mirror of
https://github.com/morluto/rea.git
synced 2026-10-31 20:06:14 +00:00
f93a0335a8
Project large local Evidence captures through the authenticated file-source boundary and add guarded opt-in sibling-pair path inference. Preserve request cancellation, resource ownership and file-relative validation diagnostics. Validated relevant source and compiled CLI/MCP/environment boundaries, static/build/docs/catalog gates, and a real isolated Gitea 28.2.0 capture through CLI, default stdio MCP and HTTP replay. Current-head automated review passed; unrelated long-running platform jobs may still be pending. Co-authored-by: Edgar Arakelyan <edgar.arakelyan@scale.com> Co-authored-by: Kaoru0822 <208281322+kaoru0822-kitauji@users.noreply.github.com> Co-authored-by: gensokyo-dev <gensokyo-dev@users.noreply.github.com>
105 lines
6.3 KiB
Markdown
105 lines
6.3 KiB
Markdown
# Verification
|
|
|
|
Verified commit: a64367a9 on main 2dd5effe.
|
|
|
|
## Gates
|
|
|
|
Run on macOS 26.5.1 with Node 22.23.1 on that tree: `npm run typecheck`, `lint`
|
|
(oxlint plus `verify:module-boundaries`, 2172 files), `format:check`, `knip`,
|
|
`metadata:check`, `docs:check`, `jscpd` (3.43% duplicated lines), `scan:todos`
|
|
and `mcp-catalog:check` all exit 0. oxlint on the 12 changed TypeScript files
|
|
reports no finding that main does not already report for the same files.
|
|
|
|
`npm run test:prepare`, then `npm run test:changed` (the tests related to the
|
|
changed files): 84 files, 1531 passed, 3 skipped, 0 failed. They include:
|
|
|
|
- the service: an Evidence file projects exactly what the same Evidence projects
|
|
inline; a capture whose response text exceeds the inline budget is refused
|
|
inline with `resource_constraint` and projected from its file; a relative
|
|
path, a missing file and JSON that is not Evidence are refused; the
|
|
synchronous API refuses an unread file source;
|
|
- the compiled CLI: a request naming an Evidence file projects the same
|
|
manifest as the same Evidence inline;
|
|
- MCP: inline, retained and file sources give identical manifests, outputs
|
|
validate against the advertised schema, and a relative path is refused;
|
|
- the file source module: the size estimate, the authentication headroom
|
|
refusal, a cancelled request refused before reading, and a read within the
|
|
caller's resource scope; the service returns typed cancellation for a
|
|
cancelled file request; a request aborted while it runs, even after its read, returns typed
|
|
cancellation because the service yields before each check;
|
|
- the pair rule: three owner/repo pairs merge for collection and numeric detail
|
|
routes, while two pairs, or three pairs sharing one namespace, stay literal;
|
|
with `sibling-variance` also selected the pair still merges, because pairs
|
|
are merged first. Each pair parameter keeps its role when
|
|
`sibling-variance` later templates the segment before the pair.
|
|
|
|
## A rich capture
|
|
|
|
A Gitea 28.1.0 instance seeded through its API with an invented world (12
|
|
people, 4 organisations, 16 public repositories, 320 issues with about 600
|
|
comments, labels, milestones, commits, pull requests and releases) was read by
|
|
one `capture-browser-scenario` run of 1,217 API reads: 1,218 transactions in
|
|
24 s, 10.7 MB of Evidence, nothing dropped. Scenario parents are not on main,
|
|
so this run used this commit merged onto the scenario-capture change
|
|
(morluto/rea#5182) and packed for the host.
|
|
|
|
| | Inline source, previous rules | Evidence-file source with `sibling-pair-variance` |
|
|
| --------------------- | ------------------------------------- | ------------------------------------------------- |
|
|
| Manifest | `resource_constraint` (96 MiB budget) | built in 4.1 s, 459 MB maximum RSS |
|
|
| Tools | 446 (with the budget raised by hand) | 63 |
|
|
| Recorded exchanges | 1,218 | 1,218 |
|
|
| Tools with one record | 77% | 49% |
|
|
|
|
The pair rule templated 25 route families as `{repo_namespace}/{repo_name}`.
|
|
A model provider that accepts at most 128 tools refused every request for the
|
|
446-tool manifest with HTTP 400.
|
|
|
|
The 63-tool environment was then exported, and in the AgentEnv runner its
|
|
smoke task passed and twelve generated retrieval tasks scored 12/12 with
|
|
`openai/gpt-oss-120b` and 12/12 with `zai-org/GLM-5.3-Flash`, with no provider
|
|
errors. The 446-tool manifest had scored 0/12 and 8/12 on its own generated
|
|
tasks, the 0 being the provider's tool limit rather than the model.
|
|
|
|
Review then found that the rule as run here also merged one id's
|
|
sub-resources (`/users/alice/followers`, `/users/bob/following` and
|
|
`/users/carol/starred` became one tool), and a pair now qualifies only when a
|
|
literal follows it. The Gitea capture above predates that guard and has not
|
|
been projected with it: whether any of the 63 tools merged sub-resources this
|
|
way, and the tool count with bare `/repos/{owner}/{repo}` routes kept one per
|
|
pair, are not established here.
|
|
|
|
## Linux follow-up (2026-10-20)
|
|
|
|
Based on `dd2bf3565`, including main `5fa4defcb`, with Node 24.18.0 and npm
|
|
11.16.0. The file reader now owns authentication; invalid Evidence returns the
|
|
selected file path and file-relative `content_issues`. Service and advertised
|
|
MCP error-envelope regressions cover this boundary.
|
|
|
|
- Domain analysis views, application analysis views and error projection:
|
|
16 files, 298 tests passed.
|
|
- Compiled analysis-view CLI/MCP and environment serve/export boundaries:
|
|
5 files, 48 tests passed.
|
|
- Serial Turbo run: typecheck, lint, format, knip, package metadata and required
|
|
runtime/contract artifacts: 11 tasks passed. Documentation/catalog validation:
|
|
8 tasks passed. Architecture guards passed across 1,595 source files.
|
|
- Real Gitea 28.2.0 official Linux binary (published compressed SHA-256 checked),
|
|
isolated SQLite database, two users and three public repositories with large
|
|
issue bodies. Actual Chromium navigation captured nine API transactions,
|
|
45 events, no drops, 12,584,312 bytes of Evidence. Inline manifest projection
|
|
was refused with `resource_constraint`; file CLI and default stdio MCP
|
|
projections agreed on five tools with all template rules selected. The
|
|
three trailing user sub-resources stayed separate. Malformed file diagnostics
|
|
named the file's JSON root. HTTP SDK replay verified six complete small
|
|
responses and twelve exact 256 KiB prefixes of large responses, each marked
|
|
truncated. Owned Gitea, browser and replay-server processes were stopped.
|
|
|
|
One CPU and one test worker were used. Sampled sum of descendant RSS every
|
|
50 ms (shared pages can count twice): focused tests 15.834 s / 15.327 CPU s /
|
|
764.69 MiB peak; boundaries 194.916 s / 190.121 CPU s / 1,795.15 MiB;
|
|
static/build 266.311 s / 259.658 CPU s / 4,677.44 MiB; real Gitea workflow
|
|
31.631 s / 24.999 CPU s / 2,146.29 MiB. These are individual runs, not a
|
|
cold/warm comparison; model token usage was not measured. The first external
|
|
probe stopped at the expected CLI refusal because its driver rejected exit 1;
|
|
the driver was corrected and the complete workflow passed. No new macOS,
|
|
Windows, model-task scoring or AgentEnv runner verification is claimed here.
|