Improve context transparency and review references; validate semantic navigation #913

Problem

The community audit identified three possible improvements, but package feature lists do not establish either an OpenPI gap or an acceptable implementation. Current baseline is OpenPI 1120cafcd3f5c1ad466df1c7b823976a3e911f72, with Pi SDK 1.2.0.

Scope and value

Advance three bounded investigations and implement the smallest changes supported by the evidence:

• Context transparency: show the composition of the context observed at Pi's existing per-call hook through the existing Web session usage view. Keep observation, provider-reported usage, estimates and unknowns distinct.
• Semantic navigation and differential diagnostics: compare real cross-file cases and controlled diagnostic changes with the current tools before deciding whether an integration is worth its dependencies and lifecycle cost.
• Review source references: preserve the displayed review identity and selected excerpt when feedback enters the normal composer; investigate the same mechanism for plan excerpts.

Ownership and limits

Pi continues to own Sessions, context construction, tools, provider requests and lifecycle. The context hook is an observation point, not proof of the final network request or provider receipt. Existing Web trajectory records are Session projections, not captured request payloads.

The first implementations add no model-facing tool, forced review loop, model call, configuration toggle, telemetry export, raw prompt archive or parallel comment database. Semantic navigation was initially experimental; its bounded native Skill implementation now has local real-server and package evidence, with all final CI jobs now passed. Review selection and feedback submission do not imply approval.

Related work: #194 (historical context-surface research), #454 (exact request/timing observation), #903 (separate opt-in diagnostic export initiative). This issue does not expand those initiatives.

Acceptance evidence

• Fixed source and runtime identities for experiments; source-based expected results independent of tool output.
• Context measurement does not modify its input, honestly labels its measurement and observation boundary, and stays isolated across Session lifecycle changes.
• Review references preserve old/new side, original excerpt and the displayed source identity; selecting a reference neither sends feedback nor approves work.
• Focused meaningful checks, bun run check, bun run test, and matching UI/runtime smoke evidence are recorded separately.
• Reusable findings, rejected approaches and remaining limitations are recorded in the research record.

Starting references

• pi-context-view
• pi-lens
• pi-lsp and context management
• pi-review-loop
• Plannotator

This tracks evidence and delivery, not adoption of a new architectural constraint.

Delivery status

• Context composition: PR #914; local full checks/tests and isolated native Web browser acceptance passed; all 19 final-head CI jobs passed; independent GitHub approval is still required.
• Review and plan source references: PR #915; local full checks/tests and native plan-ready/browser acceptance passed; all 19 final-head CI jobs passed; independent GitHub approval is still required.
• Semantic navigation: actual project TypeScript 7 stdio server resolved alias and same-name definitions. A controlled diagnostic-delta case exposed a false negative, and server termination returned a nonzero exit. These initial results are preserved in the research record. A second TS7 pull-diagnostics probe retains same-signature occurrences as ambiguous rather than silently cancelling replacements. The bounded native Skill is now PR #916, head 1919dccd18e6641224bd9be3aa468245e54f24f7. Full local check/test passed (Node 2469 pass / 8 skip, UI 1632 pass), as did actual TypeScript navigation, six diagnostic comparisons, native Bash cancellation and extracted-package discovery. The explicit three-Skill manifest preserves object-source users; string sources now discover the existing two orchestration entries too. All 19 final-head CI jobs passed, including actual native loader tests on Pi minimum/latest and Windows. Independent GitHub approval is still required. See the semantic research record.

The three implementation PRs are independent and target main; none is a native stack. A local combined-tree check passed at a07d1b21691efc689082ba7388f617547c2c0df1: full check and test passed (Node 2471 pass / 8 existing skip, UI 1638 pass). #914/#915 conflict only in generated app.js/build-info.json; both were rebuilt from combined sources. #916 merged cleanly. This local proof is not a remote merge. Selection and feedback still use normal editable drafts and explicit user submission.

Completed delivery

All three PRs (#914, #915, #916) were merged with explicit user-authorized administrator review bypass. Final main is f50cd7d30138ca0b13d2494ba2e20750ac612650; its tree exactly matches the independently validated combined tree a8680192d7b684f26579c3ef4281377f79bb838c. Full combined check/test passed (Node 2471 pass, 8 existing skips; UI 1638 pass). The conflict-resolution head of #915 also passed all CI jobs in run 42317790629 before merge. Earlier delivery-status paragraphs preserve the investigation history; this completion entry is current.