Weekly Status Reports: aggregate a week's entries into a manager-facing report

[artcashin]

Aggregates a week's journal entries into a status report a developer can send to their manager — narrative, accomplishments, per-project breakdown, and items still outstanding. Generated from a user-editable prompt template, stored with version history, exported as markdown. Triggered from the CLI, a weekly launchd job, or a button in the web UI.

Draft — stacked. This branch is based on the OpenCode/model-provider work, so its diff against main currently shows ~79 commits. Only 13 are this feature. It contains #19 and #22; it needs llm.ts from #20. The diff collapses to those 13 commits once those merge. Review commits from 75d0b0c onward, or wait for the base to land.

Why the prompt is a file, not a constant

Different managers want different things, and the person best placed to say what a status report should contain is the manager receiving it. report_template_url points at a markdown file; edit it once and a whole team's reports change shape. Placeholders: {{week_label}}, {{week_start}}, {{week_end}}, {{entries}}, {{open_questions}}, {{project_list}}. Unknown placeholders are left untouched, so a typo degrades one line rather than killing the run.

Every successful fetch is cached. If the URL is unreachable the cached copy is used. If there is no cache, generation fails rather than silently falling back to the shipped default — a differently-shaped report going out under someone's name is worse than no report.

Design decisions worth a reviewer's attention

• Versioned, never overwritten. A report may already have been sent, so the record of what was sent matters. Each generation is a new row; latest wins by version DESC.
• Markdown is the source of truth. A ## heading parse sits on top as a convenience and is allowed to come back empty — a user-authored template is under no obligation to use headings.
• Outstanding items are judged, not just collected. The model weighs open questions against the week's later entries and drops ones that were resolved, then lists what it dropped under "Resolved This Week" so the judgement is auditable.
• Single model call. The busiest measured week is ~11k characters; chunking would be premature.
• 05:15 Monday, not midnight. With day_start_hour at 5 the previous logical week does not close until 05:00 Monday, so a midnight job would report on an incomplete week and freeze it that way.

Verified against real data, not just tests

265 tests pass, typecheck clean. Beyond that:

• A live model call produced a real five-section report for 2026-W31 from actual journal entries.
• Both trigger paths exercised independently — v1 from the CLI, v2 from the web button — which also demonstrates versioning appends rather than overwrites.
• The launchd job is installed and was triggered: it computed the last completed week on its own, ran under launchd's minimal environment, exited 0, wrote nothing to stderr.
• The template URL chain was exercised against a real HTTP server in all four states: fetch succeeds and caches; server down with cache falls back; server down without cache fails hard with exit 1; and a custom remote template visibly reshaped the model's output. The database records which was used per report (template_source).

Known limitations, stated rather than discovered

• resolveTemplate has no fetch timeout. A URL that accepts the connection and never responds blocks the run. AbortSignal.timeout() would close it.
• The status endpoint returns HTML, so the React button in the follow-up plan will need JSON or a second route.
• Journal text is interpolated into the prompt without delimiting. Those strings are LLM output over session transcripts that can contain third-party text; this is the first feature that forwards them to someone else.
• The job registry is an in-memory Map, deliberately. The report commits before the job is marked done, so a restart loses job status but never a report.

The React button is not in this PR. It depends on the React frontend and is a separate plan.

🤖 Generated with Claude Code