promissory

PROMISSORY — Non-Green Promise Assault

The canonical, always-current procedure is docs/fable/2026-07-01-promissory-nongreen-assault-runbook.md (repo root). Read it first, in full, every invocation — this skill is the launcher, not the law. If the skill and the runbook disagree, the runbook wins.

Argument parsing

• No args → supervisor mode: rank the pool, then either dispatch a fleet wave (if fleet capacity is ready) or personally assault the top target.
• A promiseId (e.g. khala_code.forum_hotbar.v1) → single-target mode on that promise. This is explicit owner direction: it overrides the runbook's "already mapped by other work" exclusion (note the override in the claim issue).
• A number N (e.g. 10) → supervisor mode, dispatch a wave of N fleet workers at the top N unclaimed targets.
• Anything else (e.g. a campaign hint like throughline: mobile) → treat as a temporary throughline override for scoring this run only.

Scoreboard (mandatory, start AND end)

Print the registry scoreboard before any work starts and again after the run finishes, so the user sees states flip. Compute it from the canonical source (one-off script from the openagents repo root):

bun -e "
const { publicProductPromisesDocument } = await import('./apps/openagents.com/workers/api/src/product-promises.ts');
const d = publicProductPromisesDocument();
const c = {};
for (const p of d.promises) c[p.state] = (c[p.state] ?? 0) + 1;
console.log('registry', d.version, '| total', d.promises.length, '|', JSON.stringify(c));
"

• Opening scoreboard: registry version, total records, and the count per state (green / yellow / red / planned / degraded / withdrawn), plus the target(s) selected and their current states.
• Closing scoreboard: the same counts recomputed from the merged main, presented as a before → after diff — explicitly name every record whose state or blocker set changed during the run (e.g. "foo.v1 planned → yellow. bar.v1 cleared 2 of 3 blockers, still red") and the new owner-decisions-ready count in NEEDS_OWNER.md.
• In fleet/wave mode the supervisor prints the opening scoreboard once at wave start and the closing one after the last closeout merges. Long waves should also emit interim scoreboards as each PR lands.

Operating summary (details and exact rules live in the runbook)

1. Snapshot + rank. Load the registry via publicProductPromisesDocument() from apps/openagents.com/workers/api/src/product-promises.ts (the scoreboard script above already does this — reuse its output). Apply the eligibility filter (runbook §2 — including steer clear of promises already mapped by open issues/epics/roadmap lanes: PROMISSORY hunts the hidden and overlooked, unless the user directed a specific target) and the scoring formula (§3, throughline-weighted. Current campaign: Khala Code launch).
2. Claim atomically. One promiseId per claim. GitHub issue titled PROMISSORY: <promiseId> — search open AND recently-closed first. lower issue number wins races (§4).
3. Assault ladder (§5): audit the record → decompose every blockerRef into BUILD / EVIDENCE / OWNER / EXTERNAL → implement fully in a fresh worktree from clean origin/main with tests and dereferenceable evidence → update the registry record + one note + version bump in the same PR per the concurrent-safe edit protocol (§7) → verify (record's own verification, relevant suites, check:deploy, promise-test pins) → merge to main, close the issue, write owner-gated residue to the workspace NEEDS_OWNER.md, release the claim, take the next target.
4. Fleet dispatch (supervisor/wave modes): use the runbook §8 pinned $PYLON khala request --workflow codex_agent_task template (or the khala_fleet MCP equivalent), one worker per target, refill freed slots with the next-ranked unclaimed target on every closeout, keep the dispatch ledger and exact token-row verification per docs/fable/EXECUTION.md.

Hard guardrails (never violate, even under time pressure)

• Never flip a promise to green — green is owner-signed, receipt-first. The success metric is owner-decisions-ready: promises one owner action from green, batched in NEEDS_OWNER.md.
• planned→yellow / red→yellow only when the record's own verification criteria are met with cited evidence. Honest downgrades encouraged.
• Never weaken a gate/test/policy to clear a blocker. Never edit the green-count test pin to silence a failure. Never broaden public copy.
• One promise per claim/PR. Registry edits touch only your record + one note + the version constant.
• Isolated worker homes always. Never touch ~/.codex or the live ~/.claude. Public-safe prompts and evidence only.