| name | using-aitp |
| description | HIGHEST PRIORITY - Use for ANY theoretical-physics research, topic continuation, exploratory physics discussion, old-knowledge Q&A tied to project topics, prior topic progress/status inquiries, idea steering, paper learning, derivation work, validation planning, or study of physical systems. Classify request intensity before deciding read-only or write-capable flow. |
Using AITP v5 - Kimi Code
Hard Gate
Use this skill before substantial theoretical-physics work. This includes
brainstorming that may become a route, literature exploration, derivation,
validation planning, prior-topic progress/status inquiries, old-knowledge Q&A
tied to a project topic, and long-running theoretical-physics work.
Do not treat chat summaries, Markdown notes, or generated hook config as scientific truth. AITP v5 truth comes from typed records, execution briefs, validation results, promotion packets, and approved memory entries.
First-Run Setup
Before research calls in a new installation, check whether the AITP MCP server
is in setup mode:
aitp_config_status()
If that tool exists and reports configured=false, do not proceed with
research calls yet. Ask the user for the local AITP-Research-Protocol
checkout path and a topics root (default ~/.aitp/topics), then call:
aitp_configure(repo_root=<repo path>, topics_root=<topics path or empty>)
After successful configuration, tell the user to run /reload or open a new
Kimi Code session so the full aitp_v5_* surface loads. If the user has no
checkout, offer to clone https://github.com/bhjia-phys/AITP-Research-Protocol.git
into a user-chosen directory first.
Environment
- AITP v5 runs through the native MCP entrypoint at
<repo-root>/brain/v5/native_mcp.py, resolved by the plugin launcher (env AITP_REPO_ROOT -> ~/.aitp/kimi-plugin-config.json -> ~/.aitp/install-record.json -> vendor/AITP-Research-Protocol). Pass base="" to AITP v5 tools when the topics root is unknown; the MCP server resolves the empty base to its configured AITP_TOPICS_ROOT.
- Kimi Code should expose the MCP server as
aitp; typed tools are named
aitp_v5_*.
- Project installs may also expose legacy-friendly discovery aliases
aitp_list_topics, aitp_get_execution_brief, and aitp_bootstrap_topic
from the same v5 native MCP server. Treat these aliases as discovery or
bootstrap compatibility only; they are not the execution contract for v5 work.
- Legacy L0-L4 write tools are read-only guards by default. If an old
aitp_*
write call returns legacy_aitp_writes_disabled, do not retry it; continue
through v5 migration/binding and typed aitp_v5_* writes.
- Current AITP project installs use
.kimi/config.toml and .kimi/skills/.
- Newer Kimi Code installs may use
.kimi-code/config.toml, .kimi-code/mcp.json, and .kimi-code/skills/.
- If the local Kimi CLI supports explicit paths, load project assets with
--config-file, --mcp-config-file, and --skills-dir.
- On Windows terminals, set
PYTHONIOENCODING=utf-8 and PYTHONUTF8=1 before kimi mcp test if the CLI crashes while printing Unicode status symbols.
Entry Procedure
Request Intensity
Classify the request before deciding how much AITP to load:
- Status or prior-progress inquiry: read-only recovery, execution brief, and
claim relation map. Do not write unless the user asks for a handoff or an
unresolved human checkpoint must be answered.
- Old-knowledge/textbook Q&A: answer normally if generic. If tied to a known
topic or claim, restore brief plus relation map; write only durable sources,
gaps, routes, or corrections.
- Light exploratory discussion: restore known topic context, but record only
when a durable route, question, source, artifact, result, or gap emerges.
- Continuation, derivation, source reading, code/numerical work, validation,
contradiction, final synthesis, trust update, or L2 promotion: restore the v5
session and follow typed runtime gates.
Intent Matrix
Use the lightest AITP path that preserves truth.
| User intent | AITP read depth | Recording default | Escalate to write when |
|---|
| Generic textbook or old-knowledge Q&A | None, unless the answer names an existing topic or claim | No write | The answer corrects project memory, finds a durable gap, or introduces a reusable source |
| Project-linked old-knowledge Q&A | Recovery audit, brief, relation map | No write | The answer changes a claim boundary, source role, route, or proof/validation obligation |
| Prior progress/status inquiry | Recovery audit, brief, relation map, summaries | No write | User asks for a handoff/status artifact or resolves a human checkpoint |
| Light exploratory discussion | Topic context if known; classifier only after a durable moment appears | No write | User accepts a route/question, identifies a source, or exposes a reusable gap |
| Topic continuation or derivation | Brief, relation map, lightweight recording navigation; process graph only when needed | Write at durable moments | Source, artifact, result, proof obligation, route decision, or validation state changes |
| Code/numerical/literature execution | Brief, relation map, source/code context, recording navigation | Write provenance and outputs | Tool run completes, artifact appears, validation passes/fails, or anomaly is observed |
| Final claim/trust/L2/memory action | Brief, relation map, trust/promotion preflight | Human-gated write only | Explicit v5 gate and user decision allow it |
When unsure between two rows, choose the read-only row first and let the
recording classifier decide whether a durable moment exists. Do not create a
new topic, claim, session, or binding merely because a conversation is
interesting.
- If
ResearchAction is available, open a WorkFrame before substantive AITP
reads:
ResearchAction(open_work_frame, topic=<topic>, goal=<restore-or-research-goal>).
After the execution brief and claim relation map are loaded, call
ResearchAction(compile_context_pack, work_frame_id=<frame-id>) before final
synthesis. If ResearchAction is not available, continue with the AITP MCP
steps below.
- If the request might belong to theoretical physics, call
aitp_v5_get_execution_brief for the active session before doing
substantive work.
Also call aitp_v5_get_claim_relation_map for the same session before
interpreting failures, blockers, support, limitations, or next actions.
- If only a topic slug is known, first call
aitp_v5_build_workspace_recovery_audit for that topic. If the row is
recovery_ready, use the selected session_id and active_claim_id; then
call aitp_v5_get_execution_brief and aitp_v5_get_claim_relation_map for
that session. Do not migrate, create, bind, or update claim status during
recovery when a ready v5 session already exists.
- Use
aitp_list_topics and aitp_get_execution_brief only to orient legacy
topics after the v5 recovery audit has failed. Before substantive research,
migrate/bind a v5 session with aitp_v5_migrate_curated_legacy_topic_to_v5
for known curated topics, aitp_v5_migrate_legacy_topic_to_v5 for generic
preservation, or create a new v5 topic/claim/session with
aitp_v5_create_topic, aitp_v5_create_claim, and
aitp_v5_bind_session.
Do not write new progress back into old L0/L1/L3/L4 files.
- Read the execution brief and follow its risk, claim, evidence, validation, and next-action fields.
Read the claim relation map as the read-only conclusion-boundary layer; it
cannot update claim trust.
- For every meaningful result, use typed writes:
aitp_v5_record_physics_object
aitp_v5_record_object_relation
aitp_v5_record_evidence
aitp_v5_record_tool_run
aitp_v5_create_validation_contract
aitp_v5_record_validation_result
aitp_v5_record_sensemaking_report
- Before trust changes or L2 memory promotion, use the v5 trust/promotion gate. Never promote from a summary alone.
Progressive Recording Navigation
Do not write AITP records at every chat step. Trigger navigation at durable
moments: known-topic session start, active claim creation/change, durable source
identity/location, completed tool run, produced artifact, observed result or
anomaly, negative result, proof or validation gap, route pivot, final answer
about an active claim, trust/promotion request, or session-end handoff.
Use this sequence, mapped to the available Kimi/Hakimi tool names:
aitp_v5_build_workspace_recording_audit(base="") # read-only, if placement is unclear
aitp_v5_classify_recording_candidate(base="", ...) # read-only
aitp_v5_get_recording_navigation_state(base="", session_id=<session-id>, claim_id=<claim-id>) # read-only
aitp_v5_expand_recording_slot(base="", session_id=<session-id>, slot=<slot>, claim_id=<claim-id>) # read-only
<existing typed write or preflight tool named by the slot expansion>
aitp_v5_verify_recording_effect(base="", session_id=<session-id>, expected_refs=[...]) # read-only
Audit, classifier, navigation, slot expansion, and verification surfaces cannot
update claim trust. Only the deepest typed write or preflight tool may mutate
kernel state, and trust still requires explicit trust/human gates.
Do not run the progressive navigator for generic explanation, vague
brainstorming, duplicate status summaries, or source/file scans that do not
change a claim, route, gap, artifact, or validation state. If a light
discussion becomes research, first restate the durable moment in one sentence,
then classify that candidate.
Kimi Hook Installation
Generate or merge project-local Kimi hooks from the v5 kernel:
python -m brain.v5.cli --base <workspace> adapter install-hooks kimi-code <session-id> --settings <workspace>/.kimi/config.toml
python -m brain.v5.cli --base <workspace> adapter install-hooks kimi-code <session-id> --settings <workspace>/.kimi-code/config.toml
Audit the installed Kimi config:
python -m brain.v5.cli --base <workspace> adapter install-audit kimi-code --settings <workspace>/.kimi/config.toml
python -m brain.v5.cli --base <workspace> adapter install-audit kimi-code --settings <workspace>/.kimi-code/config.toml
The installed hooks are lifecycle guards only. They can block risky pre-tool actions and append post-tool trace events, but they cannot update claim trust.
Working Style
- Talk naturally with the researcher, but record durable scientific content as typed records.
- Keep definitions, claims, evidence, tool runs, validation contracts, validation results, and sensemaking separate.
- Record negative results and failure modes immediately.
- If Kimi hooks are unavailable, continue through MCP tools and note the hook gap as runtime metadata only.
Red Flags
Stop and re-enter AITP v5 if you catch yourself saying:
- "I can make substantial recovery calls before opening the available WorkFrame."
- "I can answer this research question directly without a brief."
- "This summary is enough to promote memory."
- "The hook config says this happened, so the claim is validated."
- "This runtime/application failure proves the algorithm works or fails."
- "I need to bind or update claim status just to restore an existing ready topic."
- "I can enable AITP_LEGACY_ENABLE_WRITES during normal research."
- "I'll record the tool run later."