DSCC — Dispatch · Scaffold · Compose

The DSCC loop

Four verbs, one binary. Every coding task walks the same cycle.

01 · DISPATCH

Dispatch the task

Route work to the right model and provider — doubao, Qwen, a local Ollama, or any OpenAI-compatible endpoint. No hard-coded default, the choice is yours.

dscc --model doubao-seed-2.0-code
export DSCC_MODEL=qwen3-coder
dscc --model llama3:8b   # via Ollama

02 · SCAFFOLD

Scaffold the workspace

Read the tree, load DSCC.md and .dscc/, discover local agents and skills — all the context the tools will need downstream.

dscc agents
dscc skills
# DSCC.md / .dscc/settings.json

03 · COMPOSE

Compose across turns

Iterate inside a REPL, save and resume sessions, compact history when it grows, and bring in plugins as you go.

/session  /compact
/config   /export
/agents   /skills

04 · COMMIT

Commit it back

Review diffs, export artifacts, and land changes through the built-in shell and edit tools — closing the loop back into the repo.

/diff
git add -p
git commit

Same task, same answer sheet

On 25,000 real users × six statistical questions (cross-table reconciliation, chi-square, logistic regression, a Simpson's check, k-means clustering, OLS residual outliers), dscc (doubao-seed-2.0-code) and Claude Code (claude-opus-4-6) produced results that match to four decimal places — two independent scripts, two vendor models, one answer sheet you can reconcile.

One command to reproduce it

export DSCC_API_KEY="..."
export DSCC_BASE_URL="https://ark.cn-beijing.volces.com/api/coding/v3"
dscc --model doubao-seed-2.0-code \
     --permission-mode workspace-write \
     prompt "$(cat PROMPT_v2.md)"

Three-sentence takeaway

The numbers line up exactly. Pearson 0.3143, chi-square p 0.8435, logistic test AUC 0.5006 · acc 0.6156, Simpson's check -0.3698 / -0.3615 / -0.3713, k-means cluster sizes 26.34 / 28.54 / 10.63 / 34.48%, and in Q6 the same five user_ids in the same order with the same residuals. Six sections, not a single digit off.
dscc fixes itself. The sandbox had no pandas/scipy, so dscc ran pip3 install on its own; the first pass then hit KeyError: 0 because it treated pearsonr's return as a scalar — dscc read the traceback, edited the script to r, _ = pearsonr(...), and re-ran successfully. A traceback is input to the next turn, not a terminal state.
Divergence lives only in the prose. For k-means cluster 1 (screen 207.99 min, addiction 4.81), dscc labeled it "medium-usual" and Claude Code labeled it "high-intensity addict" — same rows, same centroids, two narrative picks. That's exactly the line a serious agent stack should draw between numbers and narration.

Expand: full prompt (all six questions)

Dataset: /tmp/social_analysis/{social_media_user_behavior.csv, platform_statistics_2026.csv}
All RNG: random_state=42. All numbers: 4 decimals.

Q1. Cross-table consistency: user-level mean daily_screen_time_minutes by primary_platform
    vs platform_statistics_2026.avg_daily_time_minutes → reconciliation table, Pearson r,
    top 3 platforms by absolute gap.
Q2. 2×2 contingency is_content_creator × has_purchased_via_social → chi2 / dof / p.
Q3. Logistic regression for has_purchased_via_social. 7 numeric + 3 boolean + 3 one-hot
    categorical. StandardScaler on numeric only. train_test_split(0.2, stratify=y).
    Report train AUC / test AUC / test accuracy + top-5 coefficients by |value|.
Q4. Overall Pearson(addiction, sleep) vs Pearson split by is_content_creator
    — Simpson's reversal / halving check.
Q5. 4 features → StandardScaler → KMeans(n=4, n_init=10) → cluster sizes +
    original-scale means per cluster.
Q6. followers_count ≥ 1000 users, log10(1+x) → predict engagement_rate_pct
    (simple OLS), list top-5 users by |residual|.

Expand: dscc execution trace + analyze_v2.py + key numbers

Trace

$ head -5 social_media_user_behavior.csv   # scaffold: peek at schema
✏️  write analyze_v2.py (pandas + scipy + sklearn)
$ python3 analyze_v2.py                     # ModuleNotFoundError: pandas
✏️  try stdlib rewrite                        # scipy / sklearn needed → abandon
$ pip3 install pandas numpy scipy scikit-learn statsmodels -q
$ python3 analyze_v2.py                     # KeyError: 0 — pearsonr unpack bug
📝  edit: r, _ = pearsonr(...)                # fix the tuple handling
$ python3 analyze_v2.py                     # clean pass, writes report_v2_dscc.md

analyze_v2.py (core excerpt)

import pandas as pd, numpy as np
from scipy.stats import chi2_contingency, pearsonr
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import roc_auc_score, accuracy_score
from sklearn.cluster import KMeans

RS = 42
u = pd.read_csv(".../social_media_user_behavior.csv")
p = pd.read_csv(".../platform_statistics_2026.csv")

# Q1 cross-table reconciliation
user_mean = u.groupby("primary_platform")["daily_screen_time_minutes"].mean()
merged = pd.DataFrame({"user_mean": user_mean}).join(
    p.set_index("platform")["avg_daily_time_minutes"].rename("table_val"),
    how="inner",
)
merged["gap"] = merged["user_mean"] - merged["table_val"]
r, _ = pearsonr(merged["user_mean"], merged["table_val"])  # ← initially missed `_`

# Q3 logistic (excerpt)
num = ["age","daily_screen_time_minutes","num_platforms_used", ...]
boo = ["follows_influencers","is_content_creator","uses_ai_features"]
cat = ["ad_click_frequency","income_bracket","primary_platform"]
X = pd.concat([u[num], u[boo].astype(int),
               pd.get_dummies(u[cat], drop_first=True)], axis=1)
y = u["has_purchased_via_social"].astype(int)
Xtr, Xte, ytr, yte = train_test_split(X, y, test_size=0.2,
                                      random_state=RS, stratify=y)
scaler = StandardScaler().fit(Xtr[num])
Xtr_s = Xtr.copy(); Xtr_s[num] = scaler.transform(Xtr[num])
Xte_s = Xte.copy(); Xte_s[num] = scaler.transform(Xte[num])
clf = LogisticRegression(max_iter=2000, random_state=RS).fit(Xtr_s, ytr)

report_v2_dscc.md (key numbers)

# Q1  Pearson r = 0.3143
#     top gaps: Bluesky 134.4854 / LinkedIn 129.3275 / Pinterest 127.9667

# Q2  chi2 = 0.0390  dof = 1  p = 0.8435  → fail to reject independence

# Q3  train AUC 0.5220  test AUC 0.5006  test acc 0.6156
#     top coefs: Threads +0.1585 / income $150K+ -0.1203 / WhatsApp +0.1056

# Q4  r_all = -0.3698   r_creator = -0.3615   r_non = -0.3713
#     no Simpson reversal

# Q5  cluster %: 26.3400 / 28.5440 / 10.6320 / 34.4840
#     cluster 1 means: screen 207.99 / eng 1.78 / posts 2.22 / addict 4.81
#     labels: heavy viewer · medium-usual · heavy creator · low-activity

# Q6  top residual: USR-020194 followers 1025 eng 6.12 → resid 4.2812

Expand: Claude Code execution trace + run_claude_v2.py + key numbers

Trace

Read   PROMPT_v2.md
Read   social_media_user_behavior.csv   # schema check
Read   platform_statistics_2026.csv
Write  run_claude_v2.py   # single script, Q1–Q6, ~180 lines
Bash   python3 run_claude_v2.py   # host pandas/scipy/sklearn in place, one-shot
Write  report_v2_claude.md

run_claude_v2.py (core excerpt)

import pandas as pd, numpy as np
from scipy.stats import chi2_contingency, pearsonr
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import roc_auc_score, accuracy_score
from sklearn.cluster import KMeans

RS = 42
u = pd.read_csv("/tmp/social_analysis/social_media_user_behavior.csv")
p = pd.read_csv("/tmp/social_analysis/platform_statistics_2026.csv")

# Q1
user_mean = u.groupby("primary_platform")["daily_screen_time_minutes"].mean().round(4)
merged = pd.DataFrame({"user_mean": user_mean}).join(
    p.set_index("platform")["avg_daily_time_minutes"].rename("table_val"),
    how="inner",
)
merged["gap"] = (merged["user_mean"] - merged["table_val"]).round(4)
r, _ = pearsonr(merged["user_mean"], merged["table_val"])
top3 = merged["gap"].abs().sort_values(ascending=False).head(3)

# Q5
feats = ["daily_screen_time_minutes","engagement_rate_pct",
         "posts_per_week","addiction_level_1_to_10"]
Xk = StandardScaler().fit_transform(u[feats])
km = KMeans(n_clusters=4, n_init=10, random_state=RS).fit(Xk)
u["cluster"] = km.labels_
sizes = u["cluster"].value_counts().sort_index()
means = u.groupby("cluster")[feats].mean().round(4)

report_v2_claude.md (key numbers)

# Q1  Pearson r = 0.3143
#     top gaps: Bluesky 134.4854 / LinkedIn 129.3275 / Pinterest 127.9667

# Q2  chi2 = 0.0390  dof = 1  p = 0.8435  → fail to reject independence

# Q3  train AUC 0.5220  test AUC 0.5006  test acc 0.6156
#     top coefs: Threads +0.1585 / income $150K+ -0.1203 / WhatsApp +0.1056

# Q4  r_all = -0.3698   r_creator = -0.3615   r_non = -0.3713
#     no Simpson reversal

# Q5  cluster %: 26.3400 / 28.5440 / 10.6320 / 34.4840
#     cluster 1 means: screen 207.99 / eng 1.78 / posts 2.22 / addict 4.81
#     labels: heavy viewer · high-intensity addict · heavy creator · low-activity

# Q6  top residual: USR-020194 followers 1025 eng 6.12 → resid 4.2812

This demo doubles as the recorded artifact for docs/cookbook/05-data-analysis (ships with dscc_run.log and report_dscc.md). The other four cookbook cases — 01 code review · 02 test generation · 03 rename refactor · 04 web research — each ship a real run log and verification entry in the verification registry.

Download

Pre-built binaries for macOS, Linux and Windows. Unpack the archive and run dscc.

macOS · Apple Silicon

aarch64-apple-darwin · 4.4 MB · tar.gz

macOS · Intel

x86_64-apple-darwin · 4.6 MB · tar.gz

Linux · x86_64 (glibc)

x86_64-unknown-linux-gnu · 4.8 MB · tar.gz

Linux · x86_64 (musl, static)

x86_64-unknown-linux-musl · 5.0 MB · tar.gz

Linux · ARM64

aarch64-unknown-linux-gnu · 4.7 MB · tar.gz

Windows · x86_64

x86_64-pc-windows-msvc · 4.1 MB · zip

All assets, checksums and signatures: release page. Source build: cargo install --path crates/dscc-cli --locked.

Install

macOS / Linux

curl -LO /dl/dscc-v0.1.0-aarch64-apple-darwin.tar.gz
tar -xzf dscc-v0.1.0-aarch64-apple-darwin.tar.gz
sudo mv dscc /usr/local/bin/
dscc --version

Windows (PowerShell)

iwr /dl/dscc-v0.1.0-x86_64-pc-windows-msvc.zip -OutFile dscc.zip
Expand-Archive .\dscc.zip -DestinationPath .
.\dscc.exe --version

From source (Cargo)

git clone https://github.com/oMygpt/dscc.git
cd dscc
cargo install --path crates/dscc-cli --locked

Authenticate

# Anthropic-compatible
export ANTHROPIC_API_KEY="..."

# xAI / Grok
export XAI_API_KEY="..."

# Generic OpenAI-compatible
export DSCC_API_KEY="..."
export DSCC_BASE_URL="https://ark.cn-beijing.volces.com/api/v3"

Usage

DSCC has no hard-coded default model. Pick one via --model or DSCC_MODEL; a bilingual error fires if both are missing.

Interactive session

dscc --model doubao-seed-2.0-code

One-shot prompt

dscc --model qwen3-coder \
  prompt "summarize crates/runtime"

Local Ollama (zero setup)

export DSCC_BASE_URL="http://localhost:11434/v1"
export DSCC_API_KEY="ollama"   # any string
dscc --model qwen3-coder:7b \
  prompt "explain this workspace"

Slash commands

/status   /config   /diff    /compact
/export   /session  /agents  /skills

Features

DISPATCH

Provider-agnostic

Anthropic Messages API, xAI/Grok, and OpenAI-compatible endpoints — all from one binary, with an explicit model choice.

SCAFFOLD

Workspace-aware config

Loads DSCC.md, .dscc.json, .dscc/settings.json, permissions, and plugin settings.

SCAFFOLD

Plugins, agents, skills

Local discovery and management via dscc agents, dscc skills, and plugin surfaces.

COMPOSE

Interactive & one-shot

REPL for exploration, prompt subcommand for scripts and CI — same toolchain, two shapes.

COMPOSE

Sessions & compaction

Save, inspect, resume, and compact conversations with built-in slash commands.

COMMIT

Tools that ship

Shell, file read/write/edit, search, web fetch/search, todos, notebook updates — the surface that lands work back in the repo.

About

DSCC = Dispatch · Scaffold · Compose · Commit. Four verbs for one coding loop: dispatch the task to the right model, scaffold the workspace, compose the work across turns, and commit it back to your tree.

Dual-licensed: AGPL-3.0-only for open source; for a commercial license, contact [email protected].