CogniRelay Continuity Authoring

Use this skill when an agent runtime needs to maintain CogniRelay continuity responsibly.

Responsibility Split

CogniRelay is the substrate for continuity storage, startup reads, bounded retrieval, graph orientation, schedule orientation, trust signals, and validation limits. CogniRelay is not the semantic author of a capsule.

The running agent authors semantic fields through explicit judgment before any continuity save. Semantic fields include stance, priorities, open loops, constraints, negative decisions, rationale, durable preferences, retrieval hints, and next-step meaning.

Hooks and adapters must not infer semantic continuity from prompts, transcripts, tool output, stale plans, git history, logs, or schedule items. They may gather mechanical facts, provide templates, validate payloads, show diffs, submit an agent-authored payload, and read back stored state.

Startup And Retrieval

At startup or pre-prompt time, use agent-assets/hooks/cognirelay_retrieval_hook.py for read-only orientation. It reads POST /v1/continuity/read with view="startup" and may call POST /v1/context/retrieve only when explicitly enabled and a task is supplied.

Do not use retrieval output as a continuity write. Do not persist prompt text, transcript text, tool chatter, shell output, or copied retrieval snippets.

Graph and schedule sections are read-only orientation adjuncts. They can help the agent decide what to do next, but they are not capsule fields to copy mechanically.

Runtime Help Gate

Before every continuity-authoring mutation handled by this skill, consult the live runtime help contract for the exact write tool about to be called. This mandatory gate is limited to mutation tools that the shipped runtime help system currently exposes: continuity.upsert, schedule.create, schedule.update, schedule.acknowledge, and schedule.retire.

If a mutation tool is not exposed by system.tool_usage or GET /v1/help/tools/{name}, do not call it from this skill. Stop and report that no shipped runtime-help contract exists for that mutation. Proceed only if the user explicitly approves that specific unsupported-help mutation.

For any mutation, first query exact tool usage:

{"jsonrpc":"2.0","id":1,"method":"system.tool_usage","params":{"name":"<exact.mutation_tool>"}}

GET /v1/help/tools/<exact.mutation_tool>

Replace <exact.mutation_tool> with the exact shipped help tool name, for example continuity.upsert or schedule.create.

For continuity payloads, also query the runtime limits index and targeted limits for every bounded or schema-sensitive field being authored or changed:

{"jsonrpc":"2.0","id":2,"method":"system.validation_limits","params":{}}

{"jsonrpc":"2.0","id":3,"method":"system.validation_limit","params":{"field_path":"continuity.session_trajectory"}}

GET /v1/help/limits
GET /v1/help/limits/continuity.session_trajectory

Use MCP when operating through the runtime tool protocol:

{"jsonrpc":"2.0","id":10,"method":"system.tool_usage","params":{"name":"continuity.upsert"}}

{"jsonrpc":"2.0","id":11,"method":"system.tool_usage","params":{"name":"schedule.create"}}

Use HTTP when operating through REST:

GET /v1/help/tools/continuity.upsert
GET /v1/help/tools/schedule.create

Query targeted limits for every bounded field you plan to change. Common continuity authoring fields include:

• continuity.stance_summary
• continuity.top_priorities
• continuity.active_concerns
• continuity.active_constraints
• continuity.open_loops
• continuity.drift_signals
• continuity.session_trajectory
• continuity.retrieval_hints.must_include
• continuity.relationship_model.preferred_style
• continuity.relationship_model.sensitivity_notes
• continuity.attention_policy.presence_bias_overrides
• continuity.capsule_serialized_utf8

If the runtime help lookup is unavailable or the target mutation is unsupported by shipped help, do not guess and do not mutate. Stop and report that runtime help is unavailable for that mutation. A degraded write is allowed only with explicit user approval for that specific mutation; when approved, name the skipped help calls, run the shipped hook facts output and local dry-run when applicable, then treat the mutation result as authoritative. If a mutation fails, query system.error_guide or GET /v1/help/errors/{code} before retrying.

Save Flow

Use agent-assets/hooks/cognirelay_continuity_save_hook.py after the agent has enough context to author a durable update.

1. Run facts for mechanical subject/config/runtime facts and help links.
2. Query runtime help for continuity.upsert and runtime limits for every field you will author or mutate.
3. Run template for a generic full continuity.upsert skeleton.
4. Author semantic fields explicitly in the payload while staying inside the live field limits.
5. Run dry-run to reject placeholders and inspect a candidate-only semantic diff.
6. Run write only after the explicit agent-authored payload exists and the runtime help gate has been satisfied.
7. Run readback or doctor to verify warnings, trust signals, and stored state.

Example savepoint flow:

system.tool_usage(name="continuity.upsert")
system.validation_limits()
system.validation_limit(field_path="continuity.stance_summary")
system.validation_limit(field_path="continuity.active_concerns")
system.validation_limit(field_path="continuity.session_trajectory")
system.validation_limit(field_path="continuity.capsule_serialized_utf8")
repeat system.validation_limit for every bounded or schema-sensitive field in the candidate payload
agent-assets/hooks/cognirelay_continuity_save_hook.py template
agent authors payload
agent-assets/hooks/cognirelay_continuity_save_hook.py dry-run --input payload.json
agent-assets/hooks/cognirelay_continuity_save_hook.py write --input payload.json
agent-assets/hooks/cognirelay_continuity_save_hook.py readback

Example pre-compaction flow:

GET /v1/help/tools/continuity.upsert
GET /v1/help/limits
GET /v1/help/limits/continuity.open_loops
GET /v1/help/limits/continuity.top_priorities
GET /v1/help/limits/continuity.active_constraints
GET /v1/help/limits/continuity.stance_summary
GET /v1/help/limits/continuity.capsule_serialized_utf8
repeat GET /v1/help/limits/{field_path} for every bounded or schema-sensitive field in the candidate payload
author compact payload from agent judgment
dry-run
write
doctor

Unsupported mutation example:

system.tool_usage(name="continuity.patch")
runtime returns unsupported tool
stop and report: no shipped runtime-help contract exists for continuity.patch
do not call continuity.patch unless the user explicitly approves this unsupported-help mutation

Scheduling

Agents may create one-shot reminders or task nudges through schedule.create or POST /v1/schedule/items only when the user or an explicit work plan needs future follow-up.

• Before any schedule mutation, call system.tool_usage for the exact schedule tool or GET /v1/help/tools/{name} for the HTTP equivalent.
• For schedule creation, query schedule.create; for changes, query schedule.update; for completion or acknowledgement, query schedule.acknowledge; for no-longer-relevant items, query schedule.retire.
• Use UTC timestamps only.
• Use kind="reminder" for general follow-up.
• Use kind="task_nudge" only when linked to a task, thread, or subject.
• Scheduling does not execute work; it surfaces future orientation through schedule_context, schedule.list, and /ui/schedule.
• Do not auto-create reminders from every open loop.
• Do not infer due dates.
• Do not acknowledge or retire schedule items unless the work was actually handled or made irrelevant.

Write Discipline

Before any mutation handled by this skill, verify through runtime help that the operation is allowed, bounded, durable, and agent-authored. Reject prompt dumping, transcript dumping, copied retrieval text, and automatic semantic inference. Treat warnings and degraded trust signals as operational input for the agent, not as hook-authored meaning. If a write fails, consult runtime error/help guidance before retrying; do not enter a blind edit/write loop.