Measure how long work in Claude Code ACTUALLY took, from your own session transcripts, and forecast or SCHEDULE the next one from measured history instead of vibes. Use BEFORE GIVING ANY TIME ESTIMATE OR DEADLINE, when someone asks "how long will this take" or "can we ship by <date>", when you want to know how many hours a past project really consumed, when tracking progress mid-build, and (optionally, after `--setup`) when pricing billable work or checking whether an engagement made money. Triggers "how long will X take", "how many hours did X take", "can we do X by <date>", "is that enough time", "what % are we through", "how much time have I spent on X", "what did this cost", "what should I quote", "am I underpricing".
Capture current session state to a curated artifact at .agent-docs/now/handoff.md plus an archive entry at .claude/handoffs/<timestamp>-<slug>.md. Updates now/status.md, now/work-plan.md, now/open-questions.md, log.md to current reality; sweeps the inter-party obligations ledger both directions (owed-by / owed-to, each receivable carrying a trigger + a default-if-silent) — the standalone now/obligations.md when it exists, else a handoff Obligations section; on multi-party installs, sweeps live inter-agent room threads into a Room-threads continuity block (participants/roles, message anchors, standing hooks, the re-read rule). Accepts optional session notes as arguments. Use at natural checkpoints, before compaction, before session-end, or when context exceeds ~80%. Does NOT auto-commit.
Read .agent-docs/now/handoff.md and surface current session state. Use at session start, after time away, or when reorienting ("where are we?"). Pulls live git state at invocation, verifies the handoff isn't stale against git reality (code ahead of docs → trust the code), runs a quick state-file health sweep (the folded-in /status check), surfaces obligations-ledger deltas — landed / came due / overdue / dangling — from now/obligations.md when it exists, else the handoff's Obligations section, and on multi-party installs runs the Room-threads resume step (read the handoff's continuity block, refresh from the comms log since the handoff timestamp, apply the re-read rule before any post).
Reconcile an installed Fieldbook copy against a NEWER kit zip. Reads .agent-docs/.kit-manifest.json, 3-way diffs ONLY kit-owned files (skills/hooks/rules/templates/schema — never the colleague's content dirs), and uses sha256-vs-manifest so a file the colleague edited is treated as "theirs" (shows a diff, never overwrites). Bumps the recorded kit-version. Reconciles the obligations-ledger form (ADR-0012) against re-detected comms posture — promoting a handoff `## Obligations` section into a standalone now/obligations.md, or retiring an empty file back to a section, content-preserving and manifest-recorded. When the manifest is MISSING but the tree is kit-lineage-shaped (hand-seeded / pre-kit port), emits a retro-adoption plan and backfills the manifest instead of stopping. Use when the author sends a newer Fieldbook zip, when a new kit's kit-version.txt is ahead of the installed one, or when upgrading/repairing an existing install. Consent-gated; shows a plan before writing.
Guided concierge that installs Fieldbook into your repo — interview, tailor, install, verify, teach. Use when setting up Fieldbook, adopting the .agent-docs context system, or running a kit install / upgrade / repair / add-module / uninstall.
Mid-session housekeeping. The main agent updates .agent-docs/now/{status,work-plan,open-questions}.md + log.md to current reality, then continues working in the same session. Does NOT regenerate handoff.md, does NOT archive, does NOT auto-commit. Use when work has progressed and now/* has drifted but you're NOT ending the session.
Re-runnable Fieldbook health check. Verifies the Claude Code version against the kit's assumptions, that the hooks fire and core.hooksPath is set, that every hook script has a tracked settings.json registration (and every registration a script), that jq/python3/stat are present, that the doc-schema + index lint are clean, and that the install manifest matches what's on disk (flagging drift/missing files). Reports findings and SUGGESTS fixes but mutates nothing without consent. Use when hooks seem inert, after a Claude Code update, after a fresh clone, or any time you want to confirm the install is still live and wired.
Run a tiered, adversarially-gated research investigation under .agent-docs/research/<R-id>-<slug>/. Codified tiers — 00-landscape → NN-<track> deep-dives → MANDATORY adversarial gate (a clean-context skeptic re-checks decision-critical claims against PRIMARY sources BEFORE synthesis) → synthesis.md (the only Tier-2-loaded surface). Use for any multi-source investigation whose conclusion will feed a decision (an ADR, a build choice, a dependency adoption). Currency-check every decision-critical claim against current-year primary docs, never training data.