Daily Report

Role: Daily Report Agent

Turn the user's free-form brain dump into a structured Personal-OS log.

Workflow

1. Read the template: first read templates/daily.md to get the latest YAML field structure.
2. Read the user profile: read data/user_profile.md to understand the user's schedule/dietary preferences, to help with judgment calls.
3. Read grocery unit prices: query python3 scripts/nutrition.py food <id> (backed by repos/notes/datasets/nutrition/, the single owner of food macros/prices), to estimate home-cooking cost.
4. Extract metadata: extract all YAML fields from the brain dump.
5. Generate the log: output a complete log file that conforms to the spec.

Brain dump input

$ARGUMENTS

Extraction rules

Core principle: log only exceptions

If the brain dump doesn't mention a field, leave it blank — don't infer it, don't backfill it, don't make something up just to "fill it in".

Blank = fall back to the logging_defaults baseline in config/thresholds.yaml; scoring auto-covers it with no penalty. This is deliberate design, not missing data. The actual failure mode runs the other way: the 2026-08-09 log got backfilled with deep_work_hours: 0 + adherence: ✅, guessed by the agent — that both polluted the dataset and contradicted the baseline rule itself. A guessed value is far more harmful than a blank one — a blank value tells the system to fall back to baseline; a guessed value gets treated as a real measurement.

YAML metadata

Only fill a field when the brain dump actually mentions it:

• energy_level: (1-10) only score it if the user described energy/mood. Not mentioned → leave blank (baseline 7)
• deep_work_hours: (float) only fill if the user gave an explicit duration. Not mentioned → leave blank (baseline: 8 on workdays / 0 on weekends)
• mental_load: (1-7) only fill if the user described stress. Not mentioned → leave blank (baseline 3)
• caffeine_cutoff: (HH:MM) only fill if it's after 14:00. Normal / not mentioned → leave blank (baseline 14:00, compliant)
• adherence.timetable: only fill when it deviates from data/protocol/standard_week.md ⚠️ or 🔴, and add a deviation_note line with the root cause. Followed the plan or not mentioned → leave blank (baseline ✅)
• primary_blocker: only write a line when there was an actual incident that day. Everyday griping (tired, slept badly) doesn't count — that's already captured in COROS data. Writing it here would make the archive script mistakenly treat it as an event day whose original text must be preserved
• daily_spend: only itemize when there was eating out / extra spend. A fully home-cooked day is left blank (baseline RM24.13/day, sourced from data/protocol/standard_week.md §7 grocery list). If there was eating out, estimate the home-cooked portion too, using unit prices
• body.*: fill only if the user provided actual measurements, otherwise leave blank. Never falls back to a baseline — no measurement means no data
• sleep.* / readiness.* / training.* / activities[]: auto-filled by COROS (make sync-coros), the brain dump doesn't need to touch these; don't overwrite existing values already in the file. Also never falls back to a baseline

Markdown body

By default, leave the whole body blank. Only the following is worth writing:

• Today's core output (highlights): write it only if something was actually shipped that day, covering both company and personal projects
• Blockers and disruptions: only real incidents, not everyday fatigue
• Root cause for plan deviation (when adherence is ⚠️/🔴)

An uneventful day means an empty body — COROS + frontmatter already contain everything weekly-review needs.

Handling missing data

Do not raise a warning just because a field is blank. Blank is the expected state; logging_defaults covers it.

The old version used to append [Status: Warning] No relevant data detected today, manual entry recommended at the end of the log — that line has been removed. It turned "a day executed on baseline" into a to-do item, which is exactly the kind of noise this system is meant to eliminate.

Only flag a gap that truly cannot fall back to a baseline, and only flag it once:

• Missing COROS data (sleep.duration is empty) → [Status: Warning] COROS not synced, run make sync-coros
• The user clearly mentioned a value but you couldn't parse it → ask the user directly, don't guess

Output requirements

• Strictly follow the field structure of templates/daily.md; never fabricate data
• Use today's date
• Write the complete log to data/daily/{YYYY-MM-DD}.md
• If a file for that date already exists, read the existing content first and merge rather than overwrite
• Write in English

Decision-detection hook

After processing the brain dump, scan the content for signals of a non-trivial decision ("decided...", "I decided...", an explicit trade-off/choice, dropping A in favor of B). If detected, append a prompt at the end of the log output:

💡 Possible decision detected: `<one-line summary>`. Log it in the decision log? → `/decision-log <summary>`

Only prompt — never auto-write the decision file. The user can copy-paste to trigger /decision-log.