{"slug":"davidondrej-agentic-productivity-setup","name":"agentic-productivity-setup","description":"Build or rebuild a private, local-first macOS system that measures daily Git commits, active AI-agent sessions, and instruction prompts, then sends aggregate charts to Discord. Use when someone asks to set up an agentic productivity tracker, install daily agent reports, recreate this architecture on a Mac, or understand how to build it safely. Differentiator: creates a fresh system from generic contracts and never copies another person''s data, paths, credentials, repositories, prompts, or private configuration.","long_description":"---\nname: agentic-productivity-setup\ndescription: 'Build or rebuild a private, local-first macOS system that measures daily Git commits, active AI-agent sessions, and instruction prompts, then sends aggregate charts to Discord. Use when someone asks to set up an agentic productivity tracker, install daily agent reports, recreate this architecture on a Mac, or understand how to build it safely. Differentiator: creates a fresh system from generic contracts and never copies another person''s data, paths, credentials, repositories, prompts, or private configuration.'\n---\n\n# Agentic Productivity Setup\n\nBuild a deterministic system that measures whether AI agents increase a\nperson's output over time. Keep collection local. Send one private Discord\nreport each morning with exactly three aggregate charts.\n\n## Privacy rules\n\nTreat these as hard requirements:\n\n1. Build a fresh implementation. Never copy another installation, runtime\n database, log, session store, credential, machine path, repository list, or\n private configuration.\n2. Never persist or transmit prompt text, responses, tool output, repository\n names, file paths, identities, raw session IDs, or commit messages.\n3. Read native agent stores only to count events. Keep temporary identifiers in\n memory, then discard them. Persist a one-way hash only when a collector needs\n a cross-run baseline.\n4. Render charts locally by default. Only Discord should receive the daily\n totals and chart images. Explain the exact data flow before enabling any\n optional remote chart service.\n5. Store the Discord webhook in the current user's macOS Keychain.\n Never put it in source code, a command argument, an environment file, the\n database, logs, tests, or the scheduler definition.\n6. Use synthetic fixtures in tests. Never use copied session data, real prompts,\n real repository metadata, or live credentials.\n7. Open native databases read-only. Collectors must not modify agent state.\n8. Report missing, unreadable, or unsupported sources as coverage problems.\n Never turn collection failure into a silent zero.\n\n## Confirm the scope\n\nFirst determine whether the user wants an explanation, a new project, or an\ninstallation. Do not create or install anything when they only asked how it\nworks.\n\nFor a build, confirm or safely default these values:\n\n- Platform: macOS. Ask before adapting the design to another operating system.\n- Code root: `$HOME/code`.\n- Timezone: the Mac's configured timezone.\n- Report time: 08:00 local time.\n- Report window: 90 days ending yesterday.\n- Agent harnesses: only tools the user wants measured.\n- Project location and reverse-DNS LaunchAgent label.\n\nDo not ask the user to paste a webhook into chat. Configure it through secure\nterminal input after the application is built.\n\nAsk before changing metric definitions, storage boundaries, report\ndestinations, or data sent off the Mac. Record consequential choices in a short\nADR inside the generated project.\n\n## Architecture\n\nUse this data flow:\n\n```text\nGit reflogs + native agent registries\n -> read-only collectors\n -> daily aggregate counts\n -> local SQLite database\n -> local chart renderer\n -> one Discord webhook request\n```\n\nKeep the source checkout separate from the installed runtime:\n\n- Copy application code into `~/Library/Application Support//app/`.\n- Store aggregate state beside it in `metrics.sqlite3`.\n- Store logs in `~/Library/Logs//`.\n- Install the plist in `~/Library/LaunchAgents/`.\n- Give directories mode `0700` and sensitive files mode `0600`.\n- Preserve aggregate state across reinstalls and normal uninstalls.\n\n## Metric contracts\n\nImplement three independent metrics. Do not create a combined productivity\nscore.\n\n### Unique local commits\n\n- Discover Git repositories recursively under the configured code root.\n- Deduplicate primary checkouts and linked worktrees by Git common directory.\n- Read local creation events from reflogs for the requested date range.\n- Include only commits whose author or committer email matches an identity\n configured in that repository.\n- Count each commit hash once across all refs and worktrees.\n- Exclude fetched commits, pushes, and branch movement by themselves.\n- Keep hashes in memory only. Store the final daily count.\n\n### Active agent sessions\n\n- Count one session on each local calendar day where its native registry records\n activity.\n- Include GUI, CLI, headless, resumed, parent, subagent, delegated, and automated\n sessions.\n- Deduplicate with the harness's native session identity in memory.\n- Keep each harness separate in storage and reports.\n\n### Instruction-bearing prompts\n\n- Count stored user, system, and developer inputs that contain instructions.\n- Include human prompts, automation, setup context, delegation, and subagent\n instructions.\n- Exclude assistant responses, tool results, empty inputs, and duplicated storage\n copies.\n- Inspect content only long enough to classify the event. Never store or log it.\n\nConvert every timestamp into the configured timezone before assigning a day.\n\n## Collector design\n\nUse one adapter per harness. Prefer native registries over process inspection,\nshell history, window titles, or guessed file timestamps.\n\nSupport common source shapes:\n\n- JSON or JSONL session records.\n- Read-only SQLite registries.\n- Editor global or workspace state databases.\n- Authenticated native CLI export when no readable local registry exists.\n\nEach adapter must return:\n\n- daily unique session identities in memory;\n- daily prompt counts;\n- `full`, `partial`, `unavailable`, `absent`, or `error` coverage;\n- a short detail string containing counts and limitations, never private data.\n\nSome stores expose a session total but no timestamp for each turn. For those\nstores, save a hashed source key and total, then attribute only positive deltas\nobserved after the local baseline. Mark older attribution as partial.\n\nDiscover installed harnesses at runtime. Keep unsupported tools absent from the\nchart, not mislabeled as zero activity.\n\n## Aggregate database\n\nUse SQLite in WAL mode. Keep the schema small:\n\n- `daily_metrics`: day, metric, harness, count, collection time.\n- `collector_health`: installation and coverage state per harness.\n- `source_snapshots`: hashed source key, last total, observation time.\n- `collector_baselines`: first reliable observation per harness.\n- `deliveries`: report day, sending state, sent time, and a short safe error.\n\nUse monotonic upserts for daily counts. A later incomplete scan must not erase a\nhigher value already stored.\n\nNever create tables for raw events, messages, prompts, responses, repository\nnames, paths, or identities.\n\n## Report and Discord delivery\n\nBuild exactly three charts for the selected window:\n\n1. Daily unique commits as a line and area chart.\n2. Daily active sessions as stacked bars split by harness.\n3. Daily instruction prompts as stacked bars split by harness.\n\nAdd a straight ordinary least-squares trendline to each chart. Use combined\ndaily totals for the session and prompt trendlines. Render PNG files locally\nwith a pinned charting dependency. Never silently fall back to a network\nrenderer.\n\nThe Discord message should contain only:\n\n- the report day;\n- yesterday's three totals;\n- a short collector coverage summary;\n- the three PNG attachments.\n\nSend one multipart POST. Disable allowed mentions. Accept only valid Discord\nwebhook hosts. Use a short timeout. Never log the request URL or body.\n\nClaim the report day in SQLite before network work. Mark it sent only after a\nsuccessful Discord response. Normal runs must send once per report day. Require\nan explicit force option to resend.\n\n## Scheduling\n\nUse a user LaunchAgent. Do not run an LLM on the schedule.\n\nConfigure:\n\n- `StartCalendarInterval` for the chosen daily report time.\n- `StartInterval` of 300 seconds when a collector needs prompt deltas.\n- `RunAtLoad` for login and wake catch-up.\n- `ProcessType` set to `Background`.\n- `LowPriorityIO` enabled.\n- A restrictive `077` umask, represented as decimal `63` in the plist.\n- stdout and stderr paths inside the private log directory.\n\nEvery interval may observe delta-only sources. Before the report time, stop\nafter that local observation. At or after the report time, collect yesterday,\nbuild the full window, and send only if that day is not already marked sent.\n\nThe installer must render absolute runtime paths into the plist, validate it\nwith `plutil`, replace the existing job safely, and start it. Keep credentials\nout of the plist.\n\n## Project shape\n\nUse a small, readable project:\n\n```text\nagentic_productivity/\n cli.py\n collectors.py\n database.py\n model.py\n reporting.py\nbin/\n agentic-productivity\nlaunchd/\n