backfill-journal-body-context

name: backfill-journal-body-context description: Walks every daily journal entry in a date range (default this year) and appends a "Body track" section BELOW the original verbatim content. Pulls health-mcp data for each date (HRV, RHR, sleep, cycle phase, lab status, recovery/sleep/strain scores) and weaves a Floor-paired interpretation. Idempotent (skips entries that already have the section). Use when user says /backfill-journal-body-context, asks to enrich journals with body data, says "backfill my journals with health" or wants existing journal entries paired with their Apple Health / Oura / Fitbit data retroactively.

Reads existing daily journals, pulls health-mcp data for each entry's date, and appends a "Body track" section BELOW the original verbatim content. The original entry text is NEVER modified — the rule from feedback_journal_verbatim_words.md is non-negotiable.

When to use

User says /backfill-journal-body-context
User wants their existing journals enriched with body data (HRV, sleep, recovery, cycle phase) retroactively
User wants to see what /weekly and /monthly would surface if they had been pulling body data all along
After running /health-setup for the first time and importing a backfill window of biometric data

Do NOT use for:

Brand-new journal entries (daily-journal already pulls body context via health-context skill)
Editing journal content beyond appending the body section
Anything that touches the original verbatim journal body

How it works

Determine the date range. Default: --year <current-year> (Jan 1 to today).
Find journal entries in that range using [VAULT_PATH]/Meta/journal-index.json (rebuild if stale).
For each entry:
- Read the file
- Check if it already has a ## Body track (health-mcp, backfilled YYYY-MM-DD) section — if yes, skip (idempotent)
- Call health_journal_context(date, voice_profile="warm") for the data + rendered prose
- Call health_cycle_context(date) if cycle data exists (for women's cycle awareness)
- Call health_recovery_score(date) + health_sleep_score(date) for the scores
- Call health_lab_panel(date, lookback_days=180) for any out-of-range markers active that period
- Pair the body data with the entry's floor tag (from frontmatter floor_level + floor)
- Render the body-track section using the template below
- Append BELOW the original content with a clear divider
Print summary: N entries processed, M backfilled, K skipped.

Body track template

Append below the original journal content, after a blank line + horizontal rule:


---

## Body track (health-mcp, backfilled {{today_iso}})

*Auto-generated context. Original journal entry above is preserved verbatim.*

**Floor that day:** {{floor_name}} ({{floor_level}})

**Cycle phase** (if cycle data exists): {{phase}}, cycle day {{cycle_day}}{{ — irregularity flag if any}}

**The body that day:**
- HRV: {{hrv_ms}} ms ({{hrv_delta_pct}}% vs 30-day baseline)
- RHR: {{rhr_bpm}} bpm
- Sleep: {{sleep_asleep_min}} min ({{sleep_efficiency}}% efficiency, REM {{rem_min}}min, deep {{deep_min}}min)
- Steps: {{steps_total}}
- Workouts: {{workout_count}} ({{workout_min}}min)
- Mindful: {{mindful_min}}min

**Scores:**
- Recovery: {{recovery_score}}/100 ({{confidence}} confidence)
- Sleep: {{sleep_score}}/100

**Floor-paired interpretation** (1-3 sentences, see synthesis rule below): {{interpretation}}

**Lab markers** (if any out-of-range result from the prior 180 days, paired with today): {{lab_flags}}

Synthesis rule (the helpful-not-just-cool part)

The interpretation line must follow the same shape as the /weekly section 0d synthesis. Each line:

Names the pattern (cycle phase explaining HRV, under-fueling masking recovery, anniversary coupling, etc.)
Hypothesizes what it might mean
Suggests a specific next-action when applicable

Examples (template SHAPES, not for verbatim use):

Floor was Fear and HRV ran 22% below baseline on a luteal day. The body and the mind both registered the threat. Without the journal entry the recovery score would have called this "rest more"; pairing with Fear shows the actual signal was "the worry is doing work the rest can't fix."

Floor was Joy after a strong gym week. HRV at baseline, sleep efficiency 94%. This is the high-water mark — note what conditions produced it.

Floor was Apathy and Vitamin D 25-OH came back at 26 ng/mL (below range) the month before. The mood floor may have had a metabolic floor under it. Worth a re-test after 3 months of supplementation.

Banned shapes (from /weekly section 0d, same rules apply):

Listing body numbers without an interpretation
Generic "rest more / eat better / hydrate" advice
Treating recovery score as ground truth when cycle phase or under-fuel explains the dip
Pretending in-range labs cause symptoms

Model routing

The interpretation line is grunt-work prose. Use the cheapest model that produces a helpful sentence. Order of preference:

Python template (zero cost) — for high-confidence cases (HRV in normal range during follicular, no out-of-range labs, no anniversary signal). The script scripts/backfill-journal-body-context.py covers this.
MiniMax (~$0.06/M tokens, very cheap) — for cases where the data shows a real pattern that deserves a synthesized sentence. Invoke via "⚙️ Meta/scripts/minimax.sh" if the vault has it, otherwise skip the LLM step and use a fallback template.
Haiku (cheap, fast, reliable) — for cases where MiniMax is unavailable.
Sonnet — only if explicit --high-quality flag passed and the user accepts the cost.

The default is Python template + MiniMax fallback. Do NOT default to Sonnet for hundreds of journal entries.

Invocation

The actual work runs in scripts/backfill-journal-body-context.py. The skill assembles arguments and hands off to the script.

When invoked:

Parse arguments:
- --year YYYY (default: current year)
- --start YYYY-MM-DD and --end YYYY-MM-DD (override year)
- --vault-root PATH (default: $VAULT_ROOT or autodetect from cwd)
- --llm-model {python,minimax,haiku,sonnet} (default: python with minimax fallback)
- --dry-run (print what would change without writing)
- --force (overwrite an existing body-track section)
Sanity checks:
- health-mcp must be registered. If not, abort with setup instructions.
- Run health_status() to confirm there's biometric data in the DuckDB. If the count is zero, abort and suggest /health-setup first.
- The vault must have a Meta/journal-index.json and a journal folder. Rebuild the index if stale.

Run the script:

/usr/bin/python3 "[REPO_ROOT]/scripts/backfill-journal-body-context.py" --year 2026 --vault-root "$VAULT_ROOT"

Surface the summary: N entries processed, M backfilled, K skipped, plus the date range covered.

Output

The skill does NOT write to the vault itself — the Python script does the file mutations. The skill only:

Validates inputs
Calls health-mcp tools to verify data exists
Invokes the Python script
Reports the result

Idempotency

The Python script checks each journal file for an existing ## Body track (health-mcp, backfilled line. If present, skip unless --force. Re-running the skill on the same range is safe.

Ongoing daily run

After the initial backfill, the same script runs daily for yesterday's entry via a scheduled task (use the /schedule skill). Suggested cadence: 7am local, after Apple Watch has uploaded the previous night's sleep data.

Scheduled-task body:

/usr/bin/python3 "[REPO_ROOT]/scripts/backfill-journal-body-context.py" --start "$(date -v-1d +%F)" --end "$(date -v-1d +%F)" --vault-root "$VAULT_ROOT"

That's "backfill yesterday, every morning, forever." Set it up after the initial backfill completes successfully.

Voice rules

The auto-generated body-track section is in warm register (narrative sentences, not clinical exact-number dumps)
For Spanish journals, render the section in Spanish (the script detects via the journal's frontmatter language: field if present)
Floor names use the appropriate language alias ([[Joy]] in English / [[Alegría]] in Spanish)
Never modify the original entry's content. Only append.

Privacy

The script reads journal files + the local DuckDB. It writes only to journal files (appending body-track sections). No data leaves the machine. The interpretation step (if LLM-backed) sends ONLY the structured body data + floor tag to the chosen model — never the journal body content.