Blocks archie-core-fwmp. Two review passes over cp/shape (9 angles each) and independent host checks agree the bead cannot be delivered by the approach taken, and the mechanism used makes the named hazard worse rather than removing it.
EVIDENCE (verified on cp/shape at b73e3e76, and reproduced on the host):
RegisterStepTypeProvider (internal/domain/workflow/step_vocabulary.go:97) has no production caller: grep returns only its own definition/comments plus _test.go. So ShippedStepVocabulary() is content-identical to BuiltinStepRegistry() in every binary, and PRD:95 'Plugins can add step types' is still false for anything an operator can install.
go test ./internal/domain/workflow/... ./internal/app/controlplane/... -count=2 FAILS ('plugin "probe" step type "probe.registered" is already claimed by probe'). The registrar is process-global mutable state mutated by tests with no cleanup, so the suite is green only under -count=1 — which is what the gate and task check run, so the gate cannot see it.
The seam is a domain-package global carried by the Go import graph. The validating side (archie-state-store) and the executing side (archie-agent) are separately deployed binaries, so 'resolved identically on both sides' now rests on both happening to link the same provider package, with no build check and no test. Before this change, both sides hardcoded the same function and divergence was impossible by construction. The review's bead/boundary/deadfield angles call this BLOCK for that reason.
This repo's only plugin mechanism is Yaegi interpreted plugins loaded by the daemon (plugin.LoadDir at internal/app/archied/bootstrap.go:761), reachable only through pluginextract.Symbols, which extracts internal/plugin alone. archie-agent and archie-state-store never load plugins, so interpreted plugins cannot reach the seam at all.
Scope conflict: docs/architecture/plugins-and-extensions.md requires functionality to be exposed through narrow typed capability contracts with an owning registry/manager in the capability family, and ARCHITECTURE.md#plugin-engine-rule-strict governs a new plugin engine surface. The change adds a second plugin-contribution mechanism beside the plugin host instead of a composing-root registrar.
WHY THIS IS A DECISION, NOT A FIX: a second patch on the same block is zero-patch stacking (CLAUDE.md), and no settled design exists for this question. The open question is how a compiled-in step vocabulary is provisioned into separately deployed binaries and how a mismatch is detected.
OPTIONS (for the maintainer):
A. Compiled-in provider set, injected at the composition root: providers live in their own package, the app registers them at startup before any resolution (not a domain global), and a composition-root test asserts all three binaries resolve identical vocabularies.
B. Vocabulary identity on the wire: carry a vocabulary version/hash in the control-plane resource or task brief so the executing side detects a mismatch and fails loudly instead of silently refusing a definition.
C. Reject the premise: with a Yaegi-only plugin mechanism, plugins cannot contribute typed Go step factories; keep the step vocabulary a closed set and re-scope fwmp accordingly.
A+B together is the shape that matches the settled plugin model; C is the honest answer if plugin step types are not actually wanted yet.
cp/shape is parked unmerged (3 commits: 215bd363, c1b74861, b73e3e76) and must not be merged as is.
Notes
Decided in docs/prds/workflow-step-vocabulary.md (Draft, awaiting the maintainer): a step type is a stable identifier plus a Go factory, contributed as a whole named provider through the Workflow family's typed contract; the manager validates the entire contribution atomically before applying any of it; internal/infrastructure/workflowsteps is the single bundling point; every root (State Store validating side, daemon definitions client, agent worker) builds its manager from that one provider set at its composition root, so agreement is per build and deploy skew is closed by a matching deploy, not a wire field; an unknown step type is refused at save by ParseDefinition and at run by Compile, never degraded; Yaegi plugins cannot contribute step types. Rejected: a vocabulary hash on the wire, a Yaegi-to-StepFactory bridge, a process-global registrar, and degrading on unknown types.
Close reason
Closed
Imported from beads archie-core-m7oc · GitHub #1049 · type decision · priority P1 · created 2026-09-21 · closed 2026-09-29
Blocks archie-core-fwmp. Two review passes over cp/shape (9 angles each) and independent host checks agree the bead cannot be delivered by the approach taken, and the mechanism used makes the named hazard worse rather than removing it.
EVIDENCE (verified on cp/shape at b73e3e76, and reproduced on the host):
- RegisterStepTypeProvider (internal/domain/workflow/step_vocabulary.go:97) has no production caller: grep returns only its own definition/comments plus _test.go. So ShippedStepVocabulary() is content-identical to BuiltinStepRegistry() in every binary, and PRD:95 'Plugins can add step types' is still false for anything an operator can install.
- go test ./internal/domain/workflow/... ./internal/app/controlplane/... -count=2 FAILS ('plugin "probe" step type "probe.registered" is already claimed by probe'). The registrar is process-global mutable state mutated by tests with no cleanup, so the suite is green only under -count=1 — which is what the gate and task check run, so the gate cannot see it.
- The seam is a domain-package global carried by the Go import graph. The validating side (archie-state-store) and the executing side (archie-agent) are separately deployed binaries, so 'resolved identically on both sides' now rests on both happening to link the same provider package, with no build check and no test. Before this change, both sides hardcoded the same function and divergence was impossible by construction. The review's bead/boundary/deadfield angles call this BLOCK for that reason.
- This repo's only plugin mechanism is Yaegi interpreted plugins loaded by the daemon (plugin.LoadDir at internal/app/archied/bootstrap.go:761), reachable only through pluginextract.Symbols, which extracts internal/plugin alone. archie-agent and archie-state-store never load plugins, so interpreted plugins cannot reach the seam at all.
- Scope conflict: docs/architecture/plugins-and-extensions.md requires functionality to be exposed through narrow typed capability contracts with an owning registry/manager in the capability family, and ARCHITECTURE.md#plugin-engine-rule-strict governs a new plugin engine surface. The change adds a second plugin-contribution mechanism beside the plugin host instead of a composing-root registrar.
WHY THIS IS A DECISION, NOT A FIX: a second patch on the same block is zero-patch stacking (CLAUDE.md), and no settled design exists for this question. The open question is how a compiled-in step vocabulary is provisioned into separately deployed binaries and how a mismatch is detected.
OPTIONS (for the maintainer):
A. Compiled-in provider set, injected at the composition root: providers live in their own package, the app registers them at startup before any resolution (not a domain global), and a composition-root test asserts all three binaries resolve identical vocabularies.
B. Vocabulary identity on the wire: carry a vocabulary version/hash in the control-plane resource or task brief so the executing side detects a mismatch and fails loudly instead of silently refusing a definition.
C. Reject the premise: with a Yaegi-only plugin mechanism, plugins cannot contribute typed Go step factories; keep the step vocabulary a closed set and re-scope fwmp accordingly.
A+B together is the shape that matches the settled plugin model; C is the honest answer if plugin step types are not actually wanted yet.
cp/shape is parked unmerged (3 commits: 215bd363, c1b74861, b73e3e76) and must not be merged as is.
### Notes
Decided in docs/prds/workflow-step-vocabulary.md (Draft, awaiting the maintainer): a step type is a stable identifier plus a Go factory, contributed as a whole named provider through the Workflow family's typed contract; the manager validates the entire contribution atomically before applying any of it; internal/infrastructure/workflowsteps is the single bundling point; every root (State Store validating side, daemon definitions client, agent worker) builds its manager from that one provider set at its composition root, so agreement is per build and deploy skew is closed by a matching deploy, not a wire field; an unknown step type is refused at save by ParseDefinition and at run by Compile, never degraded; Yaegi plugins cannot contribute step types. Rejected: a vocabulary hash on the wire, a Yaegi-to-StepFactory bridge, a process-global registrar, and degrading on unknown types.
### Close reason
Closed
---
_Imported from beads `archie-core-m7oc` · GitHub #1049 · type decision · priority P1 · created 2026-09-21 · closed 2026-09-29_
sam
added the highdecision labels 2026-10-27 13:40:24 +00:00
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Blocks archie-core-fwmp. Two review passes over cp/shape (9 angles each) and independent host checks agree the bead cannot be delivered by the approach taken, and the mechanism used makes the named hazard worse rather than removing it.
EVIDENCE (verified on cp/shape at b73e3e76, and reproduced on the host):
WHY THIS IS A DECISION, NOT A FIX: a second patch on the same block is zero-patch stacking (CLAUDE.md), and no settled design exists for this question. The open question is how a compiled-in step vocabulary is provisioned into separately deployed binaries and how a mismatch is detected.
OPTIONS (for the maintainer):
A. Compiled-in provider set, injected at the composition root: providers live in their own package, the app registers them at startup before any resolution (not a domain global), and a composition-root test asserts all three binaries resolve identical vocabularies.
B. Vocabulary identity on the wire: carry a vocabulary version/hash in the control-plane resource or task brief so the executing side detects a mismatch and fails loudly instead of silently refusing a definition.
C. Reject the premise: with a Yaegi-only plugin mechanism, plugins cannot contribute typed Go step factories; keep the step vocabulary a closed set and re-scope fwmp accordingly.
A+B together is the shape that matches the settled plugin model; C is the honest answer if plugin step types are not actually wanted yet.
cp/shape is parked unmerged (3 commits: 215bd363, c1b74861, b73e3e76) and must not be merged as is.
Notes
Decided in docs/prds/workflow-step-vocabulary.md (Draft, awaiting the maintainer): a step type is a stable identifier plus a Go factory, contributed as a whole named provider through the Workflow family's typed contract; the manager validates the entire contribution atomically before applying any of it; internal/infrastructure/workflowsteps is the single bundling point; every root (State Store validating side, daemon definitions client, agent worker) builds its manager from that one provider set at its composition root, so agreement is per build and deploy skew is closed by a matching deploy, not a wire field; an unknown step type is refused at save by ParseDefinition and at run by Compile, never degraded; Yaegi plugins cannot contribute step types. Rejected: a vocabulary hash on the wire, a Yaegi-to-StepFactory bridge, a process-global registrar, and degrading on unknown types.
Close reason
Closed
Imported from beads
archie-core-m7oc· GitHub #1049 · type decision · priority P1 · created 2026-09-21 · closed 2026-09-29