remora v0.1.8 — phase-aware Baton composition
Phase-aware Baton composition for GPT-5.6
remora v0.1.8 adds a complete plan-first orchestration lifecycle to the existing session-scoped GPT-5.6 role map:
Discovery → main-session Plan → explicit approval → Execution → fresh verification
The launcher remains isolated: model routing, gateway credentials, agents, and orchestration policy exist only inside the remora child process. Native Claude settings and login state remain untouched.
Included changes
| Area | Behavior |
|---|---|
| Discovery | Luna Explore and scout roles may gather bounded evidence before the implementation outcome is known. |
| Plan | The Sol verifier can challenge a material Plan using READY or REVISE; the main session owns synthesis and revisions. |
| Approval | Large, architectural, risky, or explicitly plan-first work waits for explicit approval before implementation. |
| Execution | Luna handles mechanical volume work; Sol remains available for local engineering judgment and security-sensitive work. |
| Outcome verification | A fresh Sol verifier returns CONFIRMED or REFUTED after implementation. |
| Baton composition | Baton chooses questions, topology, ownership, sequence, budgets, and stop conditions while remora retains role and model authority. |
| Dispatch brake | Small repository scans and tightly coupled debugging stay in the main session; stable repetition and substantial independent evidence remain eligible for delegation. |
| Model routing | Named roles continue to omit invocation-level model; the session-scoped role definition is the sole model source. |
| Packaging | Baton compatibility evidence is included in the release archive. |
This release keeps remora's existing six-role, session-scoped architecture. It does not install global agents or modify native Claude state.
Public compatibility gate
A fresh two-turn Gate ran through remora, Calico Claude, CLIProxyAPI, GPT-5.6, and the installed Baton skill:
| Turn | Wall time | Client-reported cost | Result |
|---|---|---|---|
| Discovery + Plan | 294.102 s | $1.233720 | Two background Luna discovery agents; Sol readiness review returned REVISE, then READY; zero writes before approval. |
| Approved execution + verification | 248.304 s | $0.915572 | Baton correctly rejected a needless writer; the main session wrote only REPORT.md; tests passed; fresh Sol verifier returned CONFIRMED. |
| Total | 542.406 s | $2.149292 | Complete lifecycle passed without startup resend. |
All four named Agent calls omitted invocation-level model. Exploration ran on GPT-5.6 Luna and verification ran on GPT-5.6 Sol.
The rejected first candidate is also disclosed: it used the wrong verdict vocabulary during Plan review and later exposed citation gaps. It is preserved as rejected evidence rather than counted as a successful compatibility run.
Exact prompts, normalized Agent calls, model routing, hashes, timing, startup-stall criteria, and limitations are published under benchmarks/baton-compatibility.
These are single-run observations, not population estimates or provider invoices.
Verification
- 34 Python tests passed.
- Installer isolation and offline bootstrap tests passed.
- Python and shell syntax checks passed.
- Agent rendering and sanitized dry-run checks passed.
- The v0.1.8 release archive was rebuilt by GitHub Actions.
checksums.txtand build-provenance attestation were published.- Codex review found no major issues.
This release does not modify CLIProxyAPI or gateway configuration.
Upgrade
Use the pinned, approval-gated v0.1.8 runbook:
Read and follow this installation runbook:
https://raw.githubusercontent.com/Nanako0129/remora-cc/v0.1.8/install/AGENT-INSTALL.md
Perform only the read-only preflight first. Show every proposed filesystem
change, trust boundary, download source, and verification step. Do not write
anything until I explicitly approve.
After upgrading, start a fresh remora session. /resume may restore an older session's previous agent definitions and orchestration policy.
Full Changelog: v0.1.7...v0.1.8