STAGE 1.0.0
Status: released, published, and installed on 2026-09-27
Prepared: 2026-09-27
Plugin: 1.0.0 · Method: 1.0
Reviewable Result
This release carries the previously unreleased 0.7 redesign and correctness repairs forward into the owner-accepted 1.x compatibility contract. The owner calls Salvageist's pilot successful; its nine-change retrospective and workflow-cost follow-up provide the practical basis for proceeding. An intermediate 0.7 publication is unnecessary. The readiness review explains the recommendation; compatibility owns the accepted terms.
Salvageist's applied lessons keep verification proportional in the actual runner: include the whole task delta, stop costly dependent stages after prerequisite failure, use compact completion signals, and reuse conclusive evidence only when its relevant inputs and policy permit. Keep recovery notes current and assign shared integration/release one owner when delegation is useful and authorized. Instructions remain model-neutral; no mandatory agent chain or new artifact is introduced. Unity guidance already carries direct-scenario, output-protection, and comparable cold/steady-state measurement lessons from the pilot.
The agent has prepared package versions, both Node lockfiles, the method version, current references, changelog, compatibility inventory, and ADR 0066. Local preparation preceded the owner's September 27 instruction to release, publish, and update the installed plugin. The canonical source remains private; public documentation continues through the scoped mirror.
Compatibility And Support
Preserve S1–S8, optional-contract semantics, plugin identity and four routes,
project-map schemas 0.1/0.3, and case-study schemas 0.1/0.3/0.4.
No supported schema or executable helper changes in the 1.0 preparation.
Successful command fields, arguments, failure semantics, first-parent/root
inventory, and protected-state behavior retain the corrected 0.7 contracts.
Installing 1.0 creates no game-project artifacts and requires no migration.
The accepted stable interface lasts throughout 1.x; prose, templates, reference layout, optional boards, and profiles remain maintainable. Support follows the exercised owner environment and declared runtime dependencies. The full inventory and safety-correction policy are in the compatibility guide. Human acceptance, model performance, and cross-project transferability remain separate claims.
Evidence Coordinates
- Starting STAGE checkpoint:
ae10b88, following the original 0.7 candidate77d8afc6166391fb066c22ded097d8316d98238dand two Salvageist documentation commits. - Applied guidance:
b864760(docs: apply Salvageist workflow-cost lessons). - The September 25 retrospective and September 26 follow-up retain their original evidence and limitations. The owner's September 27 success statement is additional informal readiness evidence, not a reconstructed revision-bound playtest.
- Read-only inspection of Salvageist's current workflow documents observed
dfb6728993c4b8e759b60622f29fd8537ffc6d4ewith one pre-existing dirty path. Source-state snapshots matched before and after. No engine run, new gameplay benchmark, game edit, or deployment was performed for this proposal.
The original paired 14-run GPT-6 Astra evaluation remains in the dated 0.7 record. It is historical support for unchanged delivery boundaries, not a new 1.0 model comparison. Private source and raw model traces remain outside the public site.
Candidate Verification
Observed on September 27, using Python 3.11.8 and Node 26.9.0 on the owner's macOS:
| Check | Result |
|---|---|
| Focused package/release-policy tests | All 9 passed |
| Change gate | Passed: 264 tests across 27 modules, plus maps, studies, links, documentation discovery, lint/format, and diff checks |
| Complete release gate | Passed from the clean pinned Python environment: 432 tests across 38 modules, including all 168 Solo Unity tooling tests and supported contracts |
| Clean dependency installation | Fresh Python environment installed the pinned graph and passed pip check; root and website npm ci --ignore-scripts passed in separate empty directories |
| Direct Codex validators | All four skills and the plugin passed |
| Documentation | Production build passed with broken links/anchors treated as errors; public-artifact screen passed for 281 generated files |
| Candidate package and unfinished markers | All 31 files in the behavior-probe package match source; no unfinished markers in the 31 changed paths |
| Preservation | S1–S8 byte-identical to ae10b88; no schema, executable helper, profile, trial, or test-source changes; Salvageist source-state snapshots matched |
| Website dependency advisory report | Valid full-graph report: 32 findings, comprising 7 high, 24 moderate, and 1 low; no critical findings |
The full gate includes the Solo Unity tooling suite; it does not run the Unity Editor or certify a game. Dependency selections are unchanged: both lockfile diffs update package identity only. The website advisory policy remains informational maintainer review, including browser-bundled dependencies; counts do not establish runtime exposure or resolve the findings. Hosted CI uses Node 24 and remains a separate check after an authorized push.
Focused Candidate Behavior
Two fresh Codex CLI 0.155.1 processes used GPT-6 Astra at high reasoning with the same explicitly supplied 1.0 package copy. User config, host skill discovery, plugins, and multi-agent execution were disabled for isolation. Both first runs are retained with exact prompts, commands, source hashes, traces, fixtures, and semantic review; neither was retried.
| Disposable-project case | Observed result |
|---|---|
| Documentation-only task | Corrected the typo, ran the declared whole-task docs gate once, preserved protected inputs, and created a task-only commit |
| Earlier committed runtime edit followed by a documentation fix | Used the original task-start revision, included the committed runtime change, reported the required engine check as unavailable, preserved protected inputs, and left the new checkpoint pending |
These exercise scope selection and the checkpoint boundary, not real engine verification, installed-plugin discovery, model comparison, or general compliance. At that candidate checkpoint, the owner's installed package remained 0.7.0. The authorized 1.0 installation and fresh-task verification are recorded below.
The retained private archive candidate-behavior-evidence.tar.gz is 98,618 bytes,
SHA-256 badfb4f296360356b029f59f09db28f132d03f390bdf53dfe842d98667780597.
Local release evidence also retains the complete gate and dependency logs,
package-validator results, preservation checks, and source-state snapshots.
The final checkpoint is bound to those receipts in local candidate.json.
At that checkpoint, only release-documentation prose and legacy heading anchors
changed after the full suite; their link checks and production rendering were
checked separately. The later acceptance and tag-policy correction received
fresh local change and release gates before integration.
Limits And Owner Decision
Salvageist supports personal usefulness; it does not establish a causal time or token saving, independent adoption, formal product acceptance, or model equality. Earlier owner-rejected dogfoods and all pilot regressions remain recorded. Missing attention measurements and a prospective fresh-agent game takeover do not block this bounded compatibility release.
The owner accepted the finished contract and authorized release, publication, and plugin refresh on September 27. The existing maintainer path still requires hosted quality for integrated current main before consequential publication.
Publication Receipts
Completed on September 27, 2026:
| Operation | Receipt |
|---|---|
| Accepted integration | PR #7, merged as 2a5124ce28f5f623192551ac1d60004db111e49f; its tree exactly matches the reviewed proposal |
| Hosted proposal checks | Quality run 36283270148: commit policy, complete release gate, and documentation build/screen/advisory report passed |
| Hosted release-source checks | Quality run 36283387508: all three jobs passed on current main; the shared eligibility guard passed again before tagging |
| Tag and release | Annotated v1.0.0 at the accepted merge; STAGE 1.0.0 GitHub Release published as the latest stable release |
| Public documentation | Publication run 36283475904 passed and fast-forwarded the public mirror to 91ff2fe, identifying the accepted source SHA; the live release and compatibility pages returned HTTP 200 with the accepted 1.0 content |
| Installed package | Local marketplace refresh reports enabled 1.0.0; all 31 installed files match the released plugin/ subtree |
| Release Please recovery | Manual run 36283623447 passed, found v1.0.0 at the accepted merge, and correctly produced no new proposal with zero subsequent commits |
The canonical source remains private. Public documentation uses the existing scoped mirror and Cloudflare deployment. These receipts identify the initial release publication; documentation can receive factual receipt updates after the immutable release tag.
Release preparation also corrected the manifest's tag policy: Release Please
had searched for stage-game-engineering-v0.6.11, while historical STAGE tags
use v<version>. Explicit include-component-in-tag: false and
include-v-in-tag: true now make proposal discovery match the published tags.
A structural regression protects that policy; draft-only permissions remain
unchanged. The pinned upstream configuration schema
defines these settings. The existing 0.7 and focused candidate evidence above
remains dated; release follow-through does not relabel it as new evidence.
Installed Fresh-Task Verification
A fresh Codex CLI 0.155.1 process with GPT-6 Astra at high reasoning used the
normal installed-plugin configuration, with multi-agent execution disabled.
The ordinary game-fix prompt did not name a skill. The trace shows discovery
and reading of the installed 1.0.0 delivery skill, reproduction of negative
health, regression coverage, and repair through the canonical game function.
All five tests and both production CLI scenarios passed. The local commit
contains only the code and test files; staged story content, unfinished dialogue,
an untracked private draft, and the protected index entries remained unchanged.
There was no repeated approval request or external publication.
This single disposable Python game checks installed routing and delivery within that fixture. It does not certify engine integration or general model behavior. The trace contains an unnecessary per-command hook/signing override; the fixture had no active Git hooks or configured hook path, so enforcement in hook-bearing or signing-required repositories was not exercised. The first run and this limitation are retained without a rerun.
The private installed-release-evidence.tar.gz archive is 30,585 bytes,
SHA-256 703235aa4fdf0c685a2e9a2bcc924ee1450532339910f466ed6dada984088984.
It retains the exact prompt, command, trace, fixture, pre/post state, diff,
post-checks, and semantic review. Existing chats retain their loaded instructions;
a fresh chat loads the installed 1.0.0 package.