Problem Description
VibeGuard already pushes agents to verify before saying “done”, but the current runtime evidence is mostly hook feedback and dirty-tree checks rather than structured task state.
Examples in the repo:
hooks/stop-guard.sh gates on dirty source files and logs feedback, but does not emit a durable completion artifacthooks/learn-evaluator.sh records signals, but not a task lifecycle verdictscripts/codex/app_server_wrapper.py forwards hook feedback, but not a reusable verification ledgerdocs/how/memory-files.md documents .omx/ runtime memory, yet the repo does not define a concrete lifecycle schema for active/cancelled/completed/resumable work
That makes continuation and completion more heuristic than they need to be.
Proposed Solution
Persist structured lifecycle and verification state in repo-local runtime artifacts.
Suggested scope:
- add a completion artifact such as
.omx/state/<scope>/completion.json
- add a verification ledger such as
.omx/state/<scope>/verification-log.jsonl
- add a canonical current-plan pointer / resume hint for continuation flows
- distinguish
cancelled, failed, incomplete, and completed instead of inferring state from chat text or dirty trees
- require completion claims to reference structured verification evidence rather than prose alone
Example fields to capture:
modescopestatusstarted_atupdated_atcompleted_atcancelled_atcurrent_stepverification_statusverification_commandsknown_failuresnext_required_action
Alternatives Considered
-
Keep using hook messages + dirty-tree heuristics
Lower implementation cost, but not durable enough for continuation, auditing, or machine-checkable “done” semantics.
-
Keep lifecycle state only outside the repo
Simpler for install/runtime management, but weak for repo-local continuation and shared execution artifacts.
Additional Context
Relevant repo evidence:
hooks/stop-guard.shhooks/learn-evaluator.shscripts/codex/app_server_wrapper.pydocs/how/memory-files.mddocs/command-schemas.mdworkflows/plan-mode/SKILL.mdworkflows/plan-flow/references/execplan-template.md
This would let VibeGuard move from “remind the agent to verify” toward “persist a reusable verification contract that continuation and orchestration can trust.”
Problem Description
VibeGuard already pushes agents to verify before saying “done”, but the current runtime evidence is mostly hook feedback and dirty-tree checks rather than structured task state.
Examples in the repo:
hooks/stop-guard.shgates on dirty source files and logs feedback, but does not emit a durable completion artifacthooks/learn-evaluator.shrecords signals, but not a task lifecycle verdictscripts/codex/app_server_wrapper.pyforwards hook feedback, but not a reusable verification ledgerdocs/how/memory-files.mddocuments.omx/runtime memory, yet the repo does not define a concrete lifecycle schema for active/cancelled/completed/resumable workThat makes continuation and completion more heuristic than they need to be.
Proposed Solution
Persist structured lifecycle and verification state in repo-local runtime artifacts.
Suggested scope:
.omx/state/<scope>/completion.json.omx/state/<scope>/verification-log.jsonlcancelled,failed,incomplete, andcompletedinstead of inferring state from chat text or dirty treesExample fields to capture:
modescopestatusstarted_atupdated_atcompleted_atcancelled_atcurrent_stepverification_statusverification_commandsknown_failuresnext_required_actionAlternatives Considered
Keep using hook messages + dirty-tree heuristics
Lower implementation cost, but not durable enough for continuation, auditing, or machine-checkable “done” semantics.
Keep lifecycle state only outside the repo
Simpler for install/runtime management, but weak for repo-local continuation and shared execution artifacts.
Additional Context
Relevant repo evidence:
hooks/stop-guard.shhooks/learn-evaluator.shscripts/codex/app_server_wrapper.pydocs/how/memory-files.mddocs/command-schemas.mdworkflows/plan-mode/SKILL.mdworkflows/plan-flow/references/execplan-template.mdThis would let VibeGuard move from “remind the agent to verify” toward “persist a reusable verification contract that continuation and orchestration can trust.”