Days 1–5: application comparison baseline, scope design and roadmap

[pengfei-threemoonslab]

Scope

Days 1–5 of the owner-approved roadmap. Refs #778, #868, #580, #655, #610, #795, #812, #780, #369. These implementation epics remain open.

• Select application-agent comparison as primary, bounded host maintenance, named accountable owner and technical executor; mark prior recruitment-first scheduling historical while retaining Value gate before any outreach: measure automatic author-actionable value on a released build #830.
• Rebaseline 27 pinned PRs on PyPI 1.1.0/contract 40 and main d9a6d0e/contract 41: 54 scans; separately run eight fresh exact-ref verifies across four representative PRs.
• Record exact wheel digest, dependencies, source/config identity and limits. No qualifying ten-case or adoption claim.
• Specify independent scopes, truthful absence/relocation, provenance and advisory versus reviewed-base boundaries for Verifier: bind absent and relocated scoped base inputs without inventing an empty agent surface #580/shipgate diff v1 across Route A: manifest-free base synthesis; the capability delta is never unavailable on first adoption; verdict only as a footer #655.
• Reproduce Preflight: empty plan exposes merge/completion permissions without verifier identity #610's empty-plan permission defect; reconcile remaining Show concrete field changes so an unfamiliar reviewer can act on the first advisory result #795/Show what each host comparison checked, partially checked and could not establish #812/Add an exercised host-only advisory PR recipe with no manifest or saved baseline #780/feat(control): carry structured argv on the operational control contracts #369 criteria without false closure.

Engineering review

Two coding-agent self-review/address rounds, documented in docs/research/application-days1-5/review.md. Corrected verifier-artifact projection, truncated error data and scan/verify denominator ambiguity. No remaining blocking design findings. This is not independent human qualification.

Validation

• Ruff: pass. git diff --check did not pass at 8a0dbbd: README.md had a trailing space and replay.csv used CRLF line endings. Both were fixed in 523725f, and it passes now.
• CI recipe / adoption pin targeted tests: 16 passed.
• Full hosted CI passed on 8a0dbbd: all three suite shards, core test job, Windows/clean-checkout launchers, both MCP dependency jobs and both verifier checks. Release-tag consistency is intentionally skipped on PRs. The first local run failed at sandboxed wheel-build setup; the network-enabled duplicate local run was stopped after hosted full validation passed. The local run also exposed fixture commits blocked by the machine-wide Git ignore rule for .claude/settings.local.json; all 41 tests in that file pass with core.excludesFile=/dev/null scoped to the test process. User Git settings were not changed. No local full-suite pass is claimed.
• Documentation/research change only; no manifest, policy, instruction or runtime implementation edits.

Review follow-up (2026-09-27)

Review fixes are in 523725f:

• a dated Status as of 2026-09-27 note in ROADMAP and in the research README: Non-GitHub CI recipes run python -m pip from the untrusted checkout, so a PR-supplied pip/ package can execute before install #787, Verifier: bind absent and relocated scoped base inputs without inventing an empty agent surface #580 and Google ADK: resolve repository-local imported functions and module-qualified tool bindings #864 are closed, and feat: compare application agent wiring without prior setup #871/fix(diff --application): key SDK identity on imports; unobserved agents are not removals #873/fix: compare application wiring past unrelated links and submodules; read google.adk.Agent #877/Resolve repository-local imported tools for Google ADK and OpenAI Agents SDK (#864) #879 are merged but not yet released;
• a superseded-in-part banner on comparison-design.md (feat: compare application agent wiring without prior setup #871 shipped a new --application flag with its own advisory schema);
• <a id="lead-wedge-focus"> kept above the renamed ROADMAP heading, so docs/engineering/v1-release-readiness.md still links to it;
• the whitespace and line-ending fixes above.

Review and fixes by a coding agent (Claude Code), not an independent human approval.

[pengfei-threemoonslab]

Days 1–5 validation closeout

The two coding-agent review/address rounds are recorded in this PR. Full hosted CI passed on the published head; the release-tag-only check is intentionally skipped on PRs. The PR is ready for maintainer review.

Local validation: lint passes; the initial sandboxed wheel-build setup needed network. A separate fixture failure was traced to the existing global Git ignore rule for .claude/settings.local.json; all 41 tests in that file pass when only core.excludesFile is isolated to /dev/null for that process. No user Git setting was changed and no local full-suite pass is claimed.

#869 records the baseline and design; #870 delivers #787's bounded repair with 14 focused tests and a current, complete verifier result. #610 remains a reproduced contract defect; #580/#655/#867 implementation and the ten-case proof remain in the next phase. Selected issues and milestones have been reconciled without closing those outstanding acceptance criteria.