journal

name	journal
description	Generate the daily work journal — collect today's AI usage across Claude Code, Copilot CLI, and Codex CLI plus git activity into a digest, then write the narrative, technical, and next-day-plan summaries. Use when the user asks for their work journal, daily summary, "what did I do today", or to plan tomorrow.
allowed-tools	Bash, Read, Write

Daily work journal

Resolve the plugin root before shelling out: use ${CLAUDE_PLUGIN_ROOT} if it is set; otherwise resolve ../.. relative to this SKILL.md to an ABSOLUTE path (bash cwd is not the skill dir). Call that root $ROOT below.

Collect the digest

Run the collector to build today's digest from local, read-only sources:

python3 "$ROOT/bin/journal_collect.py"

Useful flags: --date YYYY-MM-DD for a specific day (default: today), --repo PATH to add a git repo to scan (repeatable), --print to echo the digest to stdout as well as writing it. The collector is deterministic and read-only over ~/.claude, ~/.copilot, and ~/.codex — it never calls a model and never touches the network. It writes only under the gitignored journal/<date>/ directory (digest.json), nowhere else. The digest now also deepens Codex with the day's rollout files and, when tokens are found, carries a clearly-labeled API-equivalent relative-burn proxy under sources.codex_cli.extra.codex_proxy — never a bill (billed_usd stays null), never counted into usd_priced.

Persist today's telemetry (right after collecting)

Capture the day's analytics into the durable store — the transcript dirs the digest reads from ROTATE, and this snapshot is what survives:

python3 "$ROOT/bin/telemetry_snapshot.py"

It imports the analytics modules and calls their builders directly — read-only over ~/.claude, ~/.codex, and ~/.copilot, spawning no CLI and writing only under the gitignored telemetry/ (dated envelopes, one per source per day; same-day re-runs overwrite, except that a failed re-run never replaces a good envelope). The envelope's filename date is the LOCAL day — the same day this journal is keyed on, so an evening run files under today, not tomorrow (captured_at inside each envelope stays UTC). Best-effort: if it fails, note the failure and continue with the journal — the snapshot never blocks the summaries.

Write the summaries (in this session — default)

Get the exact prompts for today's digest:

python3 "$ROOT/bin/journal_summarize.py" --date <date> --dry-run

This prints the three prompts (narrative, technical, next-day-plan) without dispatching any model — the current session is already paid for, so write the documents yourself instead of spawning a nested call. Read journal/<date>/digest.json for the facts, follow each printed prompt's required headings exactly, and use only digest facts — no invented details, no transcript text. Save your three drafts as journal/<date>/narrative.md, journal/<date>/technical.md, and journal/<date>/next-day.md. When done, summarize the three documents for the user and link their paths — do not paste full drafts back into chat. When signals.harness is present, follow the next-day prompt's ## Harness plan section — per-task harness + model tier + ready-to-paste command + one-line WHY, advisory only.

Or run it headless

For an unattended run (e.g. from the nightly schedule) let the script dispatch a model itself:

python3 "$ROOT/bin/journal_summarize.py" --date <date>

This tries a routed cheap/mid model tier first, with at most one escalation to the next tier, and exits with code 3 if any of the three documents still failed after that escalation. Use --start-tier to pick the starting tier and --model to pin one model id explicitly (this skips escalation).

Inbox & schedule

Drop meeting notes or email to-dos into journal/inbox.md as plain lines; the next collection folds them into the digest so they resurface as next-day to-dos. To run the collector and summarizer automatically every night, install the schedule:

python3 "$ROOT/bin/journal_schedule.py" install

This writes a launchd plist (default 22:00) and prints the launchctl bootstrap command needed to activate it — the installer never runs launchctl itself, so loading the schedule is the user's own later, manual step. Use uninstall to remove the plist, status to check whether it is loaded, and run (with --collect-only or --dry-run) for a manual one-shot in between scheduled runs.

Next-day runbook

Once a digest exists, build tomorrow's runbook — a dated, checkable plan of next-day tasks:

python3 "$ROOT/bin/journal_plan.py" build

It writes journal/plan/<date>.md — one dated, checkable card per planned task (drawn from open kit tasks, WIP repos, the inbox, and journal/plan/seed.md), each with a What/How, an ideal-harness line, and ready-to-paste commands for Claude Code, Copilot CLI, and Codex CLI with cost estimates from the pricing files (Codex figures are API-equivalent proxies, never a bill). Unchecked cards from earlier days carry forward automatically.

Enrich the What/How steps in this session (the current session is already paid for — the summaries precedent): run

python3 "$ROOT/bin/journal_plan.py" prompt

follow the printed prompt exactly (rewrite ONLY the What/How bodies, keep every other line byte-identical — the H1, checkboxes, field lines, and Harness blocks), and save the revised document back over the same journal/plan/<date>.md path.

On the day tasks are due, check and track them:

python3 "$ROOT/bin/journal_plan.py" check
python3 "$ROOT/bin/journal_plan.py" done <id> --date <file-date>
python3 "$ROOT/bin/journal_plan.py" defer <id> --to <date> --date <file-date>

(or just edit the checkboxes in the markdown by hand — the file is yours). On quiet days, hand-add lines to journal/plan/seed.md (same plain-line format as the inbox) and rebuild.

Advisory only: the runbook prepares and tracks — it never schedules or executes anything. Every command above is text you choose to run, and there is no scheduler here by design (the user tabled it).

External tools (Teams / Outlook / Copilot Studio)

The journal has no Graph/OAuth/MCP connectors and never will by default — instead it generates an offline ask-the-tools pack you run yourself, in your own Microsoft tools, two passes:

python3 "$ROOT/bin/journal_askpack.py" --date <date> --print

This writes journal/<date>/ask-the-tools.md with one ready-to-paste prompt per tool (Copilot Studio, Teams, Outlook) and also prints them to stdout. Run each printed prompt inside that tool's own AI, then paste the bullet results it gives you back into journal/inbox.md. Re-run the collector and redo the summaries so the enriched inbox flows into the digest. This is offline text generation only: the journal never adds network, OAuth, Graph, or MCP calls to fetch this content — you carry it over by hand. Each prompt asks for at most 15 subject-level bullets per tool (titles, people, decisions, action items — no message bodies).

Privacy

Writing the summaries sends the digest — project and repo names, commit subjects, kit task titles, and any inbox text — to a model. The digest itself is metadata-only: it never carries transcript or message text, only counts, ids, titles, and short strings like the above. Everything the journal produces stays under the gitignored journal/ directory; nothing here is committed to git. Pasted ask-the-tools bullets become inbox text, so they also travel to the model when you write the summaries.