Compare commits

..

4 Commits

Author SHA1 Message Date
javis-bot
07615ee7a1 chore(app): version 0.1.21
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-16 21:13:39 +09:00
javis-bot
4d0ae91e8a perf(reply): v2 tool turns get the tool guidance in short, and a shorter reply-language default
Strata reads every reply prompt from the start in 2,048-token steps. The v2
compact prompt carried v1's full tool guidance and constraints (about 900
Qwen3.6 tokens) on every turn with a data tool, and the reply-language
default (about 120 tokens) twice.

COMPACT_TOOL_GUIDANCE keeps the same rules (act instead of refusing, facts
only from the result, look up named things, auto-derived and self-contained
arguments) in about a quarter of the tokens, without English examples; the
reply-language default says the same in half. v1's prompts are unchanged.

Measured on Strata (Qwen3.6, 20 lines of recent conversation): a weather
turn's prompt 3,436 -> 2,722 tokens, a search turn's second call
4,096 -> 3,383. Evals (compact prompt + intent judge, Qwen3.6): 24/24
before, 30/30 twice after (weather location, follow-up arguments, tool-skip
confabulation, search language, reply language, Possessor fidelity, reply
length, multi-turn context).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-16 21:13:08 +09:00
javis-bot
7da5bca783 fix(reply): a one-step turn whose model skips the tool twice is planned and guarded
After the one nudge, a second answer without the tool no longer goes out
as it is: the turn takes the planned path (the planner runs then, one
extra call only on this failure) and the skipped-tool guard runs the
planned tool, so a made-up "24 degrees" answer cannot reach the user. A
plan that needs no tool (a question back to the user) still lets the
answer go out.

Evals (qwen3:8b, EVAL_COMPACT_PROMPT=1 EVAL_INTENT_JUDGE=1, 3 runs):
the five tool-skip/weather/screen/follow-up/router files 72/84 (71/84 on
0.1.20, same known failures); Possessor director-then-filmography 0/3,
also 0/3 on 0.1.20 (qwen3:8b stops after one search; webSearch turns are
still planned).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-16 19:09:11 +09:00
javis-bot
a6932d9b16 perf(reply): v2 turns with one one-step tool skip the planner
When the router's only pick is a tool that does its job in one call
(getWeather, logMeal, fetchMeals, screenView) and the chat model calls
tools natively, the planner no longer runs: on Strata it averaged 2.5 s
per tool turn (2026-11-16 16:45-16:58 engine log, 9 calls) before the
answer could start. Search, browsers, schedule, notes, chat channels and
picks of two or more tools still plan. The rule reads the router's pick,
not the words, so it is language-agnostic.

Without a plan the skipped-tool guard has nothing to run, so these turns
get one nudge instead: an answer given before any tool ran is dropped and
the model is asked to call the tool (or ask for a missing detail).

Evals (qwen3:8b, EVAL_COMPACT_PROMPT=1 EVAL_INTENT_JUDGE=1, 3 runs of
tool_skip_confabulation, weather_autoderive_location, screen_view_tool,
followup_supplies_missing_tool_arg, tool_router_implicit): 71/84 before,
72/84 after, same known failures; planner calls 14 -> 3.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-16 19:00:36 +09:00
12 changed files with 302 additions and 37 deletions

2
app/Cargo.lock generated
View File

@@ -1520,7 +1520,7 @@ checksum = "3bd1574e617a9a6c1f8f8181bd7e841e1f20e4277607ef28945765188d6284d4"
[[package]]
name = "jarvis-app"
version = "0.1.20"
version = "0.1.21"
dependencies = [
"async-trait",
"base64 0.22.1",

View File

@@ -1,6 +1,6 @@
[package]
name = "jarvis-app"
version = "0.1.20"
version = "0.1.21"
edition = "2021"
rust-version = "1.85"
description = "Jarvis v2 Windows app core: Discord voice, voice worker supervisor, chat channels"

4
app/shell/Cargo.lock generated
View File

@@ -2637,7 +2637,7 @@ checksum = "3bd1574e617a9a6c1f8f8181bd7e841e1f20e4277607ef28945765188d6284d4"
[[package]]
name = "jarvis-app"
version = "0.1.20"
version = "0.1.21"
dependencies = [
"async-trait",
"base64 0.22.1",
@@ -2666,7 +2666,7 @@ dependencies = [
[[package]]
name = "jarvis-shell"
version = "0.1.20"
version = "0.1.21"
dependencies = [
"jarvis-app",
"serde_json",

View File

@@ -1,6 +1,6 @@
[package]
name = "jarvis-shell"
version = "0.1.20"
version = "0.1.21"
edition = "2021"
rust-version = "1.85"
description = "Jarvis v2 management app window (Tauri 2 over the app core's AdminApi)"

View File

@@ -15,7 +15,7 @@ Every distinct LLM call in Jarvis, what feeds it, what consumes it, and how it i
- Redacted user query
- Recent dialogue (last 5 minutes), including in-loop tool-call + tool-role messages from prior replies within the active conversation (tool carryover, `DialogueMemory.record_tool_turn` / `get_recent_turns_with_tools` in [src/jarvis/memory/conversation.py](src/jarvis/memory/conversation.py); per-prompt cap via `cfg.tool_carryover_max_turns` / `tool_carryover_per_entry_chars`; storage cap `_tool_turns_max_storage = 16`; cleared on `stop` signal AND on new-conversation entry; UNTRUSTED WEB EXTRACT fence markers preserved on truncation; both `content` and `tool_calls[*].function.arguments` scrubbed on write)
- Unified system prompt from [src/jarvis/system_prompt.py](src/jarvis/system_prompt.py) + ASR note + tool-protocol guidance. Reply language is resolved by `reply_language_directive(lang, cfg.tts_engine)` where `lang = _resolve_output_language()` — the single source of truth that prefers the settings-web UI value (config JSON `output_language`) over the compose `OUTPUT_LANGUAGE` env, so changing the language in the settings page takes effect. An explicit lock wins (forces "reply only in `<language>`", also forbidding other scripts so small models stop leaking trailing CJK/Hanja); else a Piper/Chatterbox TTS forces English (English-only voices); else (multilingual TTS, no lock) the assistant replies in the user's own language. The v2 voice worker's brain (`src/voice_worker/brain.py`) runs with `tts_engine="melo"`, so v1's English-only Piper default never locks v2 voice or chat answers to English, and passes `default_language` (worker `--reply-language`, default `Korean`): with no lock, `default_language_directive()` makes replies come in that language unless the user explicitly asks for another one, and `build_system_prompt()` rewrites the persona's "in the user's language" clause to "in the reply language" (eval `evals/test_reply_language_default.py`). The directive is inserted near the FRONT of the guidance list so a small model gives it primacy, and the SAME resolved `lang` feeds `build_system_prompt()`, which rewrites the persona's "in the user's language" clause to the locked language so the persona cannot contradict the directive (previously the persona read the raw env while the directive read the config value, so a settings-UI change was honoured by one and ignored by the other). Gated in `_build_initial_system_message()` at [engine.py](src/jarvis/reply/engine.py).
- **v2 compact prompt** (`compact_prompt=True`, the v2 brain): `build_voice_persona()` replaces the v1 persona (about 1.3k characters instead of 7.4k: no single-user long-term memory or on-screen browser rules, answers as short as possible), and of the size-specific components only the ASR note stays; the full tool guidance (`tool_guidance` + `tool_constraints`) is added only when a tool other than `stop`, `toolSearchTool` and `chatChannels` is allowed, else the one-line `COMPACT_ACTION_RULE`. Measured on Strata (Qwen3.6, 2026-11-12): a greeting's reply prompt went from about 4,500 to 1,350 tokens and its prompt read from 1.07 s to 0.36 s (Strata reads in 2,048-token steps). The v2 brain also turns `stream_browser` off. The `[Context: ...]` time carries the 12-hour form beside the 24-hour one (`10:37 (10:37 AM)`): with the short prompt Qwen3.6 read 10:37 as 오후 (pm) 4 of 4 times.
- **v2 compact prompt** (`compact_prompt=True`, the v2 brain): `build_voice_persona()` replaces the v1 persona (about 1.3k characters instead of 7.4k: no single-user long-term memory or on-screen browser rules, answers as short as possible), and of the size-specific components only the ASR note stays; the tool guidance is added only when a tool other than `stop`, `toolSearchTool` and `chatChannels` is allowed, as `COMPACT_TOOL_GUIDANCE` (v1's `tool_guidance` + `tool_constraints` rules in about a quarter of the tokens), else the one-line `COMPACT_ACTION_RULE`; the reply-language default (`default_language_directive`, at the front and the end) is short too. Measured on Strata (Qwen3.6, 2026-11-16, four lines of recent conversation): a weather turn's reply prompt went from 3,075 to 2,360 tokens and a time question's from 2,115 to 2,031, inside one 2,048-token step (Strata reads about 0.44 s up to 2,048 tokens, 0.86 s up to 4,096, 1.25 s up to 6,144). Measured on Strata (Qwen3.6, 2026-11-12): a greeting's reply prompt went from about 4,500 to 1,350 tokens and its prompt read from 1.07 s to 0.36 s (Strata reads in 2,048-token steps). The v2 brain also turns `stream_browser` off. The `[Context: ...]` time carries the 12-hour form beside the 24-hour one (`10:37 (10:37 AM)`): with the short prompt Qwen3.6 read 10:37 as 오후 (pm) 4 of 4 times.
- **Tool honesty rules**: `TOOL_HONESTY_RULES` inside the persona, plus `tool_honesty_reminder(allowed_tools)` appended after every context block (just before the closing reply-language lock) naming the turn's tools. Added after M0 (2026-11-10) caught models answering past an available tool.
- **Operator instructions** (two sources, both framed "Additional instructions from the operator:" and appended near the end of the guidance list): the settings-UI `llm_instructions` config field, and every `*.md` file in `AGENTS_DIR` (default `/app/agents`, bind-mounted from `./agents`). The file-based set is read once per turn by `load_agent_instructions()` in [src/jarvis/system_prompt.py](src/jarvis/system_prompt.py) and concatenated in filename order, so dropping/editing a `.md` applies on the next reply with no rebuild/restart; fail-open to `""` when the folder is absent/empty/unreadable.
- **Warm profile block** (query-agnostic User + Directives excerpt from the knowledge graph, composed by `build_warm_profile()` / `format_warm_profile_block()` in [src/jarvis/memory/graph_ops.py](src/jarvis/memory/graph_ops.py) at Step 3.5 of `reply()`; no LLM call, pure SQLite read; injected unconditionally so personalisation is the default; result cached in `DialogueMemory._hot_cache` under `DialogueMemory.WARM_PROFILE_CACHE_KEY` for the lifetime of the active conversation. Invalidated on `stop`, on new-conversation entry, AND on User/Directives graph mutations via the listener registered in [src/jarvis/daemon.py](src/jarvis/daemon.py) against `register_graph_mutation_listener` in [src/jarvis/memory/graph.py](src/jarvis/memory/graph.py); World-branch writes are ignored)
@@ -161,7 +161,7 @@ Every distinct LLM call in Jarvis, what feeds it, what consumes it, and how it i
## 12. Task-list Planner (pre-flight decomposition, gates the whole turn)
- **File**: [src/jarvis/reply/planner.py](src/jarvis/reply/planner.py) — `plan_query()`.
- **Trigger**: once per reply, **after the tool router and before memory search**. **v2 brain** (`intent_judge=True`): runs only on turns where the router picked a tool (not for the sentinels `stop` or `toolSearchTool` alone; a router-picked `chatChannels` counts), so a turn without tools has one LLM call before its answer. Skipped when `cfg.planner_enabled = False`, when the query is shorter than `MIN_QUERY_CHARS` (4), or when no model / base URL is available.
- **Trigger**: once per reply, **after the tool router and before memory search**. **v2 brain** (`intent_judge=True`): runs only on turns where the router picked a tool (not for the sentinels `stop` or `toolSearchTool` alone; a router-picked `chatChannels` counts), so a turn without tools has one LLM call before its answer. Also skipped when the router's only pick is a one-step tool (`_ONE_STEP_TOOLS` in `engine.py`: `getWeather`, `logMeal`, `fetchMeals`, `screenView`) and the chat model calls tools natively; the chat model then gets the tool directly, and an answer it gives without the tool is dropped and followed by one nudge to call it; a second answer without the tool runs the planner after all (one extra call, only then) and the skipped-tool guard applies (planner.spec.md, "Intent judge (v2)"). Skipped when `cfg.planner_enabled = False`, when the query is shorter than `MIN_QUERY_CHARS` (4), or when no model / base URL is available.
- **Model / gating**: resolution chain `planner_model (override) → ollama_chat_model`. The planner tracks the chat model so upgrading the chat model (via setup wizard or config) automatically upgrades plan quality.
- **Inputs**: user query, dialogue context, the live time/location line (`LIVE CONTEXT:`, so a location-dependent ask plans `searchMemory topic='where the user lives'` only when the location is unknown or disabled), **router-narrowed** tool catalogue (names + one-line descriptions + argument names, built by `planner_tool_catalog()`) — not the full 30+ list. When the carry-over guard from #7 fires, the previous turn's failed tool name is unioned into this catalogue before the planner sees it, so the planner can plan a re-call without `toolSearchTool` round-tripping. **No** memory context — the planner decides *whether* memory is needed.
- **System prompt**: `_PROMPT_TEMPLATE` in `planner.py`. Teaches the `searchMemory topic='...'` directive for prior-conversation lookups, short imperative tool steps, angle-bracket entity placeholders, final synthesis step, same-language output (the words of search and play arguments too, repeated in the user message's closing line), no numbering.
@@ -215,7 +215,7 @@ Every distinct LLM call in Jarvis, what feeds it, what consumes it, and how it i
| 10 | Graph extraction | ~1/session | No (background) | LARGE |
| 11 | Graph best-child | 0-N | No (background) | SMALL (via router chain) |
| 11b | Graph node merge | 0-N (per node, batched) | No (background) | SMALL (via router chain) |
| 12 | Planner (plan_query) | 1 | yes (planner_enabled) | LARGE/SMALL (tracks chat model) |
| 12 | Planner (plan_query) | 0-1 (v2: tool turns except one-step tools, which plan only when the model skips the tool twice) | yes (planner_enabled) | LARGE/SMALL (tracks chat model) |
| 13 | Plan step resolver | 0-N (SMALL only) | auto by size + plan | SMALL (via router chain) |
| 14 | Tool-specific | per-tool | n/a | LARGE |
| 15 | Voice call judgements | 0 (per utterance / tie, outside the reply) | call mode / policy | LARGE (chat model) |

View File

@@ -13,6 +13,7 @@ from ..system_prompt import (
build_system_prompt,
build_voice_persona,
COMPACT_ACTION_RULE,
COMPACT_TOOL_GUIDANCE,
load_agent_instructions,
reply_language_directive,
tool_honesty_reminder,
@@ -1095,6 +1096,23 @@ def _attach_images(message: dict, result) -> None:
message["images"] = list(images)
def _starts_with_text_tools(cfg) -> bool:
"""Whether the reply starts in text-tool mode: a small model outside the native tool-call
families (see the loop's comment where it is used)."""
model_l = (cfg.ollama_chat_model or "").lower()
native_capable = any(k in model_l for k in (
"qwen2.5", "qwen2", "qwen3", "llama3.1", "llama3.2", "llama3.3",
"mistral", "hermes", "command-r", "firefunction",
))
return detect_model_size(cfg.ollama_chat_model) == ModelSize.SMALL and not native_capable
# Tools that do their job in one call (planner.spec.md, "Intent judge (v2)"): a v2 turn whose only
# router pick is one of them needs no plan. Tools a request may use several times or chain (search,
# browsers, schedule and notes lists before a change, chat channels) are not here.
_ONE_STEP_TOOLS = frozenset({"getWeather", "logMeal", "fetchMeals", "screenView"})
def _offered_builtins(cfg, chat_available: bool, schedule_available: bool = False,
notes_available: bool = False, screen_available: bool = False) -> dict:
"""The built-in tools this turn can offer: the enabled ones (``enabled_builtin_tools``),
@@ -1400,6 +1418,14 @@ def run_reply_engine(db: "Database", cfg, tts: Optional[Any],
# aside; the chat tool counts here, its every-turn offer is added below); a turn without one
# goes straight to its answer.
_turn_needs_tools = any(t not in ("stop", "toolSearchTool") for t in routed_tools)
# Intent judge (v2), native tool calls: a lone one-step tool goes to the chat model without a
# plan; the nudge in the loop below stands in for the plan-based skipped-tool guard.
_picked_tools = [t for t in routed_tools if t not in ("stop", "toolSearchTool")]
_one_step_tool = (
_picked_tools[0]
if intent_judge and len(_picked_tools) == 1 and _picked_tools[0] in _ONE_STEP_TOOLS
and not _starts_with_text_tools(cfg) else None
)
# In screen-share mode, always offer setBroadcast. "Turn the broadcast
# on/off" is language-agnostic intent the embedding/keyword router won't
@@ -1447,7 +1473,9 @@ def run_reply_engine(db: "Database", cfg, tts: Optional[Any],
try:
# Intent judge (v2): the router decides whether the turn needs tools; only then does the
# planner order the steps (planner.spec.md, "Intent judge (v2)").
if not intent_judge or _turn_needs_tools:
if _one_step_tool:
debug_log(f"intent judge: one-step tool {_one_step_tool}, no plan", "planning")
elif not intent_judge or _turn_needs_tools:
action_plan = plan_query(
cfg=cfg,
query=redacted,
@@ -1866,12 +1894,7 @@ def run_reply_engine(db: "Database", cfg, tts: Optional[Any],
# call, so getWeather/webSearch/controlBrowser never run. Use native for the
# tool-capable families (native still auto-falls-back to text on HTTP 400);
# only genuinely non-tool small models (e.g. gemma) default to text.
_model_l = (cfg.ollama_chat_model or "").lower()
_native_capable = any(k in _model_l for k in (
"qwen2.5", "qwen2", "qwen3", "llama3.1", "llama3.2", "llama3.3",
"mistral", "hermes", "command-r", "firefunction",
))
use_text_tools = (model_size == ModelSize.SMALL) and not _native_capable
use_text_tools = _starts_with_text_tools(cfg)
prompts = get_system_prompts(model_size)
debug_log(f"Model size detected: {model_size.value} for {cfg.ollama_chat_model} (use_text_tools={use_text_tools})", "planning")
@@ -1919,9 +1942,10 @@ def run_reply_engine(db: "Database", cfg, tts: Optional[Any],
# The persona already covers voice style, inference and tool use; the tool guidance
# (named entities, auto-derived and self-contained arguments) matters only when a
# data tool is on offer. The sentinels and the always-offered chat tool are not.
# It is v1's guidance in short (COMPACT_TOOL_GUIDANCE): every token is read each reply.
guidance.append(prompts.asr_note)
if any(t not in ("stop", "toolSearchTool", "chatChannels") for t in allowed_tools):
guidance.extend(c for c in (prompts.tool_guidance, prompts.tool_constraints) if c)
guidance.append(COMPACT_TOOL_GUIDANCE)
else:
guidance.append(COMPACT_ACTION_RULE)
else:
@@ -2281,6 +2305,9 @@ def run_reply_engine(db: "Database", cfg, tts: Optional[Any],
# synthesise from the real result. Fires at most once per reply; a
# non-empty list means it fired. See planner.spec.md.
_forced_plan_steps: list[str] = []
# One-step turns have no plan: an answer given without their tool is dropped and the model is
# asked once to call it (planner.spec.md, "Intent judge (v2)").
_one_step_nudged = False
while turn < max_turns:
turn += 1
@@ -3002,6 +3029,36 @@ def run_reply_engine(db: "Database", cfg, tts: Optional[Any],
candidate_reply = content
malformed_fallback = False
_one_step_skipped = bool(
_one_step_tool and _one_step_tool in allowed_tools and not use_text_tools
and not _forced_plan_steps and not malformed_fallback
and sum(1 for m in messages if m.get("tool_name")) == _plan_steps_baseline
)
if _one_step_skipped and not _one_step_nudged:
messages.pop()
messages.append({"role": "user", "content": (
f"[Before answering, call {_one_step_tool}: this request needs its result. "
"If a detail it needs is missing, ask the user for it instead.]")})
_one_step_nudged = True
debug_log(f"one-step turn: answered without {_one_step_tool}; asking for the call", "planning")
continue
if _one_step_skipped and not action_plan:
# Asked once and still no tool: this turn takes the planned path after all, so the
# skipped-tool guard below runs the tool the plan names (or the plan says no tool is
# needed, a question back to the user say, and the answer goes out).
try:
action_plan = plan_query(
cfg=cfg, query=redacted, dialogue_context=_dialogue_ctx,
tools=_planner_tool_catalog, live_context=_live_time_location_string(cfg),
)
except Exception as _plan_exc: # pragma: no cover — defensive
debug_log(f"planner step failed (non-fatal): {_plan_exc}", "planning")
action_plan = []
if action_plan:
_turn_state.trace("plan", steps=list(action_plan))
debug_log(f"one-step turn: still no {_one_step_tool} after the nudge; planned "
f"{len(action_plan)} step(s)", "planning")
_guard_steps = [] if (use_text_tools or _forced_plan_steps or malformed_fallback) else (
named_tool_steps(
action_plan,

View File

@@ -222,6 +222,29 @@ needs the planner to name the channel); its every-turn offer, added after the ro
to its answer with one LLM call before it; a tool turn keeps the planner's ordered steps, which
multi-step asks need ("who directed Possessor, and what else did they make").
**One-step turns.** When the router's only pick (sentinels aside) is a tool that does its job in one
call, `_ONE_STEP_TOOLS` in `engine.py` (`getWeather`, `logMeal`, `fetchMeals`, `screenView`), and the
chat model calls tools natively, the planner does not run: its steps would only repeat the one call
the chat model makes itself. The rule reads the router's pick, never the words, so it is language-
agnostic. Tools a request may use several times or chain stay planned: search and page fetches,
the browsers, `schedule` and `personalNotes` (a list before a change), `chatChannels`. Two or more
picked tools are planned as before. Text-tool mode (small non-native models) always plans, since its
direct execution runs the plan.
Without a plan the skipped-tool guard above has nothing to run, so these turns get a nudge instead:
when the chat model answers in words before any tool ran in this reply, the answer is dropped (no
later turn sees it) and a note asks the model to call that tool first, or to ask the user for a
missing detail. The nudge is sent at most once per reply. When the model answers without the tool
again, that turn falls back to the planned path: the planner runs then (one extra planner call, only
on this failure), and the skipped-tool guard above treats the second answer as on any planned turn.
A plan naming the tool runs it and the made-up answer never goes out; a plan that needs no tool (a
question back to the user, say) lets the answer go out, the planner judging that as before.
Measured on Strata (Qwen3.6, 2026-11-16 16:45-16:58, the engine log): the planner averaged 2.5 s
(9 calls), one after the other before the answer on every tool turn, while the router runs during
speech (`prepare_intent`). A planner call that does not run also leaves the engine's second session
to the chat model, whose prompt read is the largest cost.
Measured on Strata (Qwen3.6, 2026-11-12): the router took 0.42 s and the planner 0.49-0.94 s, one
after the other, before every answer; each call costs at least one full prompt read. Tried first
and dropped: the planner alone over the whole catalogue (one call) planned "reply with the

View File

@@ -106,7 +106,7 @@ Design principles enforced by the engine:
- When detected, the engine falls back to the standard "I had trouble understanding that request" error reply (model-size-aware). The malformed content is never shown to the user.
Skipped-tool guard (native tool calling):
- When a native-tool model answers in prose before any tool has run in this reply, and the plan names a callable tool as a concrete call, the engine discards that answer and direct-executes the plan's tool steps, then the model answers from the results. Only literal, schema-valid steps run (no LLM repair). When the planned call cannot run, the held-back answer is still discarded: the planned call is recorded as a failed tool result and the model answers from that (an honest "could not check or do it"). See `planner.spec.md` ("Skipped-tool guard").
- When a native-tool model answers in prose before any tool has run in this reply, and the plan names a callable tool as a concrete call, the engine discards that answer and direct-executes the plan's tool steps, then the model answers from the results. Only literal, schema-valid steps run (no LLM repair). When the planned call cannot run, the held-back answer is still discarded: the planned call is recorded as a failed tool result and the model answers from that (an honest "could not check or do it"). See `planner.spec.md` ("Skipped-tool guard"). v2 one-step turns (no plan) get one nudge to call the tool instead, and when the model skips it again the turn is planned and guarded as above (`planner.spec.md`, "Intent judge (v2)").
Task-list planner (all model sizes, strongest impact on small models):
- The planner runs at the **front** of the reply flow (see step 3 above), not after tool selection. By the time the agentic loop starts, the plan already exists, the memory block has either been run or skipped based on the plan's `searchMemory` directive, and the tool allow-list has been derived from the tool names the plan referenced. See `planner.spec.md` for the prompt contract and fail-open semantics.

View File

@@ -209,6 +209,27 @@ COMPACT_ACTION_RULE: str = (
)
# Compact prompts, turns with a data tool: v1's tool guidance and constraints (about 900 Qwen3.6
# tokens) in a quarter of the words. Strata reads prompts in 2,048-token steps of about 0.4 s each,
# so these tokens decide whether a tool turn with recent conversation stays inside one step fewer.
# The same rules as v1's (acting instead of refusing, facts only from the result, looking up named
# things, auto-derived and self-contained arguments), without English examples.
COMPACT_TOOL_GUIDANCE: str = (
"Using tools: when a tool here covers the request, call it straight away, without announcing it "
"or asking permission; when none fits an action or a live fact, call toolSearchTool with a short "
"description before saying you cannot. Answer from tool results, never about them: every "
"specific fact in the reply (names, dates, numbers, places, cast, plot) must come from the tool "
"result or the user's own words, not from your prior knowledge; if the result lacks it, say so. "
"When a search result has a 'Content from top result' section, give its facts, not the links. "
"A specific named thing (a film, song, game, product, person, place or event) you cannot state "
"verified facts about is looked up before you answer. When a tool fills an argument itself (for "
"example the current location), call it with what you have instead of asking; a value the user "
"already gave in the conversation is passed. Free-text arguments such as search queries must "
"stand on their own, because the tool does not see earlier turns: resolve pronouns and "
"references into a short keyword phrase."
)
def tool_honesty_reminder(tool_names) -> str:
"""End-of-prompt restatement of the tool honesty rules for this turn.
@@ -307,14 +328,12 @@ def default_language_directive(language: Optional[str]) -> Optional[str]:
lang = (language or "").strip()
if not lang:
return None
# Short on purpose: the reply engine puts it at the front and the end of every v2 prompt.
return (
f"REPLY LANGUAGE: write your reply in {lang} by default, whatever language the user's "
f"message or any reference text is in. The one exception overrides this default: when "
f"the user explicitly asks, in any language, for the answer in another language (for "
f"example 'answer in English'), reply in that language, for just this reply or from now "
f"on in this conversation, as they asked. Do not mix in other scripts: no Chinese/Hanja characters, "
f"Japanese kana or Cyrillic letters inside a {lang} reply. (Numerals and unavoidable "
f"proper-noun brand names are fine.)"
f"REPLY LANGUAGE: reply in {lang}, whatever language the message or any reference text is "
f"in, unless the user explicitly asks, in any language, for another language: then use that "
f"one, for this reply or from then on, as they asked. No Chinese/Hanja, Japanese kana or "
f"Cyrillic inside a {lang} reply; numerals and brand names are fine."
)

View File

@@ -267,9 +267,10 @@ which carry Discord's own time.
Two reply options keep the time to the first word short (Strata reads every prompt from the
start, about 0.36 s per 2,048 tokens, and each LLM call before the answer adds its whole time):
the short voice persona (`compact_prompt=True`: the same butler in a tenth of v1's words, without
v1's single-user memory and on-screen browser rules, answers as short as possible, the full tool
guidance only on turns that offer a data tool), and the intent judge (`intent_judge=True`: the tool
router judges whether the turn needs tools and which, and the planner runs only when it does;
v1's single-user memory and on-screen browser rules, answers as short as possible, v1's tool
guidance in short only on turns that offer a data tool), and the intent judge (`intent_judge=True`: the tool
router judges whether the turn needs tools and which, and the planner runs only when it does
and the pick is not a lone one-step tool;
`planner.spec.md`, "Intent judge (v2)"). v1's screen-share browser (`stream_browser`) is off: it
needs v1's Node bot (M6 brings the v2 one), so its tools are not offered.
- Tools: the brain offers and runs no built-in tool in the jarvis settings' `disabled_tools` (the

View File

@@ -2,7 +2,7 @@
Strata reads every prompt from the start, so prompt length is reply latency. Behaviour: the compact
prompt is far shorter than v1's for the same turn, still carries the tool honesty rules and the
reply-language default, and brings the full tool guidance only on turns that offer a data tool.
reply-language default, and brings the (short) tool guidance only on turns that offer a data tool.
"""
from unittest.mock import Mock, patch
@@ -13,7 +13,9 @@ from src.jarvis.memory.conversation import DialogueMemory
from src.jarvis.memory.history import HistoryContext
from src.jarvis.reply.engine import run_reply_engine
from src.jarvis.reply.prompts.model_variants import TOOL_CONSTRAINTS_LARGE
from src.jarvis.system_prompt import COMPACT_ACTION_RULE, TOOL_HONESTY_RULES, default_language_directive
from src.jarvis.system_prompt import (
COMPACT_ACTION_RULE, COMPACT_TOOL_GUIDANCE, TOOL_HONESTY_RULES, default_language_directive,
)
from tests.test_ollama_num_predict import _mock_cfg
@@ -46,14 +48,23 @@ def test_the_compact_prompt_is_far_shorter_and_keeps_the_rules(routed):
@pytest.mark.unit
def test_the_full_tool_guidance_comes_only_with_a_data_tool():
def test_the_tool_guidance_comes_only_with_a_data_tool():
chatty = _system_prompt(True, [])
assert COMPACT_ACTION_RULE in chatty and TOOL_CONSTRAINTS_LARGE not in chatty
assert COMPACT_ACTION_RULE in chatty and COMPACT_TOOL_GUIDANCE not in chatty
weather = _system_prompt(True, ["getWeather"])
assert TOOL_CONSTRAINTS_LARGE in weather and COMPACT_ACTION_RULE not in weather
assert COMPACT_TOOL_GUIDANCE in weather and COMPACT_ACTION_RULE not in weather
def _intent_turn(prepared, routed=("getWeather",)):
@pytest.mark.unit
def test_a_data_tool_turn_gets_the_short_tool_guidance_not_v1s():
# v1 keeps its full guidance; the compact prompt says the same rules in a fraction of it.
assert TOOL_CONSTRAINTS_LARGE in _system_prompt(False, ["getWeather"])
weather = _system_prompt(True, ["getWeather"])
assert TOOL_CONSTRAINTS_LARGE not in weather
assert len(COMPACT_TOOL_GUIDANCE) < len(TOOL_CONSTRAINTS_LARGE) / 2
def _intent_turn(prepared, routed=("getWeather",), first=None):
cfg = _mock_cfg(512)
cfg.stream_browser = False
cfg.tool_selection_strategy = "llm"
@@ -62,7 +73,9 @@ def _intent_turn(prepared, routed=("getWeather",)):
patch("src.jarvis.reply.engine._extra_config", lambda key, default="": default), \
patch("src.jarvis.reply.engine.extract_text_from_response", return_value="네."), \
patch("src.jarvis.reply.engine.chat_with_messages") as mock_chat:
mock_chat.return_value = {"message": {"content": "네."}}
answer = {"message": {"content": "네."}}
mock_chat.side_effect = [first, answer] if first else None
mock_chat.return_value = answer
run_reply_engine(db=Mock(), cfg=cfg, tts=None, text="날씨 어때?", dialogue_memory=DialogueMemory(),
history=HistoryContext(), legacy_memory=False, intent_judge=True,
prepared_tools=prepared)
@@ -74,9 +87,16 @@ def _intent_turn(prepared, routed=("getWeather",)):
def test_the_intent_judge_plans_only_turns_that_need_a_tool():
# No tool needed: one router call, then the answer.
assert _intent_turn(None, routed=())[:2] == (0, 1)
# A tool turn: the planner orders the steps.
planner, router, tools = _intent_turn(None)
assert (planner, router) == (1, 1) and "getWeather" in tools
# A tool turn that may take several steps: the planner orders them.
planner, router, tools = _intent_turn(None, routed=("webSearch",))
assert (planner, router) == (1, 1) and "webSearch" in tools
# A lone one-step tool the model calls needs no plan (tests/test_engine_one_step_turn.py).
weather_call = {"message": {"content": "", "tool_calls": [
{"id": "c1", "type": "function", "function": {"name": "getWeather", "arguments": {}}}]}}
with patch("src.jarvis.reply.engine.run_tool_with_retries",
return_value=Mock(success=True, reply_text="13C", error_message=None, images=None)):
planner, router, tools = _intent_turn(None, first=weather_call)
assert (planner, router) == (0, 1) and "getWeather" in tools
@pytest.mark.unit

View File

@@ -0,0 +1,145 @@
"""One-step tool turns in the v2 brain skip the planner (planner.spec.md, "Intent judge (v2)").
When the router's only pick is a tool that does its job in one call (the weather, a meal log, a
look at the screen), the planner's ordered steps add nothing but a full LLM call before the answer.
The engine then gives the chat model the tool directly. Its skipped-tool guard, which reads the
plan, is replaced for these turns by one nudge: an answer given without the tool is dropped and the
model is asked once to call it.
"""
from __future__ import annotations
from unittest.mock import patch
import pytest
NATIVE_MODEL = "gpt-oss:20b"
FABRICATED = "서울은 지금 24도에 맑습니다."
GROUNDED = "서울은 지금 13도이고 맑습니다."
def _content(text: str) -> dict:
return {"message": {"role": "assistant", "content": text}}
def _native_call(name: str, args: dict) -> dict:
return {"message": {"role": "assistant", "content": "", "tool_calls": [
{"id": "call_1", "type": "function", "function": {"name": name, "arguments": args}}]}}
def _run(mock_config, db, dialogue_memory, *, routed, chat_replies, plan=("Reply to the user.",)):
from jarvis.memory.history import HistoryContext
from jarvis.reply import engine as engine_mod
from jarvis.tools.types import ToolExecutionResult
mock_config.ollama_chat_model = NATIVE_MODEL
mock_config.evaluator_enabled = False
mock_config.stream_browser = False
seen: list[list[dict]] = []
replies = iter(chat_replies)
def fake_chat(*args, **kwargs):
seen.append([dict(m) for m in (kwargs.get("messages") or args[2])])
return next(replies)
invoked: list[str] = []
def fake_tool(db, cfg, tool_name, tool_args, **kwargs):
invoked.append(tool_name)
return ToolExecutionResult(success=True, reply_text="Seoul: 13C, clear sky.", error_message=None)
with patch.object(engine_mod, "run_tool_with_retries", side_effect=fake_tool), \
patch.object(engine_mod, "chat_with_messages", side_effect=fake_chat), \
patch.object(engine_mod, "select_tools", return_value=list(routed)), \
patch.object(engine_mod, "plan_query", return_value=list(plan)) as planner:
reply = engine_mod.run_reply_engine(
db=db, cfg=mock_config, tts=None, text="서울 지금 우산 필요해?",
dialogue_memory=dialogue_memory, history=HistoryContext(), legacy_memory=False,
intent_judge=True)
return reply, invoked, seen, planner.call_count
@pytest.mark.unit
@pytest.mark.parametrize("routed", [["getWeather", "stop"], ["logMeal", "stop"], ["screenView"]])
def test_a_lone_one_step_tool_needs_no_plan(mock_config, db, dialogue_memory, routed):
_, _, _, planned = _run(mock_config, db, dialogue_memory, routed=routed,
chat_replies=[_native_call(routed[0], {}), _content(GROUNDED)])
assert planned == 0
@pytest.mark.unit
@pytest.mark.parametrize("routed", [
["webSearch", "stop"], # searches chain (who directed it, what else they made)
["schedule", "stop"], # list, then cancel the one asked about
["getWeather", "webSearch", "stop"], # two tools
], ids=["search", "schedule", "two tools"])
def test_turns_that_may_take_several_steps_keep_the_plan(mock_config, db, dialogue_memory, routed):
_, _, _, planned = _run(mock_config, db, dialogue_memory, routed=routed,
chat_replies=[_content(GROUNDED)])
assert planned == 1
@pytest.mark.unit
def test_a_tool_less_answer_on_a_one_step_turn_is_dropped_and_the_tool_called(mock_config, db, dialogue_memory):
reply, invoked, seen, _ = _run(
mock_config, db, dialogue_memory, routed=["getWeather", "stop"],
chat_replies=[_content(FABRICATED), _native_call("getWeather", {"location": "Seoul"}),
_content(GROUNDED)])
assert invoked == ["getWeather"]
assert reply == GROUNDED
# The made-up answer never reaches a later turn, and the retry names the tool.
assert not any(FABRICATED in (m.get("content") or "") for m in seen[1])
assert "getWeather" in (seen[1][-1].get("content") or "")
@pytest.mark.unit
def test_a_second_made_up_answer_falls_back_to_the_plan_and_its_tool(mock_config, db, dialogue_memory):
# Asked once, the model still answers without the tool: the turn takes the planned path, the
# skipped-tool guard runs the planned call, and no made-up answer reaches the user.
reply, invoked, seen, planned = _run(
mock_config, db, dialogue_memory, routed=["getWeather", "stop"],
plan=["getWeather location='Seoul'", "Reply to the user with the weather."],
chat_replies=[_content(FABRICATED), _content(FABRICATED), _content(GROUNDED)])
assert planned == 1
assert invoked == ["getWeather"]
assert reply == GROUNDED
assert not any(FABRICATED in (m.get("content") or "") for m in seen[-1])
@pytest.mark.unit
def test_a_question_back_goes_out_when_the_plan_needs_no_tool(mock_config, db, dialogue_memory):
# The late plan judges that no tool is needed (a question back to the user): it is delivered.
question = "어느 도시 날씨를 알려 드릴까요?"
reply, invoked, seen, planned = _run(
mock_config, db, dialogue_memory, routed=["getWeather", "stop"],
chat_replies=[_content(question), _content(question)])
assert planned == 1
assert invoked == []
assert reply == question
assert len(seen) == 2
@pytest.mark.unit
def test_a_model_that_calls_the_tool_is_not_asked_again(mock_config, db, dialogue_memory):
reply, invoked, seen, _ = _run(
mock_config, db, dialogue_memory, routed=["getWeather", "stop"],
chat_replies=[_native_call("getWeather", {"location": "Seoul"}), _content(GROUNDED)])
assert invoked == ["getWeather"]
assert reply == GROUNDED
assert len(seen) == 2
@pytest.mark.unit
def test_v1_turns_still_plan_a_one_step_tool(mock_config, db, dialogue_memory):
from jarvis.reply import engine as engine_mod
mock_config.ollama_chat_model = NATIVE_MODEL
mock_config.evaluator_enabled = False
with patch.object(engine_mod, "chat_with_messages", return_value=_content(GROUNDED)), \
patch.object(engine_mod, "select_tools", return_value=["getWeather", "stop"]), \
patch.object(engine_mod, "extract_search_params_for_memory", return_value={"keywords": []}), \
patch.object(engine_mod, "plan_query", return_value=["Reply to the user."]) as planner:
engine_mod.run_reply_engine(db=db, cfg=mock_config, tts=None, text="서울 지금 우산 필요해?",
dialogue_memory=dialogue_memory)
assert planner.call_count == 1