ddflow

A portable, agent-agnostic work-queue kernel for AI coding agents.

You keep a queue of phases and tasks with declared dependencies. You say "implement phase P2". Independent tasks fan out to parallel agents in isolated git worktrees; dependent ones wait. Every task passes a quality pipeline whose gates cannot be passed by assertion. If an agent crashes, its work is found rather than lost. If everything except the log is destroyed, the project's decision history rebuilds from the log alone.

One dependency beyond python3 and git (Jinja2, for the prompt templates; see Extending it by writing text, not code). Works with Claude Code, Gemini CLI, Codex, Copilot, Cursor, Kimi, opencode, Aider, a CI job, a Makefile, or a human at a terminal — over a CLI and an MCP server that are the same implementation.

Introduction

The problem it solves

An AI coding agent is good at a task and weak at a project. One agent in one session mostly works. Run it for weeks, or run three at once, and the same failures come back:

• Work disappears. A session crashes or is closed mid-task, and the half-finished change sits in a directory nobody remembers.
• Agents collide. Two of them edit the same file, and the second merge quietly undoes the first.
• Checks that never ran look like checks that passed. The linter was missing, the reviewer endpoint was down, the tests were "run" in a summary. The agent reports done, and nothing on record says otherwise.
• The project forgets. Last week's hard-won lesson, the reason behind a design choice, the bug that was already fixed once — gone at the next session, or at the next context compaction in this one.
• Nobody can say what happened. Which instruction led to which change, and which review looked at it, lives in a chat transcript that no longer exists.

ddflow is the layer between you and your agents that makes those failures structurally hard rather than a matter of discipline. It does not write code and it is not an agent. It is a queue, a set of rules the tools enforce, and a log of everything that happened.

What changes for you

Who it is for

• One developer with one agent. A plan that survives the session, a memory that survives compaction, and a record of which checks really ran. The queue is useful even with no parallelism at all.
• Several agents in parallel — subagents, several terminals, several vendors. The dependency graph says which tasks are independent, worktrees keep them apart, and the merge step lands them without anyone switching the main checkout's branch.
• A team or a CI pipeline. The log is committed with the code, so a fresh clone knows the queue and its history. Read commands have a machine-readable --json form and every command returns the same four exit codes, so a Makefile or a CI job can drive it exactly as an agent does.

It works with the agent you already use, because everything it does is reachable both ways: as a shell command and as an MCP tool. Use whichever your agent, script or CI job has. Your workflow is text, not code — the gate pipeline, the reviewer instructions and the agent-facing prompts are files in your repository that you can edit.

What it is not

• Not an agent or a model. Your agent does the work; ddflow decides what may start, checks what was claimed, and remembers.
• Not a hosted service. Everything is files in your repository and a disposable local cache. No account, no server to run beyond the local MCP process.
• Not a replacement for your tests or CI. It runs the commands you configure and records their real exit codes and output.

A first run

# ddflow-mcp is not on PyPI yet; until the first release, install from the repository:
uv tool install git+https://github.com/delian/ddflow-mcp   # or: pipx install git+https://github.com/delian/ddflow-mcp
cd /path/to/your/project
ddflow adopt --launch python        # registers the MCP server with your agents, writes .ddflow/ and a block in AGENTS.md
ddflow phase add P1 --title "Password reset"
ddflow task add P1.T1 --phase P1 --title "Reset-token endpoint" --globs 'src/auth/**'
ddflow next                          # what can start now, and why the rest is blocked

--launch python writes the full path of the interpreter inside the environment you just installed into (plus its PYTHONPATH), not a bare python. Without it, adopt registers uvx ddflow-mcp, which fetches the package from PyPI and so cannot start until the first release is published.

Then tell your agent "implement phase P1". The driver adopt installed tells it to start with ddflow_brief, claim the task, work in its own worktree, satisfy each gate and land the change. ddflow cannot make an agent follow instructions, but it makes skipping them visible: the commit hook adopt installs flags a commit made without a lease (or refuses it, if you set [enforce].commit_without_lease = "block"), and ddflow complete refuses an item whose gates carry no outcome. Watch it with ddflow board, and ask ddflow doctor at any point whether the project is healthy.

How do I…?

Every row is a command you can run in a terminal and a tool an agent can call over MCP — the same implementation, so neither drifts from the other.

Every read command takes --json. Every exit code means the same thing everywhere: 0 healthy · 1 real failure · 2 could not run / nothing to do · 3 coordination refused. 2 is never collapsed into 0 — "nothing is ready" and "everything is fine" are different facts, and an agent that cannot tell them apart invents work.

Table of contents

• Introduction
• How do I…?
• Help: what it can do, and the workflow
• Wiring it into your agent
• From plan mode to the queue
• When a companion is missing
• What is automated, and what is not
• Two ways to drive it
  • Standalone: a terminal, a Makefile, CI
  • As an MCP server
  • What goes in AGENTS.md / CLAUDE.md
• The workflow, and changing it
• The rules, and where each came from
• Why it is built this way
• Install into any project
  • Docker — for operators with no Python toolchain
  • Extending it by writing text, not code
  • Publishing and registry
  • Cutting a release
  • What CI checks
  • Any LLM as a reviewer — local, remote, SaaS, or a CLI
  • Companion tools
• Adopting a project that already has history
  • Verifying an import, at any time
• The model: phases, tasks, dependencies, globs
• Work that changes shape while you do it
• Architectural decisions
• Recall — "have we been here before?"
• Status, progress, and loops
• The task pipeline
  • Proving a gate can fail at all
• The phase pipeline
• Human approval: a gate the agent cannot clear
• Parallelism and coordination
• Gitflow, pull requests and version tags
  • Several release lines: fixes to older majors
  • Environment branches: promoting downstream
  • Workflow choices: asked, recorded, defaulted on the record
• Many agents, one server: identity, state and sharing
  • Is it stateless?
  • Who is calling?
  • Can one server serve several projects?
  • Locking, contention and measured cost
• Crash recovery
• Reconstruction from logs alone
• Lessons that check themselves
• Checking that the checks are working
• Reading the log, and why it is never compacted
• Lessons, research and bugs
• Cadences
• Keeping session-start cost flat
• Agent portability
• Keeping the two surfaces honest
• Command reference
• Configuration
• Testing
• Documentation index

Help: what it can do, and the workflow

$ ddflow help                 # what this is, the loop, every capability grouped
$ ddflow help workflow        # workflow · import · gates · parallel · memory · recovery · config

Reachable as ddflow_help over MCP, and that is the point: an agent connecting had 59 tool descriptions and a state-aware handshake, neither of which answers "what is this, and how am I meant to work here". A tool description explains one tool to someone who already picked it; the handshake describes this repository right now.

Two halves, deliberately:

• The narrative is a template under ddflow/templates/prompts/help/, so ddflow prompts eject-style overriding applies — put your own .ddflow/prompts/help/workflow.md in place and the tool teaches your workflow.
• The capability inventory is generated from the live tool table. A hand-kept command list in a second place is the documentation-drift class, and this project has paid for it twice.

Three ratchets keep the prose honest, because a page recommending a flag that was renamed is worse than no page — whoever finds nothing reads the code, and whoever finds a wrong answer trusts it. Every command a page names must exist as a CLI leaf or an MCP tool; every topic the index offers must resolve; and every tool must fall into a group, so a new capability has to be classified rather than quietly dropped from an inventory that claims to be complete.

Two ways to drive it

The CLI is the whole product. The MCP server is a second surface over the same commands, and tests/test_mcp_parity.py fails if the two diverge — every subcommand has a tool, every flag is reachable, and each exemption carries a written reason.

Standalone: a terminal, a Makefile, CI

$ ddflow init
$ ddflow config --set gate.unit_tests.command "python -m pytest -q"
$ ddflow phase add P1 --title "Billing" --globs "src/billing/**"
$ ddflow task add P1.T1 --phase P1 --title "Tax rules" --globs "src/billing/tax.py"

$ ddflow next                          # exit 2 = nothing actionable
$ ddflow claim P1.T1                   # exit 3 = refused, with the reason
leased P1.T1 · worktree .ddflow-worktrees/P1.T1 · branch ddflow/P1.T1

$ cd .ddflow-worktrees/P1.T1 && ...    # do the work
$ ddflow gate status P1.T1             # what the pipeline wants next
$ ddflow gate run P1.T1 unit_tests     # runs it; the exit code IS the evidence
$ ddflow gate record P1.T1 implement --outcome passed --evidence "added tax.py"
$ ddflow complete P1.T1                # exit 3 lists whatever is unsatisfied
$ ddflow merge P1.T1

You get everything except the judgement. Command gates run themselves; agent gates wait for a human to record an outcome, and ddflow gate skip <id> <gate> --reason "..." is the escape hatch — recorded as a skip, never as a pass.

In CI, the exit codes are the interface:

check:
	ddflow doctor        # 1 = integrity problems, each named
	ddflow workflow      # 1 = the pipeline does not hang together
	ddflow cadence       # 2 = no periodic pass is due

2 is never "no problem". A job that treats it as success reports a green build for a suite that never ran.

As an MCP server

ddflow mcp speaks newline-delimited JSON-RPC over stdio. You rarely run it by hand — ddflow adopt writes the launch entry into each agent's own config and leaves existing servers alone:

22 agents are supported. The full table, with what each one gets, is in Wiring it into your agent.

It also copies the driver to docs/ddflow/drivers/, and installs the pre-commit hook that enforces claim-before-you-edit.

What an agent sees the moment it connects, with no call to make:

• Instructions, returned inside the initialize result itself — and state-aware: what is ready, what is in flight, which setup is missing, whether this project has history worth importing, whether an import was left unfinished.
• Tools — one per CLI command.
• Resources — ddflow://board, ddflow://brief, ddflow://lessons, ddflow://research.
• Prompts — which a client turns into slash commands. Tools are things an agent calls; prompts are things you invoke.

Two tools exist so an agent can orient itself without being told: ddflow_help (what is this, what is the loop) and ddflow_workflow (what are the rules here).

What goes in AGENTS.md / CLAUDE.md

ddflow adopt writes it as a managed block between <!-- DDFLOW:BEGIN --> and <!-- DDFLOW:END -->. Your own prose around it is preserved; re-running updates only what is inside. If you write it by hand, four things have to be in it:

1. Start every session with ddflow_brief (or ddflow brief in a shell).
2. Claim before you edit — ddflow_next → ddflow_claim → work in the worktree it creates.
3. The loop — ddflow_gate_status → satisfy each gate → ddflow_complete → ddflow_merge.
4. The exit codes, and that 2 is not success.

Without that block an agent sees the tools and has no reason to reach for them before editing. The block is what makes the queue authoritative rather than optional — and it is 232 words, because an instruction file nobody finishes reading is one nobody follows.

The workflow, and changing it

$ ddflow workflow
# The workflow this project runs

   1. research      agent
   2. rules         agent
   3. implement     agent       (required)
   4. lint          command     (required, NOT proven able to fail)
      $ ruff check .
   ...

## The rules, and where each came from

  gates.require_outcome                  True              [default]
  gates.enforce_order                    block             [file]
  schedule.max_parallel_tasks            4                 [default]

One answer to "what are the rules here": every gate in order, which are commands and which you perform, which are required, which need evidence, which need a different-family reviewer, which have been proven able to fail — plus the completion rules, the caps, the reviewers, and where each value came from, so a deliberate choice is distinguishable from a default nobody touched.

Changing it

$ ddflow workflow gate lint --command "ruff check ." --into task --after implement --required
$ ddflow workflow pipeline task research,implement,lint,unit_tests,merge
$ ddflow workflow drop dedupe

All four reach MCP — ddflow_workflow, ddflow_workflow_pipeline, ddflow_workflow_gate, ddflow_workflow_drop — so an agent can change the workflow with the operator's agreement. Their descriptions say to ask first and offer dry_run, because a pipeline governs every future item, not the one in hand.

Nothing is written until it is checked, and the order is the point: compose the change, validate the result, then replace the file atomically.

• A pipeline naming an undefined gate is refused, naming the near miss. That one is otherwise silent and permanent: the outcome folds to empty, completion refuses it forever, and gate record rejects the id as unknown — so the item can never be completed at all, and nothing says why.
• An unknown section or knob is refused, with a suggestion. [gatez] is valid TOML and used to be written happily, breaking every later command — the write path validated the merged text for syntax and then validated the config already on disk, which is a writer checking the state it is replacing.
• Dropping a gate takes it out of required too, or it becomes a requirement that quietly requires nothing.

ddflow workflow and ddflow doctor both re-run those checks against what is on disk. Everything is a file you can also edit by hand: gates in [gate.<id>], reviewers in [[reviewer]], companions in .ddflow/companions.toml, and every prompt — including the instructions your agent receives at connect — under .ddflow/prompts/.

One caveat with MCP: the connection instructions are computed once, when the server starts. A workflow changed mid-session is live for every tool call immediately, but the text the agent was handed is stale. Tell it to call ddflow_workflow, or restart.

Why it is built this way

The append-only event log is the source of truth; everything else is a projection that can be deleted and re-derived. The SQLite index, the markdown boards, the search index, the recovery bundle — all disposable, all rebuilt by ddflow rebuild.

That single inversion is what makes the four hard properties fall out for free rather than needing to be engineered:

The design decisions, with the probes that settled each, are in docs/RESEARCH.md. The two that most shaped it:

• SQLite-on-NFS is correct here but 36× slower than local (measured, 12 processes × 40 increments). So the log is authoritative and the database is a disposable cache — which also happens to be the choice that stays correct on filesystems where locking is broken.
• An expired lease must never be reclaimed automatically. A crashed agent's worktree is sometimes irreplaceable work and sometimes a superseded draft, and nothing in the metadata distinguishes them. Recovery measures and advises; it never deletes.

Install into any project

One line in your agent's MCP config. Nothing else.

Until the first release is on PyPI, uvx ddflow-mcp has nothing to fetch: install from the repository and run ddflow adopt --launch python, as in A first run.

{ "mcpServers": { "ddflow": { "command": "uvx", "args": ["ddflow-mcp"] } } }

uvx fetches and runs the published package in an ephemeral environment on first use — no clone, no virtualenv, no PYTHONPATH, no install step for an operator to forget, and no vendored copy to drift from upstream. ddflow needs one runtime dependency beyond python3 and git (Jinja2), which is what lets it install inside sandboxes, CI images and other tools' ephemeral containers.

Then, from the agent, with no shell at all:

That is the whole adoption. The per-project instruction text is 232 words — a managed block in AGENTS.md, because the MCP tool descriptions already carry the how, and a second copy of that would drift from the one the model actually reads.

uv tool install ddflow-mcp        # or: pipx install ddflow-mcp
cd /path/to/your/project
ddflow adopt            # every supported agent
ddflow adopt --agents claude,cursor,vscode,kimi   # or name the ones you use

Docker — for operators with no Python toolchain

{ "mcpServers": { "ddflow": { "command": "docker", "args": [
    "run", "-i", "--rm",
    "-v", "${workspaceFolder}:/repo",
    "--add-host=host.docker.internal:host-gateway",
    "ghcr.io/delian/ddflow-mcp:latest" ] } } }

ddflow adopt --launch docker writes exactly that. The image is 107 MB (Alpine; ddflow is pure standard library, so there is no compiled dependency to worry musl about) and behaves identically on Linux, macOS and Windows.

Four things go wrong when a containerised tool touches a bind-mounted git repo. All four are silent, one of them loses work, and all four are handled:

And one that cannot be fully handled, so it is reported: 127.0.0.1 inside a container is the container. A model server on your own machine is not reachable from there. ddflow rewrites loopback reviewer URLs to host.docker.internal, and ddflow doctor tells you that on Linux you must also pass --add-host=host.docker.internal:host-gateway, because unlike Docker Desktop the Linux engine does not provide that name.

The related portability fix: worktree paths are stored in the event log relative to the repo root. The log is committed and shared, so an absolute path is true only on the machine that wrote it — false for a teammate who cloned elsewhere, for CI, and for a container where the repo is /repo. Pinned by test_the_event_log_carries_no_absolute_paths.

Extending it by writing text, not code

Every prompt is an external template, resolved config → project → shipped:

ddflow prompts list              # where each template currently comes from
ddflow prompts eject             # copy the shipped ones into .ddflow/prompts/
$EDITOR .ddflow/prompts/review_system.md

Adding a mode of your own: [[macro]]. Overriding a shipped workflow needs no code, and neither does adding one. A macro is a named, parameterised prompt — "enter debugger mode" — that appears everywhere the shipped workflows do: prompts/list and prompts/get over MCP, which is what a client turns into a slash command, and ddflow prompts list|show in a terminal.

# .ddflow/config.toml   (or .ddflow/macros.toml, if you prefer to split it out)
[[macro]]
name = "debugger"
title = "Enter debugger mode"
description = "Reproduce first, then bisect. No fix without a failing probe."
params = ["symptom"]                                  # required, not optional
tools  = ["ddflow_bug_found", "ddflow_gate_run", "ddflow_bug_fixed"]
prompt = """
You are debugging: {{ symptom }}

Reproduce it before you theorise. Paste the command and its output.
"""

Use prompt_file = "docs/modes/debugger.md" instead for anything long enough that TOML quoting gets in the way.

When to reach for a macro rather than a gate. A gate is a step every item passes through, recorded against that item and blocking its completion. A macro is a MODE an operator enters, belonging to no item and recorded nowhere — "audit this release", "handle this incident". If the thing should hold up a task until it is done, it is a gate; if it is a way of working you want to name and re-enter, it is a macro. Putting a mode in the pipeline makes every task wait for something that was never about that task.

tools is declarative, not a sandbox. It is rendered into the prompt as the ordered set the mode expects, so the agent is told what the mode is for and the next reader can tell what it was supposed to do. It does not restrict what the agent may call — MCP has no mechanism for that, and claiming a security property this cannot honour would be worse than not having it. This is the deliberate departure from dx-zero/mcpn, whose toolMode: situational lets the model pick freely from a bound set with no recorded ordering: a session you cannot replay is a session you cannot review, which is the property the event log exists to give you.

Three things a macro refuses, because each alternative fails quietly: a missing parameter (a prompt with a hole in it reads as a complete instruction), a name that belongs to a shipped command (silent shadowing leaves you editing a block that does nothing), and both prompt and prompt_file (two sources for one body means one is dead and looks live).

Including the one the agent actually reads first. mcp_instructions.md is the block an MCP client injects into the model's context on connect — the workflow, the reporting duties, and which companion tools to reach for. It is the file to edit when you want this project to work differently:

ddflow prompts eject mcp_instructions
$EDITOR .ddflow/prompts/mcp_instructions.md      # or [prompts] mcp_instructions = "..."

It renders against the live state — adopted, task_pipeline, setup_todo, companions, missing_companions, gate_gaps, recoverable, loops — so the instruction is the next concrete action rather than a fixed blurb the model learns to skip. A broken override says so in the instruction block itself instead of falling back to the default: this is the one surface where nobody would ever notice their edit was not live.

Templates render with Jinja2, which is ddflow's one runtime dependency, and with a strict standard-library renderer when it is absent — a stripped deployment with no reachable package index still starts. The shipped templates use the subset both engines agree on, and tests/test_template_engines.py walks the template REGISTRY, rendering every entry through both engines and asserting the outputs are byte-identical.

That test is iterated rather than hand-listed for a reason. Its predecessor named three templates in a dict, mcp_instructions.md was never added, and in 0.1.1 the largest and most important template rendered correctly under Jinja2 and failed under the fallback — so the entire MCP handshake for an unadopted repository, the first thing a new user ever sees, degraded to ddflow's instruction template could not be loaded. Jinja2 was not a declared dependency at the time, so developers had it and the project venv did not: python -m pytest was green and uv run pytest was red on the same commit.

The fallback now raises on any construct it does not implement rather than copying it through. The old regex engine emitted what it could not parse, so a condition as ordinary as {% if a or b %} — which its single-name pattern never matched — reached the client as literal template source.

Both renderers are strict about undefined variables: a prompt silently missing the diff it was supposed to carry is the vacuous review in template form — the model dutifully reviews nothing and reports no findings.

The rest is TOML: gates and their pipelines ([gate.*], gates.task_pipeline), reviewers ([[reviewer]]), companions ([[companion]]), enforcement ([enforce]), cadences, and the rest of the 120 knobs. ddflow config --set <key> <value> edits one key in place, preserving comments.

Publishing and registry

Nobody should have to paste JSON into an IDE to use this. server.json is the MCP registry manifest (io.github.delian/ddflow-mcp), and publishing it is what makes ddflow findable in the VS Code and Cursor marketplaces rather than something you configure by hand. It offers three ways to run the same server, so a client picks whichever it supports:

The image is built for amd64 and arm64, because an Apple-silicon operator running it under emulation pays that cost on every tool call, and tool calls are all this server does.

How CI authenticates — four mechanisms, one stored secret:

Docker Hub is the only one that needs a long-lived credential, because it has no OIDC equivalent. Use an access token scoped to this repository, never an account password. If that is one secret too many, delete the Docker Hub login and its two tags — ghcr.io alone satisfies the OCI entries a marketplace needs, and server.json lists both so a client picks whichever resolves.

environment: release on the publishing jobs is a control worth knowing about: point it at a GitHub environment with required reviewers and every release waits for a human, with no change to the workflow.

Order matters and the workflow encodes it. mcp-publisher validates that every package named in the manifest exists, so the registry step runs after both PyPI and Docker — publishing the manifest first would advertise a version nobody can fetch.

Four things gate a release, and each exists because the failure it catches is public and irreversible:

• the tag, pyproject.toml, server.json's version and every OCI identifier's tag must agree — a :0.1.0 left behind while version moved on publishes a manifest pointing at the previous image, installable and wrong;
• the full suite, plus the slow end-to-end scenarios, which -m 'not slow' otherwise excludes from every ordinary run;
• the wheel must install into a clean venv and run, and carry its templates — uv build succeeding proves the metadata parses, not that ddflow help works;
• the image must answer initialize over stdio. A built image that cannot is a broken release every marketplace will happily offer.

Cutting a release

$ scripts/bump.sh patch          # 0.1.0 -> 0.1.1, in all FIVE places that declare it
$ scripts/release.sh             # build + verify everything locally; publishes nothing
$ git commit -am 'release 0.1.1' && git push origin main

That push is the whole release. CI publishes PyPI, Docker Hub, ghcr.io and the MCP registry, then creates v0.1.1 and a GitHub release — last, and only once every publish succeeded, because a tag pointing at a half-release is worse than no tag: it looks authoritative.

The version bump is the release decision, and it is deliberate on purpose. A push to main publishes exactly when that number changes. Publishing on every push is arithmetic that does not work — PyPI refuses to re-upload a version, so the second push fails and every one after it — and deriving a unique version per commit instead would mean an irreversible release for a README typo. So one reviewable line in a diff decides, and everything after it is automatic. workflow_dispatch with force: true is there for the case where you need to republish deliberately.

The version lives in five places — pyproject.toml, server.json's version, its per-package version, the tag inside every OCI identifier, and SERVER_INFO, which is what the server tells every client it is. scripts/bump.sh moves all five and then re-reads them to check it did; tests/test_packaging.py fails if they ever drift. (That test caught the bump script missing SERVER_INFO on its first run.)

scripts/release.sh runs all of that locally and publishes nothing. It is dry by default, needs no credentials, and exists because a tag is not reversible: PyPI refuses a re-upload, :latest is on someone's disk before you notice, and a registry manifest is what an IDE offers people. If it fails on your laptop, the tag was going to fail an hour later in public. --publish is the escape hatch for when CI is unavailable, and it makes you type the version to confirm.

What CI checks

.github/workflows/ci.yml runs on every push and pull request, in four jobs that fail for different reasons so you can tell at a glance which:

Two of those exist because of specific failures. The wheel check is there because uv build succeeding proves the metadata parses, not that ddflow help works — a wheel missing its templates fails on the user's machine. And gitleaks is there because this project has already committed a live API key: a secret in git is a leaked secret, rotation is the only remedy, so the check that matters is the one that runs before every push.

bandit deliberately skips tests/, which use subprocess and temporary paths constantly and by design. A scanner that cries wolf on every fixture is a scanner nobody reads.

Any LLM as a reviewer — local, remote, SaaS, or a CLI

The critic and rubber_duck gates are run by ddflow, not claimed by the agent. Point them at whatever you have:

ddflow reviewers presets            # 19 ready-made provider settings
ddflow reviewers add --preset ollama --model qwen3:8b
ddflow reviewers detect --write     # probe local ports and register what is serving
ddflow reviewers test               # send a known-buggy diff, check the reply

Four backends, because "any LLM" means four wire formats in practice:

command is the escape hatch that makes the answer to "can it use X?" always yes: a model with no HTTP API, behind a corporate gateway, or wrapped in an in-house tool is still usable, with no SDK and no dependency.

[[reviewer]]
name   = "local-qwen"
kind   = "openai"
base_url = "http://127.0.0.1:11434/v1"
model  = "qwen3:8b"
family = "alibaba"              # must differ from the author's family
gates  = ["critic"]
# Optional: start it if it is not already running.
launch = { command = "ollama serve", ready_url = "http://127.0.0.1:11434/v1/models" }

[[reviewer]]
name    = "claude-via-cli"
kind    = "command"
command = "claude -p --model {model}"
model   = "claude-sonnet-5"
family  = "anthropic"
gates   = ["rubber_duck"]

Auto-launch is opt-in per reviewer — starting a multi-gigabyte model server as a side effect of asking for a code review is a surprise nobody wants by default. When it fails it never leaves a half-started process behind, because a reviewer stuck "starting" forever is indistinguishable from one that is down except that it also holds a process.

Keys are never written to the config. Only api_key_env, the name of an environment variable — the config file is committed, and a key in git is a leaked key.

Every way of not reviewing is reported distinctly, with its remedy: no key names the variable, a missing CLI names the binary, a dead server names the launch block you could add, a non-zero exit shows stderr, and empty output on exit 0 is UNAVAILABLE rather than "no findings" — the vacuous pass arriving by the most innocent-looking path there is.

Reasoning models need a large max_tokens. Default 32000, measured not guessed: on Qwen3.8-Flash-Next over a 30 KB diff, a 6000-token budget produced zero characters of content — the whole budget went to reasoning and the reply was truncated. That case is reported as TRUNCATED with the remedy named, never as an empty completion and never as a clean review.

Companion tools

ddflow imposes the order and demands the evidence. It does not perform the judgement inside most of its gates: standards wants an automated standards review, research wants documentation to check a claim against, rules wants memory of the last time somebody hit this. A project that installs ddflow and stops has those gates wired to nothing — and because an agent gate passes on an assertion, that gap is invisible in exactly the way the rest of this design exists to prevent.

So the gap is named:

$ ddflow companions
Companion tools

  [x] context7   Current library documentation
       gates: research, standards
       registered for: claude, cursor
  [x] roborev    Automated second-opinion code review
       gates: standards, bug_hunt, dedupe
       installed (roborev 0.9.1). A cli tool — the agent shells out to it, so
       there is nothing to register.
  [+] codeguide  Language and framework coding standards
       gates: standards
       installed (…) but no agent is configured to launch it.
       -> ddflow companions add --id codeguide
  [ ] sequential Structured step-by-step reasoning
       gates: research, rubber_duck, bug_hunt
       not here: `npx --no-install @modelcontextprotocol/server-sequential-thinking` exited 1
       -> ask the operator, then: npx -y @modelcontextprotocol/server-sequential-thinking

Gates in this project's task pipeline with no companion behind them:
  rules, implement, rubber_duck, critic, unit_tests, bug_hunt, dedupe, merge

Three states, reported separately because the remedies differ: registered, installed but not wired up (one command away), not installed (with the command and the URL). ddflow adopt prints the same summary, so the gap is visible at adoption rather than discovered six tasks later. Exit 2 when a default companion is missing — "no data", never collapsed into "no problem".

Servers and command-line tools are different things, and the registry says which: kind = "mcp" is registrable into an agent's config, kind = "cli" is a tool the agent shells out to. OptMem is the live example — a real tool with no MCP mode, so companions add refuses it and says why instead of writing a launch entry that would fail its first handshake. A cli companion counts toward its gate's coverage once it is installed; registered is a state it cannot reach.

Your stack needs servers this registry cannot know about. ddflow prompts show research-companions walks an agent from the pipeline's uncovered gates, through the repository's actual manifests, to candidates checked against their primary sources — provenance, maintenance, what they execute, what credential they want — and produces [[companion]] blocks you can read and delete. It proposes; you install. A rejection is part of its report, so the next session does not re-research it.

Registering is previewable. ddflow companions add --dry-run (and ddflow_companions_add with dry_run=true) reports the exact config entry it would write and writes nothing — not the file, not even its parent directory. The handshake tells an agent to dry-run first and show the operator the actual entry rather than a description of it, because registering changes which processes their agent launches. The preview is asserted to match what the real write produces; a preview that drifts from the write is worse than none, since the operator has now signed off on it.

ddflow never installs anything itself — running an install command on someone's machine is the operator's decision. What it does instead is instruct the agent to ask: the MCP instruction block lists each missing companion with the gates it serves and the exact command that would install it, and tells the agent to put that to the operator early, install it if they agree, and record the affected gates unavailable if they decline. Never on its own word.

companions add also refuses to register a server that is not present: that writes a launch command which fails mid-task, at the moment a gate told the agent to reach for it. Detection is read-only and bounded — and when it has not run, the state is reported as unknown, not as absent. ddflow companions probes; the MCP handshake does not, because making an agent wait on npx before it can do anything is the wrong trade.

Adding a fifth is a TOML block in .ddflow/companions.toml, not a patch:

[[companion]]
id      = "my-linter"
title   = "House linter"
gates   = ["standards"]
detect  = ["my-linter", "--version"]
command = "my-linter"
args    = ["mcp"]
install = "cargo install my-linter"

Wiring it into your agent

ddflow adopt --agents claude,cursor,codex writes everything below. This table is what it writes, so you can check it or do it by hand.

7 more are supported with no MCP file to write — a verified absence, not an unresearched gap. adopt writes the delta doc and the AGENTS.md block and names the one manual step. Inventing a path would be worse: ddflow would write a file the agent never reads, and you would believe it was wired up.

One set of rules, every agent

AGENTS.md is the cross-agent convention and most of the 22 read it. Seven do not read it first, or at all, so adopt writes the same managed block into their own surface too:

The rules are inlined, not pointed at. A one-line "see AGENTS.md" stub was the obvious design, and ddflow's own notes had already refuted it: a link is only followed if the agent chooses to follow it. A rule that binds only when the model feels like opening a file is not an enforced rule.

That means several copies of one text, and the answer is that a check owns them: one generator, a managed DDFLOW:BEGIN/END block in each, and ddflow doctor comparing every copy against the generator. Five kinds of break are reported and each fails doctor:

$ ddflow doctor
note:    QWEN.md's ddflow section is from an older version and has drifted
note:    .clinerules/ddflow.md exists but its ddflow section was removed
PROBLEM: .goosehints does not exist — the agent has no project rules at all
PROBLEM: .cursor/rules/ddflow.mdc exists but does not bind: `alwaysApply` is not
         true, so the agent may never load it
PROBLEM: .aider.conf.yml exists but does not bind: it does not list `AGENTS.md`
         under `read:`, and Aider loads no instruction file it was not told to load

Files the project already owns — QWEN.md, replit.md, .goosehints — get a block merged into them; your own content stays. Aider's read: list is extended, not replaced. Adopting is idempotent: re-running never appends a second block.

Every other agent reads AGENTS.md directly, which is the point of it being canonical.

Cursor gets its own rules file because its precedence puts project rules above AGENTS.md — writing only AGENTS.md there would be writing to a file the agent outranks. adopt merges into these files rather than overwriting: they hold your other servers and your other rules, and a tool that stomps them is a tool you run once.

The MCP entry is one line in any of them:

{ "mcpServers": { "ddflow": { "command": "uvx", "args": ["ddflow-mcp"] } } }

uvx fetches and runs it in an ephemeral environment on first use — no clone, no PYTHONPATH, no install step to forget. Prefer Docker? docker run -i --rm -v "$PWD:/repo" ghcr.io/delian/ddflow-mcp, which needs the repo bind-mounted because ddflow operates on your actual git checkout.

Standalone, with no MCP at all, is a first-class mode rather than a fallback. Add to AGENTS.md / CLAUDE.md:

This project's work is a queue managed by ddflow. Before doing anything, run
`ddflow brief`. Claim before you edit (`ddflow claim <id>`), satisfy every gate
(`ddflow gate status <id>`), then `ddflow merge` and `ddflow complete`.
Never pass a gate you did not perform — record `unavailable` with the reason instead.

That is the whole integration. An agent with nothing but a shell can drive the entire workflow, which is why MCP is a convenience layer here and never a requirement.

From plan mode to the queue

Agents plan well and forget reliably. A plan that lives in a chat transcript is gone at the next session; a plan in the queue survives, fans out to parallel agents, and carries its own gates.

Tell the agent, at the end of planning:

Put that plan in ddflow before you build any of it. One phase for the whole plan, one
task per independently-shippable step. Declare each task's globs — the files it will
write — and its needs, the tasks that must finish first. Then show me `ddflow next`.

What the agent does with that:

$ ddflow phase add P3 --title "Rate limiting"
$ ddflow task add P3.T1 --phase P3 --globs 'limiter/**'      --title "token bucket"
$ ddflow task add P3.T2 --phase P3 --globs 'api/middleware/**' \
      --needs P3.T1 --title "wire it into the request path"
$ ddflow task add P3.T3 --phase P3 --globs 'docs/**' --needs P3.T2 --title "document it"
$ ddflow next
Ready (1 ready, 0 running, 2 blocked):
  P3.T1  token bucket
      writes: limiter/**
  (blocked) P3.T2: deps — P3.T1 is open
  (blocked) P3.T3: deps — P3.T2 is open

The two fields that do the work are --globs and --needs. Globs are how two agents are stopped from editing the same file: claim refuses an item whose writes overlap one already held, and names what to take instead. Needs are how ordering is enforced without anyone remembering it. A plan whose tasks declare neither is a list, not a queue — it will look parallel and then two agents will fight over one file.

ddflow split <id> --into a,b,c exists for when a task turns out to be three, which is the normal case rather than a failure of planning.

When a companion is missing

ddflow imposes the order and demands the evidence. It does not perform the judgement inside most gates — that is what the companion tools are for. So the honest question is what happens when one is absent, and the answer is deliberately never "the gate passes".

The rule, and it is enforced: a gate whose tool could not run is recorded unavailable with the reason, never passed. ddflow complete reports those as a coverage gap on the completion event, so a finished item never silently implies that a check happened. Set [gates].unavailable_is_failure = true and a gap blocks completion outright.

ddflow companions reports four states, and the difference between the last two is the whole point: registered, installed but not wired up (one command away), missing (with the install command and the URL), and not checked — because the MCP handshake does not probe, and "nobody looked" must never render as "not there".

ddflow never installs anything. Detection is read-only and the report is advice. ddflow companions add --dry-run shows the exact config entry it would write, so an agent can show you the change before making it.

Adopting a project that already has history

A queue that starts empty tells the next agent "nothing is in flight" about a repository with three branches in flight and forty open items in a todo file — and the agent believes it, because the tool said so. That is worse than having no tool at all.

$ ddflow import                      # looks; writes nothing
What this project already has (nothing written yet):

  314 phase(s):
    [ ] 142.A     the scaling-law advisor is wrong (P0; CONFIRMED)   docs/todo.md:26517
  1170 task(s):
  442 lesson(s):
  ...
  47 memory(s):
    [ ] M-0002    Hardware: 8x H200 GPUs on this box, usually idle.  .agent_memory/LOG.txt:3

  NOTE: 3631 already-ticked task(s) were NOT imported. They are history, not a queue.
  NOTE: 32 phase heading(s) say the work is finished while their checkboxes are still
        unticked: 99 (4 open), 103 (3 open), ... Ask the operator which is stale.

$ ddflow import --apply              # writes them, each recording its source line

Seven sources, all optional, all in the places projects actually keep them:

What it will and will not decide for you

Mechanical, and verifiable: a ticked checkbox is a fact, a ## heading is a section, a branch with unmerged commits is work. The id in ### 142.A — … or - [ ] **WFOPT.4.6** — … is read, not invented, so the imported queue uses the ids the project has been writing in commit trailers for months.

Judgement, and yours: which open items are actually live, what each task writes, what depends on what. The /import-existing-project prompt walks an agent through that with the operator. It is not automatable, and a confident guess produces a wrong queue the scheduler then hands out.

Four guard rails, each of which exists because the alternative is silent:

• Dry run by default. --apply writes. Looking is free and never a side effect.
• Finished work stays out — it is history, not a queue — except a completed item that open work depends on, which comes along as done so the open item is not stranded on an id the queue has never heard of.
• [importer] max_tasks (default 200) refuses a whole history. An import writes events into a log that is committed to git; one real repository yielded 4,799 checkboxes. Over the cap it proposes none and says so — the phases are withheld with them, because a queue of empty phases is not a smaller import, it is a misleading one.
• Idempotent. Ids derive from the source, so re-running after you edit the todo adds what is new and leaves the rest alone. A second run over an unchanged project exits 2.

What an unticked box means, and where things live

An open box is not always work. The import reads the project's own dispositions — the vocabulary and positions are those of the picker ddflow was extracted from, and agree with it on 1,166 of that repository's 1,170 open boxes. Of the four, two are the picker's own false positives ("cells run / skipped" in plain prose) and two are recorded non-findings with an unclosed "(… refuted it" aside, which ddflow closes:

[importer] archive_globs names plan files that are history until a section is named (a 20,000-line legacy docs/todo.md): their open boxes import blocked, and ddflow unblock <phase> releases a whole section at once.

Each source family's location is a knob — todo_globs, lesson_globs, lesson_summary_globs, decision_globs, research_globs, journal_globs, memory_globs — and a set knob replaces the defaults, because the same filename means opposite things in different projects (docs/LOG.md is one repository's whole journal and another's generated index of it; an Index section is never imported).

Lessons are split at the level they actually live at: ### L100. … entries grouped under ## <date> headings import one per lesson with their own ids, so [L147] cross-references still resolve. A lesson's **Compressed:** paragraph becomes its summary; a hand-written lessons-summary.md bullet that cites exactly one lesson becomes that lesson's summary, and every other bullet becomes a consolidated lesson tagged summary. A GENERATED summary file is skipped. ddflow render writes them all back out as docs/ddflow/LESSONS-SUMMARY.md (also ddflow://lessons-summary).

Verifying an import, at any time

The import's weak spot was never the parsing. It is everything after --apply: 1,170 tasks arrived in the real-corpus run, and the workflow prompt tells an agent to give each one globs and declare its dependencies. Nothing checked whether that ever happened — and an imported queue nobody finished misrepresents the project exactly as an empty one does, believed harder because a tool produced it.

$ ddflow import --verify
Imported between 2026-09-25 and 2026-09-25:

       4 decision(s)
    1727 journal(s)
     442 lesson(s)
      47 memory(s)
     314 phase(s)
      72 research(s)
    1170 task(s)

Left to decide or fix:
  - 1078 imported task(s) declare no globs, so the conflict detector cannot protect
    them and two agents can be handed the same file: OPIK.1b, OPIK.2, ...
  - 27 phase(s) say the work is finished while a task under them is still open: 99,
    103, 115.D.2, ... Ask the operator which is stale before anyone claims from them.

Three answers, three exit codes, because collapsing them loses the one that matters:

It reports status (what is imported, per kind, and when), whether it is still true (what a re-run would add, which sources yielded nothing, which source files have since vanished), and whether anyone finished it (tasks with no globs; phases whose heading claims SHIPPED over an open task).

It deliberately does not repeat ddflow doctor, which already reports unresolved dependencies, duplicate globs and cycles. Two commands reporting one defect in different words is how an operator learns to read neither.

Provenance is a field, not prose. Item.source is docs/todo.md:41; the body still says "Imported from docs/todo.md:41." for a human reading ddflow show. Answering "which items came from the import" by regexing that sentence would mean the day someone rewords it, the count silently becomes zero and the verification passes.

The connection handshake follows through. The offer to import stops once the queue has anything in it — but if imported work is still missing globs, or a phase still claims SHIPPED over open tasks, the MCP instructions say so and tell the agent to run ddflow_import_verify before handing any of it out. That check is computed from the already-folded queue, so it costs nothing; the source re-scan (~0.65 s) stays out of every session start and happens only when someone asks for it.

Re-running is a first-class path. /import-existing-project opens by checking what is already imported and switches to finishing and refreshing rather than repeating — fix the globs it names, ask the operator about the SHIPPED drift, re-run ddflow import for sections added since.

What it reports rather than fixes

Three kinds of drift it can see and must not resolve on its own, because either answer could be the wrong one:

• A phase heading that says SHIPPED over unticked checkboxes (32 of them in the repository this was measured against). One-sided risk: if the heading is right, the queue is about to hand out work that is already done.
• A dependency on an id nothing produced. Kept and treated as unmet — deliberately, so a typo surfaces as blocked work rather than as work that starts early — but named, because "never offered" otherwise looks exactly like "nobody has got to it yet".
• A file that matched a source pattern and yielded nothing, which usually means an unusual format rather than an empty file.

The model: phases, tasks, dependencies, globs

Plan ──► Phase ──► Task

A phase is a unit of review: its own research, its own whole-suite test pass, its own live smoke run, merged as one coherent feature. A task is a unit of execution: one agent, one worktree, one pipeline, one merge. Both carry needs (dependencies, which may cross phases) and globs (the files they will write).

ddflow phase add P2 --title "Billing" --needs P1
ddflow task add P2.T1 --phase P2 --title "invoice model"  --globs "src/billing/invoice.py"
ddflow task add P2.T2 --phase P2 --title "tax rules"      --globs "src/billing/tax.py"
ddflow task add P2.T3 --phase P2 --title "checkout wiring" --needs "P2.T1,P2.T2" \
                                                            --globs "src/checkout/*"

Declare globs. They are what lets two agents work at once safely. A task with no declared globs is a task the conflict detector cannot protect.

Dependencies are inherited. A phase is never claimed — only its tasks are — so P2 needs P1 has to govern everything inside P2, or it governs nothing that anyone picks up. The readiness rule therefore consults an item's ancestors as well as itself:

$ ddflow next
Ready (1 ready, 0 running, 1 blocked):
  P1.T1  money
  (blocked) P2.T1: deps — phase P1 has 3 open task(s) (inherited from P2)

The refusal names where the dependency came from, because an operator told only "P2.T1 needs P1" goes looking for a declaration that is not written there. The one dependency not inherited is one pointing into your own subtree: an umbrella that declares a dependency on its own child would otherwise make the child wait for itself, turning a plan typo into a permanent hang.

ddflow claim asks the same predicate ddflow next does. They used to disagree — next withheld a task on its dependencies and claim handed out a worktree for it a second later — so an agent picking work by id rather than by asking bypassed the dependency graph entirely.

There is deliberately no third level of kind: a sub-task is a task whose parent is a task, so depth is unlimited while the rules stay one set.

Work that changes shape while you do it

Tasks can be added at any time, including while their parent is being worked — mid-task discovery is the normal case, not an exception, and a queue that cannot absorb it pushes the work into someone's head.

Sub-tasks are just tasks whose parent is a task. Not a separate concept with its own rules: a sub-task declares its own globs, carries its own dependencies, is claimed by its own agent, and runs in parallel with its siblings when nothing links them — exactly like any other task.

ddflow task add P1.T1a --parent P1.T1 --globs "src/parse.py"
ddflow split P1.T1 --into "P1.T1a=parse input" --into "P1.T1b=write records"

split works in place: the original keeps its id, its lease history and everything recorded against it, and becomes an umbrella that completes when its children do. Closing it and opening two new ones instead would lose the thread between what was planned and what happened — which is exactly what ddflow replay needs.

An umbrella is never offered as ready (its children are), and cannot complete while any descendant at any depth is unfinished. An abandoned child counts as settled, so a sub-task you decide against does not hold its parent open forever.

Becoming an umbrella releases the lease, however you get there — by split, or by adding the first sub-task to a task you are already working. An umbrella holding a live claim on globs that overlap every child's means a second agent cannot take one of those children, and crash recovery points at a worktree where nothing further will happen. split already did this; task add --parent did not, which is the shape of bug worth naming: one transition, two ways in, guarded on one.

Architectural decisions

The code shows what was built and never why, nor what was rejected on the way. So decisions are recorded as events, and reach the person writing the code:

ddflow decision add --title "Storage is SQLite with WAL" \
  --decision "One file, WAL mode, BEGIN IMMEDIATE for writes." \
  --context "Three call sites were each opening their own connection." \
  --alternatives "Postgres — rejected: no server allowed in this deployment." \
  --globs "src/storage/*" --by operator

--globs is what makes a decision consulted rather than merely filed. ddflow brief and ddflow decision applicable <item> surface the decisions governing an item's declared files automatically — the agent does not have to suspect they exist.

Decisions are never edited or deleted. A reversal is a new decision naming the old one (--supersedes), so the history of how the architecture got here survives, and a superseded decision is shown with a pointer to its replacement rather than silently withheld.

Recall — "have we been here before?"

ddflow recall "how should durations be represented"

One search across everything the project remembers: architectural decisions, lessons, operational memories, research verdicts, past bugs, similar tasks, and the operator's own earlier prompts. Results are labelled by kind, because a binding decision, a transferable lesson and a prompt from three weeks ago should change what you do in different ways.

It exists so the operator does not have to say the same thing twice and the agent does not have to learn the same thing twice. Both failures are invisible in the moment and obvious in the log.

Operational memory

ddflow memory add "8x H200 on this box; check nvidia-smi before a GPU test" --tags gpu
ddflow memory list                       # newest first; --query to rank, --all for forgotten
ddflow memory forget M-0003 --reason "the box was upgraded"

One fact about this machine, repository or working state — not a rule (lesson), not what happened (session note), not how the software is built (decision). Capped at [memory] max_chars (280) and refused, not truncated, when longer. The newest [memory] brief_items appear in every ddflow brief, right after the binding decisions, and recall searches them — the job an OptMem store beside the repository used to do, now in the log, so every worktree sees a memory the moment it is written. A memory that stopped being true is forgotten with a reason, never deleted: "we thought X until Y" is what stops the next agent re-learning X. An OptMem LOG.txt imports as memories dated when they became true. The log is committed: never put a secret in one.

Resources and long-running jobs

Globs keep two agents out of one file. Work that runs on something — GPUs, a model server, a shared fleet — declares that too, and the same refusal applies:

ddflow update TRAIN.3 --resources gpu:6            # or **Resources:** gpu:6 in the plan
ddflow claim TRAIN.3                                 # refused (exit 3) if it does not fit

[schedule] resources = ["gpu=8", "vllm-fleet=1"] sets capacities; a resource named nowhere is exclusive. Every live claim counts, the claimant's own included — one agent starting two 8-GPU runs overcommits the box just the same — and next withholds what does not fit, saying who holds what.

The run itself is a job:

ddflow job run TRAIN.3 "uv run main.py train -c cfg.toml"   # detached; survives you
ddflow job list        # running | exited N (from its log) | gone (killed) | elsewhere
ddflow job end J3f2 --note "loss 0.12, ckpt in out/"         # refused while it runs

A launched job runs in the item's worktree, in a session of its own (it outlives the agent, the MCP server and a restarted remote-control service), and appends its exit code to its log so a run nobody watched still says how it ended. Liveness is computed, not stored: a zombie is not alive, and a reused pid is caught by the process start time. ddflow job add --pid registers a process started some other way. Every brief lists jobs not yet recorded as ended — "WAIT, do not start it again" for a running one.

Dependencies on another repository

[schedule]
repos = ["run_nemo_run=../run_nemo_run"]

ddflow update GEN.4 --needs run_nemo_run:132.D then waits for item 132.D there to be done. ddflow external sync (and every session-start hook) reads the sibling's log — never writing it — and records what it observed in this one, only when it changed. Readiness is decided from that dated fact, so the fold stays pure and "why was this started?" is answerable later. An unobserved external dependency is unmet; one naming a repository that is not configured is a doctor problem, since it can never be met.

Status, progress, and loops

ddflow status      # what is done, in flight, ready, blocked — one answer
ddflow progress    # attempts, hours held, gate runs, commits, per item
ddflow loops       # circular references and runtime loops (exit 2 = none)

Dependency cycles are the easy case. The expensive ones are runtime loops, where the graph is perfectly acyclic and the work still never finishes:

Every threshold is a [loops] knob, and on_detect = "block" makes ddflow claim refuse an item that is already looping — a warning is read by a human later, a refused claim is read by the agent now.

The task pipeline

Ten gates, in order, configurable per project:

Four things are enforced rather than requested:

Silence is not a pass. Every gate in the pipeline must carry some outcome before an item completes — passed, failed, unavailable, partial, or an explicit ddflow gate skip <id> <gate> --reason "...". Without this, gates.required held only implement, unit_tests and merge, so six of the ten steps could be omitted with no trace at all. gates.require_outcome = false makes the pipeline advisory again; gates.enforce_order ("warn" by default, or "block") reports a gate recorded before an earlier one has run, because a rubber-duck review recorded before implement reviewed an empty diff.

UNAVAILABLE is never a pass. A reviewer whose endpoint was down approved nothing; a linter that is not installed found nothing. Each gets its own outcome and shows as a coverage gap. (The inverse matters too: this codebase's first version classified a missing binary — shell exit 127 — as failed, so an uninstalled linter looked like a linter reporting problems. Fixed, with a mutation-verified regression test.)

Evidence or it did not happen. Gates in gates.evidence_required reject a bare pass; they want the command, its exit code and its output digest.

And evidence says WHICH tree and HOW MUCH. Every gate that produces an OUTCOME — command gates, and agent gates recorded with gate record — carries a tree_sha: a fingerprint of the working tree it ran against, covering committed state, uncommitted changes to tracked files, and the content of untracked ones (a new module is untracked until its first commit, which is the ordinary state of agent work). ddflow's own .ddflow/ is excluded, or recording a gate's outcome would invalidate the gate that just recorded it. If the tree moves afterwards, complete warns that the pass describes source nobody is shipping — a warning, not a block, because refusing on a comment-sized change is how a check gets switched off.

Beside it, diff_stat records files, insertions, deletions and untracked count — including the lines in untracked files, because a new module is untracked until its first commit and a task that is entirely new files would otherwise report zero insertions. The fingerprint answers which tree and is opaque; this answers how big, and that is what makes a pass auditable later — a review gate that passed over 4,000 changed lines in two minutes is a different claim from one that passed over 12.

Neither is recorded for a skip (nothing was reviewed, so a magnitude would imply an inspection that did not happen) nor for an unavailable gate that never ran.

Hashing untracked content is capped by MAX_UNTRACKED_HASHED (512). Above it the fingerprint falls back to file names and says so inside the digest, because a check that quietly stopped covering content would go silent for exactly the repositories that need it most.

Reviewer independence is checked, and an unidentified reviewer establishes nothing. Same-family reviewers share the author's blind spots, so their agreement measures shared priors rather than correctness. complete refuses unless one reviewer came from a different pretraining family — and a reviewer whose model is not in [agent].families counts as unknown, never as different. (It used to count as different: gate record defaults the reviewer to the agent id, so a standards gate recorded with no --model arrived as family "host-12345", compared unequal to "anthropic", and satisfied the independence requirement on its own.)

$ ddflow complete P1.T1 --model claude-opus-5
cannot complete P1.T1 — 1 unmet condition(s):
  - reviewer independence not satisfied: every reviewer (rubber_duck) was family
    'anthropic', the same as the author. Same-family agreement is not independent evidence.

Every unmet condition is listed at once — a refusal that reveals one problem at a time trains an agent to reach for --force.

Proving a gate can fail at all

$ ddflow gate verify T1 unit_tests
  OK   src/calc.py: detected

unit_tests CAN fail: every registered mutation was caught.

A gate that cannot go red is worse than no gate — it reports success on every change, and everyone downstream reads that as evidence. gate verify breaks what the gate guards, using the mutations registered beside it, and requires the gate to notice:

[gate.unit_tests]
command = "python -m pytest -q"
cwd = "repo"
mutations = [ { file = "src/calc.py", old = "return a + b", new = "return a - b" } ]

Four things make it honest rather than ceremonial, and the last one is this feature's own bug, found by a cross-family review of it:

• A mutation that did not apply is a FAILURE, not a skip. If old is absent — or present twice, so the edit is ambiguous — the check fails. Skipping turns "the mutation never happened" into a green run, which reads as the opposite of the truth.
• The source is restored whatever happens, including on exception, or a failed verification leaves the tree broken and the next gate reports the verifier's fault.
• A gate with no registered mutations is reported as unproven. Declaring a check nobody has shown can fail is what this exists to catch. An agent gate says plainly that it has no command to mutate and rests on its evidence contract instead.
• A green baseline is required first. A gate already red for an unrelated reason — one pre-existing failing test, a tool that stopped being installed, a flake — reports failed for every mutation, so every mutation reads as detected and the gate is certified as able to fail when nothing has shown any such thing. The check written to catch the vacuous-pass class contained it. It now runs unmutated first and refuses without a pass.

When the exit code is not the verdict

Some tools say "I could not run" or "I only did part of it" with an exit code, and some exit 0 whatever happened. A command gate can say which:

[gate.critic]
command = "uv run scripts/critic_review.py --dirty -c configs/review_critic.toml"
unavailable_exits = [2, 143]      # endpoint down / SIGTERM: UNAVAILABLE, not failed
partial_exits = [3]               # reviewed part of the diff: PARTIAL
require_output = '^STATUS:'       # exit 0 without it = the tool did not do its job
fail_output = '^\s*- \[(HIGH|MEDIUM)\]'   # exit 0 WITH findings = failed

ddflow gate run also renews the caller's lease every [lease] heartbeat_s while the command runs, so a 25-minute suite does not outlive a 30-minute lease and read as abandoned work.

The phase pipeline

research → [ task, task, task … ] → unit_tests → bug_hunt → dedupe
         → live_test → corrections → merge

live_test is the one most often skipped and the one most worth keeping: a green unit suite and a working feature are different claims. Run the real thing on a small input and paste what it printed.

Human approval: a gate the agent cannot clear

Every other gate here is satisfied by the agent — it runs a command, or it asserts it did the thinking. That is right for work whose correctness is checkable afterwards, and wrong for a plan: by the time an agent has built the wrong thing, the cost is already paid.

A gate marked human = true is where the operator says yes, build that before the compute is spent.

# .ddflow/gates.toml
[gate.plan_approved]
title  = "Operator approves the plan"
human  = true
prompt = "Show the operator what you intend to build, then ask."

$ ddflow gate run T1 plan_approved
gate 'plan_approved' is a HUMAN-APPROVAL gate. It is not something you can run or
record — it is where the operator decides whether this work should proceed.
  ddflow approve T1 plan_approved
  ddflow approve T1 plan_approved --reject --reason '...'

$ ddflow gate record T1 plan_approved --outcome passed --evidence "looks fine"
'plan_approved' is a human-approval gate: it is cleared by a person, not by an agent
recording that it happened.                                            # exit 3

$ ddflow approve T1 plan_approved --note "read the plan, ship it"
T1.plan_approved approved by delian — read the plan, ship it

A rejection is a first-class outcome, not the absence of an approval: "the operator looked and said no" and "nobody has looked yet" are different states, and an item sitting in the second forever is how a checkpoint becomes a silent stall. --reject requires --reason.

There is deliberately no MCP tool for this, and tests/test_mcp_parity.py records the exemption with that reason. A human checkpoint reachable from the MCP surface is not a human checkpoint — it is a second gate record with a longer name. gate skip is refused too: "the operator does not need to approve this" is not the agent's call.

What this is, precisely. An audit trail and a speed bump, not a security boundary. An agent with shell access can run ddflow approve itself, and no design here changes that — the tool does not control the machine.

The guarantee, as narrowly as it holds: no MCP tool records a human outcome, and a clearance carries the OS user and a human flag, so a forged one is visible in the log rather than indistinguishable from a real one. (gate.<id>.human is also refused by the config writer, because two MCP calls — flip the flag, then record — used to clear the gate with no shell involved. Declare human gates in .ddflow/gates.toml, which no tool writes.)

Opt-in: the shipped pipeline has no human gate, and a test keeps it that way.

Parallelism and coordination

$ ddflow next --phase P1
Ready (3 ready, 0 running, 1 blocked):
  P1.T1  persistent store
      writes: shortener/store.py, tests/test_store.py
  P1.T2  base62 encoder
      writes: shortener/encode.py, tests/test_encode.py

These are independent — run them in parallel worktrees.
  (blocked) P1.T3: deps — P1.T1 is open; P1.T2 is open

ddflow claim <ID> leases the item and binds it to a worktree.

If you are already in one, it adopts that one. Agent harnesses — Claude Code, Cursor — often isolate the agent themselves. Claiming from inside a linked worktree binds the item to that tree and branch rather than building a rival and telling you to leave the one holding your uncommitted work:

$ ddflow claim T1            # run from inside the harness's own worktree
claimed T1 (lease 1800s, renew every 300s)
  worktree: /work/agent-tree  (adopted — you were already in it)
  branch:   agent-work
  Carry on where you are.

ddflow never needed to have created the tree — it needs to know which tree an item is worked in, so recover can find stranded work and merge knows what to merge. An adopted tree is recorded as adopted, not created, so remove_on_merge will never delete something ddflow did not make. A tree already bound to another open item is refused: two items in one tree cannot be merged or recovered separately. worktree.adopt_existing = false restores the old behaviour; --no-worktree skips binding entirely.

A second agent is refused, and told what to take instead:

$ ddflow claim P1.T4 --agent gamma
P1.T4 writes 'shortener/store*.py' which overlaps 'shortener/store.py' held by alpha on P1.T1

You could take instead: P1.T2, P1.T5

Exit codes are the contract, and agents branch on them:

"Nothing is ready" and "everything is fine" are different facts. An agent that cannot tell them apart invents work.

The critical path is reported, because it, not the task count, sets the wall-clock floor — adding a fifth agent to a phase whose runtime is a four-deep chain buys nothing.

Gitflow, pull requests and version tags

Two independent axes in [flow], because teams combine them freely:

Which workflow that makes, and what is not covered:

Each of these is a workflow choice: the operator sets it, or an agent records it, and when nobody does the default is applied at first use and followed from then on.

The agent's loop does not change. next → claim → work → gates → merge. What merge means changes with the repository's policy:

• merge in PR mode pushes the branch and opens (or updates) a pull/merge request, then releases the lease and parks the item in REVIEW. The agent is free at once and takes the next task — nobody waits for a human. It refuses to open a request for work whose own pipeline is unfinished: a reviewer's time is the scarce resource, and a refusal after the merge could no longer stop anything.

• pr sync turns what reviewers did back into queue state. next runs it for you while anything is in review (sync_on_next):

  the forge says	ddflow does
  merged	passes the merge gate on the forge's evidence (URL, merge sha), completes the item, retargets anything stacked on it, removes its tree
  changes requested	returns the item to the queue with the review text (bodies and line comments); brief leads with it; a re-claim resumes the same tree and a re-merge updates the same request
  closed	parks it for a person — a "no" is not something to retry
  approved, checks green	merges it (pr_merge = "on_approval", pinned to the approved head)

• Stacking keeps work moving through review. While T1 waits in review, a task that needs it may start on top of T1's branch (stack = true); its request targets T1's branch and is retargeted to the real base when T1 merges. ddflow never merges a stacked request first — that would land it unreviewed inside T1's merge. A task depending on two unmerged branches waits: one branch cannot sit on two.

• Who presses merge is pr_merge: on_approval (default — a person's approval is still required, and branch protection still applies), auto (ask the forge to auto-merge when its own rules are met), or human.

Gitflow. Tasks fork from develop as feature/ or bugfix/ branches (by tag); a task tagged hotfix forks from production and lands on production and develop. Merges never switch a checkout: a target that is not checked out is merged in a throwaway worktree, and one checked out in someone else's tree is refused.

Versions. version show reads the highest v1.2.3 tag reachable from the release branch and computes the next version from Conventional Commits (feat → minor, fix → patch, !/BREAKING CHANGE → major; below 1.0.0 a breaking change bumps minor) and from the tags of items finished since (breaking, feature, bug, hotfix). version cut tags it — annotated, with generated release notes. Under gitflow it cuts release/X from develop, merges it into production, tags it and merges the tag back into develop; in PR mode it opens the release request instead, and pr sync tags the merge commit once a person merges it and opens the back-merge request.

ddflow holds no token: it drives gh or glab, logged in as the operator, so every permission question is answered by the forge. A forge that cannot be reached is exit 2 — "could not ask" is never reported as "nothing changed".

[flow]
model = "gitflow"          # or "trunk"
integration = "pr"         # or "merge"
pr_merge = "on_approval"   # or "auto" | "human"
pr_reviewers = ["alice"]

The research behind this is RESEARCH R16.

Several release lines: fixes to older majors

Projects that keep older majors alive — main is 3.x while 2.x and 1.x still get fixes — declare them as release lines, oldest first. The newest line is always the current one and follows model as above; a maintenance line lands straight on its branch.

[flow]
current_line = "3"
port_strategy = "cherry-pick"      # or "forward-merge" (the default)

[flow.lines]                        # oldest first
"1" = "maint/1.x"
"2" = "maint/2.x"

• An item belongs to a line: task add T --line 2, or phase add P --line 2 and every task in it inherits it. A line that does not exist is refused, never read as "the current one".

• A fix for several lines is one command: task add FIX --lines 1,2,3. ddflow writes it where the strategy says and generates a port task FIX@<line> for each other line — an ordinary task with its own branch, gates and merge or pull request. A port starts only once what it carries has landed (review is not enough):

  port_strategy	the fix is written on	each port…	ports run
  forward-merge (default)	the oldest line	merges the previous line's branch into its own (and passes through every line in between — a merge cannot skip one)	one after another
  cherry-pick	the newest line	applies exactly what the fix landed (the target's before→after range, whatever the merge strategy) with a three-way apply	in parallel

• A conflicting port is work, not a failure. The claim leaves the conflict markers in the port's tree and names the files; the agent resolves, commits and carries on.

• Lines never collide. The same file on 2.x and on 3.x is two branches, so two agents may hold them at once; on the same line the glob check applies as always.

• Versions per line. version show --line 2 reads the highest tag reachable from maint/2.x; version cut --line 2 tags it there, and refuses a bump that would leave the 2.x major — a breaking change belongs on the current line.

Environment branches: promoting downstream

GitLab flow's environment branches — each mirroring what is deployed there — are declared in order, downstream of the current line's target (the base branch; under gitflow, production, so an environment receives what was released):

[flow]
environments = ["pre-production", "production"]
auto_promote = ["pre-production"]   # optional: continuous delivery to staging

• Work reaches an environment only by promotion, one step at a time. ddflow promote add pre-production files a task that merges main into pre-production; promote add production merges pre-production into production. So production only ever receives what the environment before it already has ("upstream first"). Nothing else targets an environment branch.
• A promotion is an ordinary task: claim it and its tree is made from the environment branch with the upstream branch already merged in (a conflict is left for the agent, like a port's). It runs gates.promotion_pipeline — unit_tests and merge by default, since what it carries already passed its own pipeline; add a human gate there for a person's sign-off on each deploy. It needs no cross-family reviewer: it authors nothing.
• With integration = "pr" the promotion lands through a merge request into the environment branch — the approval is the deploy approval — and pr sync completes it.
• One open promotion per environment, and none when there is nothing to carry (exit 2).
• auto_promote lists environments ddflow next promotes to by itself when the branch upstream moves. Empty by default: a deploy is the operator's call.
• ddflow promote status shows each environment's head, how many commits it is behind the branch upstream of it, and any open promotion.

Workflow choices: asked, recorded, defaulted on the record

ddflow supports several ways of working and never picks one silently. Each decision — model, integration, pr_merge, on_changes_requested, stack, port_strategy — is a choice, and its value comes from, in order:

1. the operator's config (.ddflow/config.toml or env), which always wins;
2. a recorded choice — ddflow flow choose port_strategy cherry-pick --reason "2.x has diverged", by the operator or by an agent the operator left it to, attributed in the log;
3. the default — applied the first time the choice matters (the first claim, the first pull request, the first fix filed across lines) and recorded, so the project keeps following it even if a later ddflow ships a different default.

Until then, a relevant choice nobody made heads ddflow brief under Open workflow choices, so an agent asks at the start rather than discovering at the end that the project wanted something else. ddflow flow show lists every choice with its value, its options, and who decided — config, a named agent or person with their reason, or "DEFAULT (nobody chose)". A recorded choice that the config file overrides is shown as such, never silently ignored.

Research: RESEARCH R17.

Many agents, one server: identity, state and sharing

Several agents and subagents sharing one queue is the case this tool is for. Here is exactly how that works, because each of these has a wrong answer that looks right.

Is it stateless?

The queue is. The connection is not, in exactly one respect.

The append-only event log is the sole source of truth, and every read re-derives state from it — fold(read_all()), from scratch, on every call. Nothing is cached between requests, so there is no stale projection, no invalidation, and no divergence between two agents' views. Restart the server mid-task and nothing is lost: it never had anything the log did not.

The one piece of per-connection state is who you are (below). It is deliberately not in the log, because it is a property of the caller, not of the work.

Who is calling?

By default, identity is derived from the working tree. That is right for one agent per worktree, and silently wrong for several agents in one tree — they all resolve the same path to the same name, their events merge into one stream, brief answers with a sibling's task, and reviewer-independence compares an agent with itself and passes. Nothing errors. There is no signal that can tell them apart, so identity is declared:

A name you set yourself is never suffixed, so the same DDFLOW_AGENT in two clones is still one agent to ddflow; ddflow doctor notes a shard whose clock goes backwards, which is what that leaves behind after a merge.

Innermost wins. ddflow_identify is idempotent, persists for the connection, and refuses a name that could not be a log filename — it becomes one, and refusing at declaration time means the caller reads the reason rather than discovering it at the first write.

If more than one agent works one tree at once, declare identity. Everything that attributes work depends on it. A subagent must not call ddflow_identify on a shared connection — that renames its parent — and passes as_agent instead; without it, two subagents claiming the same file are one holder, and a holder's own leases never conflict with each other.

Can one server serve several projects?

No — one server process serves one repository, fixed at start from --repo, DDFLOW_REPO, or the working directory. No tool takes a repo argument, and a test asserts none ever does. Point a second agent at a second project by running a second server; they are cheap, and the isolation is the point.

One project shared by many agents is the supported case — and the one that needs no special setup beyond declaring identity:

• Writes never conflict, and they never block readers. Each agent appends to its own log shard, so there is no shared file to overwrite and no merge conflict to resolve. Writers do serialise briefly: one repo-wide lock is held across the clock allocation, the append and its fsync. Short, but not nothing — per-agent shards remove file contention, not lock contention.
• Reads take no lock at all, so a read-heavy agent cannot be starved by a write-heavy one, and a reader can never block a writer.
• File ownership is coordinated by globs. claim refuses an item whose writes overlap one already held, and names what to take instead — exit 3, not a failure.
• Lessons, decisions, research and bug history are shared by construction: they are events in the same log, so one agent's finding is immediately visible to every other.

Locking, contention and measured cost

Measured on this machine, single process, full read_all() + fold():

Linear, converging on ~8.7 µs/event; the higher figure at small sizes is fixed per-call overhead, not the fold. A project with 20,000 events pays ~175 ms for a state-reading call. Search and recall do not pay this — they run off a SQLite projection rebuilt only when the log's head moves.

tests/test_mcp_load.py runs 12 concurrent agents through the real MCP surface and asserts no deadlock, no lost append, no repeated Lamport value within an agent, and correct attribution for every event — not for a sample. Its thresholds are environment variables (DDFLOW_LOAD_AGENTS, DDFLOW_WRITE_LATENCY_BUDGET_S, DDFLOW_GROWTH_TOLERANCE, …) because a load test with a hardcoded budget either flakes on a shared runner or is too loose to fail.

The deadlock bound is a hard timeout: a wedged lock does not fail, it hangs, and an unbounded hang reads as a broken CI runner rather than as a bug.

Crash recovery

An agent is killed. Nothing is cleaned up, because in a real crash nothing runs.

$ ddflow recover
1 recoverable situation(s); 1 may contain work:

!! P1.T1  [expired_lease]  was: delta
     worktree /repo/../.ddflow-worktrees/P1.T1
     INSPECT FIRST — 1 uncommitted file(s), 1 unmerged commit(s).
     `git -C .../P1.T1 diff main` then salvage,
     then `ddflow release P1.T1 --note salvaged`.

Four behaviours, each chosen against a specific way this goes wrong:

• While the lease is live, nothing happens. A dead agent is indistinguishable from a slow one until the lease expires, and guessing is how two agents end up in one tree.
• Recovery measures the tree — uncommitted files, unmerged commits — rather than trusting the recorded state. "Is there work in here?" is the only question that decides the remedy.
• An expired lease is never stolen silently, and recover --apply expires only trees it measured as empty.
• Adoption, not duplication: an agent resuming a recovered item gets the existing worktree back, not a second one beside it.

Reconstruction from logs alone

$ ddflow replay --out ./recovery-kit
wrote:
  recovery-kit/RECONSTRUCTION.md
  recovery-kit/QUEUE.md
  recovery-kit/LESSONS.md

RECONSTRUCTION.md is written as instructions to a fresh agent, not as a report about the past: every operator prompt in order, every research verdict, every lesson, the queue's shape — and every approach already tried and rejected, with the measurement that killed it.

It states its own limit, in the document: it reproduces the decisions, not the bytes. Model outputs are not deterministic, so replaying prompts will not recreate the original source. What it recreates is every input that produced it, which no other artefact holds.

The demo destroys an entire repository and rebuilds from 3.9 KB of JSONL, then checks nine specific fragments are present — including the operator's stated reason for a constraint, and the probe output behind a rejected design.

Secrets are redacted on the way in, not on the way out — the log is committed, so a scrub at read time is a scrub that git show walks straight past.

Lessons that check themselves

A lesson can name the mistake in code, not just in prose:

$ ddflow lesson add --title "Never swallow a bare OSError" \
      --rule "Catch the specific error; a broad except turns a loud failure into a silent one" \
      --pattern "except OSError" --globs "*.py"
lesson La0c745cc recorded — inventory: 2 site(s) now

$ ddflow lesson verify          # later, after somebody adds a third
La0c745cc: 1 NEW site(s): c.py: except OSError:

Exit 1 names the file. That is the whole design, and it comes from a failure worth repeating: on the project ddflow was extracted from, a count-based clone ratchet sat red for ~350 commits. It was advisory so it never blocked, it reported a number so every reader learned to skip it, and 24 new clones arrived through that gap.

A count says "worse" and never "which".

A number cannot be acted on or reviewed. A list can: a new entry is a line somebody opens, and a disappeared entry is progress — reported, and never a failure, because the inventory may only shrink.

Three details that decide whether a ratchet survives contact with a real repository:

• A site is <path>: <matched text>, not path:line. Line numbers churn on every edit above a site, which would invent a matching pair of "new site" and "fixed site" findings out of an unrelated change — and a ratchet that cries wolf is one that gets switched off.
• An uncompilable pattern is refused, not stored. An empty inventory reads exactly like a clean repository, and would ratchet every real occurrence away the first time it ran.
• Exit 2 when no lesson declares a pattern. Not a pass. A corpus with zero ratchets should not be able to report "all clear".

Vendored and untracked files are never sites — matches in code nobody owns are findings nobody will act on. Most lessons stay prose, and a prose lesson produces no findings at all.

Checking that the checks are working

Three questions ddflow asks about itself, all derived from the log and all reported by ddflow doctor. They exist because an unmeasured mechanism is indistinguishable from a missing one.

Can the queue's work actually be picked up? Every other check counts the items that are present; this one asks whether any of them can be started. An open phase with no task under it is work ddflow next will never offer — a note by default (schedule.empty_phase, since a project that files phases before breaking them down lives there on purpose) — and a phase whose tasks are all finished while the phase stays open is always a problem, because that is a queue held open by an item nobody can act on.

Tasks cannot go missing here, and that is a property rather than an untested gap: two hypotheses about how one could were probed and both refuted, and a test now pins the invariant so a future filter cannot quietly reintroduce it.

Does a gate ever say yes? A gate that fails on everything is worse than no gate: it trains the next reader to skip it. A gate at or above gates.rate_max_fail once it has gates.rate_min_runs decisive runs is reported as flaky or as measuring a moving target — re-running it will not converge. A skipped gate is not a run, because counting skips as failures would make an unconfigured gate look like a broken one.

Did the periodic passes ever fire? The mechanism you did not measure is the one that is not running. Because a cadence here counts completions rather than wall-clock, this is exact rather than estimated: since is the completions elapsed since the pass last fired, which is the same quantity ddflow cadence uses to decide due-ness — deliberately, because two measures of "is this behind" that can disagree is a situation nobody can reason about. A pass more than cadence.max_missed scheduled runs behind is reported. Being merely due is not a finding (ddflow cadence already says that), and running early is not one either.

All three are notes, not problems: a defect in the machinery that checks the work must not block the work.

Reading the log, and why it is never compacted

The log only grows, so every state-reading call used to re-read and re-parse all of it. Measured at 20,000 events, that read costs 115 ms — and the breakdown is the whole design argument:

Parsing dominates, and an append-only file guarantees the bytes already parsed have not changed. So EventLog.read_all re-parses only the appended tail, and re-hashes the bytes it is re-using to prove they are still the same bytes:

A command like ddflow doctor — which reads four times — pays the full cost once instead of four times.

Two knobs, [log]:

The validity check is a content check, and that is the whole design. The consumed prefix is re-hashed on every read — 1.1 ms to read plus 4.3 ms to digest, against the 97 ms of parsing it avoids. The first version used st_ino instead, on the reasoning that "a git merge writes a temp file and renames, so the inode changes". That is false:

$ git checkout -q other && stat -c %i .ddflow/events/a1.jsonl
218500670
$ git checkout -q main  && stat -c %i .ddflow/events/a1.jsonl
218500670

Git rewrites tracked files in place. So switching between two branches that had diverged left a warm server serving events from the branch you left, silently losing the ones actually on disk, with the tail read starting mid-line — and because Store.rebuild takes its fingerprint from the real file while taking its events from the cache, that wrong state was written into the SQLite index stamped as current, which a fresh process would not rebuild away. A content digest makes a rewrite, a truncation, a git checkout, a git merge, a delete-and-recreate and a torn tail all one case, so there is no list of mechanisms to keep current.

A torn final line from an append that died mid-write is reported by ddflow doctor and re-read until the writer completes it, never marked consumed. Every guarantee here is mutation-verified in tests/test_log_read_cache.py — including that the digest covers the whole prefix rather than a trailing window of it, which a smaller fixture cannot tell apart.

The compaction that was declined

An event kind log.compacted was reserved for a retention pass that shrank the log. It has been removed, because the recipe it was reserved for cannot be implemented without breaking two shipped commands. Three probes:

1. A compaction survives a merge=union merge. One branch compacts, the other appends; the deletions stick. So union is not the obstacle.
2. Two divergent compactions merge to neither side's result, and out of Lamport order — the case union cannot resolve.
3. The decisive one. ddflow progress and ddflow loops read raw events, not folded state: progress.work pairs each lease.acquired with the next release across the whole history. Leases and gate outcomes are not PROVENANCE_KINDS, so keeping "the last state-bearing event per subject" leaves a lease.released with no acquire to pair with. A queue whose loop detector fires repeat_claims before compaction reports nothing after it, and six attempts become zero — and under [loops] on_detect = "block" that is a behaviour change, not just a lost report.

Growth is addressed by making the read cheap rather than the log short, which keeps it append-only and auditable. Beyond ~100k events the right answer is an on-disk state snapshot, not a shorter history.

Lessons, research and bugs

ddflow lesson add --title "Truncating a slug can leave a trailing separator" \
                   --rule "Strip separators AFTER slicing to length, not before."
ddflow lesson search "cutting a url short leaves a dangling hyphen"

Retrieval is BM25 over FTS5 and finds that entry despite no shared keyword. Probed against embeddings and found sufficient at lesson-corpus scale (R4); lessons.search_backend exists for when that stops being true.

Research entries must carry a verdict, and CONFIRMED/REFUTED are refused without a probe:

$ ddflow research --question "is it fast?" --verdict CONFIRMED
CONFIRMED requires a --probe (and ideally --probe-output): a verdict with no probe behind
it is an opinion. Use THEORETICAL and say why no probe was possible.

And a bug cannot be closed without the test that would catch it again:

$ ddflow bug fixed B1
a bug may not be closed without --regression-test naming the test that would catch it
again. Write the test, watch it FAIL against the unfixed code, then close.

That refusal is the whole mechanism by which the same bug does not ship twice.

Cadences

Periodic whole-repo passes a per-task gate structurally cannot do. Due-ness is derived from completed work, so there is no state file to drift:

$ ddflow cadence
DUE: integration_tests — 5 tasks since last (every 5)
DUE: mutation_tests — 3 phases since last (every 3)

Record one with: ddflow cadence --ran <name>

Configurable: integration tests, architecture review, mutation testing, duplication sweep, lessons compression.

Keeping AGENTS.md true

ddflow adopt writes a managed block into AGENTS.md (and CLAUDE.md, and each agent's native rules file). That block is what tells an agent it must claim an item before editing — and every coordination guarantee here rests on that, because an agent that does not claim has its work destroyed by a parallel one.

Nothing used to check it again. Adoption is judged by .ddflow/config.toml existing, so a deleted AGENTS.md, a block someone stripped, or a block written by an older ddflow all left the agent reading rules that were absent or wrong while every surface reported the project as adopted. Adoption is a config file; the instructions are a separate fact.

Cursor does not really follow AGENTS.md, and it is not alone. Its precedence is Team Rules > Project Rules > User Rules > .cursorrules > AGENTS.md, so .cursor/rules/ddflow.mdc is what actually binds — which is why adopt writes it. That file is checked too, for every agent the project was adopted for (read from the driver deltas on disk, so a Claude-only project is never asked for a Cursor rule).

It carries the same block with binding frontmatter, and alwaysApply: true is part of what is verified: a rule with alwaysApply: false exists, reads perfectly, and may never be loaded — which for claim-before-you-edit is the same as not having it, and strictly worse than drifted text. It is reported at the severity of missing, not of stale.

Five states are detected — current, stale (drifted from what this version writes), no_block (file there, block gone), not_binding (native rule that will not apply), missing — and reported on three surfaces:

The two surfaces repair it differently, on purpose.

• From a shell, the operator is right there: ddflow adopt rewrites the block. It replaces only what is between the DDFLOW:BEGIN/DDFLOW:END markers and leaves the rest of your file alone, and re-running it is a no-op. ddflow init reports the problem and does not write — writing prose into your AGENTS.md is not what init was asked to do.
• Over MCP, ddflow does not touch it. The handshake tells the agent to show the operator what is wrong and call ddflow_setup only if they agree. It is a file in their repository, usually with their own prose around the block, and rewriting it is not a decision a tool gets to make on their behalf — the same rule as companions ("propose; never install") and the human-approval gate.

Surviving a compaction

The instruction block reaches the model once, at connect. After a context compaction it may retain none of it, and MCP has no server-to-client primitive for injecting context — the three that exist (roots/list, sampling/createMessage, elicitation/create) all go the other way or ask a question. Three things already survive:

• the AGENTS.md / CLAUDE.md sections ddflow setup writes, plus each agent's native rules file — the client re-reads its own rules, so this is the durable channel;
• the commit hook, which refuses a commit with no item trailer and says what to add. Enforcement at the moment of the act needs no context at all;
• ddflow help <topic>, which the agent can ask for — if it thinks to.

What none of those do is speak up unprompted. A footer on tool results is the only channel that is guaranteed to be heard again, because an agent driving ddflow calls tools continuously:

ddflow: left undone in this project —
  · 1 bug(s) still open: B1 — close with `ddflow_bug_fixed` (it requires the regression test) or say why not
  · 2 gate(s) skipped, not run: T4.critic, T4.standards — run them, or leave the skip on the record deliberately

It is not a banner, and the difference is the whole design. A fixed reminder appended to 63 tools is trained out inside a session and costs tokens on every call. This one:

• names what happened, never restates a rule — an id, a count, and the call that discharges it;
• stops once the thing is dealt with, so it cannot be trained out by repetition;
• says nothing at all when the project has nothing outstanding — not a cheerful "all clear", which is the same thing readers learn to skip;
• is cadenced: at most once every every_calls calls and every_seconds seconds, so a burst of calls is not a burst of footers;
• cannot break the call it rides on. It is a courtesy on top of an answer, appended after the body, and a failure inside it is swallowed. content[0] is still the structured result.

[reinstruct]
enabled      = true   # false silences it entirely
every_calls  = 12
every_seconds = 240
max_items    = 3

What it currently notices: bugs found and never closed, gates skipped and never revisited, and work finishing with no lesson ever recorded (after [lessons] reflect_after_items, so one task is not reported — the pattern is, and the threshold is a knob because where the line sits is a judgement).

Keeping session-start cost flat

ddflow brief --phase P2

Returns, inside session.brief_max_tokens (default 1200): recoverable work first, then the current item and its remaining gates, then what is ready, then why everything else is blocked, then the handful of past lessons ranked against this task's text.

This replaces reading the project's rule and lesson corpora. The budget is enforced by truncating from the bottom, so the safety-critical head survives a squeeze — and a project's opening cost stays roughly constant as its lesson corpus grows.

Agent portability

One canonical driver, templates/drivers/implement-phase.md, plus a delta per agent covering only what genuinely differs: how iteration continues, how to ask the operator, how to spawn a subagent, file-reference syntax.

Deltas rather than copies, for a measured reason: on the project this was extracted from, a reworded per-agent duplicate of the driver silently accumulated three instructions that were false at the time of writing while missing four gates the canonical file had gained. A delta removes the surface that can drift instead of policing it.

Both surfaces are one implementation — the MCP server maps each tool onto the same cli.main() call in-process, and two tests plus a demo step assert they cannot diverge.

Keeping the two surfaces honest

Every CLI command is reachable over MCP — that is the point of the tool list, and it is the requirement that an operator in a chat window, possibly driving a remote agent, can do everything a shell can. Three ratchets keep it true, and each one was added after the previous one turned out to be too shallow:

The flag ratchet derives its own input from the parser rather than a hand-written list — its first version carried eleven tools and was blind to remove --force for exactly that reason. Omissions are allowed, but each must be an entry in FLAG_EXEMPTIONS with its reason, so "we chose not to expose this" and "nobody noticed" stop looking alike.

A fourth pins something subtler: whether a tool returns JSON or prose is a decision, not an accident. Some tools deliberately return prose — brief, gate status and replay exist to hand the model an instruction or a narrative, and JSON-encoding a paragraph so the client can decode it again helps nobody. But decision add returned JSON while task add returned prose for no reason either could state. Each prose tool now carries its justification in PROSE_TOOLS.

Command reference

ddflow adopt [--agents ...]     install into a project, for one or more agents
ddflow init                     create .ddflow/ only

ddflow phase add <id> [...]     add a phase
ddflow task add <id> --phase .. add a task
ddflow update <id> [...]        change title/body/needs/globs/tags/priority

ddflow next [--phase P]         what may start now       (2 = nothing actionable)
ddflow claim <id> [--globs ..]  lease + create worktree  (3 = refused)
ddflow heartbeat <id>           renew a lease
ddflow release <id>             give it up

ddflow gate status <id>         pipeline position + the next gate's instruction
ddflow gate run <id> <gate>     execute a command gate, record its evidence
ddflow gate record <id> <gate>  record an agent gate    (--outcome, --reason, --model)
ddflow gate skip <id> <gate>    skip, with a mandatory reason
ddflow approve <id> <gate>      a PERSON clears a human gate  (no MCP equivalent)
ddflow approve .. --reject      ...or refuses it, with --reason
ddflow gate verify <id> <gate>  prove the gate CAN fail  (1 = it cannot)

ddflow merge <id>               merge from the primary checkout, no checkout
                                ([flow].integration=pr: push + open/update a PR instead)
ddflow pr sync [--item]         what reviewers did: complete / reopen / park / merge (2 = forge unreachable)
ddflow pr status               every item's request, from the log (no forge call)
ddflow version show            current and next version, why, release notes (2 = nothing new)
ddflow version cut [--push]    tag it (gitflow: via release/X, or a release PR)
ddflow version show|cut --line L    the same, for a maintenance line (keeps its major)
ddflow task add <id> --lines 1,2,3  a fix for several release lines: ports generated
ddflow promote add <env>        file a promotion one step downstream (2 = nothing to carry)
ddflow promote status           each environment: head, behind upstream, open promotion
ddflow flow show                how this project works: model, lines, every choice + who made it
ddflow flow choose <knob> <v>   record a workflow choice, with --reason
ddflow complete <id>            finish        (3 = unmet conditions, all listed)
ddflow block <id> --reason ..   mark blocked

ddflow brief [--item|--phase]   budgeted session-start pack
ddflow board / show <id>        human views
ddflow render                   regenerate docs/ddflow/*.md

ddflow help [topic]             what this is, what it can do, the workflow
ddflow workflow                 the rules this project runs by  (1 = incoherent)
ddflow workflow pipeline ...    set the gates a task or phase passes
ddflow workflow gate ...        define or change one gate
ddflow workflow drop <id>       take a gate out of the pipelines
ddflow import [--apply]         propose an existing project's work  (2 = nothing)
ddflow import --verify          is the import still true, and did anyone finish it?
ddflow history [--item|--kind]  one timeline of everything that happened (2 = nothing)

ddflow lesson add|search        capture and retrieve lessons
ddflow research --verdict ..    record a finding (probe required for CONFIRMED/REFUTED)
ddflow bug found|fixed          regression test required to close

ddflow session start|prompt|note|end     provenance logging
ddflow replay [--out DIR] [--verify]     reconstruct from the log

ddflow recover [--apply]        find crashed agents' work   (2 = nothing)
ddflow doctor                   integrity + health
ddflow rebuild                  re-derive the index
ddflow cadence [--ran NAME]     which periodic passes are due  (2 = none)
ddflow config --explain         every knob, its value, its source and its docs
ddflow config --append-toml ..  add config without a shell editor (validated first)
ddflow reviewers detect|list|test   find and check cross-family review endpoints
ddflow review <id> --gate ..    run the configured reviewer, record the evidence
ddflow mcp                      run the MCP stdio server

Every one of these is reachable over MCP, and a test enforces it. One tool goes the other way and has no CLI equivalent, because it has nothing to mean there:

ddflow_identify(agent=...)      declare who you are ON THIS CONNECTION (MCP only)

A CLI invocation is one process that exits, so it says who it is with --agent and the question does not outlive the command. An MCP connection is a session, so identity is declared once and persists — see Who is calling?.

Configuration

120 knobs across 17 sections, every one documented in place:

$ ddflow config --explain --filter lease
lease.ttl_s = 1800   [default]
    Seconds a lease stays valid without a heartbeat. After this it is EXPIRED and
    reclaimable. Longer = fewer false expiries when an agent is deep in a slow gate;
    shorter = faster recovery after a crash.

Resolution: dataclass defaults → .ddflow/config.toml → DDFLOW_<SECTION>_<KNOB> env. An unknown knob is an error, never a silent drop. A test asserts every knob carries documentation, so the reference cannot rot.

What is automated, and what is not

The honest split, because a tool that claims to automate judgement is lying about the part that matters.

Automated — happens without anyone remembering it:

• The handshake briefs the agent. On connect, the MCP server injects the live state: the pipeline every task must pass, work recoverable after a crash, what is ready, which companions are missing, and what to do about each. It is a template (ddflow prompts eject mcp_instructions), so the workflow is text you edit, not code you fork.
• Gates are enforced, not suggested. ddflow complete refuses on a required gate that has not passed, on open sub-tasks, on a silent gate under require_outcome, on a requirement no pipeline runs, and on a reviewer from the author's own family. Refusals list every unmet condition, not the first — an agent that cannot see how many more are coming reaches for --force.
• Conflicts are refused at claim time, by glob overlap, with an alternative named.
• Dependencies gate readiness. ddflow next withholds a task whose needs are open and says which.
• Bugs are offered before features. A task tagged bug/fix/hotfix (the [flow] bugfix and hotfix tags), or named by an open bug record, comes ahead of every feature in ddflow next and gets a free slot first, so a standing bug is fixed before more work is built on it. Priority orders each group; [schedule] bugs_first = false orders by priority alone.
• Gate evidence records which tree and how much — a working-tree fingerprint plus files/lines changed — so a pass names what it passed on. If the tree moves afterwards, complete warns that the evidence describes source nobody is shipping.
• Crash recovery: ddflow recover finds worktrees whose lease expired, so an interrupted agent's work is found rather than lost.
• Cadences (ddflow cadence) tell you which periodic passes are due — bug hunts, dedupe, lesson compression — from the log, counted in completed work; and, for a rule like "a bug hunt every week", by the calendar ([cadence] every_days = ["bug_hunt=7"]: never run means due now). The SessionStart hook lists what is due.
• Post-merge review: ddflow review <item> --commit <sha> reviews one landed commit against its first parent — a merge as what it brought in — when the branch is gone.
• A commit hook (ddflow hooks install) can refuse an unclaimed edit outright, and refuses a staged ddflow render view that the log no longer regenerates byte-for-byte — hand-edited, or stale ([enforce] generated_views). It also reports a doc line still naming an identifier, file or default the commit removes or renames ([enforce] stale_docs, doc_globs, doc_exclude; warns by default). Its commit-msg sibling requires an Item: trailer when [enforce] require_item_trailer is on — or the project's own keys (item_trailer_keys = ["Phase", "Phase-ships"]); merges are exempt.
• A Claude Code SessionStart hook (ddflow hooks install --claude) puts the brief — crashed work to recover, ready items, binding decisions, operational memory — into every session, including after a context compaction, whether or not the agent remembers to ask. In a worktree behind its base branch it says so, and names the rulebooks that changed there. It is added beside the project's own hooks in .claude/settings.json, removed alone, and always exits 0.

Not automated, on purpose:

• Installing anything. Detection is read-only; the report is advice.
• The judgement inside an agent gate. ddflow records that you claim to have hunted bugs; it cannot check that you did. What it can do — and does — is make silence visible: a gate never run and never skipped blocks completion, so the failure mode is a refusal rather than a quiet omission.
• Deciding whether a plan is right. That is what a human = true gate is for.
• Pushing, releasing, or anything outward-facing.

The design assumption is that an agent's honesty cannot be verified, so the system is built to make an unverifiable claim expensive to make and easy to see: evidence contracts, mutation-verified gates, coverage gaps recorded on completion, and an exit code that distinguishes "could not" from "did not need to".

Testing

uv run pytest tests/ -q -n auto      # unit/integration tests, in parallel (pytest-xdist)
python3 demos/run_all.py             # 6 end-to-end scenarios, 219 assertions

The demos invent whole projects and drive them for real — real git worktrees, real pytest and npm test runs, real merges, real concurrent processes:

The scenarios and the stress test have found most of the bugs this project fixed; the unit tests found few of them. The full-lifecycle scenario was written to exercise the requirements rather than the code, and found four defects before it passed once — every one of them a CLI/MCP divergence that command-level parity could not see. The composed MCP run alone found eight that 213 unit tests and four other scenarios missed — including two that made core features useless out of the box. They all lived in seams: between two processes, between a read and a write, between two output surfaces, between a declared vocabulary and its callers, including one that does not reproduce below ~6 concurrent processes. They are catalogued with their regression tests in R6.

Documentation index