ashish-sadh

Web-search for new domain knowledge relevant to the current coach's user, filtered by level, current focus, and open gaps. Propose findings as vocabulary additions, session topics, or program adjustments. Triggered on user request "research for me", "what's new in [domain]?", "learn something", or automatically by the weekly Sunday research scheduled job.

Scan vault files across People, Places, Restaurants, Events, and coach folders to find missing wikilinks between related files, and propose adding them. Triggered on user request "find links", "cross-link vault", "what connects to X?", or periodically after inbox-triage runs.

Process vault root Inbox.md entries and route each one to its proper home — People, Places, Restaurants, Events, coach folders, or Ideas. Triggered on user request "triage inbox", "process inbox", "clean up inbox", or when Inbox.md exceeds ~30 lines.

Periodic structural cleanup. Detects when files are too large, themes are emerging, or content is stale, and proposes restructuring. Triggers on user request "maintain the vault" / "clean up", or auto-suggested when last maintenance > 7 days. Always proposes before acting.

End-to-end creation of a new coach. Asks for domain and name, scaffolds local files from domain template, creates Drive folders, adds shell alias, runs onboarding interview, generates initial snapshot, outputs system prompt for phone Claude.ai project. Triggers on "spawn coach" / "create [domain] coach" / "add a [domain] coach".

Per-coach self-review. Every ~10 sessions, the coach reads its own session history and proposes content extractions, vocabulary additions, profile refinements, and structural changes. Triggers on user request "evolve this coach" or auto-suggested after threshold.

Generate a comprehensive snapshot from current local canonical state and upload to Drive as a new dated file. Triggered at end of substantive coach conversations, or on user request "save", "sync to Drive", "push snapshot".

At session start, pull phone-written inbox files from Drive that are newer than the most recent local-tracked sync timestamp, and integrate their contents into local canonical coach files. Triggered automatically by coach CLAUDE.md at conversation start, or on user request "refresh from Drive" / "pull from phone".

TestFlight publish recipe. Invoked by the watchdog as its next-session slot when ≥3h since last publish AND is-clean-state passes (NOT by launchd cron — the watchdog already serializes one claude -p at a time, so slotting TF into that queue gives zero concurrency with senior/junior xcodebuild or main-branch commits). Skips quietly if not clean or <3h since last publish, otherwise bumps build, archives, uploads, stamps state, updates releases.json. Haiku model — recipe is deterministic.

Deep planning for Drift autopilot. Reads + drains feedback, triages P0 bugs, processes design-doc backlog, researches the next arc (Explore sub-agent + optional WebSearch), and emits ONE primary epic per cycle. Standalone sprint-tasks no longer filed — bugs auto-route to routine-fix; permanents continue via their lifecycle. Invoked headlessly by the watchdog as `claude -p "/planning"`. Harness phase 3c per Docs/refactor/harness-phase-2-workflows.md.

Reply to admin comments on report PRs that arrived since the last run. Sonnet model, ~5k tokens per cycle. Was bundled inside /planning step 3; now its own skill spawned by the watchdog when `~/drift-state/last-admin-replies-at` is older than 24h. Keeps planning focused on the next epic.

Junior task executor (Sonnet model) for Drift autopilot. Same shape as /senior but verifier-only-debate, no design-doc routing, no security-review. For sweeps of mechanical changes, invoked via /batch from planning. Watchdog spawns `claude -p "/junior" --model sonnet`.

Senior task executor for Drift autopilot. Claims one SENIOR sprint-task, reads its <done_when> block as ground truth, implements, runs verifier debate via debate-moderator, commits only on PASS. Abandons on context budget exhaustion (never WIP-commits to main). Invoked headlessly by the watchdog as `claude -p "/senior"`.

Evaluator-optimizer for Drift UI commits. Reads `<visual_criteria>` from the parent epic, captures simulator screenshots via mobile-mcp, scores each criterion against the rendered pixels, and emits `<visual_verdict decision="PASS|REJECT">`. On REJECT, files a refinement sub-task under the parent epic so the next senior session picks up the gap. Invoked post-commit by senior when the diff touches Drift/Views/**/*.swift; harness phase 2c per Docs/refactor/harness-phase-2-workflows.md.

Screenshot Drift on the iOS simulator across all five tabs + key sheets, dump the UI hierarchy for each, and report visible issues (invisible text, broken layout, dark holdovers). Run after any UI change before declaring done, and especially before publishing TestFlight.

Sub-skill invoked by /senior when claimed task is design-doc-labeled. State machine — pending → write doc; in-review → respond to comments (lightweight or delegated to /autofix-pr for substantive); ready-to-close → verify all impl tasks closed + close parent. Never claims independently; always called from /senior.