Files
archie/ARCHITECTURE.md
T
sam 4750ac097b feat(workflow): post line-anchored review comments with suggested fixes (#891)
* feat(workflow): post line-anchored review comments with suggested fixes

Bead archie-core-q9au (docs/prds/inline-review.md). A review's findings reached
the operator only as a flat list in the PR body, so the author had to map every
finding back to its line by hand and apply mechanical fixes by hand. Now the
line-anchored findings are also posted as inline review comments, carrying a
one-click suggestion when the reviewer could state the fix.

Domain (internal/domain/workflow):
- ReviewFinding gains Suggestion; Validate rejects one on anything but a
  confirmed finding, so the agent's record_finding call is refused with
  something it can act on. Validate splits into validateAnchor /
  validateClassification / validateDisposition, which also brings it back under
  the cyclop limit the gate enforces.
- StagePostReviewComments runs after StageOpenPR in the implement workflow. It
  posts only line-anchored findings (Line > 0); a whole-file finding has no line
  to attach to and stays in the body list, which remains the complete record.
  It is best-effort: the PR is open and the body already lists every finding, so
  a failed comment is logged, never a park -- the same reasoning as OpenPR's
  existing best-effort LinkBranch call.
- The suggestion fence is gated on Verdict == confirmed where the fence is
  written, because it is a one-click apply: a wrong suggestion breaks the code
  and the author has to notice.

Forge (internal/forge): ReviewCommentWriter + GitHub/Gitea implementations,
type-asserted like PullRequestReader so a forge that cannot is skipped rather
than faked (NoopForge is pinned as not satisfying it). GitHub loops one
CreateComment per comment and joins the failures, so one refused anchor cannot
drop the rest; Gitea carries them on a single COMMENT-state review.

Two deliberate departures from the PRD's sketch, both recorded in it:
- The interface is batch (CreateReviewComments), not one call per comment:
  Gitea's only inline shape is one submitted review, and the set then crosses
  NATS once instead of once per finding.
- The head SHA is resolved inside each forge implementation rather than by the
  caller. The stage runs in archie-agent, which holds no forge credentials and
  has no PullRequestReader; the daemon-side forgerpc handler is the only place
  that does.

Transport (internal/forgerpc): workflow.Forger gains the method, so the agent
reaches the daemon over the new create_review_comments subject; the server
type-asserts the writer and answers an incapable forge with an error the stage
logs. This is the one proxied method whose absence on an image skew does not
park a task.

The reviewer prompt and record_finding schema ask for a suggestion on a
confirmed, mechanically fixable defect only.

Regenerating the wfextract symbol table -- the documented way a new exported
workflow symbol reaches interpreted stage plugins -- also picked up symbols that
were missing from the committed table (the Status* and Source* constants,
ReviewCheck, Store, Task, and the _Store wrapper). Nil-guards were re-applied to
every wrapper afterwards, as that file's header requires.

Evidence: task check exit 0 (incl. 229 UI tests). Three mutations of the new
stage logic each fail a test that names the behaviour -- dropping the
confirmed-only gate, dropping the Line > 0 filter, and letting a posting failure
escape -- so the tests discriminate rather than passing vacuously.

* fix(build): drop the Dockerfile pin assertion

The guard fetched golangci-lint's go.mod at image build time to fail the build
when GOLANGCI_LINT_VERSION and GOFUMPT_VERSION disagreed. It only ever tripped
on self-inflicted pin drift, but it added a network fetch and a new failure
mode to the release build, which is not where that belongs.

The pins stay correct (golangci-lint 2.13.2, whose vendored gofumpt v0.11.0
equals GOFUMPT_VERSION) and the comment above them still states the invariant,
so the reasoning survives without the build-time machinery. The pin change
itself reaches the agent image through the normal release path: CI rebuilds
and pushes archie-agent on an archie/vX.Y.Z release, and a gateway-only
release deliberately does not.

* feat(workflow): add the remediate workflow for PR review remediation

Implements step 5 of docs/prds/pr-review-remediation.md: a remediate
workflow that resumes the task's existing worktree onto its PR branch
(worktree.Manager.Resume, not Prepare's refresh), runs a builder agent
against the review's actionable comments, commits and pushes to the
same branch, and replies via the forge's ReplyToReview. Task.RetryCount
doubles as the remediation round counter, capped by the repo's
max_retries; the round-cap stage parks the task and posts one comment
when it's hit.

Task gains a ReviewPayload field (JSON-encoded review unit) with full
store/proto/staterpc plumbing. workflow.Forger gains Comment and
ReplyToReview, proxied to archie-agent via a new forgerpc RPC. Resume
and its Dir lookup are added as an unexported optional Trees capability
rather than widening the exported interface, which is already at
golangci-lint's interfacebloat cap and is projected into wfextract's
Yaegi symbol table for repository-authored stage plugins.

The daemon-side reaction consumer that dispatches into this workflow
(subscribing ARCHIE_REACTIONS, grouping comments by review, enforcing
per-task serialization) is tracked separately in archie-core-8li9.4.

* feat(messaging): extract standalone messaging service binary and process composition

* feat(messaging): adapt channel frontends to consume Gateway ChatContract

* feat(messaging): remove in-process channels and Router from daemon bootstrap

Telegram, email and webhook now run only in the Messaging Service, which
reaches this process through the Gateway's ChatContract. The daemon builds
no channel and no gateway.Router, so the wiring that existed only to serve
them goes with it: setupGateways, setupEmailGateway, setupWebhookGateway,
setupTelegramGateway, buildTelegramRouter, makeTelegramReload and the
release-announcement path helper.

The seams those paths owned move to where they can still be honoured:

- Channel ValidateConfig runs in archiemessaging.compose, which now returns
  an error, so a configured-but-invalid front-end fails the service rather
  than being dropped silently.
- The inbound rate limiter moves to the Gateway process, which owns the only
  Router and therefore the only place a per-sender budget applies to every
  channel's turns.
- The daemon's gateway readiness probe becomes a ContractProbe against the
  Gateway. It cannot observe a channel lifecycle it no longer runs, and a
  channel-count probe would have reported every daemon degraded.
- telegramSetup is now chatSetup (in chat_setup.go), trimmed to the fields a
  chat turn runner actually reads.

Telegram operator features that were wired only in the daemon (/update,
release announcements, version reporting, SIGHUP reload) are not yet wired
in the Messaging Service: archie-core-8cda.6.6. Channel lifecycle reporting:
archie-core-8cda.6.8. The Gateway process builds no worktree manager, so
review_pr is unadvertised on every channel: archie-core-8cda.6.7.

* feat(messaging): re-wire the Telegram operator surface in the Messaging Service

setupTelegramGateway satisfied eight telegram.Gateway seams from the daemon's
own process. Each is resolved by where the fact it needs is owned, not by
where it used to be built; the decision table is in
docs/architecture/migration-decisions.md.

Wired here: /update (built from this service's own [chat.telegram] check and
install commands), the update-report relay (same sha256(bot_user)[:8] path the
daemon wrote, so a pending report is found rather than written afresh),
show_tool_calls, and /restart's config reload, which re-resolves the token and
allowlist from this service's own file and refuses a reload that resolves to
an empty token.

/version is sourced from the Gateway: ChatSnapshot gains a Version field
carrying the Gateway's rendered build block, read per call and bounded by the
dependency timeout. cmd/archie-messaging deliberately takes no version
ldflags, so it cannot report a partial upgrade as a matched one.

RunningVersions and ReleaseAnnouncements stay nil. Both need a component's
real running version, and the only versions this process can obtain are a
different binary's stamps -- reporting them would manufacture exactly the
false success the update check exists to catch, and the announcer without
parseable versions is a no-op that reads as configured. archie-core-8cda.6.9
carries the contract that unblocks them. Dangerous stays nil: no composition
has ever supplied a command authority.

Enrich is dropped from the update service: componentInstallTypeEnricher read
[nats], which this service must not decode, so the check command's own
reported install type stands.

* fix(workflow): guard inline review comments against head drift

Two findings from the review of q9au.

A comment's line numbers describe exactly one revision of the branch, but the
forge resolved the pull request's head at posting time -- after the review had
already read the worktree. A push in that window (a collaborator, a retry) meant
the findings were anchored to whatever now occupied those lines, mislabelling
instead of failing. The revision the review measured now travels with the
comment set: OpenPR records it from the same worktree read it already performs
(TaskContext.ReviewedHeadSHA), StagePostReviewComments threads it through
Forger to forgerpc, and the forge refuses the whole set once the head it reads
has moved off it -- which the best-effort stage logs, the PR body remaining the
record. An unmeasured revision posts unverified rather than dropping every
finding, so a worktree that cannot report one degrades instead of regressing.

Suggestion was documented single-line only, because GitHub applies a suggestion
fence up to its first line and a multi-line replacement would silently corrupt
the file. Only the verdict was validated, so a multi-line suggestion reached the
renderer and posted as that truncated fence. Validate now rejects an embedded
newline, which makes record_finding refuse it with feedback the model can act on
instead of the reviewer discovering the corruption later.
2026-09-19 23:37:23 +10:00

23 KiB

archie-core Architecture

archied is a resident daemon that discovers forge work — GitHub and Gitea issues assigned to its bot user, carrying its label, or either (dispatch mode, per identity) — by polling (the default) or, single-identity deployments only, reacting to forge webhook events the instant they arrive (Forge.Intake = webhook/both, see docs/prds/event-sources-and- reactions.md), works each matched issue in an isolated worktree through a routed workflow, and opens pull requests for human review. A single daemon can serve several identities across several forges, each with its own bot_user, label, and repository set.

Package Map

This lists the load-bearing packages, not all of them — internal/ holds ~36. The domain migration is in progress (internal/domain/, internal/infrastructure/), so the tree is authoritative over this table. See docs/architecture/organisation.md for the target structure and docs/architecture/migration-decisions.md for the current position.

Package Role
cmd/archied/ Binary entry: config→forge→store→runtime→workflows→daemon
cmd/archie-agent/ Binary entry: flags, environment, signals, and agentworker invocation
internal/app/agentworker/ Agent worker composition, lifecycle, routing, and workflow execution
internal/infrastructure/agentboot/ Worker boot-brief filesystem decoding
internal/infrastructure/agentgit/ Worker git safe-directory setup
internal/infrastructure/agenttransport/nats/ Worker core-NATS full-task handoff and RPC client construction
internal/config/ Config types only — loading lives in infrastructure/configuration
internal/daemon/ Resident loop: poll, enqueue, claim, route, process, reconcile
internal/domain/workflow/ Engine: stage lists, routing, shared steps, workflow definitions
internal/domain/workflow/task/ Task-execution vocabulary (Task, statuses, mid-run Store, Definition) importable without the engine
internal/domain/messaging/ Canonical messages plus the chat wire contracts (ChatContract, sessions, turns, streaming) shared by gateway, channels, webui, and archie-ui
internal/channels/status/ Operator-facing channel lifecycle ledger (State/Descriptor/Status/Manager); the plugin contract stays in internal/channels
internal/app/archiemessaging/ Messaging Service composition: Telegram/email/webhook channels, dialling the Gateway's ChatContract over gRPC. The daemon runs no channel and builds no gateway.Router
internal/domain/workintake/ TaskEnvelope, routing Kind, label vocabulary, task subjects
internal/agentexec/ Worker-local stage protocol and ai-sdk loop runner
internal/forge/ Forge interface: GitHub and Gitea implementations
internal/forge/webhook/ Optional forge webhook receiver: GitHub issue events → workintake.TaskEnvelope, same publish path as polling
internal/webhookguard/ Cross-cutting inbound-webhook mechanics: HMAC verification, per-source rate limiting
internal/worktree/ Git operations: clone, branch, commit, push, diff, cleanup
internal/store/ Task/capture/mapping/binding SQLite implementation; producer-owned contract interfaces live in internal/domain/storecontract (aliases kept here)
internal/domain/storecontract/ Producer-owned State Store read contracts consumed by webui, intake, and the gRPC adapters (state-store-contract rev. 2d)
internal/events/ In-process event bus: publish/subscribe for observability
internal/infrastructure/ Config loading, eventbus transport, model catalog
internal/webui/ Shared dashboard HTTP layer served by the archie-ui process
internal/gateway/ (18.7k LOC) Persistent-connection layer between archie and its users (Telegram, web UI, etc.); shared CommandRouter for dispatch
internal/tools/ (20.9k LOC) Central tool subsystem: built-in tool implementations, approval, execution contract
internal/domain/memory/ Memory engine contract: four typed scopes, CRUD + revisions, content scanner
internal/infrastructure/memory/ Builtin memory engine + its builtin/ markdown store
internal/secret/ Secret reference resolution across named engines (env, vault, etc.)
internal/plugin/ Core plugin interface: Yaegi-interpreted .go plugins loaded from ~/.config/archie/plugins/ and registered with the daemon's extension registry
internal/skill/ Parses and discovers agentskills.io SKILL.md files
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, but they no longer comment on the forge issue: the reason is recorded on the task as park_reason (cleared on requeue) and surfaced by the dashboard and /api/tasks. Retries are an explicit operator action capped by max_retries, not a label a human removes.

Adversarial review: the implement workflow runs a fresh adversarial review after the deterministic gates and before pr_open (see docs/architecture/adversarial-review.md). A surviving blocking finding — or a review that failed to run — parks the task with the findings as its park_reason; otherwise the PR opens with a findings section on its body. Opt-in per repo via repo.review_enabled.

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

Provenance by origin: a task's origin determines which enqueue path created it, though every origin converges on the same tasks table and the same Route/Run lifecycle above once queued. Forge-sourced tasks — whether discovered by polling or, single-identity only, delivered by the forge webhook receiver (internal/forge/webhook) — enqueue via the same workintake.TaskEnvelope and PublishTask path (see docs/prds/event-sources- and-reactions.md's decided contract); TaskEnvelope.IdempotencyKey is keyed on owner/repo/number, not delivery source, so a webhook-delivered issue and a later poll of the same issue dedup for free instead of double-enqueuing. Playbook-binding-dispatched tasks enqueue via EnqueueChatTask/EnqueueBindingTask (internal/store), the same direct path chat-spawned tasks use, stamped with binding_id/binding_version for provenance. See docs/architecture/bindings.md for the binding model and its threat model in full.

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. The built-in bindings are:

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

Both the closed bug/feature/bootstrap set's target workflow and arbitrary additional labels are YAML-overridable at daemon startup (internal/domain/workflow/routing.go's LoadKindWorkflowsYAML/ LoadLabelWorkflowsYAML, WorkflowRoutingFile/WorkflowLabelsFile/ PlaybookDirs config) -- see docs/prds/eda-playbook-engine.md. A label already owned by the closed kind set, or a binding declared twice across sources, is a reported load failure, not a silent override.

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 the forge. GitHub (github.go) and Gitea (gitea.go) both implement it, plus a no-op (noop.go) used when no credential resolves — see the degradation rule in docs/architecture/configuration.md. A multi-forge deployment is configured per identity; see deployments/multi-forge-github-gitea.toml.

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. They report state; they do not drive it. Removing a label no longer retries a parked task -- retries are an explicit operator action from the dashboard or chat, capped by max_retries.

Webhook intake (alternative to polling)

internal/forge/webhook is a separate HTTP receiver, not a Forge method: it verifies a GitHub webhook's HMAC signature (internal/webhookguard.VerifyHMAC), decodes an issues event, applies the same dispatch predicate the poller uses, and calls the same PublishTask the poller calls. It is opt-in via Forge.Intake (poll default, webhook, both) and single-identity only -- a deployment with [[identities]] configured keeps polling per identity instead. See docs/prds/event-sources-and-reactions.md for why this stayed a thin decode-and-reuse layer rather than a second intake mechanism.

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/gitea), host, token secret reference; intake (poll default, webhook, or both) with webhook_secret/webhook_addr when reacting to forge events instead of, or alongside, polling (single-identity deployments only)
  • [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
  • [nats] -- embedded or external broker deployment
  • [containers] -- managed archie-agent image, limits, storage, and network
  • [budgets] -- max steps, max tokens, wall clock, gate max failures
  • [web] -- dashboard listen address, bound by the standalone archie-ui process (the daemon runs no dashboard listener after the UI cutover)
  • [notify] -- webhook URL for notifications

Secrets (forge token, provider API keys, channel tokens) are not embedded in config files. internal/secret defines a pluggable Engine interface (Name, Version, Resolve(key) (string, error)) resolved through a Registry; the built-in env engine reads an environment variable by name and the built-in bws engine caches the accessible Bitwarden project at startup. Custom engines (sops, Vault/OpenBao) are Yaegi-interpreted .go plugins loaded from a directory, following the same convention as internal/plugin's daemon plugins. Config fields reference a secret as {engine, key} (secret.SecretRef), e.g. token = { engine = "env", key = "ARCHIE_GITHUB_TOKEN" }. As of this writing Forge.Token and Provider.APIKey use SecretRef; resolved provider values are exported under private, identity-scoped environment names because the SDK and container boundary consume credential names rather than plaintext wire fields. The older Provider.APIKeyEnv, NATSConfig.TokenEnv, and TelegramConfig.TokenEnv field names remain supported.

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

Plugin engine rule (strict)

“Engine” is reserved for a typed, lifecycle-managed capability family. It is not a synonym for a generic plugin hook. Every new plugin engine must satisfy all of these invariants:

  1. Typed domain contract. The engine interface exposes real capability-specific operations in addition to identity metadata. Examples are resolving a secret, delivering a message, or running an agent. A plugin that only exposes Name and Version is metadata, not an engine.
  2. Owning family manager. The capability package owns a Registry or Manager that controls provider registration and discovery, initialization, or access. Providers do not register arbitrary callbacks directly on the daemon.
  3. Lifecycle and health. When an implementation owns resources or background work, its typed family contract must expose start, health, and stop semantics. The owning manager invokes them and defines failure isolation and shutdown ordering. Stateless providers may make these operations explicit no-ops.
  4. Narrow host access. Plugins receive the typed family registrar and only the host services their declared capability requires. They never receive *daemon.Daemon, a service locator, or an untyped hook map.
  5. Metadata-only generic plugin contract. plugin.Plugin remains limited to Name() and Version(). Domain hooks belong on typed engine interfaces; adding capability methods to the generic interface is forbidden.
  6. Trust boundary stays explicit. Explicitly trusted, operator-installed plugins—including third-party secret-engine plugins the operator chooses to trust—may run in-process and have daemon privileges. Repository-supplied code runs in the task container. Integrations that are not trusted as daemon code run out of process through MCP or a versioned protocol and, when a confidentiality or integrity boundary is required, inside a real container sandbox. A subprocess boundary alone is not a security sandbox, and calling Yaegi-loaded code a plugin does not isolate it.

internal/plugin/architecture_test.go enforces the mechanically testable parts of this rule: the generic plugin method set, the shared agent-instruction source, and the requirement that engine interfaces have domain behavior plus an owning registry or manager.

Memory engine family

One engine, addressed by four typed scopes -- ScopeGlobal, ScopeAgent, ScopeUser, ScopeAgentUser (internal/domain/memory, Scope/ScopeKind). A caller names exactly the scopes it may read or write; the engine applies no access policy of its own, so isolation falls out mechanically because ScopeAgent{a} and ScopeAgent{b} are different storage keys. See docs/prds/memory-engine-unification.md for the decision this contract implements; it superseded internal/memory (MemoryProvider/Manager), deleted in that PRD's slice 5.

MemoryEngine (internal/domain/memory/contract.go) is the typed family contract: Create, Get, Query, List, Update, Forget, Revisions -- CRUD with revisions retained, not just create/read/delete. Update supersedes rather than overwrites: the superseded state is appended to the scope's HISTORY.md before the live block is replaced, and Update carries no Scope field, so a record can never change scope. The family's content scanner (prompt-injection / sensitive-data patterns) is applied inside Create and Update, the single choke point every producer crosses.

internal/domain/memory.Registry is the owning family registry: registration, start/health/stop, shutdown ordering, and failure isolation, following the plugin engine rule (ARCHITECTURE.md#plugin-engine-rule-strict). internal/infrastructure/memory.NewBuiltinEngine is the only registered engine today; it persists through internal/infrastructure/memory/builtin, a markdown store (MEMORY.md-shaped per-scope files) that owns its on-disk format end-to-end.

Two per-turn seams in internal/gateway connect a chat turn to the engine, each narrowed to exactly the operations it needs:

  • Read (turn_memory.go, MemoryStore): prepareTurn runs a scope-only Query over the resolved Subject's readable scopes, synchronously and before the prompt is built. It never returns an error -- a panicking or failing engine degrades to an empty <memory> block rather than failing the turn -- and is bounded by a per-scope record limit and a byte cap on the rendered block.
  • Write (turn_memory_tool.go, MemoryWriteStore): four tools -- memory_create, memory_update, memory_delete, memory_list -- built per turn onto the resolved Subject's writable scopes, not registered once at boot. The model chooses the scope kind; it never chooses the agent or user id, which come from the turn's resolved Subject and have no schema field to override. ScopeGlobal is never model-writable.

boot.setupMemoryEngine (internal/app/archied/bootstrap.go) registers and starts the engine and must run before setupGatewayChat, which constructs the only chat turn runner in production and captures the engine at construction time (internal/app/archied/composition_order_test.go pins the ordering), and before setupCurators, whose Registrar captures the engine registry. The session curator (internal/infrastructure/sessioncurator) writes derived records onto ScopeAgentUser, making it the engine's other live producer besides the chat write path.

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. ai-sdk/agentloop and ai-sdk/toolkit provide the common agent loop and tool set.
  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
  • Forge webhook intake as a feature-flagged alternative/backstop pairing to polling (single-identity deployments; see internal/forge/webhook)
  • Bootstrap, implement, TDD, feasibility workflows
  • GitHub and Gitea forges (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:

  • Multi-identity webhook intake (currently single-identity only)
  • More ecosystems
  • Notification integrations beyond n8n webhook
  • Skills (SKILL.md support -- see .archie/skills/)