# Ralph Specum Start > This skill should be used only when the user explicitly asks to use `$ralph-specum-start`, or explicitly asks Ralph Specum in Codex to start or resume a spec. - Skill: `tzachbon/ralph-specum-start` (Agent Skill, multi-file: 2 files) - Install (CLI): `npx skillmds add tzachbon/ralph-specum-start` - Raw SKILL.md: https://api.skillmd.com/api/skills/tzachbon/ralph-specum-start/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: tzachbon (https://skillmd.com/u/tzachbon) - Updated: 2026-09-10 - Page: https://skillmd.com/skills/tzachbon/ralph-specum-start --- # Ralph Specum Start Use this for the `start` and `new` entrypoints. Derive `RALPH_CODEX_PLUGIN_ROOT` from this loaded skill by resolving two parent directories from the `SKILL.md` directory. Never derive it from the project working directory. ## Contract - Read `.claude/ralph-specum.local.md` when present - Default specs root is `./specs` - Keep `.current-spec` in the default specs root - Keep the standard Ralph files stable - Merge `.ralph-state.json`. Do not replace the full object ## Action 1. Parse explicit name, goal, exact `--quick` or exact `--interactive`, `--resume `, commit flags, optional specs root, and optional `--tasks-size fine|coarse`. Reject both mode flags together, `-q`, variants, and natural-language substitutes. 2. Classify new versus resume intent, then resolve the classified target by explicit path, exact name, or `.current-spec`. A new-spec request does not inherit or recover the existing current spec. 3. If the same name exists in multiple configured roots, stop and require a full path. 4. Check active epic context from `specs/.current-epic` when no explicit spec was chosen. 5. For large or cross-cutting goals, route to triage instead of forcing a single spec. 6. `new` is an alias here. Create the spec directory if needed. 7. Initialize or merge state with: - `source: "spec"` - `name` - exact `goal` text used by the context gate - `basePath` - `phase: "research"` - `taskIndex: 0` - `totalTasks: 0` - `taskIteration: 1` - `maxTaskIterations: settings default or 5` - `globalIteration: 1` - `maxGlobalIterations: 100` - `commitSpec: settings auto_commit_spec or true` - `relatedSpecs: []` - `awaitingApproval: false` while the goal interview is active - preserve or set `quickMode` - preserve or set `granularity` when `--tasks-size` was supplied - preserve or set `epicName` when starting from an epic suggestion 8. Update `.current-spec`. 9. Write `.progress.md` with goal, current phase, next step, blockers, learnings, and skill discovery results. 10. On resume, prefer `tasks.md` and present files over stale state when they disagree. 11. Run `phase_gate.py mode` through `"$RALPH_CODEX_PLUGIN_ROOT/scripts/phase_gate.py"` with `STATE` and the exact supplied mode flag. No flag normalizes invalid legacy quick state to interactive. 12. Run skill discovery pass 1 against the goal. Collect plugin skills, project `.agents/skills`, project `.claude/skills`, and the current Codex harness catalog. Always select explicitly named skills and record shadowed duplicates. 13. Load `"$RALPH_CODEX_PLUGIN_ROOT/skills/interview-framework-codex/SKILL.md"`, its required algorithm and domain-modeling references, and every selected domain contract in both interactive and quick mode. In interactive mode, follow that algorithm for the start goal territory: outcome and observable success, material scope boundaries, critical constraints, and viable high-level approach. Ask no spec-path, branch, task-size, or discoverable question. 14. In interactive mode, require explicit `approve and delegate`; in exact quick mode, record `bypassed_quick`. In both modes, run `phase_gate.py check-delegation` with the current loaded-manifest identity before creating a `research-analyst` child. Apply the shared hard-transition invariant before fresh or resumed research dispatch. A failed `check-delegation` in either mode stops this invocation before phase transition, child dispatch, or target-artifact write; after a normal-mode failure, only the next explicit invocation creates a fresh manifest/interview identity. After an exact `--quick` delegation failure, the next explicit invocation reruns discovery, records a fresh `phaseSkillLoad` and interview identity, and does not reuse a terminal `bypassed_quick` interview or its discovery revision. Only a matching in-progress `collecting` or `awaiting_confirmation` interview is resumable. Exact `--quick` retains its existing question-and-approval bypass and discovery, manifest, delegation, and writer checks. 15. Keep the existing child packet, identity tuple, and receipt behavior unchanged: pass the absolute helper path, state path, identity tuple, a unique teammate dispatch identity, and the verbatim `phaseSkillLoad` manifest. The child must record its loads and pass `check-agent-write` with that unique identity before writing `research.md`. 16. Validate `research.md`, merge `phase: "research"` and `awaitingApproval: true` in interactive mode or `false` in exact quick mode, update progress, and present artifact approval when interactive. 17. In exact quick mode, record the quick bypass, generate missing artifacts in order, skip normal approval pauses, and continue into implementation in the same run. ## Prototype Reconciliation and Resume After resolving the classified target and before normal resume or quick routing. Skip existing-current-spec recovery for new-spec intent: 1. Use `"$RALPH_CODEX_PLUGIN_ROOT/scripts/resolve_spec_paths.py"` and only its resolved `basePath`. When both `basePath` and `/.ralph-state.json` exist, run `"$RALPH_CODEX_PLUGIN_ROOT/scripts/prototype_records.py" reconcile --base-path "$BASE_PATH" --state "$BASE_PATH/.ralph-state.json"`, then re-read state. 2. Treat a missing `activePrototypes` field as an empty map. Sort entries by `created`, then ID. 3. Resume an explicit active ID through `$ralph-specum-prototype --resume ` and stop this skill. In normal mode, resume the sole active entry automatically. When several remain, list deterministic IDs with question, status, blocker, `returnPhase`, and `returnTaskIndex`, then stop for an explicit ID. 4. In quick mode, ask no question. Sort entries that block design by `created`, then ID. At the post-requirements boundary, route through `$ralph-specum-prototype --quick`; the prototype skill takes over the oldest design blocker and owns every decision. Preserve earlier quick flow and unrelated entries. 5. Treat each `resume_review` candidate as recovery work, not terminal evidence. Parse the exact candidate, verify its ID and hash, and reconstruct a minimal recovery entry from the record and source pointers. Before reviewer dispatch, reserve its ID through create-only `locked_state.py upsert-prototype` with `status: reviewing`, the exact `candidateHash`, null owner and lease fields, a null or absent `harnessRun.id`, blocker and return fields, source pointers, and recovery timestamps. This is a recovery-only entry: cancellation verifies that `owner`, `leaseToken`, and `harnessRun.id` are all null or absent and that no builder is associated, then skips interrupt and release; inconsistent builder ownership fails closed. On a concurrent reservation, re-read and continue only if its `candidateHash` matches; never overwrite it. Route the restored entry through deterministic exact-candidate review and publish recovery. If reconstruction or reservation cannot proceed, stop and report the candidate ID and candidate hash. Exclude quarantined or malformed records. If no overlay exists, continue the existing start behavior without extra output. 6. Quick mode completes its automatic local route and does not reach the interactive Response Handoff below. ## Branch Isolation - If the user wants isolation, offer a feature branch in place or a worktree with a feature branch. - If a worktree is created, stop after creation and ask the user to continue from that worktree. ## Response Handoff - After creating or resuming the spec, name the resolved spec path and summarize the current state briefly. - After setup in normal mode, begin the goal grill in the same run without stopping at a setup prompt. - After final `approve and delegate`, dispatch research in the same turn. - After `research.md` in interactive mode, name the file, summarize it, and end with exactly one artifact approval prompt: - `approve current artifact` - `request changes` - `continue to requirements` - With exact `--quick`, do not show this prompt; continue directly through the remaining phases after the gates succeed. - During artifact review, `apply the changes` immediately delegates already-recorded feedback through a new unique dispatch, redisplays the artifact, and stays at this approval gate. Ask one focused change question only when no feedback is pending. Control-only `continue`, `proceed`, and `go ahead` approve nothing.