pickup

Pickup — Resume Work

Config

Vault path + GitHub task repo: read from ${CLAUDE_PLUGIN_ROOT}/brain-os.config.md (or user-local ~/.brain-os/brain-os.config.md which takes precedence). Keys: vault_path:, gh_task_repo:. In shell/hook contexts, source hooks/resolve-vault.sh and use $VAULT_PATH / $GH_TASK_REPO. In skill prose, substitute $GH_TASK_REPO below with the configured value.

Invocation modes

• /pickup — no args: interactive HITL flow over owner:human Ready list (handovers first, then top task)
• /pickup <N> — direct claim of issue #N, any owner; skips Ready-list filtering entirely (explicit override)

/pickup is HITL-only. The previous /pickup auto background spawner has been removed; AFK queue drainage now lives in /impl auto (see /impl skill). See § Mid-flow rule below for which work belongs on /pickup vs /impl.

Mid-flow rule

Some work is fundamentally human-in-the-loop and MUST stay on /pickup, never /impl. These are tasks where automation produces work the user will throw away:

• Handovers (type:handover) — resuming requires reading the doc, verifying state, possibly resuming an incomplete grill. AFK workers cannot resume an interrupted grill.
• Incomplete grills — a grill session marked INCOMPLETE / NOT completed needs the human's judgment on the open questions, not a bot guessing.
• Mid-draft writing — anything in a writing pipeline (brainstorm → fb-writer → substack-writer → checker → visual). Each stage takes human steering. Never AFK.

Rule: any of the above MUST carry owner:human. /impl auto skips them by label filter. /pickup no-args picks them up (it filters owner:human). /pickup <N> accepts any owner — use the bare-integer form when you (the human) explicitly want to resume a non-owner:human issue interactively.

/pickup <N> — Direct claim of a specific issue

When the user passes a bare integer (e.g. /pickup 100, cr /pickup 100), treat it as "claim this issue and start." Do NOT run the interactive Ready-list flow. Owner label is not checked — this is the explicit override for any-owner direct claim.

1. Claim atomically via the canonical helper (single source of truth):

   bash "${CLAUDE_PLUGIN_ROOT}/scripts/pickup-claim.sh" N
   

   Exit-code semantics:

   • 0 → flip happened (status:ready → status:in-progress); proceed.
   • 3 → already status:in-progress (no-op); for /pickup <N> continue anyway since user explicitly chose this issue.
   • 1 → not claimable (closed, blocked, backlog); abort with the helper's stderr message.
   • 2 → gh API / config error; surface and abort.

2. Fetch full body for context (the helper only inspects labels):

   gh issue view N -R $GH_TASK_REPO --json title,body,labels
   

3. Load context:

   • Labels include type:handover → extract the relative [handover](daily/handovers/…) link from the body, read that file, follow the "Handover Task Behavior" section.
   • Otherwise → treat the issue body as the spec; read any linked grill / plan docs referenced in the body.

4. Start executing immediately. Never re-grill locked decisions.

/pickup — Interactive flow (no args)

1. Query open Ready HITL tasks from GitHub. /pickup is human-in-loop only: filter to owner:human so AFK-eligible work stays in /impl's lane and doesn't show up here:

   gh issue list -R $GH_TASK_REPO --state open --label status:ready --label owner:human --json number,title,labels,body
   

   Labels carry all metadata (priority, weight, owner, type) — do not parse title for markers. To pick up a non-owner:human issue interactively, use the explicit /pickup <N> form.

2. Handovers first: filter the Ready list to issues carrying type:handover. Do NOT infer handovers from title prefixes or body wiki-links — the label IS the signal.

   • Found → claim → load linked handover doc → start
   • Multiple → show list, ask user which one
   • None → continue to step 3

3. Ready tasks remaining?

   • Yes → skip to step 4 (no groom needed)
   • No → groom Backlog: query gh issue list -R $GH_TASK_REPO --state open --label status:backlog --json number,title,body, promote unblocked + clear items to Ready via bash "$CLAUDE_PLUGIN_ROOT/scripts/gh-tasks/transition-status.sh" N --to ready (the central transitioner is the single source of truth — atomic, validated, idempotent); ask user about vague ones

4. Autogrill Ready tasks (see Autogrill section below) — skip type:handover issues (they already have a spec doc)

5. Claim and start top remaining task — for the chosen issue N:

   bash "${CLAUDE_PLUGIN_ROOT}/scripts/pickup-claim.sh" N
   

   On exit 0 → begin work. On exit 3 (already in-progress) → skip to next candidate. On exit 1 or 2 → surface error to user.

Claim protocol

Single source of truth: ${CLAUDE_PLUGIN_ROOT}/scripts/pickup-claim.sh <N>. All /pickup entry points (/pickup, /pickup <N>, plus the interactive Ready-list flow) call this helper for the atomic flip. /impl and /impl auto also call it for AFK claims. Do NOT re-implement the flip in skill prose — the recurring bug is LLMs skipping the flip when described step-by-step (worker spawned, label still status:ready, /status mis-classifies). The script is deterministic and untouched by judgment.

What the script does:

1. Resolves $GH_TASK_REPO from hooks/resolve-vault.sh.
2. Fetches issue state + status label (one gh issue view call).
3. If status:in-progress → no-op (exit 3).
4. If status:ready (or no status:* label) → atomic flip to status:in-progress by shelling out to the central transitioner (bash scripts/gh-tasks/transition-status.sh N --to in-progress), which issues a single gh issue edit removing status:ready and adding status:in-progress from one process invocation (exit 0).
5. If status:blocked / status:backlog / unknown → abort (exit 1).
6. On gh / config error → exit 2.

The flip is atomic by GH semantics — no ephemeral claim:session-* labels. They were overengineering for a single-user repo and caused "label not found" errors because gh issue edit --add-label won't auto-create.

Race note: two sessions reading status:ready simultaneously and transitioning in the same second will both succeed (GitHub has no label CAS). In practice — solo user, a handful of sessions — this is rare and benign: duplicate work at worst, and the handover / issue body converges the outputs. The TOCTOU window is the helper itself (1-3 ms between view + edit) — much smaller than orchestrator-prose timing.

On completion — close the issue via the central helper (strips status:* labels first):

bash "$CLAUDE_PLUGIN_ROOT/scripts/gh-tasks/close-issue.sh" N

Step 2: Autogrill Ready Tasks

For each non-type:handover Ready task, run vault-first Q&A:

1. Generate grill questions — the same questions you'd ask a human:

   • What's the goal?
   • What context exists?
   • What are the constraints/requirements?
   • What's the expected output?
   • Task-specific questions based on the domain
   • Aim for 5-10 questions per task

2. Search vault for each answer — grep, read wiki-links, follow references

3. Log the Q&A to {vault}/daily/grill-sessions/YYYY-MM-DD-autogrill.md:

   ## HH:MM — <task name> (#N)
   
   ### Q&A
   
   | # | Question | Answer | Source |
   |---|----------|--------|--------|
   | 1 | What is X? | Found: ... | [[path/to/source]] |
   | 2 | How does Y? | ❌ NOT FOUND | — |
   
   ### Gaps (needs human)
   - Q2: How does Y?
   
   ### Decision
   - N/M answered → auto-close / escalate
   
   ### Artifacts
   - Files created/modified
   

4. Decide per task:

   • ≥80% questions answered → auto-close: produce the artifact, bash "$CLAUDE_PLUGIN_ROOT/scripts/gh-tasks/close-issue.sh" N, log process
   • <80% answered → escalate: surface specific unanswered questions for human grill

5. Run tasks in parallel — launch multiple agents for independent tasks

6. Present results:

   Autogrill complete:
   ✅ #12 [Task A] — auto-closed (7/8 questions answered)
   ❓ #15 [Task B] — needs you on 4 questions:
      - Q3: What pricing model?
      - Q5: What's the timeline?
      - Q7: Who is the target?
      - Q8: What channels?
   

After human answers gaps:

When the user grills on the remaining questions:

1. Write answers back to vault — update the relevant context/strategy/goals files
2. Re-run autogrill on the task — should now auto-close with full answers
3. The vault gets smarter every cycle

/pickup auto — Launch tasks in background (DEPRECATED — see /impl auto)

/pickup auto has been removed. AFK queue drainage now lives in /impl auto (which wraps /ralph-loop around /impl). The pickup-auto.sh script and its launchd cron have been deleted.

If a user types /pickup auto, do NOT spawn anything. Print verbatim and exit:

/pickup auto is deprecated. Use /impl auto for AFK backlog drainage.

Migration:
  - /pickup auto         → /impl auto              (wraps /ralph-loop around /impl)
  - /pickup auto --force → /impl auto              (no time-aware weight filter; /impl ignores weight:heavy distinction)
  - /pickup auto --dry-run → gh issue list -R $GH_TASK_REPO --state open --label "status:ready,owner:bot"

/pickup is now HITL-only — see § Mid-flow rule in this skill.

Background: see daily/grill-sessions/2026-04-25-skill-process-pocock-adoption.md and parent story ai-brain#145. Rationale — /pickup and /pickup auto overlapped on the Ready-claim+spawn surface, creating confusion about whether AFK work and HITL handovers could coexist on the same skill. The split: /pickup owns HITL (owner:human only), /impl owns AFK (owner:bot, /tdd-driven, trunk-gated).

Handover Task Behavior

For each handover issue (type:handover) found:

1. Extract handover doc path from the issue body — look for a relative markdown link like [handover](daily/handovers/YYYY-MM-DD-topic.md) and read that file from the vault completely (all sections)
2. Run "Commands to Verify State" from the handover to confirm nothing changed
3. Check for incomplete grills — scan handover for "INCOMPLETE", "NOT completed", "grill session started but". If found:
   • Do NOT start executing tasks
   • Present the incomplete grill status
   • Resume the grill FIRST using /grill <topic> before any implementation
4. Present brief summary:

   Picking up: [topic] (#N)
   Status: [status]
   TL;DR: [from handover]
   
   Next steps:
   1. [step 1]
   2. [step 2]
   
   Starting with step 1.
   

5. Start executing step 1 immediately (don't wait for confirmation unless blocked)

After completing all steps:

1. Close the issue (central helper strips status:* labels then closes): bash "$CLAUDE_PLUGIN_ROOT/scripts/gh-tasks/close-issue.sh" N
2. Update handover status in the vault doc: status: completed
3. If new unfinished work emerged: run /handover to create a new handover (which opens a new issue)

Rules

Do NOT re-ask settled decisions

The handover's "Decisions Made" section = requirements. Never re-grill.

Always verify state first

Run the "Commands to Verify State" before starting. Files may have changed.

Stale handover warning

If handover is > 7 days old:

⚠️ This handover is [N] days old. Verifying current state first...

GitHub Issues are the task queue

• /handover opens an issue in $GH_TASK_REPO with type:handover + status:ready + handover doc link in body
• /pickup queries $GH_TASK_REPO for Ready issues, filters type:handover first
• State lives on the GH server (atomic labels) — no file merge conflicts across sessions
• The user just types /pickup and work resumes

Always re-query at render time

Labels change across sessions. Never cache the Ready list within a long interactive session — re-query before claiming.

Owner label hygiene (what belongs on owner:bot vs owner:human)

owner:bot means /impl auto will claim it (AFK queue drainage via /tdd, trunk-gated). owner:human means a person must grill + design before execution; /pickup no-args picks these up. Getting this wrong creates false "work in progress" signals — owner:bot auto-promotes tasks that actually need judgment, they stagnate in In Progress with no work happening.

Categories that ALWAYS need owner:human (👤):

• Seed pages (people, companies, meetings) — needs judgment about who counts, role mapping, relationship type
• Context gap fills (strategy.md, goals.md, business.md) — needs strategic input from user
• Replicate spec across N skills (outcome log markers, trace patterns) — each skill needs per-skill review of where to insert
• Airtable / external sync design — needs decisions about schema mapping, conflict handling, sync direction
• Cron schedule design for new automations — needs safety review + frequency tradeoffs
• Skill creation or modification — skills are design-heavy (behavior, edge cases, integration). Always start with /grill to align on requirements before writing any code. After /grill aligns on requirements, route the actual creation/eval work through /skill-creator — the gold standard for skill creation AND behavioral-eval generation (variance analysis, triggering accuracy, with-skill-vs-baseline subagent runs). Don't hand-edit evals.json as a shortcut for behavioral-eval upgrade; coverage-parity additions matching the existing pattern (1-2 cases mirroring new SKILL.md content, e.g. when /improve auto-edits a skill) are fine. /skill-creator and /improve are complementary: /improve is the auto-tuning loop on outcome-log signal; /skill-creator is the manual, interactive flow for creation and behavioral-eval generation. Auto-executing a "build skill" task without grilling produces work that doesn't match the user's intent.

Default: When creating new tasks, default to owner:human unless the task is a pure mechanical transform (file rewrite per spec, code refactor with clear before/after, format conversion).

Flow

Session A:
  /handover → creates handover doc + opens GH issue (status:ready, type:handover, owner:human)

Session B (new Claude):
  /pickup → queries GH for status:ready + owner:human → filters type:handover first
        → claims (status:ready → status:in-progress) → loads handover doc → starts working
  ... work done ...
  bash "$CLAUDE_PLUGIN_ROOT/scripts/gh-tasks/close-issue.sh" N

Session B with explicit issue (any owner):
  /pickup 100 → skip list, directly claim #100 → load body/handover → start

Background AFK drainage (was /pickup auto, now /impl auto):
  /impl auto → wraps /ralph-loop around /impl → queries GH for status:ready + owner:bot
            → /impl runs /tdd on each → commits + pushes + closes issue

Outcome log

Follow skill-spec.md § 11. Append to {vault}/daily/skill-outcomes/pickup.log:

{date} | pickup | {action} | ~/work/brain-os-plugin | N/A | commit:N/A | {result}

• action: pickup (interactive), pickup:N (direct claim of issue N), or auto:deprecated (user invoked /pickup auto — log the deprecation event, do not spawn)
• result: pass if issues started/completed, partial if some escalated to user, fail if unrecoverable error (e.g. gh auth failure or network)
• No eligible tasks (empty Ready column, no owner:human issues, or all already in-progress) → log pass | score=0/0, not fail
• Optional: args="{issue-number}"