Why: a fallback chain can answer with a different model than the one that was requested, and usage has to name the model that actually served.
Evidence: go test .
Impact: the field is omitted when empty, so existing usage JSON stays the same.
Refs: feature/durable-rewind
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: The product needs a session mode and a place to show tool arguments on a confirmation card without inventing a second card type.
Evidence: GOWORK=off go test . passed. New fields use omitempty, so existing golden JSON stays the same.
Impact: Empty permission mode still means the profile policy. Clients that ignore unknown fields keep working.
Refs: feature/permission-hooks
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: The final reply is composed after the work loop, and the automatic end check needs its own part and event so the timeline can show it without mixing it into the user-facing text.
Evidence: go test ./...
Impact: New RunPhase finalizing, PartType check, and run.check. Existing payloads are unchanged.
Refs: feature/final-reply-compose
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: The client only learned about a tool call once execution started, so long argument generation looked like a frozen cursor.
Evidence: go test ./...
Impact: New tool.call.delta and tool.output events; tool.result may include durationMs. Unknown event types stay ignored.
Refs: feature/stream-display-overhaul
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: The workflow executor needs a shared contract for subagent profiles, plugin scripts, and per-step plus whole-run retry caps.
Evidence: go test ./... passed, including the new workflow_def golden.
Impact: New fields are omitempty, so existing workflow JSON still decodes. Kind script is additive.
Refs: feature/agent-harness-p4
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: User overrides of a subagent profile need a place on the shared contract, and empty fields must stay omitted so older modules keep their meaning.
Evidence: go test ./... passed, including the new prompt_module_subagent golden.
Impact: Existing modules serialize the same. Frontend and UI contracts still need a copy of the type.
Refs: feature/agent-harness-p3b
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: Downstream repositories need a versioned public manifest for skills, MCP, knowledge, parsers, agents, channels, and pages without forking the base.
Evidence: go test ./... passed after omitzero on empty scope and backend so the golden plugin_manifest stays stable.
Impact: apiVersion 1.0 remains accepted. New provides fields are optional. Merge without squash so Backend can pin this commit.
Refs: Agent-System-Dev feature/base-extension-roadmap
Why: Downstream repositories need a versioned public manifest for skills, MCP, knowledge, parsers, agents, channels, and pages without forking the base.
Evidence: go test ./... passed after omitzero on empty scope and backend so the golden plugin_manifest stays stable.
Impact: apiVersion 1.0 remains accepted. New provides fields are optional. Scope and backend are omitted when zero.
Refs: docs/proposals/2026-10-01-plugin-contract-v2.md
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: The chat timeline needs a persisted thinking duration, and tool duration already lives on the tool call, so the thinking part had nowhere to put it.
Evidence: go test ./... passes; the message golden now includes durationMs on the thinking part.
Impact: Additive JSON field with omitempty. Existing clients ignore unknown fields.
Refs: feature/chat-timeline-preview
Co-authored-by: Cursor <cursoragent@cursor.com>
Why: project channel convention (pseudo-key project:<projectId>) was
duplicated across runmgr/server/store; centralize so replay envelopes
match live stamping.
Evidence: go test green.
Impact: no behavior change; enables store replay stamping.
Why: D50 approved — the project inbox needs its own SSE stream reusing the
conversation event mechanism via the pseudo-key project:<projectId>; the
envelope stamps project ownership (additive field).
Evidence: conformance golden event_project_stream; go test green.
Impact: additive only (omitempty); consumers ignore unknown fields.
Refs: documents proposal 2026-09-30-card-stream-artifact-unify-project-inbox.md (D50).
Why: the fs tool family (proposal 2026-09-30-file-preview-and-reserved-
capabilities) is implemented but must stay invisible to models and tool
listings until a runtime gate opens it.
Evidence: new bool field with json omitempty; no golden samples carry it;
go test ./... green.
Impact: registries can mark gated tools; servers and runners filter on
it (backend PR in the same task).
Refs: task feature/file-preview-caps.
Why: agents rediscovered Gitea hosts and credential locations from
scratch each session, and the retired git-panel host kept failing with
TLS broken-pipe errors before the right host was recalled.
Evidence: curl git-home...:30319/api/v1/version with token from
~/.secrets/gitea -> {"version":"1.27.3"}; ssh -p 31411 git@git-ssh...
(BatchMode) -> "Hi there, weichen!" with key ~/.ssh/home_gitea (ED25519,
SHA256:GYXaOXIMeloQ1fKSfdJ/ndgPQzs/Ix7Hk62ekfkfQ20). Same section already
merged in Agent-System-Dev (PR #73).
Impact: AGENTS.md only; records hosts and credential locations, no
secret values (baseline hard boundary 2).
Refs: Agent-System-Dev PR #73
Why: cross-session transcript reading needs per-conversation outbound
(allowCrossRead, default true) and inbound (crossPrivate, default false)
flags alongside the project-level gate; see the approved proposal.
Evidence: go test ./... passes with golden conversation.json regenerated
to exercise both wire fields.
Impact: additive contract fields (omitempty), wire-compatible for
existing clients that ignore unknown fields.
Refs: Agent-System-Documents proposals/2026-09-30-cross-session-reading.md
(Dev task feature/cross-session-reading).
Why: plugins and the host need one shared avatar primitive for member/
owner displays in cards, messages, and admin surfaces (user decision
2026-09-30: avatars, cards, settings become base capabilities plugins
reuse directly).
Evidence: go test green with regenerated golden (single avatar with
sub/tone plus a stacked group); omitempty keeps existing wire samples
byte-stable.
Impact: additive optional UINode field + type; unknown-type consumers
ignore it (forward compat). Renderers gain 'avatar' via the ui repo.
Refs: Agent-System-Documents proposal 2026-09-30-ui-redesign-and-auth.md
Why: the approved task-centric sidebar (proposal 2026-09-30, option A) needs
server-authoritative pin and archive state on conversations; client-side
storage would be fake state against the event-sourced model.
Evidence: go test ./... passes; omitempty keeps wire format and golden
samples byte-identical for zero values (backward compatible).
Impact: additive optional fields; consumers ignoring unknown fields are
unaffected (forward-compat principle). Backend consumes via submodule bump.
Refs: Agent-System-Documents docs/proposals/2026-09-30-ui-redesign-and-auth.md (P6/P8, R1)
Why: prompts are moving from code into a data library; core needs the
PromptModule/Trigger/Block contract and profiles need an ordered
promptModules reference (empty = default base+env set).
Evidence: unit-covered via backend prompts package tests over this
contract; composite make test green.
Impact: additive contract fields only; no behavior change in core.
Refs: feature/prompt-library task card.
Why: Contract foundation for the inbox/command-bus/workspaces phase (proposal
D28/D30): ActivityCard as the first-class state object behind the right-rail
inbox, card.* stream events, inline-confirm anchoring (anchorMessageId +
blocked/blockReason) on HITL, seven new UI schema vocabularies, plugin
provides.commands/workspaces, workflow card step and workspace declarations.
Evidence: go vet clean; conformance goldens regenerated and verified — new
activity_card / hitl_card_blocked / event_card_created / event_card_resolved /
ui_vocabularies samples plus extended plugin_manifest; CardStatus.Terminal
helper unit-covered via build; ts/types.ts mirrors all additions.
Impact: Additive only (omitempty fields, new types, new event types) — existing
consumers ignore unknown values per forward-compat rule; vocabulary spec types
carry a Spec suffix to avoid const-name collisions.
Refs: modules/documents docs/proposals/2026-09-30-ui-inbox-command-workspace.md
Why: the composition adopted the delivery gate (Dev repo ef89fc8); resident
entry rules must carry the gate so agents load them without reading the
guidelines submodule first.
Evidence: text-only entry change; gate mechanics validated in the Dev repo
scratch-repo suite (19/19) and merged as PR weichen/Agent-System-Dev#1.
Impact: entry doc only; no code, contract, or build change in this repo.
Refs: Development-Guidelines 0.1.0-draft.7; Agent-System-Dev ef89fc8
weichen
