feat: add four agent packages + knowing-state-sdk + AtomDirective schema by Wescome · Pull Request #74 · Wescome/function-factory

Wescome · 2026-06-05T17:40:44Z

Summary

Adds the Factory I-layer runtime governance stack as five new packages.

Packages added

@factory/knowing-state-sdk — domain-agnostic Knowing-State Prosthesis SDK. Provisional name; will move to @koales/knowing-state-sdk (DECISIONS N+3)
@factory/mediation-agent — per-repo Knowing-State Prosthesis DO. Append-only event log, Bead graph flush to ArangoDB, Verification-Process hosting. SPEC-MEDIATION-AGENT-DO-001 v2.0
@factory/commissioning-agent — per-repo spec-execution governance Worker. Commission flow, polling loop, Hypothesis formation (Claude Opus), Amendment proposal. SPEC-COMMISSIONING-AGENT-001
@factory/conducting-agent — execution substrate interface. Gas City session dispatch, AtomDirective execution, trace reporting. SPEC-CONDUCTING-AGENT-001
@factory/architect-agent — Factory-wide governance singleton DO. D1 patch governance, D2 CRP resolution, D3 vertical slicing, D4 pipeline config. SPEC-ARCHITECT-AGENT-DO-001

Schema addition

packages/schemas/src/atom-directive.ts — canonical AtomDirective schema closing the shared blocker across all four agent specs (SPEC-CONDUCTING-AGENT-001 §1.2)

Status

All implementations are stubs with // TODO: markers aligned to spec section references. Two shared blockers closed: AtomDirective schema and Divergence severity classification policy (DECISIONS N+2).

Decisions recorded

DECISIONS N+2: Divergence severity classification (blocking / advisory / informational)
DECISIONS N+3: @factory/knowing-state-sdk standalone package placement

Packages added: - @factory/knowing-state-sdk — domain-agnostic Knowing-State Prosthesis SDK Provisional name; will move to @koales/knowing-state-sdk (DECISIONS N+3) - @factory/mediation-agent — per-repo Knowing-State Prosthesis DO Append-only event log, Bead graph flush, Verification-Process hosting SPEC-MEDIATION-AGENT-DO-001 v2.0 - @factory/commissioning-agent — per-repo spec-execution governance Worker Commission flow, polling, Hypothesis formation, Amendment proposal SPEC-COMMISSIONING-AGENT-001 - @factory/conducting-agent — execution substrate interface Gas City session dispatch, AtomDirective execution, trace reporting SPEC-CONDUCTING-AGENT-001 - @factory/architect-agent — Factory-wide governance singleton DO D1 patch governance, D2 CRP resolution, D3 vertical slicing, D4 pipeline config SPEC-ARCHITECT-AGENT-DO-001 Schema addition: - packages/schemas/src/atom-directive.ts — canonical AtomDirective schema Closes shared blocker across all four agent specs SPEC-CONDUCTING-AGENT-001 §1.2 All implementations are stubs with // TODO: markers aligned to spec sections. Two shared blockers closed: AtomDirective schema, Divergence severity policy (DECISIONS N+2).

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 59afd295ac

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-06-05T17:44:12Z

@@ -0,0 +1,21 @@
+{
+  "name": "@factory/architect-agent",


Update the lockfile for the new packages

This commit adds new workspace package manifests and dependencies, but pnpm-lock.yaml has no importers for packages/architect-agent, packages/commissioning-agent, packages/conducting-agent, packages/knowing-state-sdk, or packages/mediation-agent. I checked .github/workflows/ci.yml, and both CI jobs run pnpm install --frozen-lockfile (lines 28 and 66), so the install will fail before typecheck/tests whenever this commit is built.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-06-05T17:44:13Z

+    // Read last 200 events (configurable)
+    const events = await this.readRecentEvents(200)


Replay enough events before the first snapshot

Before the first snapshot is written at the default 500 events, reconstructState() starts from emptyState() but only replays the last 200 events. In any active repo that records more than 200 traces/divergences after commissioning but before event 500, the original CommissionEvent and LifecycleEvent fall out of this window, so /dispatch reconstructs an UNINITIALIZED/empty policy state and blocks or completes instead of continuing execution.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-06-05T17:44:13Z

+  private stateFromSnapshot(s: SnapshotPayload): RepoState {
+    return {
+      ...this.emptyState(),
+      lifecycleState: s.lifecycleState as AgentLifecycleState,
+      workGraphVersion: s.workGraphVersion,
+      openDivergenceIds: s.openDivergenceIds,
+      coherenceVerdict: { value: s.coherenceVerdict as 'favorable' | 'unfavorable' | 'pending' },
+      fidelityVerdict: { value: s.fidelityVerdict as 'favorable' | 'unfavorable' | 'pending' },
+    }


Preserve commissioned state in snapshots

When a snapshot exists, this starts from emptyState() and restores only lifecycle/version/verdict/divergence fields, dropping policyBeadId, atoms, detectorSpecs, agentsMd, lastCommissionAt, and lastFlushSeq from the commissioned state. After the first snapshot, a repo can still look ACTIVE and coherent, but /dispatch sees an empty atoms array and returns complete, so execution stops even though the commission data was previously present in the event log.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-06-05T17:44:13Z

+      if (spec.exitCode !== undefined) {
+        const dr = body.detectorResults.find(r => r.detectorId === spec.detectorId)
+        if (dr?.fired) fired = true


Guard traces that omit detector results

When any detector spec with an exitCode is configured, this dereferences body.detectorResults.find(...), but the Conducting Agent trace it posts does not include detectorResults in either the normal trace object or the invalid-directive trace. In that scenario /trace throws before returning recorded, losing the trace and preventing divergence handling; either make the field optional here or include it in the emitted trace payload.

Useful? React with 👍 / 👎.

chatgpt-codex-connector · 2026-06-05T17:44:13Z

+    const crpQueue = (await this.state.storage.get<{ items: { status: string }[] }>('crp:queue')) ?? { items: [] }
+    const patches = (await this.state.storage.get<PatchRecord[]>('patches:active')) ?? []
+
+    return Response.json({
+      factoryLifecycleState: state.lifecycleState,
+      activeRepoCount: state.activeRepos.length,
+      repoHealthBreakdown: {
+        healthy: state.activeRepos.filter(r => r.healthStatus === 'healthy').length,
+        degraded: state.activeRepos.filter(r => r.healthStatus === 'degraded').length,
+        suspended: state.activeRepos.filter(r => r.healthStatus === 'suspended').length,
+        unknown: state.activeRepos.filter(r => r.healthStatus === 'unknown').length,
+      },
+      openEscalationCount: 0,  // TODO: read from ArangoDB
+      activePatchCount: patches.filter(p => p.status === 'propagating').length,
+      pendingCrpCount: crpQueue.items?.filter(i => i.status === 'pending').length ?? 0,


Count CRPs using the stored queue shape

handleCRP() stores crp:queue as a plain CRPItem[], but the health endpoint reads the same key as { items: [...] }. After any CRP is queued, crpQueue.items is undefined for the stored array, so pendingCrpCount always reports 0 and factory health hides unresolved CRPs.

Useful? React with 👍 / 👎.

…ive.ts Replace manual SuccessConditionType declaration with z.infer-based pattern. The ZodLazy wrapper's _input type was incompatible with the manually declared union because z.default() makes fields optional in input types. Using a base discriminated union + lazy recursive extension fixes the TS2322 error.

…Type<T,_,T> constraint z.default() makes fields optional in Zod input types (number | undefined) but the manual SuccessConditionType declaration requires them to be required. Removing z.default() from variant fields makes input and output types identical, satisfying the recursive ZodType<T, ZodTypeDef, T> constraint. Callers should provide explicit values (e.g. expectedCode: 0) rather than relying on defaults within SuccessCondition.

Repo tsconfig has verbatimModuleSyntax enabled. All type-only re-exports across the five new packages updated to use 'export type' syntax.

- Spread sandboxOutputRef conditionally to satisfy exactOptionalPropertyTypes - Replace lastTrace! non-null assertion with explicit guard

…PropertyTypes)

Wescome · 2026-06-10T21:19:02Z

Superseded by feat/ksp-implementation (#86). The agent stubs were ArangoDB-bound pre-KSP design with no salvageable implementation — the real CoordinatorDO, bead graph, and loop closure are now live. Route surfaces noted for future agent work.

claude added 2 commits June 5, 2026 17:22

chore: update pnpm-lock.yaml for new agent packages

a3be10b

chatgpt-codex-connector Bot reviewed Jun 5, 2026

View reviewed changes

claude added 5 commits June 5, 2026 17:45

fix: use 'export type' for type re-exports (verbatimModuleSyntax)

e769652

Repo tsconfig has verbatimModuleSyntax enabled. All type-only re-exports across the five new packages updated to use 'export type' syntax.

fix: exactOptionalPropertyTypes and null assertion in conducting-agent

274fa60

- Spread sandboxOutputRef conditionally to satisfy exactOptionalPropertyTypes - Replace lastTrace! non-null assertion with explicit guard

fix: conditional reason spread in Verdict construction (exactOptional…

31ef923

…PropertyTypes)

Wescome closed this Jun 10, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat: add four agent packages + knowing-state-sdk + AtomDirective schema#74

feat: add four agent packages + knowing-state-sdk + AtomDirective schema#74
Wescome wants to merge 7 commits into
mainfrom
feat/agent-infrastructure-packages

Wescome commented Jun 5, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

chatgpt-codex-connector Bot Jun 5, 2026

Uh oh!

chatgpt-codex-connector Bot Jun 5, 2026

Uh oh!

chatgpt-codex-connector Bot Jun 5, 2026

Uh oh!

chatgpt-codex-connector Bot Jun 5, 2026

Uh oh!

chatgpt-codex-connector Bot Jun 5, 2026

Uh oh!

Wescome commented Jun 10, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

		// Read last 200 events (configurable)
		const events = await this.readRecentEvents(200)

Conversation

Wescome commented Jun 5, 2026

Summary

Packages added

Schema addition

Status

Decisions recorded

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Jun 5, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Jun 5, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Jun 5, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Jun 5, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Jun 5, 2026

Choose a reason for hiding this comment

Uh oh!

Wescome commented Jun 10, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants