A user asked Fable to schedule "this Friday at 8am" on Wednesday 4/29
2026. The model picked 4/30 (Thursday) and confidently labeled it
"Friday." The TZ pipeline did everything correctly given the model's
date — the bug was upstream: the model was guessing weekdays from ISO
dates without an anchor, and the calendar tools had no way to verify.
Three layered fixes:
1. **System prompts now name the weekday alongside the ISO date.**
Both the journal-conversation prompt and the general chat prompt
used to say "Today is 2026-04-29 (America/New_York)." They now say
"Today is Wednesday, 2026-04-29 (...)." LLMs are unreliable at
deriving weekday names from ISO dates; supplying the name removes
the guess.
2. **`expected_weekday` parameter on create_event / update_event.**
When the model passes `expected_weekday="friday"`, the backend
computes the resolved start_date's weekday in the user's local
timezone and rejects mismatches with a self-correcting error
("Date 2026-04-30 falls on Thursday, not Friday. Recompute..."),
without creating the event. The check is local-aware: a Friday
23:00 event in Tokyo crosses midnight UTC but the local view
stays Friday, and the validator respects that.
3. **Tool descriptions instruct echo-and-confirm.** create_event and
update_event descriptions now tell the model: when the user names
a weekday, state the resolved date in the reply BEFORE calling
the tool, and pass `expected_weekday`. Costs nothing in code,
reinforces the validator.
6 new tests — match success, mismatch rejection (with create/update
not invoked), omitted-param backcompat, invalid weekday name, local-
not-UTC weekday computation, and the update_event variant. All 18
calendar-tool tests + 33 event-related tests pass; ruff clean.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Inspection showed only ONE record_moment call across the entire day's
journal — and that one had hallucinated person_ids. Multiple clear beats
went uncaptured: AP installation, going to watch a show with daughter,
decompressing-with-game.
The prior calibration said "use record_moment freely for meaningful
beats" — too soft. The model treated it as optional, especially when
already in chatbot-reply mode.
Rewritten: record_moment is now framed as the model's PRIMARY JOB. The
calibration includes an explicit checklist of what counts as a beat
(event, encounter, decision, observation, plan, feeling, accomplishment)
and an explicit instruction to call record_moment FIRST, before composing
the reply. Multiple beats → multiple calls. The ONLY skip case spelled
out: purely meta-conversational messages (acknowledgements, meta-asks
about prior tool results).
Tests on a fresh conversation will tell us if this moves the needle —
today's journal is poisoned by ten prior chatbot-flavored turns that
the model is pattern-matching against in its own history.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Per user clarification: previous over-rotation dropped the LLM-generated
prep prose entirely (just a phase greeting) and made the chat persona
extremely sparse ("you are a place where words go down"). User actually
wanted only the chat replies pulled back, NOT the prep dropped, and the
chat to behave largely like normal /chat — asking follow-ups and
verifying earlier details.
services/journal_prep.py — restored:
- _render_sections_for_prompt
- _PREP_SYSTEM_PROMPT (the direct, briefing-style prompt from 590a07b)
- _generate_prep_prose
- _fallback_prep_text
- ensure_daily_prep_message now calls _generate_prep_prose again
- removed _phase_for_now / _phase_prompt helpers (no longer needed)
services/journal_pipeline.py — persona rewritten:
- Old: "You are the user's journal. Be quiet. Listen. You are not helpful."
- New: "You are the user's assistant. Behave like the rest of the app's
chat: respond conversationally, ask follow-up questions, verify details
from earlier turns, use tools naturally."
- Calibration block reorganized: PEOPLE/PLACES (ask first), MOMENTS
(silent + use *_names), STATE-CHANGING TOOLS (confirmation flow),
OTHER, RESPONSE STYLE.
- RESPONSE STYLE keeps the no-apologizing / no-option-menus /
no-verbatim-repetition / match-user-length rules but drops the "be
quiet, one short sentence" framing.
Net behavior:
- Open journal → LLM-generated prep prose with today's tasks/events/weather
- Reply → assistant responds conversationally like /chat, asks follow-ups,
verifies details, uses tools
- Background: silently records moments via *_names, asks before creating
new people/places
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The prep was generating a multi-sentence recap of tasks/events/weather/
projects/recent moments via an LLM call. Per user direction, that's
redundant — the right-side widgets already show today's data — and the
verbosity made the journal feel chatty when the user wanted quiet.
Replaces the prep prose generator with a single phase-aware check-in
question (drawn from a static map: morning="How are you starting the
day?", midday="How's it going so far?", evening="How did the day shake
out?"). No LLM call. The structured `sections` are still gathered and
persisted on msg_metadata for provenance and possible future tooling
(e.g., search), they just don't render in the prep message.
Also pulls the journal persona way back. The prior framing pushed the
model toward stock therapy-template patterns ("I'm sorry you're feeling…"
+ numbered option lists). The new persona is "you are the user's journal —
listen, be quiet, stay out of the way." RESPONSE STYLE rules now lead the
calibration block and explicitly forbid:
- apologizing for the user's feelings
- offering to help / pitching tools
- multi-option menus
- verbatim repetition of prior replies
- padding short replies into paragraphs
Most replies should be one short sentence. Sometimes the right reply is
"Got it" + a record_moment tool call.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Three real bugs surfaced from inspecting today's journal turns:
1. record_moment was getting fed hallucinated person_ids (the LLM passed
[1, 2] instead of the IDs save_person had just returned). Result: the
moment was linked to two random old test-data notes ("test task 2",
"Tell a joke"), not the people the user actually mentioned.
2. The calibration rule "ask before save_person" was being silently
ignored — model just called save_person on first mention of Victoria
and Mother without asking the user.
3. The model produced a verbatim-identical reply to its previous turn when
the user mentioned "overwhelmed" twice — same numbered-list of 4
options, same closing line. The "warm listener / ask gentle questions"
persona was pushing toward stock therapy-template patterns.
Fixes:
services/tools/journal.py — record_moment now accepts *_names parameters
(person_names, place_names, task_titles, note_titles). Server resolves
each name to a note ID via case-insensitive title match, scoped by
note_type or task-status. *_ids parameters still exist but are now
documented as DISCOURAGED. The LLM physically cannot invent the wrong ID
when using names — names with no match are silently dropped. Resolution
happens via _resolve_entity_ids_by_name helper.
services/journal_pipeline.py — JOURNAL_PERSONA tightened (no more
"warm/curious listener" framing that pushed toward stock comfort
patterns). JOURNAL_CALIBRATION rewritten as scannable sections with
imperative language: PEOPLE/PLACES require asking before save_person;
TASK/NOTE state changes use the confirmation flow; MOMENTS are silent
but MUST use *_names not *_ids; OTHER notes the no-set_rag_scope and
no-auto-notes invariants. Added a RESPONSE STYLE section that explicitly
forbids verbatim repetition and stock multi-option menus.
After deploy, force-regenerate today's prep via fable_trigger_journal_prep
to also pick up the tighter prep prompt from 590a07b.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
bvandeusenandClaude Opus 4.7