chore: sync private v5.7.0 (1b6bb16)

[agent-kit-startup]

Summary

• Automated allowlist sync from the private source of truth.
• Release v5.7.0.
• Head branch sync/v5.7.0-1b6bb16.

Release notes

Changed

• Consumer install copy is Cursor-first: root package.json, README eyebrow, and getting-started no longer list VS Code or Windsurf as peer IDEs. Partial generators still ship; Action A7 (generator parity) stays a non-goal.
• Mission Kit landing shows the current public release version (New release pill plus a link to the public GitHub Release) and short product notes in the existing changelog field, matching the original pill-and-frame visual. Release version and product notes are static page fields, not a third-party feed or a factory-changelog dump. Operators stamp those two fields with pnpm landing:update-release (does not deploy).
• Consumer documentation rewritten with usage-first Fable copy. README hero section, getting-started guide, and CONTRIBUTING page now focus on concrete install steps, surfaces, and day-to-day workflow. Retired five-layer claim matrix from consumer storefront; factory evidence files preserved for CI.

Added

• Public changelog filter: public GitHub CHANGELOG.md, GitHub Release notes, and the landing product-notes field publish only consumer and contributor product notes. Factory-only bullets stay behind <!-- changelog-private --> fences. Preview with node scripts/public-changelog.mjs; landing stamp uses --version X.Y.Z --blurb (never pass CHANGELOG.md as --notes-file).
• /kit-staging and /kit-prod wrap native /git-staging and /git-prod (those stay git-only: branch/MR for staging, staging → main plus tag and npm for prod). When a product changelog or release actually changed, the kit bundles add a surgical public-landing field update and, after an explicit Ask, deploy the landing to staging or promote it to production. docs(memory) monitor commits do not trigger a landing deploy. Repo-only shipping stays the native git commands.
• CLI-native Mission Control TUI (agent-kit mission-control): browser-free ASCII view of Mission, Flight Log, Checklist, and Crew Monitor. Reuses dashboard/dashboard-data.mjs / semantic-model.mjs without starting the HTTP server; live refresh uses the existing PERIODIC_REFRESH_MS poll. Visual-kit tokens (HELMET_* / LABEL_MUTED) and TTY / NO_COLOR / CI / AGENT_KIT_REDUCED_MOTION gates stay as in ADR 2026-08-13_cli-visual-kit-space-chrome.md. ADR 2026-08-27_cli-native-mission-control-tui-thin-deps-third-surface.md records stay-inside-thin-deps (@clack/prompts / citty / kolorist only; no Ink/ora/chalk/figlet) and that the TUI is a third surface, not a Path C replacement. Claude Code /agent-kit renders mission-control --once (one-shot; not a live loop). Plan cli-native-mission-control-tui.
• Mission Control Git tab visual tree + new DevOps tab. The Git tab's existing markdown git log --graph block stays as-is; a new state-colored vertical tree renders next to it, reusing the .now-stepper/.now-step-marker timeline primitive built for the Current Mission panel (parses SNAPSHOT.git.graph's commit lines, flattened to one row per commit — HEAD/promoted-to-origin/main-or-origin/staging state-colored, degrading to a compact empty-state on an unparseable graph). New DevOps tab (#devops, nav dot, hash-deep-linkable) is scoped to CI/CD + deploy signal, distinct from the existing local-process-only Processes tab: a pipeline card from a new dashboard-data.mjs gh run list collector (budget-guarded, fails soft to an honest empty-state when gh is unavailable/unauthenticated or the run itself would blow the snapshot's soft time budget), and a best-effort "what shipped recently" deploy-activity card from v* git tags + the latest non-Unreleased CHANGELOG.md entry — deliberately not a live infra/hosting poll, and says so. dashboardDataVersion 1.3.0 -> 1.4.0. Plan mc-git-tab-visual-tree-and-devops-panel.
• Per-agent/subagent context-usage cap, docs + enforcement: ADR 2026-08-24_agent-context-usage-cap-window-pressure.md sets a <=50%-of-window target using the existing context_usage_percent signal from the Cursor preCompact hook — the only real local window-pressure signal, reused exactly as-is (no new token/cost telemetry, no violation of the standing 2026-08-14_mc-idle-session-token-cost-no-go.md no-go). Documents the operator's "190k tokens" framing as a window-size-class ratio (200K-class ~95%, 1M-class ~19% — hence the enforceable cap is a percentage, not a literal token count), and the coverage gap that preCompact is Cursor-only (no signal exists for Claude Code sessions today). packages/cli/src/hooks/pre-compact.ts gained CONTEXT_USAGE_CAP_PERCENT = 50 and threshold logic in buildPreCompactUserMessage: a preCompact firing at/above the cap (or with no reported percentage — auto-compaction implies high pressure) escalates the emitted user_message to name the cap and force a HANDOFF checkpoint before continuing; below-cap firings keep the original gentle wording. .cursor/commands/run-plan.md's tick-close contract gained step 3a treating an observed escalated firing as a mandatory (not optional) HANDOFF checkpoint. New packages/cli/src/hooks/pre-compact.test.ts (5 tests). Also surveyed 5 real third-party token-reduction/context-compression tools (Headroom, Caveman, Graphify, atlassian-labs/mcp-compressor, Tokalator) for scam/legitimacy risk — all rated evaluate-further (no confirmed malicious signal in any, but three show anomalous star-velocity worth independent scrutiny before reliance); none installed or adopted. Plan agent-context-cap-and-token-tool-research.

Fixed

• DevOps tab pipeline card: collectPipelineRuns()'s gh run list timeout (GH_RUN_LIST_TIMEOUT_MS) had no real headroom over this repo's observed latency — 3500ms against live-measured samples spanning 2.1-9.9s across sessions (8 fresh samples this pass: 3476-6584ms, one 7674ms outlier). Fixed in dependency order: (1) the collector now runs last among the snapshot's budget-guarded collectors (after processes, detached-audit-sessions, agentPrompts, externalReports, subagentRuns), not ahead of them, so a slow gh network call can no longer starve the local-only collectors sharing SNAPSHOT_BUDGET_MS; call cost trimmed too (--limit 10 -> 5, unused databaseId/url --json fields dropped — neither is read by the card). (2) GH_RUN_LIST_TIMEOUT_MS raised 3500ms -> 12000ms with real margin over the measured range; SNAPSHOT_BUDGET_MS default raised by the same delta (12000ms -> 20500ms) so the larger reserve the collector needs is actually available. (3) The pipeline card's empty-state copy no longer unconditionally blames "gh is unavailable or not authenticated" — the collector now returns a reason ("budget" when the shared snapshot budget ran out before gh was even attempted, "call-failed" when gh was attempted and failed/timed out) and the card copy follows it. Live 5-run + 1-diagnostic-budget verification (this sandboxed dev environment) found the dominant real failure mode is in fact "budget": a separate, out-of-scope collector (SNAPSHOT.git, ~10 unguarded execSync calls with no budget check) is consuming the entire default budget on its own here, before the snapshot even reaches the now-last-placed pipeline collector — confirmed via a diagnostic run with AGENT_KIT_DASHBOARD_DATA_BUDGET_MS=60000, which returns available: true with real runs. That collector is unguarded/unbudgeted and outside this fix's scope; noted as a residual. Plan close-mc-git-tab-devops-pipeline-timeout-residuals.

Source

• Commit: 1b6bb16
• Head branch: sync/v5.7.0-1b6bb16

[agent-kit-startup]

Superseded by a newer private→public sync.