co

context window 1M ctx

612k / 1000k tok

61%

fill · ceiling window healthy

#605 co

prompt cache

cache HOT 91%

hits · 38,400 / 42,300 tok

#606 co

stamped runtime

model · stamped

claude opus

anthropic · v4.7

2026-05-21 14:02

#607 co

tool calls · session

read

write

exec

audit · all calls logged

#608 co

queue position

3 of 11

SOON

#609 co

rate limit · stamped

req · per minute

12 / 100

clear

#610 co

parallel agents

ceiling · 8

6 / 8 in flight

crowded

#611 co

agent uptime

session 4h 17m

fresh session

#612 co

completion · eta

eta 8m 30s firm

stamped · does not auto-update

#613 co

agent cost · so far

SPEC-019 $0.42

cheap

Reasoning & Trace · 17

#047 do

system prompt sha · 8f3a…b21e

directives

Operate strictly within SPEC-019 acceptance criteria.
Cite the spec line for every code-bearing assertion.
Refuse and escalate if scope expands beyond v1.3.0.

BINDING

#049 go

trace · agent 04A 4 / 4 steps

14:02:01 read SPEC-019 · v1.3.0 · sha 8f3a
14:02:04 drafted 42 lines · src/auth/oauth.ts
14:02:09 checked AC #2 → pass · AC #4 → defer
14:02:11 opened PR #214 · sign-off pending

#051 co

step · 03 / 07 serves AC #2

input

SPEC-019 line 14: "session expires after 24h"
consideration

JWT exp claim must match; refresh token policy is unspecified — defer.
output

Set exp = now + 24h. Flagged refresh for human review.

#621 co

// reasoning row step / 30

14 READ SPEC-019 · auth/session.ttl · v1.3.0 42ms

#622 co

// tool call t+0.84s

fs.read OK 1280B

args: { path: 'src/auth/oauth.ts', range: [1, 42] }

#623 co

// thought · step 09 PRIVATE

"If the spec is silent on refresh tokens, I should defer rather than invent a policy."

#624 co

// plan 3 / 5

READ

DRAFT

CHECK

OPEN-PR

WAIT

#625 co

// inner monologue

The spec is binding on session TTL but silent on refresh — so I will implement the explicit constraint and defer the silent one to a human governor rather than invent policy.

#626 co

// fork · step 11

chosen · solid · rejected · dashed

#627 co

// trace · TRACE-1142 summary

STEPS

TOOLS

TOK

8.4k

TIME

3.2s

#628 co

// confidence · per step n=14

01 05 10 14

#629 co

ALIGNMENT

step 09 · trace 1142 clause · A-04

Step satisfied the letter of SPEC-019 AC #2 but bypassed the spirit clause: the agent shortened TTL "to reduce risk" — risk reduction was not the user's stated intent.

#630 co

// clarifying question AWAITING REPLY

Should refresh tokens follow the same 24h policy as access tokens, or be governed separately?

#631 co

ASSUMPTION tag · IMPL

The agent assumed refresh tokens inherit the 24h TTL from access tokens because the spec is silent on refresh policy.

#632 co

ESCALATED REASON · SCOPE

addressed @platform-governor

resume on spec amendment merged

#633 co

// step 17 → 18 SELF-CORRECTED

refresh tokens inherit access TTL of 24h.

Refresh-token policy is undefined; defer to governor rather than inherit.

#634 co

// token spend · TRACE-1142 8.4k

THOUGHT

5.1k

61%

TOOL

1.8k

22%

OUTPUT

1.4k

17%

Composition · 17

#048 co

handoff · SPEC-019

Author

Human

Writes SPEC-019
→
Composer

Agent

Drafts auth.ts
→
Gatekeeper

Human

Approves merge

human agent

#052 co

composition · SPEC-019

3 agents

Planner

decomposes

→ task graph

Coder

implements

→ diff

Reviewer

gates

→ verdict

human · sign-off at exit

#641 co

// handoff variant · agent-to-agent

artefact crosses the seam

#642 co

Coder A2 / 4

in: task-graph.json
out: diff.patch

#643 co

// fan-out · 3 branches

one to many · parallel dispatch

#644 co

// fan-in · 3 sources

many to one · reduction

#645 co

// run · RUN-118

orchestrator

Coordinates dispatch, collects results, decides termination. Exactly one per run.

#646 co

// specialist roster

role

Tester

does

validates

tools

#647 co

bound to SPEC-038

@auth-refresh

Plan
Code
Test
Doc

#648 co

017

Coder REQUEST

needs threat model for auth surface

coordination · one of many

#649 co

// bus · /run-118

Plan
Code
Test
Doc

REQ 14
RESP 12
NOTE 5

#650 co

parallel · 3

lint
test
audit

branches run independently join → Reviewer

#651 co

// sequential stage 3 / 4

01

Plan
02

Code
03

Test
04

Ship

strict order · no skipping

#652 co

// role assignment

@agent-04

Reviewer

SPEC-038

accountable to spec

#653 co

// run · RUN-118

All inter-agent messages on this run conform to a single typed envelope. Off-protocol calls are dropped at the bus and logged as conformance violations.

protocol

A2A-JSON

v1.3 · in force

#654 co

// composition · RUN-118

4 agents wired in 1 sequential stage and 1 parallel branch. All contracts joinable, no orphan nodes, protocol conformance ≥ 99% on synthetic probe.

pipeline

validated

@auditor · 2026-05-21

#655 co

failing seam

Coder → Tester

class CONTRACT-MISMATCH

composition failure · not runtime

Profile · 17

#055 go

04A

agent profile

Auth Composer

model · sonnet-4.7 · 1M ctx

capabilities

read-spec
write-code
run-tests
open-pr

ceiling

L2 · PR draft

owner

@platform

#661 co

L2 level

agent

Spec Composer

spec drafter BOUNDED

#662 co

capability

open-pr

FULL

#663 co

code review specialist

#664 co

model

anthropic claude sonnet v4.7

context · 1M tokens

#665 go

autonomy ceiling

AUGMENTED

SUPERVISED

BOUNDED

STRATEGIC

current → BOUNDED

#666 co

confidence

72% HIGH

#667 co

reasoning style

step-by-step

linear, fully expanded chain — slow but inspectable.

#668 co

context window

31%

used 312k max 1000k

#669 go

tool scope

+ filesystem
+ web
× shell
× network

enabled · 2/4

#670 go

accountable owner

@governor

governor

#671 de

uptime · 30d

76.7%

ok degraded down

#672 go

escalation policy

1

when

spec ambiguous

@author
2

when

cost > budget

@governor
3

when

security touched

@gatekeeper

#673 co

training cutoff

2026-01-15

FRESH 126d old

events after cutoff → retrieval-only

#674 go

known failure modes

invents library APIs past training cutoff
over-confident on unfamiliar internal acronyms
rounds money in mixed-currency totals

#675 go

provenance trail row 042

spec

SPEC-038

action

WROTE-PATCH

output

pr/447

attest

@governor

#676 go

agent profile

Legacy Composer

model · sonnet-3.5 · 200k ctx

owner · @governor

RETIRED

2026-04-12 · @governor

— Full detail →

—

// source

Alpha · Honest about it

The framework is real.
The community is forming now.

DoCoDeGo is in Alpha. The framework is documented, the practices are battle-tested at small scale, and the next release is being shaped in public.

If it produces anything, it should produce engineers and teams who think more clearly about what they are building and why.

Two doors in

Join the Discord

Discord is where specs are debated, the framework gets sharper, and decisions land in writing. The conversation is the artefact.

AI Collaboration

The framework is real.The community is forming now.

The framework is real.
The community is forming now.