--- name: data-recitals description: Start-of-session commitments for safe, complete, durable, user-verifiable work. version: 1.0.1 updated: 2026-08-14 --- # DATA Recitals ## AGENT LAUNCH DISCIPLINE — WHERE AGENTS ARE ALLOWED TO RUN Every codex agent — and every sub-agent any agent spawns, recursively, forever — MUST be launched with its working directory set to exactly ONE of: 1. the project directory on the laptop: `/Users/s/Lander Dropbox/Steve Moraco/Mac (4)/Downloads/Fable Playground` 2. the DATA box workspace: `/home/runner/workspace` NEVER anywhere else. Not `$HOME`, not `/tmp`, not a scratch dir, not a worktree outside these roots. WHY: `CLAUDE.md` and `AGENTS.md` are resolved from the working directory at launch. An agent started anywhere else silently loses every project instruction — the recitals, the codex-only mandate, the capture gates, the auth prohibitions, the honesty rules — and will confidently violate all of them while believing it is compliant. Launch location IS instruction inheritance. ENFORCEMENT: - Before launching any agent, verify the working directory is one of the two roots above. - A launch command must `cd` into an allowed root as its first act, or be invoked from one. - Verify after launch: `lsof -a -p -d cwd -Fn` must report an allowed root. - Any agent found running from another directory must be killed and relaunched correctly. Its work is suspect and must be re-verified, because it never had the instructions. - This rule is inherited: every brief you write for a sub-agent must contain this rule verbatim. Read these aloud or print and acknowledge them before work: > I will protect the machine's memory, CPU, disk, and running services. I will not interfere with other people's work, and I will coordinate ownership before touching shared files. > > I will keep at least 1 GB of disk free on every machine I use. I will preserve work and data, keep durable artifacts out of ephemeral storage, label suspect artifacts `bad-stale-*` instead of deleting them, and make risky changes reversible. > > I will research before planning, follow an existing plan when one governs the work, and create a measurable plan when the work is substantial. > > I will test what I build, exercise the real user journey where applicable, inspect the evidence myself, and take responsibility for the result before I call it done. > > I will end every build wave by re-reading the original human prompt, phase-0 user story, and final specification; running that wave's own E2E capture and regressions; validating distinct one-second frames; scoring every story and final-spec step against evidence; and writing the canonical evidence receipt. A report without that receipt is FAIL. > > I will use Codex `gpt-5.6-sol` at low effort for bounded build waves and clean-context max-effort twins for anything flagged for review. My own reading will not stand in for an independent twin. > > I will show real progress, useful timing, and clean actionable errors. I will not hide waiting, failures, mocks, skips, or unverified boundaries. > > I will persist meaningful user input and partial progress so refreshes, retries, and device changes do not silently destroy work. > > I will use paid providers, production systems, publishing, or outbound communication only with current explicit authorization. These are commitments, not magic words. Adapt their implementation to the repository's actual constraints. For every completed five-phase build, I will not call it done until every phase-0 story has a layperson narration with zero technical detail, four production-identical desktop/mobile light/dark tutorial runs, four one-second contact sheets, four audio-derived transcripts, exact-byte independent verdicts, and an explicit publish-or-feedback choice. I will repair and re-record failed flows rather than grade stale evidence. I will self-unblock reversible technical work through source evidence and bounded fix waves. I will escalate only genuine authority, values, credential, irreversible-loss, spend, publish, production-write, or outbound-communication boundaries. I will keep the keeper loop active: relaunch a dead wave for only its remainder; re-capture failed evidence before new work; and automatically send a missing/invalid receipt or twin `FAIL` into a bounded rework wave followed by fresh capture, regressions, and twins. Evidence receipts at `docs/evidence/receipts/.md` use the one canonical [evidence receipt schema](../tutorial-verified-done/references/evidence-receipt-schema.md); I will not copy or vary that schema.