distribute-tokens

${var} — Phase + target selector. Grammar: [plan:|all:][dry-run:]<target>

• `` (empty) / <label> / dry-run:<label> → send phase: distribute a list from memory/distributions.yml (empty = first list). [default — no prefix]
• plan: / plan:<week> / plan:dry-run / plan:dry-run:<week> → plan phase only: compute rewards from the repo's merged PRs and write the list into memory/distributions.yml.
• all: / all:<week> / all:dry-run / all:dry-run:<week> → plan then send in one run.

<label> is a distribution-list label (e.g. contributors-2026-W17). <week> is an ISO week (2026-W17); empty <week> = most recent completed ISO week. dry-run: previews without side effects (no yml/state writes in plan; no transfers in send).

Why this design

This skill owns the whole contributor flywheel: who deserves what (plan) and moving the money (send). It is split into two phases that can run independently or chained.

Plan phase — the wiring the project was missing. Merged PRs already name the people moving the project, but shipped work had no path to a wallet credit. The plan phase is that wiring: it ranks contributors by the PRs they merged in the target week (straight from the GitHub API), prices each eligible contributor against a tier table, and writes a labelled list into memory/distributions.yml — the exact file the send phase reads. Keeping a human-visible diff on memory/distributions.yml between plan and execution is the cheapest possible audit trail when real money is involved: the plan lands in git, and the operator (or all: mode, or a chained step) runs the send next.

Send phase — this moves real money. The biggest failure mode is double-sending (re-runs, retries after partial failures, day-rollover bypass of "skip if today" logic) or sending into a black hole (no preflight balance, deprecated API path, missing handle resolution). The send phase therefore:

1. Persists per-recipient idempotency state in memory/state/distributions.json keyed on (list, recipient, date_utc) with the txHash. A successful transfer is never re-sent within the same UTC day, even across re-runs or workflow restarts.
2. Two-phase execution: RESOLVE (validate config, key, balance, resolve all handles → addresses, build plan) → EXECUTE (send each transfer, persist state after each one). RESOLVE failures abort before any send.
3. Dry-run mode outputs the full plan with no transfers.
4. Wallet API only for actual transfers — Bankr's docs deprecate the Agent API for transfers. Agent API is used only for handle→address resolution.

The two phases stay decoupled by design: the send phase is the only sanctioned transfer path and owns the idempotency state file, so the plan phase never touches transfer state. all: mode simply runs plan then send in sequence (plan writes the list, send reads it), so nothing is re-implemented.

Config

Reads two independent config/state surfaces depending on phase:

• memory/distributions.yml — the distribution lists (read by send, written by plan).
• The plan phase computes its ranking live from the repo's merged PRs via the GitHub API — no input file (see Phase A).
• memory/state/distributions.json — per-recipient send idempotency (read/written by send).
• memory/state/contributor-reward-state.json — plan idempotency + first-PR-bonus history (read/written by plan).

If memory/distributions.yml is missing when the send phase needs it, bootstrap with a commented template (see Send Step 1) and exit cleanly with DISTRIBUTE_TOKENS_OK — bootstrapped distributions.yml; edit and re-run.

# memory/distributions.yml
defaults:
  token: USDC          # USDC | ETH (Base only)
  amount: "5"
  chain: base

lists:
  contributors:
    description: "Weekly contributor rewards"
    token: USDC
    amount: "10"
    recipients:
      - handle: "@alice_dev"      # Twitter/X — resolved via Bankr Agent API
        amount: "15"
      - handle: "@bob_builder"
      - address: "0x742d...5678"  # direct EVM address — preferred path
        label: "Charlie"
        amount: "20"

Required secrets

Token addresses on Base

• USDC: 0x833589fcd6edb6e08f4c7c32d4f71b54bda02913
• ETH (native): tokenAddress: "0x0000000000000000000000000000000000000000", isNativeToken: true

Tier pricing (plan phase)

First-PR bonus: +5 USDC, additive, applied once-ever per login (tracked in memory/state/contributor-reward-state.json). Rewards landing your first merged upstream PR — the highest-leverage signal in the leaderboard scoring.

Eligibility floor: score ≥ 10 AND the contributor must own a non-empty @handle (logins without @ prefix in the table are skipped — bots and parsing artifacts). A single merged upstream PR (+10) qualifies — the goal is to reward shipped work, not gate on volume.

Default token: USDC on Base. Operator can override per-recipient amounts in memory/distributions.yml after the plan is written if a special bonus is warranted.

Read memory/MEMORY.md and scan the last ~3 days of memory/logs/ for anything already reported (don't re-report the same signal).

Step 0 — Parse the selector and dispatch

Shell shape. Each Bash call is one allowed command (pipes are fine). The run refuses a call that sets a variable (today=$(date ...), JOB=$(./secretcurl ...), RESP=$(./secretcurl ...)), uses if / for / while or [ ... ] tests, calls sleep or sed, or joins commands with ;, && or ||, and each refusal costs a model turn. So get a date with a bare date call and write the printed value into later commands; write the repo, week bounds, logins, handles, job ids, addresses and amounts into each command literally (the <TODAY>, <JOB_ID>, <ADDR> style placeholders in the snippets); poll a Bankr job by repeating one call, never a loop; wait with a one-line python3 -c "import time; time.sleep(N)" call; create and change files under memory/ with Write/Edit.

Resolve time anchors up front with one bare call:

date -u +%F

It prints today; today_utc is the same value. Later steps write it in literally as <TODAY> (shown as ${today} / ${today_utc} in the templates below).

Parse ${var}:

1. Phase prefix. If ${var} starts with plan: → PHASE=plan, strip plan:. Else if it starts with all: → PHASE=all, strip all:. Else → PHASE=send and do not strip anything (the remaining legacy grammar is parsed by the send phase itself).
2. Dry-run. For PHASE=plan/all: if the remainder starts with dry-run (optionally dry-run:), set MODE=dry-run and strip that token; else MODE=execute. (For PHASE=send, the send phase parses dry-run: itself — see Send Step 1.)
3. Target.
   • PHASE=send: the (unstripped) var is the send target — dry-run:<label> or <label> or empty.
   • PHASE=plan/all: the remainder is an optional <week>. If it matches ^\d{4}-W\d{2}$, set TARGET_WEEK=<week>; else run date -u +%G-W%V as its own Bash call and take the printed value as TARGET_WEEK (ISO-8601 week-numbering year + week - %G/%V not %Y/%U, so Monday-anchored weeks roll over correctly across years).

Dispatch:

• PHASE=plan → run Phase A only.
• PHASE=send → run Phase B only.
• PHASE=all → run Phase C (A then B).

Selector examples: `` → send first list · contributors-2026-W17 → send that list · dry-run:contributors-2026-W17 → dry-run send · plan: → plan most recent leaderboard · plan:2026-W17 → plan that week · plan:dry-run → plan preview · all: → plan + send most recent · all:dry-run:2026-W17 → full end-to-end preview for that week.

Phase A — Plan (reward computation)

Ranks the target week's merged-PR authors (GitHub API) and turns that ranking into a tier-priced list in memory/distributions.yml.

A1. Determine the target week and repo

• REPO - the running instance's repo. Run echo "$GITHUB_REPOSITORY" as its own call; if it prints owner/repo, that is REPO. If it prints an empty line, run git config --get remote.origin.url as its own call and take the owner/repo part of the URL yourself (drop the host prefix and any trailing .git).
• TARGET_WEEK comes from the selector; empty = the most recent completed ISO week (the last full Mon–Sun). Compute its UTC bounds WEEK_START..WEEK_END as ISO datetimes (YYYY-MM-DDT00:00:00Z).

A2. Rank contributors by merged PRs in the week

Compute the ranking directly from GitHub — no upstream skill or article required.

• Fetch every PR merged inside the window, by author (one call, with REPO, WEEK_START and WEEK_END written in literally): gh api -X GET search/issues -f q="repo:<REPO> is:pr is:merged merged:<WEEK_START>..<WEEK_END>" --paginate --jq '.items[].user.login'
• Drop bot authors (*[bot], dependabot*, github-actions*). Count each remaining login's merged PRs → score. Rank by score descending; tie-break by earliest merge time, then login ascending.
• First-PR ✨ per ranked login - did they have any prior merged PR to the repo? One call per login, the login written in literally: gh api -X GET search/issues -f q="repo:<REPO> is:pr is:merged author:<LOGIN> merged:<<WEEK_START>" --jq '.total_count' → 0 means this is their first-ever merged PR (set first_pr_marker = ✨).
• If zero merged PRs in the window → log CONTRIBUTOR_REWARD_NO_MERGED_PRS — week ${TARGET_WEEK} to memory/logs/${today}.md, exit silently (no notify). Nothing shipped, nothing to reward.
• If the GitHub API is unreachable (see Network note for the gh api → WebFetch fallback) → log CONTRIBUTOR_REWARD_API_FAIL, notify the operator, exit.

A3. Load plan idempotency state

// memory/state/contributor-reward-state.json
{
  "weeks": {
    "2026-W17": {
      "written_at": "2026-04-26T09:00:00Z",
      "label": "contributors-2026-W17",
      "source": "github:merged-prs",
      "rewards": [
        { "login": "alice_dev", "rank": 1, "score": 47, "amount": "25", "first_pr_bonus": false },
        { "login": "bob_builder", "rank": 2, "score": 31, "amount": "20", "first_pr_bonus": true }
      ]
    }
  },
  "first_pr_bonus_paid": ["bob_builder", "carol_eng"]
}

Bootstrap with {"weeks": {}, "first_pr_bonus_paid": []} if the file doesn't exist.

A4. Compute the plan

For each ranked login with rank ≤ 5 AND score ≥ 1 (at least one merged PR):

• Look up base_amount from the tier table (rank 1→25, 2→15, 3→10, 4-5→5).
• If first_pr_marker == "✨" AND login ∉ first_pr_bonus_paid → set first_pr_bonus = true, amount = base_amount + 5. Otherwise first_pr_bonus = false, amount = base_amount.
• Build row: { rank, login, score, base_amount, first_pr_bonus, amount }.

If weeks[TARGET_WEEK] already exists in state → this week was already processed. Diff the current plan against state.weeks[TARGET_WEEK].rewards keyed on login:

• If diffs are empty (same logins, same amounts) → log CONTRIBUTOR_REWARD_ALREADY_PROCESSED — week ${TARGET_WEEK}, exit silently (no notify). Idempotent re-run.
• If diffs exist (leaderboard re-ran after first reward write — late tweet bumped a score, etc.) → flag RE_PROCESS. Continue but don't re-pay anyone already in state.weeks[TARGET_WEEK].rewards; add only the deltas. New entries get full reward; existing entries with bumped amounts get the delta (e.g. moved from rank 3→2 = additional 5 USDC top-up). Demoted entries are not clawed back.

If the plan is empty (zero eligible contributors after threshold + dedup) → log CONTRIBUTOR_REWARD_NO_ELIGIBLE and exit silently.

A5. Render the plan

Contributor Reward Plan — ${TARGET_WEEK} (${MODE})

Source: ${LEADERBOARD_FILE}
Tier: rank 1=25, 2=15, 3=10, 4-5=5 USDC; first-PR bonus +5 once per login.

  ✓ #1 @alice_dev      score 47  →  25 USDC                  [NEW]
  ✓ #2 @bob_builder    score 31  →  20 USDC (15 + 5 first-PR)[NEW + BONUS]
  ✓ #3 @carol_eng      score 24  →  10 USDC                  [NEW]
  ✓ #4 @dave_ops       score 18  →   5 USDC                  [NEW]
  ↻ #5 @eve_hax        score 14  →   5 USDC                  [DEDUP — already in state]

Total to write: 60 USDC across 4 new entries.
Total in state for ${TARGET_WEEK} after write: 5 entries, 65 USDC.

Next: distribute-tokens "dry-run:contributors-${TARGET_WEEK}" (preview)
      distribute-tokens "contributors-${TARGET_WEEK}"          (execute)

If MODE=dry-run (plan-only dry-run, i.e. plan:dry-run...): notify this plan with header *Contributor Reward Plan — ${TARGET_WEEK}* — DRY RUN, log to memory/logs/${today}.md, exit CONTRIBUTOR_REWARD_DRY_RUN. Do not touch memory/distributions.yml or the state file.

all: mode note: when this phase runs as part of all: with MODE=dry-run, do not notify here and do not exit — hand the computed plan rows straight to Phase B (see Phase C). When all: runs with MODE=execute, continue through A6–A8 normally but replace the trailing Next: line in the A9 notification with Distributing now (phase=all)….

A6. Update memory/distributions.yml (skipped when MODE=dry-run)

Read memory/distributions.yml. If missing → bootstrap with the standard header (matching the send-phase bootstrap style):

# memory/distributions.yml
defaults:
  token: USDC
  amount: "5"
  chain: base

lists:

Compute the new list block:

  contributors-${TARGET_WEEK}:
    description: "Weekly contributor rewards for ${TARGET_WEEK} (auto-generated from merged-PR ranking)"
    token: USDC
    amount: "5"
    recipients:
      - handle: "@alice_dev"
        amount: "25"        # rank 1
      - handle: "@bob_builder"
        amount: "20"        # rank 2 + first-PR bonus
      - handle: "@carol_eng"
        amount: "10"        # rank 3
      - handle: "@dave_ops"
        amount: "5"         # rank 4

Recipient ordering matches plan order (rank ascending). Per-recipient amount is required so the send phase picks up the tier-priced value rather than falling back to the list default.

Update strategy:

• If a list named contributors-${TARGET_WEEK} already exists in the YAML, replace it wholesale (the plan is the authoritative current state).
• Otherwise append the block under lists: (preserving existing lists — never rewrite them).
• Use a YAML-aware update: first check the existing file parses with one call, python3 -c "import yaml; yaml.safe_load(open('memory/distributions.yml'))", then make a careful block replacement with Edit keyed on the ^ contributors-<TARGET_WEEK>:$ line (or insert the block under lists: with Edit). If YAML parse fails on the existing file → log error, do not write, notify the operator (file is hand-edited; auto-edit would clobber).

Verify the write by re-reading the file and confirming the list is present and has len(recipients) == len(plan).

A7. Update plan state file (skipped when MODE=dry-run)

Atomically write the updated state JSON to memory/state/contributor-reward-state.json:

• Set weeks[TARGET_WEEK] = { written_at: now_utc, label, source: "github:merged-prs", rewards: [{login, rank, score, amount, first_pr_bonus}, ...] } (full replacement on RE_PROCESS, otherwise additive).
• Append any logins where first_pr_bonus == true to first_pr_bonus_paid (deduplicated).

Write the whole updated JSON in one Write call (never piecemeal edits or a shell redirect) so partial writes can't corrupt state.

A8. Notify (plan-only execute)

*Contributor Reward Plan — ${TARGET_WEEK}*

Wrote ${N_NEW} new entries (${TOTAL_USDC} USDC) to memory/distributions.yml as `contributors-${TARGET_WEEK}`.

Top of plan:
  #1 @alice_dev   — 25 USDC
  #2 @bob_builder — 20 USDC (✨ first-PR bonus)
  #3 @carol_eng   — 10 USDC
  #4 @dave_ops    —  5 USDC
${IF_DEDUP}

Source: ${LEADERBOARD_FILE}
First-PR bonuses awarded: ${LIST_OR_NONE}

Next: run `distribute-tokens dry-run:contributors-${TARGET_WEEK}` to preview, then drop the `dry-run:` prefix to execute.

Plan: https://github.com/${GITHUB_REPOSITORY}/blob/main/memory/distributions.yml

Suppress the ${IF_DEDUP} line when no entries were deduped. Use $GITHUB_REPOSITORY env var for the link target. Send via ./notify.

Significance gate: notify only when N_NEW ≥ 1. Re-process runs that produced zero new entries (RE_PROCESS with all rewards already paid) → silent log only. (In all: execute, skip this notify per the A5 note; the send-phase summary carries the report.)

Then log (see Log) and exit CONTRIBUTOR_REWARD_OK.

Phase B — Send (on-chain distribution)

Reads a list from memory/distributions.yml and executes transfers via Bankr Wallet API. This phase moves real money — idempotency and preflight are non-negotiable.

Read memory/state/distributions.json (if present) for send idempotency state before doing anything.

B1. Parse the send target and load config

• If the send target starts with dry-run:, set MODE=dry-run and LABEL=${target#dry-run:}. Otherwise MODE=execute and LABEL=${target}. (When entered from Phase C, LABEL and MODE are set by Phase C instead — see Phase C.)
• If memory/distributions.yml missing → Bootstrap: write the example config from the Config section (commented out so it's inert), notify DISTRIBUTE_TOKENS_OK — bootstrapped distributions.yml; edit and re-run, log, exit.
• Parse YAML. If LABEL empty, use the first list. Else find the matching list. If not found → notify DISTRIBUTE_TOKENS_ERROR — list '${LABEL}' not found, log, exit.
• (today_utc was resolved in Step 0.)

B2. Pre-flight: key, write access, balance

If BANKR_API_KEY not set → DISTRIBUTE_TOKENS_ERROR — BANKR_API_KEY not configured, log, exit.

./secretcurl -fsS "https://api.bankr.bot/wallet/me" -H "X-API-Key: {BANKR_API_KEY}"

It prints the /wallet/me body (ME); on an HTTP error, curl's -fsS prints the status code instead.

• HTTP 403 → DISTRIBUTE_TOKENS_ERROR — API key is read-only; needs wallet write scope, exit.
• HTTP 429 → DISTRIBUTE_TOKENS_ERROR — rate-limited at /wallet/me; aborting, exit.
• Network failure → use WebFetch fallback. If still failing → DISTRIBUTE_TOKENS_ERROR — Bankr /wallet/me unreachable, exit.

./secretcurl -fsS "https://api.bankr.bot/wallet/portfolio?chain=base" -H "X-API-Key: {BANKR_API_KEY}"

It prints the portfolio body (PORTFOLIO).

Extract sender's balance for the target token. Compute total_required from the recipient list (sum of per-recipient amounts, applying overrides). If balance < total_required * 1.05 (5% headroom for any failed retries) → DISTRIBUTE_TOKENS_ERROR — insufficient balance: have X, need Y ${TOKEN}, exit. Do not start a partial run.

B3. RESOLVE phase — build the plan

For each recipient, build a row: {key, type, amount, token, target_address, label, status} where key = sha256("${LABEL}|${recipient_id}|${today_utc}") and recipient_id is the handle (lowercase) or address (lowercase).

Idempotency check (before resolving): if memory/state/distributions.json contains key with status=completed → mark row SKIPPED_DEDUP, carry forward the prior txHash.

Handle resolution (@username): use Bankr Agent API to look up the linked wallet. One call per handle, the handle written in literally:

./secretcurl -fsS -X POST "https://api.bankr.bot/agent/prompt" \
  -H "X-API-Key: {BANKR_API_KEY}" -H "Content-Type: application/json" \
  -d '{"prompt":"What is the EVM address linked to <HANDLE> on Base? Respond with only the address."}' | jq -r '.jobId'

It prints JOB_ID. Then poll the job: repeat this one call, with JOB_ID written in literally, until its .status is completed or failed, at most 15 times (the old every-2s, max-30s budget):

./secretcurl -fsS "https://api.bankr.bot/agent/job/<JOB_ID>" -H "X-API-Key: {BANKR_API_KEY}"

It prints the job body (R). If .status is still pending, wait 2 s with one python3 -c "import time; time.sleep(2)" call, then repeat. After the 15th call without completed or failed, stop polling and treat the handle as unresolved (below). Extract the address from the response (regex 0x[a-fA-F0-9]{40}). If extraction fails → mark row RESOLVE_FAILED with reason NO_LINKED_WALLET. Do not abort the whole plan; let the executor skip this row.

Address resolution (0x...): validate format ^0x[a-fA-F0-9]{40}$. If invalid → RESOLVE_FAILED reason BAD_ADDRESS.

After RESOLVE, print the plan to the console (and to the dry-run notification if MODE=dry-run):

Plan for list '${LABEL}' (${today_utc}):
  ✓ @alice_dev → 0x1234... — 15 USDC          [READY]
  ✓ Charlie    → 0x742d... — 20 USDC          [READY]
  ↻ @bob_builder → 0xabcd... — 10 USDC        [SKIPPED_DEDUP] (tx 0xprev...)
  ✗ @inactive → ?                             [RESOLVE_FAILED: NO_LINKED_WALLET]

Summary: 2 to send (35 USDC), 1 deduped, 1 unresolvable. Sender balance: 100 USDC.

If MODE=dry-run: notify the plan, log, exit DISTRIBUTE_TOKENS_DRY_RUN. Do not proceed.

If 0 rows are READY (everything deduped/failed) → notify the plan, log, exit DISTRIBUTE_TOKENS_OK — nothing to send.

B4. EXECUTE phase

For each READY row, send via /wallet/transfer (the only sanctioned transfer endpoint per Bankr docs). One call per row, with that row's address, token address, amount and native flag written in literally:

./secretcurl -fsS -X POST "https://api.bankr.bot/wallet/transfer" \
  -H "X-API-Key: {BANKR_API_KEY}" -H "Content-Type: application/json" \
  -d '{"recipientAddress":"<ADDR>","tokenAddress":"<TOKEN_ADDR>","amount":"<AMT>","isNativeToken":<IS_NATIVE>}'

It prints the response body (RESP), or curl's HTTP error code under -fsS. The sleeps below are each one python3 -c "import time; time.sleep(N)" call; a retry is the same transfer call again.

Outcome handling:

• HTTP 200 + success: true → status COMPLETED, store txHash. Persist the state file immediately (write after every recipient, not at the end — survives mid-run crashes).
• HTTP 200 + success: false → status FAILED, store error field as reason.
• HTTP 403 → status FAILED reason READ_ONLY_KEY. Abort remaining rows (key won't suddenly gain write access). Persist state.
• HTTP 429 → status FAILED reason RATE_LIMIT. Sleep 60s, retry once. If still 429, abort remaining (rolling-window quota exhausted). Persist state.
• HTTP 5xx or network error → retry once after 10s. If still failing, status FAILED reason API_ERROR.
• Any other → status FAILED reason HTTP_${code}.

State file shape (memory/state/distributions.json, append/upsert):

{
  "contributors|@alice_dev|2026-04-20": {
    "list": "contributors",
    "recipient": "@alice_dev",
    "address": "0x1234...",
    "amount": "15",
    "token": "USDC",
    "status": "completed",
    "txHash": "0xabc...",
    "timestamp": "2026-04-20T12:34:56Z"
  }
}

B5. Build summary notification

Top line is a verdict: COMPLETE (all READY succeeded) / PARTIAL (some failed) / FAILED (none succeeded) / DRY_RUN / NOTHING_TO_SEND.

*Token Distribution — ${today_utc}* — VERDICT

List: ${LABEL} (${description})
Token: ${TOKEN} on Base
Sent: ${total_sent} ${TOKEN} to ${n_success}/${n_attempted} recipients
Skipped (already sent today): ${n_dedup}
Unresolvable: ${n_unresolved}

✓ @alice_dev — 15 USDC ([tx](https://basescan.org/tx/0xabc...))
✓ Charlie (0x742d...) — 20 USDC ([tx](https://basescan.org/tx/0x123...))
↻ @bob_builder — 10 USDC (already sent: [tx](https://basescan.org/tx/0xprev...))
✗ @inactive_user — RESOLVE_FAILED: NO_LINKED_WALLET

Sender balance after: ${remaining} ${TOKEN}

Suppress empty sections (no Skipped: line if n_dedup=0, etc.). Send via ./notify. Then log (see Log) and exit with the send verdict code (DISTRIBUTE_TOKENS_COMPLETE / DISTRIBUTE_TOKENS_PARTIAL / DISTRIBUTE_TOKENS_OK for nothing-to-send).

Phase C — All (plan then send)

Runs Phase A, then feeds it into Phase B in one invocation. TARGET_WEEK and MODE come from Step 0.

all: execute (MODE=execute):

1. Run Phase A in execute mode (A1–A8). It writes contributors-${TARGET_WEEK} into memory/distributions.yml + state and posts the plan notification (with the A5-note tweak: trailing line becomes Distributing now (phase=all)…).
2. If Phase A took a terminal early exit — CONTRIBUTOR_REWARD_NO_LEADERBOARD, _STALE_LEADERBOARD, _PARSE_FAIL, _NO_ELIGIBLE, or _ALREADY_PROCESSED with zero new entries — stop: there is nothing to send. Log and exit with that Phase A code.
3. Otherwise set LABEL="contributors-${TARGET_WEEK}", MODE=execute, and run Phase B (B1 reads the just-written list from memory/distributions.yml; B2–B5 as normal). The send-phase summary is the final notification.

all: dry-run (MODE=dry-run): full end-to-end preview, no writes, no transfers.

1. Run Phase A in dry-run mode (A1–A5) to compute the plan — but per the A5 note, do not notify or exit here; hold the computed plan rows in memory.
2. Enter Phase B with an in-memory recipient list built from the plan rows (handle="@${login}", amount=<tier amount>) instead of reading memory/distributions.yml; set LABEL="contributors-${TARGET_WEEK}", MODE=dry-run. Run B2 (preflight — this still calls Bankr /wallet/me + /portfolio, so BANKR_API_KEY is required for this preview; if absent, report DISTRIBUTE_TOKENS_ERROR — BANKR_API_KEY not configured for the send preview but keep the rendered plan) and B3 (RESOLVE + print). B3's MODE=dry-run branch notifies the combined preview and exits DISTRIBUTE_TOKENS_DRY_RUN. No state or yml is written by either phase.

Log

Append to memory/logs/${today}.md under one heading (the health loop parses this shape). Always use ### distribute-tokens; the Phase/Mode discriminator lines say which branch ran.

### distribute-tokens
- Phase: plan | send | all
- Mode: execute | dry-run
# --- plan phase (present when Phase A ran) ---
- Plan mode: execute | dry-run | already-processed | no-merged-prs | api-fail | no-eligible
- Week: ${TARGET_WEEK}
- Source: GitHub merged PRs for ${TARGET_WEEK}
- List label: contributors-${TARGET_WEEK}
- Entries written (new): ${N_NEW} | deduped: ${N_DEDUP} | total USDC planned: ${TOTAL_USDC}
- First-PR bonuses: [list or "none"]
# --- send phase (present when Phase B ran) ---
- List: ${LABEL} | Token: ${TOKEN}
- Verdict: ${VERDICT}
- Sent: ${total_sent} ${TOKEN} to ${n_success}/${n_attempted}; deduped: ${n_dedup}; unresolved: ${n_unresolved}
- Failures (if any): @x — REASON, @y — REASON
- State file: memory/state/distributions.json (${total_keys} entries)
- Notification sent: yes/no

Omit the block for whichever phase did not run.

Exit codes (for downstream automation)

Plan phase (Phase A):

• CONTRIBUTOR_REWARD_OK — plan written, notification sent
• CONTRIBUTOR_REWARD_DRY_RUN — plan rendered, no writes, notification sent
• CONTRIBUTOR_REWARD_ALREADY_PROCESSED — week already in state with identical plan, silent exit
• CONTRIBUTOR_REWARD_NO_MERGED_PRS — no PRs merged in the target week, silent exit
• CONTRIBUTOR_REWARD_API_FAIL — GitHub API unreachable (gh api + WebFetch both failed), notified
• CONTRIBUTOR_REWARD_NO_ELIGIBLE — zero contributors above threshold, silent exit
• CONTRIBUTOR_REWARD_ERROR — file I/O or YAML write failure, notified

Send phase (Phase B):

• DISTRIBUTE_TOKENS_OK — nothing to send (everything deduped or list empty), or bootstrap
• DISTRIBUTE_TOKENS_COMPLETE — all READY rows succeeded
• DISTRIBUTE_TOKENS_PARTIAL — some succeeded, some failed
• DISTRIBUTE_TOKENS_DRY_RUN — dry-run completed, no sends
• DISTRIBUTE_TOKENS_ERROR — preflight or config failure, no sends attempted

For all:, the terminal exit code is the send phase's code (or the Phase A early-exit code when the plan produced nothing to send).

Network note

• Plan phase (A): ranks contributors via gh api search/issues (gh handles GitHub auth internally, so no secret ever lands on the command line). If gh api fails, fall back to WebFetch on the public https://api.github.com/search/issues?q=… URL. Also reads/writes memory/state/contributor-reward-state.json, memory/distributions.yml, memory/logs/${today}.md. No postprocess scripts required.
• Send phase (B): every Bankr call is auth'd, so make it with ./secretcurl using the {BANKR_API_KEY} placeholder — never a bare $BANKR_API_KEY (the Bash permission layer refuses a secret on the command line) and never plain curl. /wallet/transfer is an irreversible money-movement write; it runs in-run as the executor's final action (Phase B4), behind the B2 balance preflight and the per-recipient idempotency in memory/state/distributions.json (persisted after every send, so re-runs never double-pay). There is no deferred/postprocess step — a failed transfer is recorded (FAILED with its reason) and the run continues to the next row. Never silently drop a transfer.

Constraints

Send (money movement):

• Idempotency is non-negotiable. Always read memory/state/distributions.json before sending; always persist after every transfer. Never batch state writes to end-of-run.
• Treat the 24h Bankr rate limit (100/day standard) as a hard cap. Lists >50 recipients should be split.
• Never send if preflight balance is < total_required * 1.05.
• Never use the Agent API for transfers (deprecated). Agent API only for handle→address resolution.
• Never abort the RESOLVE phase on a single bad recipient — collect all errors, present them, then let the executor skip.

Plan (reward computation):

• Idempotency is per-(week, login). Re-runs in the same week add only deltas; demotions never claw back already-paid amounts.
• First-PR bonus is once-ever per login. Track in first_pr_bonus_paid; never re-award even if the same person appears as ✨ in a later week (which they shouldn't, since ✨ means first ever merged PR — but defend against API drift).
• No silent overwrites of distributions.yml. If the file exists and is malformed, fail loudly rather than rewriting.
• Eligibility floor stays low (≥ 1 merged PR) by design. A single PR merged in the week qualifies — reward shipped work, not volume.

Future iterations

• Schedule all: directly (weekly, after the week closes) for full hands-off payout — the plan computes its own ranking from merged PRs, so no upstream skill or chain wiring is needed.
• Add a Bankr Agent API "wallet-linked?" pre-filter in the plan phase so contributors without linked wallets are flagged in the notification (prevents the send phase from logging RESOLVE_FAILED rows on every run).
• Tier table should become operator-configurable via memory/contributor-reward-config.yml once the first month of runs reveals the right curve. Hardcoded for v1.