Inspector cross-provider session discovery

Traceability

• Spec ID: inspector-cross-provider-sessions
• Status: Implemented

Intent

better-harness harness-inspector render currently reads sessions from exactly one --platform (default qoder). Developers who alternate between hosts such as Qoder, Claude Code, and Codex on the same repository must run the inspector once per host and never see one merged delivery picture. The intended workflow is npx @qoder-ai/better-harness inspector: discover the current project's sessions across every supported session-analysis provider automatically, merge them by recency, and render one self-contained HTML report.

Acceptance Scenarios

• AC-1: With --platform all (the new default), the inspector enumerates every platform in SUPPORTED_SESSION_PLATFORMS, skips providers whose evidence roots do not exist locally without failing, and merges discovered sessions from the remaining providers into one report.
• AC-2: A provider that throws during scope resolution, discovery, or session hydration does not fail the render; the failure surfaces as a privacy-safe report diagnostic naming the platform.
• AC-3: Merged sessions are ranked by lastSeen descending and bounded by the existing global --max-sessions limit; every rendered session keeps its platform identity through the existing platform field and <platform>/<sessionId> locator.
• AC-4: --platform <host> and --platform <a>,<b> still scope the render to the named providers; an unsupported platform name fails fast with a usage error (exit 64) that lists the supported platforms and leaks no option value.
• AC-5: better-harness inspector resolves to harness-inspector as a hidden alias, and an argv that starts with an option flag (for example inspector --out report.html) implies the render command. Empty argv and --help keep printing help without reading workspace or provider state.
• AC-6: The report model gains a providers array (platform, status, discovered and included counts) and the HTML header badge reflects the providers that contributed sessions; the report stays privacy-safe (no home paths, no raw error payloads).

Non-goals

• No cross-provider session dedupe beyond the existing per-platform identity; session ids are already namespaced by platform in locators.
• No new session-analysis subcommand and no change to its single-platform contract; aggregation lives in the collector used by the inspector.
• No provider filter panel or per-provider lanes in the workbench UI; the UI change is limited to the header badge plus the existing per-session locator.
• No package rename; the published bin stays better-harness under @qoder-ai/better-harness.
• No per-provider session quota; the global recency-ranked --max-sessions bound is the declared semantic.

Plan and Tasks

1. scripts/commit-session-link/session-source.mjs: add collectMultiPlatformSessionSummaries(\{ workspace, repoRoot, platforms, since, until, maxSessions, includeToolTrace, includeDialogue, createAnalyzer \}) returning \{ sessions, providers \}. Discovery phase per platform (resolveScope → discoverSourceRoots → short-circuit when no root exists → discoverSessions) with per-provider try/catch; hydration phase only for the global top-N candidates. createAnalyzer is injectable for tests. Export through scripts/commit-session-link/index.mjs.
2. scripts/harness-inspector/cli.mjs: parse --platform as all, a single host, or a comma list validated against SUPPORTED_SESSION_PLATFORMS (default all); route through the multi-platform collector; map provider errors and no-evidence providers to report diagnostics; treat a leading option flag as an implicit render; extend help text and stdout JSON with provider counts.
3. scripts/harness-inspector/report-model.mjs: project a providers array (safe platform text, bounded counts, status whitelist) and widen the filters.platform text bound to hold a comma list.
4. scripts/harness-inspector/render-html.mjs: derive the header badge from providers that contributed sessions, falling back to the requested filter.
5. scripts/better-harness-cli/registry.mjs: add hidden alias inspector to the harness-inspector command.
6. Tests in test/inspector-cross-provider.test.mjs plus targeted additions to existing suites (alias inventory already generically covered).

Decision rationale: aggregation belongs next to collectSessionSummaries because workspace binding is already implemented inside every platform analyzer's discoverSessions; a global recency bound (not per-provider quotas) matches the single-platform semantic and the "what happened recently in this project" reading of the report.

Test and Review Evidence

• node --test test/inspector-cross-provider.test.mjs — AC-1, AC-2, AC-3, AC-4, AC-6 via injected fake analyzers and spawned CLI usage-error checks.
• node --test test/harness-inspector.test.mjs — existing AC regression plus AC-5 help zero-side-effect guard.
• node --test test/better-harness-cli.test.mjs — AC-5 alias dispatch through the generic alias inventory paths.
• npm test — full suite before commit (1362 pass). The frozen commands --json and schema baselines in test/scripts-refactor-contract.test.mjs were revised after diffing the inventories; the only change is the declared hidden inspector alias.
• Manual smoke: node scripts/harness-inspector/cli.mjs --max-sessions 8 --commits 20 --out <tmp> on this repository enumerated all 10 providers (4 with evidence), merged the top 8 sessions from qoder and codex, and the header badge rendered qoder · codex · 8 sessions; better-harness inspector --help dispatches through the alias.
• Risk: default --platform changes from qoder to all; render output for existing users may now include other hosts' sessions for the same workspace. Mitigation: sessions remain workspace-bound by each provider adapter and the report remains local-only and privacy-projected.