Inspect session usage and context evidence

Traceability

• Spec ID: harness-inspector-usage-context
• Status: Implemented
• Refs: https://github.com/QoderAI/better-harness/pull/122
• Cursor current-context source precedence: superseded by Read current Cursor composer context usage

Intent

Make Harness Inspector report the token and context evidence that supported coding-agent sessions without confusing provider-specific accounting or embedding sensitive system/developer instructions in its portable HTML. The Inspector should expose what was actually observed, where it came from, and which fields remain unavailable for Cursor, Codex, Qoder, and Claude Code.

This change also closes two correctness gaps found during the evidence audit: Codex developer messages must not be projected as assistant dialogue, and Cursor transcripts without native event timestamps must not silently disappear from the default Inspector window.

Acceptance Scenarios

• AC-1: Codex response_item.message records retain distinct user, assistant, developer, and system roles. Only user and assistant content may enter the Inspector dialogue projection; developer/system content contributes bounded context-manifest counts without retaining text.
• AC-2: Codex event_msg.token_count records produce session usage from the cumulative total_token_usage snapshots. Repeated snapshots are not summed, and a decreasing cumulative total starts a new monotonic segment. The result keeps input, output, cached-input, cache-write, reasoning-output, and explicit total tokens without counting cached input twice.
• AC-3: The shared Inspector usage projection keeps optional cache-creation, reasoning-output, explicit-total, source, basis, and coverage fields. Claude cache-creation tokens survive adapter, session summary, report, and UI projection. Providers without an explicit total continue to show a breakdown without inventing a billing total.
• AC-4: Codex model, effort, model provider, CLI version, context-window usage, compaction count, and context-layer counts are projected as bounded metadata. Cursor Context Usage Canvas evidence may contribute window/category counts only when its composer id matches the Session. Every context projection states that raw text was omitted.
• AC-5: A Cursor transcript without native timestamps remains discoverable. Its file modification time may supply a low-authority source timestamp only when labelled source-file-mtime; it must not be presented as a native event time.
• AC-6: Standalone Inspector and native Studio Session Detail expose one docked Usage and context entry. It progressively discloses token breakdown, context-window occupancy, provenance/coverage, runtime metadata, and honest unavailable states at wide, compact, and narrow widths.
• AC-7: The self-contained report contains no raw base instructions, developer/system message content, context item text, absolute home paths, rate-limit/credit data, encrypted reasoning, or raw tool payloads.
• AC-8: Every retained model-inference event in a Turn exposes its own observed token usage. When that same event includes a context-window observation, the row also shows used tokens, window size, and percent full. Session totals, hard-coded model limits, and later snapshots must never be substituted for missing per-inference evidence.
• AC-9: Standalone Inspector and Studio render per-inference usage as compact, ordered process-trace rows at wide, compact, and narrow widths. Claude or Cursor rows without an observed context window say that it is unavailable; Cursor Context Usage Canvas remains a session-current snapshot rather than being repeated across historical responses.
• AC-10: Without an explicit time window, Cursor coverage treats every matched workspace transcript as relevant, including terminal-only and unreadable transcripts whose native event time is unobserved. A partial in-window subset must not silently reduce the coverage denominator.
• AC-11: Session Detail keeps a compact Usage and context summary in its right-hand outline. The summary leads with current context, comparable net context growth, derived Session processing, and unique model-call count; visualizes only token-weighted context categories; supplies an explicit Other remainder when categories do not cover the observed used context; and never stacks overlapping input/cache/reasoning accounting as if it were context composition.
• AC-12: A labelled View report action in that right-hand summary opens a read-only Usage report inside the existing Session Detail shell. It does not nest another modal or expand the narrow outline into the report. The report has a stable host-native URL state (inspector-view=usage in Studio and session-mode=usage in the standalone Inspector) and a labelled route back to Trace; closing Session Detail retains its existing behavior.
• AC-13: The Usage report leads with context-window occupancy, then shows observed per-inference context progression, separate processing accounting, the current token-weighted context composition when retained, provider accounting, bounded context-layer counts, runtime facts, provenance, and the raw-context omission boundary. Missing category, per-inference, runtime, or context-window evidence stays explicitly unavailable and is never reconstructed from a model-name lookup table.
• AC-14: Studio and standalone Inspector expose the same summary/report semantics, labelled controls, and keyboard-reachable navigation at wide, compact, and narrow layouts. The main Session evidence remains the primary surface until the reviewer explicitly opens the report, and neither surface gains document-level horizontal overflow or browser console/page errors.
• AC-15: Qoder assistant usage preserves its observed message.usage.context_usage_ratio as context occupancy. A Session-local contextWindow may supply the denominator and derived used-token count only when that value was actually retained; otherwise the report shows the observed percentage without inventing an absolute window. Qoder compactMetadata records contribute compaction boundaries.
• AC-16: Claude response usage exposes observed prompt-context tokens as the sum of input, cache-read input, and cache-creation input for that inference. It remains a used-token-only observation when the transcript has no context window, so the report must not infer a model limit or percentage.
• AC-17: Qoder and Claude tool, Skill, MCP, subagent, system-message, and other activity counts must not be presented as Cursor-style token composition. Category composition remains explicitly unavailable unless a host supplies token-weighted categories; partial context observations still appear in the summary, progression, and provenance views.
• AC-18: Cursor, Codex, Qoder, and Claude Code project through the same bounded Usage report contract while preserving their distinct evidence capabilities. Cursor may show native token-weighted categories from a composer-matched Context Usage Canvas; Codex may show per-response used/window occupancy and compactions; Qoder may show ratio-only or ratio plus an observed Session window; Claude Code may show observed prompt tokens without a window. A provider with weaker evidence must never inherit fields from a stronger one.
• AC-19: Claude usage observations are unique by Session-local responseId. Repeated records for the same response contribute one model call and one set of counters; exact duplicates are collapsed, synthetic or all-zero usage is excluded, and conflicting duplicates select one canonical record while retaining a bounded conflict count. The portable report never exposes the raw provider response id.
• AC-20: The Usage report keeps provider-reported totalTokens distinct from derived processed tokens. For a unique Claude response, processed tokens are the additive sum of input, cache-read input, cache-creation input, and output; the Session processed value is their sum across unique responses and carries the basis derived-accounted-usage. It is not labelled as provider total, billing, or cost.
• AC-21: Context progression reports one point per unique retained inference. Each point may expose the absolute prompt snapshot, the net delta from the previous comparable response, processed tokens for that call, and output. The first point is a baseline, a model change is an incomparable boundary, and a negative same-model delta is labelled context shrink/reset rather than negative consumption. Missing windows are stated once for the report instead of repeated on every progression row.
• AC-22: The Session Detail summary separates current context, net comparable context growth, derived Session processed tokens, and unique model-call count. The detailed report uses an absolute context-progression chart plus a separate processing-accounting visualization; it never presents cache/input/output accounting as Cursor-style current-context composition.
• AC-23: Qoder contributes exactly one Usage progression point for each retained model.response.completed event. A nearby project-transcript assistant event may enrich that response with its observed context ratio when both records agree on model and stop reason, but it never contributes a second call. Turn, fork-agent, and other usage-bearing summary events remain Session evidence and never become model calls. Unmatched assistant context stays available to the Session-current context manifest without being invented as a historical response.
• AC-24: Cursor audit evidence contributes one model call only from each usage-bearing afterAgentResponse record. Usage counters repeated on stop or tool lifecycle records remain excluded from Session token totals and progression. A composer-matched Context Usage Canvas remains available to the Session-current context manifest, but it never contributes a historical inference or increases the unique model-call count. Cursor hook counters use the basis agent-response: they describe the retained parent-agent response, not subagent usage, billing, or a reconstructed low-level inference trace.
• AC-25: Every Inspector render without an explicit --since or --until uses the latest 30 UTC calendar days, including the explicit render subcommand and the package npm run inspector path. Either explicit bound keeps the caller-owned window instead of silently adding the default range.
• AC-26: At wide desktop width, the Session Detail outline remains a bounded 330-pixel dock and never derives its grid width from long fact values. Model names and summary metrics truncate inside the dock while retaining the full model value on hover, the main evidence lane keeps at least half of the viewport, and wide, compact, and narrow views have no document-level or outline-level horizontal overflow.
• AC-27: Token-weighted context categories preserve the order retained by the host instead of sorting by token count. For native Cursor evidence this is system_prompt, tools, rules, skills, mcp, subagents, then conversation; segment widths remain token-weighted and unused context is always rendered after the observed categories.
• AC-28: Each Codex rollout's session_meta.payload.id is the accounting and context-progression stream identity. A distinct child rollout may retain the parent id in session_meta.payload.session_id, but its token snapshots, current context, model calls, and cumulative totals must not enter the parent Session. Cumulative decreases start a new accounting segment only within one rollout stream.
• AC-29: Parent-stream context compression remains observable after child-stream isolation. An explicit Codex compacted event contributes one compaction boundary, and a same-model prompt-context drop inside that parent rollout remains one context shrink/reset. Switching to or from a child codex-auto-review rollout must not create a parent compaction/reset.
• AC-30: Each retained Usage progression point preserves its observed response timestamp through derivation and portable projection. The report may bind a point to a user Turn only when that timestamp falls within the Turn's observed time boundary; missing timestamps or unmatched points remain explicitly unlinked instead of borrowing the nearest prompt or response ordinal.
• AC-31: Standalone Inspector and Studio render Context progression with an explicit response-order axis, an observed UTC time range when available, and distinct focusable markers for user prompts, context shrink/reset, and model boundaries. Hover, keyboard focus, or click opens a Tool-calls-style inspector showing response, time, context delta, Turn, and a bounded privacy-filtered prompt excerpt. Only decision-relevant markers enter the tab order, and unavailable time or prompt evidence is labelled honestly.
• AC-32: A Codex token-count snapshot whose invocation input, output, cache, and reasoning counters are all zero carries no observed model inference. The nonzero-total form emitted after compacted is compaction bookkeeping, not a model response or an absolute zero-token prompt. Its cumulative Session accounting remains available, but it does not update current context or add a progression point. The next observed post-compaction prompt snapshot carries the shrink/reset delta, while the report separately states the provider-observed compaction count.
• AC-33: Every normalized event with an explicit provider compaction boundary and an observed timestamp contributes one bounded, privacy-safe Context compaction event to the Session manifest. The standalone activity timeline renders those exact events as distinct Context compressed markers and a text legend. Token-only context shrink/reset points never become compaction markers without explicit provider evidence.
• AC-34: When an explicit provider compaction boundary has a preceding observed absolute Context token snapshot, its bounded manifest event retains that token count and the snapshot timestamp. The Usage report presents the latest Context and each retained historical compaction snapshot in the primary decision tile, keeps supporting facts visually subordinate, and never labels the snapshot sum as Session consumption or total processed tokens.

Provider evidence matrix

Provider	Current context evidence	Token-weighted composition	Progression	Compaction evidence
Cursor	Native Canvas used/window snapshot, matched by composer id (host-context-snapshot)	Native Canvas categories only	Parent-agent response counters when retained; Canvas occupancy remains current-snapshot evidence	Unobserved
Codex	last_token_usage.input_tokens plus model_context_window (prompt-tokens)	Unavailable; bounded layer counts are not token categories	Per token_count response	compacted events
Qoder	context_usage_ratio, optionally paired with a retained Session contextWindow	Unavailable	Per canonical model response, enriched by a matched assistant observation	compactMetadata records
Claude Code	Input + cache-read input + cache-creation input prompt tokens	Unavailable	Per unique assistant response	Unobserved

Non-goals

• Treating local Cursor, Codex, Qoder, or Claude Code files as stable public host APIs.
• Showing or exporting raw system prompts, developer messages, AGENTS.md, CLAUDE.md, rules, skills, tool schemas, or conversation/context item text.
• Adding billing claims, account balances, rate limits, or estimated cost.
• Making Cursor transcript modification time authoritative Session chronology.
• Reconstructing historical per-response context occupancy from a current Cursor Context Usage Canvas snapshot or a model-name lookup table.
• Unifying provider transcript/event IRs beyond the bounded shared usage and context-evidence projection consumed by Inspector.

Plan and Tasks

1. Correct Codex role normalization and add bounded usage/runtime/context metadata events for current rollout records.
2. Extend session summarization with a provider-neutral usage evidence shape, monotonic cumulative-snapshot aggregation, context manifest counts, and explicit timestamp authority.
3. Preserve Claude cache-creation usage and attach a matching Cursor Context Usage Canvas projection while keeping raw context text out of the report.
4. Extend HarnessInspectorReportV1, the standalone report, and native Studio Session Detail with the same progressively disclosed facts.
5. Add behavioral fixtures for role privacy, cumulative usage resets, Claude cache creation, Cursor timestamp fallback/context matching, report escaping, and UI projection.
6. Preserve per-inference usage/context snapshots in Turn order, render them as compact process evidence in both Inspector hosts, and keep missing provider fields explicitly unavailable.
7. Keep Cursor coverage denominators independent of timestamp observability when no time filter was requested, and retain the cross-platform fixture as the behavioral guard.
8. Replace the right-outline fact dump with a compact usage/context summary and a View report action, while keeping provider accounting and context composition as separate visual dimensions.
9. Add a Session-scoped Usage report view with context composition, observed per-inference progression, bounded runtime/provenance details, URL state, and an explicit return to Trace in both Inspector hosts.
10. Normalize Qoder ratio/window/compaction evidence and Claude prompt-token observations without adding a model-window lookup or estimating category tokens from activity counts.
11. Generalize context projection and UI formatting for percentage-only, used-token-only, and complete used/window observations.
12. Verify the shared report contract against all four providers and retain a provider-capability matrix so stronger Cursor/Codex evidence never leaks into Qoder/Claude unavailable states.
13. Own response identity, synthetic exclusion, duplicate collapsing, and additive processing derivation in one adapter-facing session-analysis module (usage-records.mjs). Claude collapses with a latest-payload canonical record and bounded diagnostics; WorkBuddy collapses with a first-observation canonical record and no diagnostics; session-efficiency shares the same identity keys. An adapter whose counters overlap simply does not opt into the additive derivation.
14. Derive the usageReport exactly once, in usage-progression.mjs, from the complete normalized event stream. Report and UI layers only project it: projectUsageReport bounds and validates but never counts, and a missing report projects to the shared EMPTY_USAGE_REPORT so no renderer carries a local default. Deriving the metrics a second time from a display-bounded dialogue is prohibited — it would answer a different question under the same field names.
15. Replace the unbounded progression list with an accessible context chart and a bounded detail table, then add a separate processing breakdown for hosts whose additive accounting basis is known.
16. Surface every retained metric or drop it: baseline context, context shrink/reset and model-boundary counts, processing coverage, and whether the retained progression is a sample must all be visible in both Inspector hosts.
17. Canonicalize Qoder multi-lane usage evidence before report derivation: enrich a retained model response from a one-to-one nearby assistant context observation, keep non-response summary events out of the call series, and cover the real logs-session/project-jsonl/turn-finished shape with a behavioral fixture.
18. Canonicalize Cursor audit usage before report derivation: retain afterAgentResponse as the audit call boundary, keep repeated stop and tool-lifecycle counters out of token accounting, and mark the native Canvas observation as Session-current-only evidence.
19. Apply the 30-day UTC default when no explicit render bound is supplied, while preserving one-sided and two-sided caller-owned bounds.
20. Bound the Session Detail outline grid and its children so long model names and usage metrics cannot widen or horizontally scroll the desktop dock.
21. Preserve adapter-retained category order through Session summarization and both report renderers; verify the sequence against a real Cursor Canvas.
22. Treat the Codex rollout id as the source Session identity and retain payload.session_id only as parent/root linkage, so discovery and hydration never merge child cumulative counters into the parent stream.
23. Add a behavioral fixture with one parent rollout, one real parent compaction, and two interleaved codex-auto-review child rollouts. Pin the parent current context, net growth, cumulative total, model-call count, and single reset/compaction while verifying both child streams stay separate.
24. Preserve normalized inference timestamps in the shared Usage progression contract, then bind projected points to dialogue Turns only through observed time containment.
25. Reuse the Tool calls chart interaction vocabulary for Context progression: response-order and UTC range metadata, semantic boundary shapes, bounded prompt excerpts, and one keyboard-accessible detail inspector in both hosts.
26. Exclude Codex all-zero invocation bookkeeping snapshots from current context and model-call progression, then surface explicit provider compactions alongside observed post-compaction shrink/reset markers.
27. Preserve bounded timestamps for explicit provider compaction boundaries in the Session manifest and project them onto the standalone activity timeline without deriving event times from context-token shrink/reset observations.
28. Retain the latest absolute Context snapshot observed at or before each explicit compaction boundary, then expose those per-boundary token values in the standalone and Studio Usage report hierarchy.

Test and Review Evidence

• AC-1/AC-2/AC-4: focused Codex adapter tests in test/sessions/session-analysis.test.mjs and session-summary tests in test/sessions/commit-session-link.test.mjs.
• AC-3/AC-5: provider fixtures in test/sessions/session-analysis-providers.test.mjs plus Inspector report projection coverage in test/reporting/harness-inspector.test.mjs.
• AC-6/AC-7: generated-report assertions, Studio component tests, Playwright at 1440x900, 1024x768, and 390x844, console/page-error inspection, and saved screenshots.
• AC-8/AC-9: Turn-folding tests for Codex invocation usage and Claude response usage, report-projection privacy assertions, and the same three-width Playwright/standalone visual contract with multiple usage rows expanded.
• AC-10: Cursor facts distinguish absent, terminal-only, and unreadable transcripts must pass on Windows as well as POSIX hosts; the Windows PR job is the authoritative receipt for filesystem behavior.
• AC-11/AC-13: component/report assertions verify Other remainder math, non-overlapping accounting labels, honest unavailable states, and omission of raw context. Category widths are derived only from contextManifest category estimates plus the observed used/window totals.
• AC-12/AC-14: focused Studio Playwright and standalone Inspector checks open the right-side View report action, verify their host-native Usage route state, return to Trace, exercise keyboard focus, and save wide/compact/narrow screenshots with console/page-error and overflow checks.
• AC-15: Qoder provider fixtures cover ratio-only usage, an observed Session window, derived used tokens, and compaction metadata. A local schema audit checks the same bounded fields without retaining transcript text.
• AC-16: Claude provider and Session-summary fixtures verify prompt-context addition across input/cache-read/cache-creation and preserve an unavailable window through report projection.
• AC-17: report/UI assertions distinguish the three partial-context states and keep category composition unavailable rather than substituting activity counts.
• AC-18: one report-projection fixture covers Cursor native categories, Codex full used/window occupancy, Qoder percentage-only occupancy, and Claude Code used-only prompt context. Local replay counts are recorded separately from fixture-backed capability evidence.
• AC-19/AC-20: Claude fixtures include repeated identical response ids, synthetic/all-zero responses, and one conflicting duplicate. Assertions cover canonical selection, collapsed/conflict counts, unique call count, additive processed totals, and unchanged provider-total semantics. The shared collapsing and derivation helpers carry their own unit coverage for both canonical strategies and for the additive opt-in.
• AC-21/AC-22: buildUsageReport tests cover baseline, growth, zero delta, same-model shrink, and model-change boundaries, plus partial processing coverage and bounded sampling that keeps Session totals complete. A report-model test pins the projection to the derived report while the fixture dialogue is deliberately shorter, so a reintroduced second derivation fails. Studio and standalone visual checks verify the shared summary metrics, chart labelling, bounded detail rows, keyboard reachability, and separation between context composition and processing accounting.
• AC-23: Qoder provider tests retain the real multi-lane ordering of logs-session model responses, project-jsonl assistant context observations, and usage-bearing turn/fork summaries. Assertions verify one progression point per canonical response, one-to-one context enrichment, unmatched context fallback, and exclusion of non-response summary events.
• AC-24: Cursor provider fixtures retain one usage-bearing afterAgentResponse, two matching usage-bearing stop records, and a composer-matched native Context Usage Canvas. Assertions verify one model call, one set of Session token counters, current Context Usage availability, and exclusion of both repeated stops and the Canvas snapshot from progression.
• AC-25: a frozen-clock CLI fixture invokes explicit render without bounds and inspects the embedded report filters; companion fixtures keep explicit one-sided bounds unchanged.
• AC-26: standalone and Studio browser checks measure the Session Detail grid, outline, fact rows, and usage summary at 1440x900, 1024x768, and 390x844. The outline has no horizontal scroll, the document has no horizontal overflow, and the long model value remains available through title.
• AC-27: a Session-summary fixture retains intentionally non-size-sorted Cursor categories and asserts the provider sequence in the report projection and rendered composition legend. A local replay records the same seven category ids from the matched native Canvas without retaining item text.
• AC-28/AC-29: a Codex adapter fixture writes a parent rollout plus child rollouts whose payload.session_id points back to the parent. Assertions pin distinct discovered source refs and parent-only hydrated events, then the Session summary and Usage report must retain the parent's own cumulative total, current context, model calls, and one real compaction/reset without counting child snapshots.
• AC-30: shared Usage progression tests retain a native timestamp through derivation and projection. A report-model fixture places points inside and outside observed Turn windows and asserts that only contained points receive a bounded Turn index and user prompt.
• AC-31: Studio Playwright and standalone browser checks focus and click prompt and context-boundary markers, inspect their response/time/Turn/prompt detail, verify bounded tab stops and visible focus, and retain the existing 1440x900, 1024x768, and 390x844 overflow and console-error checks.
• AC-32: the parent/child Codex rollout fixture places an all-zero invocation snapshot immediately after compacted. Assertions retain its cumulative Session total but exclude it from model calls and context points, keep the next real prompt as one shrink/reset, and pin the provider compaction note.
• Cross-platform evidence: focused tests must use node:path and temporary directories, then the full Node 22/24 macOS, Windows, and Ubuntu PR jobs remain authoritative for target-platform acceptance.
• Privacy risk: local host schemas may contain prompts, instructions, account data, and raw tool payloads. Tests assert on normalized shapes and forbidden secrets; portable output receives only enumerated metadata fields.
• Accounting risk: providers disagree on whether cached tokens are a subset of input. The UI uses an explicit provider total when present and never adds cached input to input as a generic fallback.
• Timestamp risk: Cursor source-file mtime is labelled as low-authority evidence and never changes eventTimestampCoverage from unobserved.

Validation Evidence

• The first follow-up Windows Node 22 PR run exposed AC-10 by reporting one relevant Cursor transcript where the fixture contained two. The initial denominator fix passed locally but exact-main run 33227758857 showed that the internal facts freeze was still being mistaken for a caller-supplied until on Windows. The repair now labels that injected boundary explicitly, and the fixture deterministically places transcript mtimes on both sides of it. The focused facts/provider suite passed 93 tests and npm run check passed with 1,594 root tests, 173 Harness tests, 31 Harness UI tests, 494 Studio tests, generated-source verification, and package/archive verification. A refreshed Windows job remains the authoritative platform receipt.
• npm run check passed on supported Node 24.19.0: root Vitest reported 1,549 passed and 2 skipped; Harness reported 172 passed; Harness UI reported 31 passed; Studio reported 293 passed; generated-source and package/archive verification also passed.
• The focused Studio Playwright scenario passed at 1440x900, 1024x768, and 390x844 with no console/page errors or document-level horizontal overflow.
• node scripts/harness-inspector/visual-contract-check.mjs passed all 15 standalone Inspector surface/viewport combinations with no clipped facts, below-floor targets, page errors, or horizontal overflow.
• A live local Inspector render over Codex, Cursor, and Claude evidence produced 58 sessions and 8,322 tool calls. It projected 5,341 per-inference usage rows for 44 of 45 Codex sessions, all with same-event context-window evidence, and 981 per-inference usage rows for all 13 Claude sessions, whose events supplied no context-window field. The current workspace supplied no matching Cursor session, so Cursor per-inference/context matching remains fixture-backed.
• A local Qoder/Claude replay after AC-15 through AC-17 read all 523 discovered Qoder Sessions and all 13 Claude Sessions without a read failure. Qoder produced 27 bounded context manifests: 3 with observed used/window totals, 18 percentage-only, and 6 compaction-only. The first Claude replay counted 1,741 raw response observations, but the AC-19 audit found only 966 actual model calls after collapsing 768 repeated records and excluding seven synthetic or all-zero observations. AC-19 through AC-22 now use those deduplicated model calls for Session processing and progression. Neither provider produced a token-weighted category manifest.
• The focused provider, Session-folding, and Inspector suites passed 176 tests; the root suite passed 1,552 tests with 2 skipped. Studio built successfully, its focused Playwright file passed 23 tests, and the standalone visual contract passed all 15 surface/viewport combinations.
• The Canvas preview smoke returned HTTP 200 for both /health and /canvas-module.js, confirming the TSX transform and SDK runtime endpoint remained available after the UI changes.
• A four-provider local replay read 93 of 93 Codex Sessions and 8 of 8 Cursor Sessions without a failure. Codex produced 93 bounded context manifests, 91 with complete used/window evidence, from 16,352 usage/context observations. No current-workspace Cursor Session matched a Context Usage Canvas. A bounded global Cursor schema audit found 2 valid native snapshots, both with composer, used/window, category, and item fields (14 categories and 177 items total), so current-workspace Cursor context remains fixture-backed rather than borrowed from an unrelated composer.
• The AC-19 through AC-22 focused provider, Session-folding, and report suites passed 133 tests. A display-bounded 1,100-response fixture retained complete Session metrics and a 1,000-point progression sample with both endpoints.
• A real four-provider render of the current workspace discovered 475 Qoder, 13 Claude Code, 8 Cursor, and 66 Codex source Sessions; the workspace filter retained 77 Qoder, 4 Claude Code, no matching Cursor, and 19 Codex Sessions. One long Claude Code Session projected 312 unique model calls after collapsing 177 duplicate records, with a 47,153-token baseline, 370,640-token current context, 323,487-token net growth, one shrink/reset, and 73,088,320 derived processed tokens.
• The focused native Studio Playwright scenario passed after opening the right-outline View report entry and checking the detailed Usage report at 1440x900, 1024x768, and 390x844. The standalone visual contract passed all 15 surface/layout combinations with zero below-floor text, unreachable clipping, page/console errors, or document-level horizontal overflow.
• The final repository Vitest run passed 1,557 tests with 2 skipped; the native Studio suite rebuilt successfully and passed all 293 tests. JavaScript syntax, TypeScript/build output, git diff --check, and the Canvas preview /health and /canvas-module.js endpoints also passed.
• AC-23 replayed the comparable retained Qoder usage Sessions before and after canonicalization. Their reported model-call total fell from 127 multi-lane observations to 80 canonical responses, exactly matching the retained model.response.completed count in every Session; the representative Session changed from 51 to 31 calls while preserving all 18 matched ratio progression points and its 11.6% Session-current occupancy. Five Claude Sessions were unchanged, and 18 stable Codex Sessions were unchanged while the active Codex Session continued to grow during validation. The focused AC-23 tests passed 109 assertions, the final root suite passed 1,572 tests with 2 skipped, Harness/Harness UI/Studio passed 172/31/293 tests, package verification passed, and the 15-surface visual contract plus focused Studio Playwright scenario completed without overflow or page errors. A privacy scan of the real portable report found no absolute home path, raw response id, encrypted content, rate-limit data, or credit data.
• AC-24 replayed all 96 retained Cursor project Sessions across 10 local project roots. The normalized report retained 30 afterAgentResponse observations, exactly matched 30 reported calls, and retained zero usage-bearing stop records after excluding 56 repeated local stop payloads. The composer-matched slice-compiler Session changed from 10 observations to 3 canonical parent responses while preserving its 56,860 / 300,000 (19%) current-context snapshot and all 7 categories, whose token estimates still sum to 56,860. A multi-response local audit sample had decreasing output counters across responses, confirming that Cursor hook counters are response-local rather than cumulative Session snapshots. The focused provider/Session/progression suites passed 115 tests. The final npm run check passed 1,573 root tests with 2 skipped, Harness/Harness UI/Studio passed 172/31/293 tests, generated sources remained clean, and package verification passed for both npm and the runtime archive.
• AC-25 added a frozen-clock explicit-render fixture. With no bound it projected 2026-07-16T00:00:00.000Z through 2026-08-14T23:59:59.999Z; --since-only and --until-only invocations retained a null opposite bound. The focused report/provider run passed all 89 tests.
• AC-26 used Chrome against the real slice-compiler Cursor Session. At 1440x900, 1024x768, and 390x844 the Session outline had zero horizontal overflow and widths of 330, 292, and 390 pixels; the primary evidence lane used 77%, 71%, and 100% of the viewport. The long four-model value truncated with its full title, the focused native Studio Playwright scenario passed, and the standalone 15-surface visual contract reported zero document or outline overflow, clipping, below-floor text, and page/console errors.
• AC-27 read the matched native Cursor Canvas as seven categories in the order system_prompt, tools, rules, skills, mcp, subagents, and conversation. The regenerated Session summary and Chrome Usage report kept that sequence, preserved the 56,860 / 300,000 (19%) totals, and appended the unused 243,140-token window after the seven observed segments.
• The final npm run check passed 1,574 root tests with 2 skipped, Harness/Harness UI/Studio passed 172/31/293 tests, generated sources remained clean, and package verification passed for 597 npm and 867 runtime entries. The Canvas preview returned HTTP 200 for /health and /canvas-module.js.
• AC-28/AC-29 isolated parent and child Codex rollout streams while retaining the parent's own compaction/reset evidence. The focused fixture and the real parent rollout no longer admit child cumulative counters into the parent model-call series.
• AC-30 through AC-32 replayed the current parent rollout through the package better-harness inspector CLI. One observed replay retained 340 real model responses, 7 time-contained user-prompt markers, and 3 provider compaction boundaries. The three post-compaction points remained shrink/resets, while all-zero bookkeeping snapshots contributed zero chart/model-call points; the first reset resolved to 35.6K context and a -204.5K delta instead of a false zero-token prompt.
• AC-33: focused summary and report-projection tests retain explicit compaction timestamps, while the standalone browser gate checks visible timeline markers, legend copy, and the absence of false markers for shrink-only data.
• AC-33 replayed the real 492-call Codex Session with four explicit compaction timestamps at 07:17, 08:56, 09:46, and 11:29 UTC. The browser activity gate rendered four focusable Context compressed markers and the text legend; each corresponding context shrink/reset snapshot remained a separate usage point several seconds later.
• Final AC-33 regression evidence: npm run check passed 1,604 root tests with 2 skipped, Harness/Harness UI/Studio passed 173/31/512 tests, and package verification passed for 625 npm and 895 runtime entries. Preview health and /canvas-module.js both returned HTTP 200.
• AC-34 replayed the retained 81.1K-context Codex Session and preserved 236.2K and 223.0K snapshots immediately before its two explicit compaction boundaries. The standalone Usage report rendered 81.1K + 236.2K + 223K with a descending current/history/supporting-fact type hierarchy and retained exact snapshot times in hover text. The visual contract passed wide, compact, and narrow layouts with zero overflow, clipping, below-floor text, or browser errors. Focused source/report tests passed 89 cases, Studio typecheck passed, and the final npm run check passed 1,604 root tests with 2 skipped, Harness/Harness UI/Studio at 173/31/512 tests, and package verification at 625 npm and 895 runtime entries. The focused Studio browser scenario was blocked before Usage by a pre-existing fixture assertion expecting 5 calls while the current dirty-worktree fixture exposes 3.
• The focused provider/progression/report suites passed 98 tests. Native Studio TypeScript/build and the focused interactive Playwright scenario passed, including marker focus, click, time/Turn/prompt detail, and wide/compact/narrow screenshots. The standalone visual gate exercised all 15 surface/layout combinations with zero overflow, clipping, below-floor text, or page/console errors.
• The final Node 24 npm run check passed 1,582 root tests with 2 skipped, Harness/Harness UI/Studio passed 173/31/494 tests, generated sources remained clean, and package verification passed for 605 npm and 867 runtime entries.