Files
archie/ARCHITECTURE.md
T
sam f6d51416f0 chore: replace 607 em-dashes with double hyphens project-wide
Includes MCP crash-after-init test fix (crash moved from handle() to serve()
entry point so transport sees immediate subprocess death).
2026-07-26 09:53:33 +10:00

9.2 KiB

archie-core Architecture

archied is a resident daemon that polls GitHub for issues labelled archie, works each one in an isolated worktree through a routed workflow, and opens pull requests for human review.

Package Map

Package Role
cmd/archied/ Binary entry: config→forge→store→runtime→workflows→daemon
cmd/archie-agent/ One-invocation autonomous stage worker (JSON stdin/stdout)
internal/config/ TOML config, per-repo gates, budgets, providers, ecosystems
internal/daemon/ Resident loop: poll, enqueue, claim, route, process, reconcile
internal/workflow/ Engine: stage lists, routing, shared steps, workflow definitions
internal/agentexec/ Versioned agent-stage protocol, worker, and in-process/subprocess runners
internal/forge/ GitHub interface: issues, labels, PRs, comments, reactions
internal/worktree/ Git operations: clone, branch, commit, push, diff, cleanup
internal/store/ SQLite: task rows, status transitions, crash recovery
internal/events/ In-process event bus: publish/subscribe for observability
internal/webui/ SSE dashboard: live event stream, task status
ai-sdk/runtime External: provider-agnostic LLM runtime
ai-sdk/agentloop External: gated agent loop with budgets and tool jails
ai-sdk/core External: tool definitions, generate options

Task Lifecycle

queued → running(workflow:stage) → pr_open → merged | rejected
                 ↓
           waiting_human → (human approves) → queued(implement)
                 ↓
            parked

Crash recovery: tasks left in running are re-queued on startup. Parks are never silent -- every park posts a comment with the reason and gate output.

PR reconciliation: the daemon polls open PRs, checks GitHub state (merged/closed), transitions the task, and cleans up worktrees.

Workflow Engine

Core types (workflow.go)

  • Workflow = named list of Stages
  • Stage = Name + Run(ctx, *TaskContext) error
  • TaskContext = mutable bag: Task, Repo, Config, Forge, Store, Agent runner, worktree dir/branch, scratch fields
  • Outcome = terminal decision (status + detail) -- ends the workflow immediately
  • Registry = map[string]Workflow

Routing (Route)

Label-driven, no LLM involved. Pre-assigned workflow wins; then labels decide:

  • bug → tdd
  • feature → feasibility
  • bootstrap → bootstrap (diagnostics, no LLM spend)
  • Default → implement

Execution (Run)

Stages execute sequentially. After each stage:

  • Task is persisted to SQLite
  • Error → park the task (comment on issue, keep worktree)
  • Outcome set → finish (transition task, update labels)
  • Neither → continue to next stage

A workflow that ends without an outcome parks -- definition bugs must not vanish silently.

Workflows

Bootstrap (diagnostics, no LLM)

prepare → apply(marker file) → commit-push → diff-cap → open-pr

Proves the plumbing: clone, push, PR. Label any test issue bootstrap to verify end-to-end.

Implement (default)

prepare → baseline-gate → plan(planner, read-only) → build(builder, gated) → commit-push → diff-cap → open-pr
  • baseline-gate: runs the repo's gate at the base commit before anything is touched
  • plan: read-only agent explores codebase, produces implementation plan, posts it as a comment
  • build: full toolset agent implements against the plan behind the repo's quality gate
  • diff-cap: parks oversized changes (>400 lines default)
  • open-pr: builder's summary becomes the PR body

TDD (bug label)

prepare → baseline-gate → analyse(planner) → repro-tests(builder, inverted gate) → capture-proof → commit-repro → fix(builder, normal gate, test files protected) → commit-push → diff-cap → open-pr + repro-evidence
  • repro-tests: the quality gate is inverted -- the test command must FAIL. Proven failure is required.
  • capture-proof: runs the test command deterministically, captures the failing output
  • commit-repro: commits the failing tests as the first commit
  • fix: test files are environmentally write-protected (not prompt-ruled). The agent fixes code, not tests.
  • open-pr + evidence: posts both the PR and the captured failure proof

Feasibility (feature label)

prepare → assess(planner, with decide tool) → [close_if_bad_fit | prd(planner)] → deliver → waiting_human
  • assess: reads AGENT.md, ROADMAP.md, README. Calls a decide tool (structured output, not parsed prose). Bad fit → close with reasons.
  • prd: read-only agent writes a PRD: problem, solution, files affected, acceptance criteria, non-goals, diff estimate
  • deliver: posts PRD as issue comment + webhook notification (n8n → email). Sets state to waiting_human.
  • The daemon watches for human replies, judges them with a triage LLM (APPROVE/REJECT/UNCLEAR), requeues or closes.

Forge Interface

A thin abstraction over GitHub (Gitea planned):

Method Purpose
AcceptInvitations Auto-accept repo invites
AssignedIssues Poll assigned issues (excluding PRs)
IssuesWithLabel Poll labelled issues
Comment Post issue/PR comment, return ID
RepliesAfter Find human replies after a comment ID
CreatePR Open pull request
PRState Check merged/open/closed
CloseIssue Close with final comment
React Emoji reaction (pickup acknowledgement)
SetStateLabel Replace old state label with new one
VerifyPush Confirm push permission at startup

State labels (archie:queued, archie:working, etc.) are created on demand, removed when transitioning.

Worktree Manager

Every task gets a fresh clone -- no persistent state between runs. The model never runs git.

~/.local/share/archie/work/<owner>-<repo>/issue-<N>/

Token auth via a generated GIT_ASKPASS helper script -- the token never appears in .git/config or process argv.

Config

Daemon-level TOML (~/.config/archie/config.toml):

  • [forge] -- type (github), host, token env var
  • [dispatch] -- trigger (assignee/label/either), labels, ack reaction
  • [[repos]] -- owner, name, base branch, gate commands, protected paths, ecosystem
  • [models] -- planner, builder, triage (provider/model refs)
  • [providers] -- LLM provider configs
  • [agent] -- execution mode (inprocess or subprocess), worker command, environment allowlist
  • [budgets] -- max steps, max tokens, wall clock, gate max failures
  • [web] -- dashboard listen address
  • [notify] -- webhook URL for notifications

Subprocess mode is a migration transport boundary, not a security sandbox: the worker still runs under the daemon UID and can access daemon-readable host resources. The "daemon never runs untrusted code" boundary requires the later container runner with a separate user, a restricted filesystem and process namespace, and explicit mounts.

Per-repo gates are command lists: [[repos.gate]] = ["go", "vet", "./..."]. The last command is the test runner (by convention -- TDD inverts it).

Observability

Every lifecycle event flows through an internal event bus:

  • SQLite event log (every event stamped with a row ID)
  • Live dashboard via SSE (server-sent events)
  • Bounded subscriber buffers -- slow consumers drop events, never block the daemon

Key Design Decisions

  1. Environmental constraints over prompt rules. The gate, test-file protection, and diff cap are code enforcement -- not "please don't do that" in the system prompt.
  2. The model never runs git. Worktree operations are deterministic steps, not agent tools.
  3. Tokens as cost, not features. Token usage is tracked per task and emitted in observability events but never budgeted beyond the agent loop's own limits.
  4. Label-driven routing. Workflow selection is a switch statement, not an LLM call. A triage LLM is used only for the human reply judgment in waiting_human.
  5. Shared code with tau at the module level. ai-sdk/agentloop and ai-sdk/toolkit are extracted from tau's coordinator and promoted to shared libraries.
  6. No daemon-internal cron. Polling is a ticker loop, not cron. The feasibility PRD notification uses an external webhook (n8n).
  7. Fresh worktree per task. No persistent state between runs. Crash recovery just re-queues.
  8. Agent execution is a data boundary. Autonomous stages receive a versioned request and return a fenced result; SQLite, forge operations, git, workflow sequencing, and outcome handling stay daemon-owned.

What's Implemented vs Planned

Implemented:

  • Daemon poll loop, enqueue, claim, process
  • Bootstrap, implement, TDD, feasibility workflows
  • GitHub forge (issues, PRs, labels, comments, reactions)
  • Worktree isolation (clone, branch, commit, push, diff, cleanup)
  • SQLite store with transition audit
  • Config with ecosystem support (go, python, node, rust)
  • Agent stages with gates, budgets, preflight, protected paths
  • Observability (event bus, SQLite log, SSE dashboard)
  • Crash recovery (re-queue stale running tasks)
  • Human reply judgment (triage LLM for waiting_human)

Planned:

  • Gitea forge implementation
  • More ecosystems
  • Notification integrations beyond n8n webhook
  • Skills (SKILL.md support -- see .archie/skills/)