feat(product): complete retained multi-session terminal

.agent/autopsies/2026-08-28-controller-projection-cache.md

@@ -0,0 +1,19 @@
+# Controller projection cache autopsy
+
+The `pnpm run test` command at commit `a3f07f4` passed 827 functional tests and failed 1 of 16 performance tests.
+
+The failed test expected stable controller reads to remain below a 10 ms p90.
+The assertion did not print the measured p90.
+
+`ApplicationUiController.view()` cached the cloned application state but called `#project()` on every read.
+The source check was `src/adapters/tui/application-ui-controller.ts:80` before the fix.
+
+The host also ran multiple compiler and model processes during the failure.
+Two isolated repetitions passed, so host saturation determined when the repeated work crossed the threshold.
+
+Classification: infra-bug.
+
+The fix caches the immutable final view by state, failure signals, and local presentation inputs.
+The deterministic check now requires the same view object for every unchanged revision.
+
+Fix: cache the projection in `src/adapters/tui/application-ui-controller.ts`, then rerun the performance and full test commands.

.agent/critical-audit/2026-08-28T21-44-00Z/findings.jsonl

@@ -0,0 +1 @@
+

.agent/critical-audit/2026-08-28T21-44-00Z/manifest.json

@@ -0,0 +1,19 @@
+{
+  "scope": "LIVE-07 and LIVE-08 provider-bound cloud proof changes",
+  "base": "1dca6db491f9a03f898a0437394d5903598c7434",
+  "head": "9671907",
+  "files": [
+    "docs/components/README.md",
+    "docs/components/cloud-proof-observations.md",
+    "scripts/live-required/contracts.mjs",
+    "scripts/live-required/tangle-sandbox-braid-interactive.mjs",
+    "scripts/live-required/tangle-sandbox-braid-soak.mjs",
+    "scripts/live-required/tangle-sandbox-braid-stress.mjs",
+    "test/tangle-live07-wiring.test.mjs",
+    "test/tangle-sandbox-braid-soak-script.test.mjs"
+  ],
+  "projectType": "Node package with live release scripts",
+  "reviewers": ["A-correctness-security", "B-architecture-performance", "C-real-system-coverage"],
+  "mode": "serial",
+  "notInspected": "Unrelated dirty source, package, and verification artifacts were excluded from this commit."
+}

.agent/critical-audit/2026-08-28T21-44-00Z/summary.md

@@ -0,0 +1,31 @@
+# Audit: LIVE-07/LIVE-08 provider-bound proofs — 1dca6db..9671907 — n=8 files, 0 findings
+
+**Verdict:** APPROVE — changed proof paths reject unverified provider evidence · 0 CRITICAL / 0 HIGH / 0 MEDIUM / 0 LOW
+**Worst:** none — no reproducible defect remained in the changed scope.
+**Next:** run protected Sandbox LIVE-07/LIVE-08 provider probes before release.
+
+## Scope
+
+| Field | Value |
+|---|---|
+| Files | n=8 via commit diff and explicit changed-file scope |
+| Base..head | `1dca6db491f9a03f898a0437394d5903598c7434..9671907` |
+| Project type | Node package with live release scripts |
+| Reviewers | A,B,C · serial |
+| Not inspected | Unrelated dirty source, package, and verification artifacts. |
+
+## Findings — 0 of 0, ranked
+
+No changed-scope finding had a reproducible failure scenario after the focused proof suite, syntax check, lint check, diff check, and credential-shaped literal scan.
+
+## Assumptions & unverified
+
+| Assumption | Finding it would flip | Check that settles it |
+|---|---|---|
+| A protected provider run is available for this branch. | Live provider behavior remains unverified. | Run the protected LIVE-07 and LIVE-08 scripts and inspect the redacted receipts. |
+| The public Sandbox API preserves the retained workspace files and usage fields. | Provider readback or telemetry could remain unavailable. | Execute the scripts against Sandbox and require their provider-bound checks. |
+
+## Self-gate
+
+9/9 passed — failed: none.
+1 verdict = decision + 1 number · 2 every finding has file:line · 3 concrete failure scenario · 4 status label · 5 evidence is a pointer · 6 cost both sides · 7 fix + verification per row · 8 zero adjectives standing in for counts · 9 words ≤600 outside tables.

.agent/pursuits/2026-08-28-orchestrate-braid-10.md

@@ -0,0 +1,87 @@
+# Braid completion workflow
+
+## Objective
+
+Ship Braid as a production terminal for concurrent, durable agent work.
+
+The release must start, stream, switch, branch, fork, ask, analyze, supervise, and control parallel sessions through current shared contracts.
+
+Every component must have an implementation note that describes the smallest general design which satisfies its contract.
+
+## Structure
+
+The workflow is a pipeline with one integration barrier.
+
+Three isolated implementation tracks run in parallel before integration.
+
+1. Stream safety owns incremental sanitation, durable event safety, replay, cancellation, and output limits.
+2. Multi-run product owns concurrent branch runs, focused-run state, controls, terminal workflows, and responsive presentation.
+3. Runtime continuity owns retained reconnect, native continuation, exact worker control, and any missing shared runtime contract.
+
+The integration branch accepts a track only after its narrow tests pass.
+
+Shared package changes merge and publish before Braid consumes them.
+
+## Coordination policy
+
+- Three Luna-max workers write independent implementations in isolated worktrees.
+- The operator owns architecture boundaries, integration, live checks, commits, releases, and final claims.
+- Workers must extend current contracts and must not add provider-specific execution paths to Braid.
+- Every durable event change includes idempotency, restart, and migration coverage.
+- Every streaming change includes disconnect, replay, duplicate-event, cancellation, and finalization coverage.
+- Every interaction change includes accept, decline, timeout, cancel, and restart coverage.
+- Every visible workflow includes keyboard proof and captures at 40x12, 80x24, 120x40, and 200x60.
+- A worker can request an upstream change, but the owning package must implement it.
+- No track opens a separate Braid pull request.
+
+## Expected worker count
+
+Three concurrent workers run in each implementation wave.
+
+The expected total is 12 to 18 worker turns across implementation, integration repair, live proof, and adversarial review.
+
+## Waves
+
+### Wave 1: foundations
+
+- Make streamed output safe before it enters the journal or reducer.
+- Replace the global active-run slot with branch-scoped active work and one explicit focus.
+- Reproduce and repair retained Tangle reconnect with preserved failure evidence.
+- Adopt the current runtime package only after its required API is proven.
+
+### Integration barrier
+
+- Integrate passing commits into one Braid branch.
+- Resolve the state, migration, and adapter interfaces once.
+- Run contract, type, dependency, performance, security, and full local checks.
+
+### Wave 2: complete workflows
+
+- Complete exact native continuation and same-session recovery.
+- Complete portable conversation handoff and workspace fork through owning packages.
+- Complete trace analysis actions from run details.
+- Complete supervisor worker navigation, attach, steer, and acknowledged cancellation.
+- Complete the conditional work strip, capability-driven selectors, help navigation, and truthful route copy.
+- Create component notes and a complete source-to-component index.
+
+### Wave 3: proof and delivery
+
+- Run the current live CLI Bridge and Tangle matrix from packed artifacts.
+- Prove concurrent sessions, retained reconnect, interactions, fork, analysis, attach, steer, cancel, cleanup, and replay.
+- Capture dark and light terminal evidence at all required sizes.
+- Record the changed user flows.
+- Run three independent adversarial reviews after all implementation checks pass.
+- Fix every valid finding and repeat the affected checks.
+- Open one pull request per changed repository.
+- Merge and publish dependencies in order.
+- Install the published Braid package in a clean directory and repeat the release checks.
+
+## Synthesis rule
+
+Prefer one general owning contract over a Braid-only special case.
+
+Prefer explicit identifiers and immutable projections over inferred global state.
+
+Reject an implementation if it cannot prove restart, replay, cancellation, capability failure, and bounded resource use where those conditions apply.
+
+The product is complete only when the published package passes the required live matrix and leaves no owned resource behind.

.agent/release-progress.md

@@ -0,0 +1,103 @@
+# Release Progress
+
+## Release record
+
+This file tracks one candidate and its matching public release cohort.
+Populate every `not recorded` value from the same release artifact and verification run.
+Do not copy a version, integrity value, or live result from an earlier cohort.
+
+- Candidate Braid package and version: `not recorded`
+- Candidate Braid commit: `not recorded`
+- Runtime package and version: `not recorded`
+- Sandbox package and version: `not recorded`
+- Agent Interface package and version: `not recorded`
+- Tangle provider package and version: `not recorded`
+- Runtime artifact commit and tarball integrity: `not recorded`
+- Braid artifact tarball integrity: `not recorded`
+- Runtime rollback package and version: `not recorded`
+- Braid rollback package and version: `not recorded`
+- Registry and live-service environment: public npm and public Tangle services
+- Release credentials: protected release environments and `gh-drew`
+
+## Public dependency cohort
+
+- Agent Interface: `not recorded`; registry version, owning-repository checks, and publication proof are pending.
+- Tangle provider: `not recorded`; registry version, owning-repository checks, and publication proof are pending.
+- Sandbox: `not recorded`; registry version, owning-repository checks, and publication proof are pending.
+- Runtime: `not recorded`; registry version, owning-repository checks, and publication proof are pending.
+- Runtime package integrity: `not recorded`.
+- Braid has no file, link, or workspace production dependencies: verify against the final lockfile before publication.
+
+## Prior cohort record
+
+The entries below describe the 2026-08-28 checkpoint only.
+They are historical evidence and do not describe the current release candidate.
+
+- Agent Interface `1.7.1` passed 478 of 478 tests before publication.
+- Tangle provider `0.14.1` passed 170 of 170 tests before publication.
+- Sandbox `0.34.0` came from merged commit `c2abc576520f6303fb26528391c706a150419a33`.
+- Runtime `0.177.0` came from merged commit `4b6ff007`.
+- Runtime package integrity was `sha512-6UYAtUEmhzluJYe6VL4ECKyNCQD0HAx9sch2W5BTcA5R/ddrlANy5jJ3Sq7TPmm07JeeMq3iolTEfrGgbDPQEw==`.
+- Runtime provenance, tag, source, workflow, and isolated installation were verified after publication.
+
+## Completed Product Work
+
+- The Work Strip supports concurrent runs, explicit focus, switching, and targeted cancellation.
+- Native continuation resumes the selected provider session.
+- Cross-runner and cross-workspace forks create new branches without changing agent identity.
+- Interaction views support typed answers, acceptance, decline, timeout, cancel, and restart recovery.
+- Analysis freezes one run, records cited findings, and preserves promotion decisions.
+- Supervisor views project the Runtime worker tree and route steer, cancel, and exact terminal attach.
+- Braid suspends its screen during worker attach and delegates terminal transport to Runtime.
+- All 51 exported terminal components map to 25 component documents.
+- Every component document states its best simple general implementation.
+- The UI captures 16 product states and 21 keyboard frames at all four required terminal sizes.
+
+## Current proof record
+
+Record each result with its candidate commit, command, date, and artifact path.
+Use `not recorded` until the check runs against the release record above.
+
+- Local package and terminal checks: `not recorded`.
+- Provider and Sandbox live checks: `not recorded`.
+- Restart, concurrent-session, and interaction checks: `not recorded`.
+- Package-to-candidate comparison: `not recorded`.
+- Dependency audit and source graph: `not recorded`.
+
+## LIVE-10 status record
+
+- Current LIVE-10 result: `not recorded`.
+- Current evidence path and artifact: `not recorded`.
+- Previous observation from 2026-08-28: Base Sepolia chain `84532` had one active blueprint-12 binary and zero blueprint-12 services.
+- Previous observation from 2026-08-28: The advertised operator endpoint did not complete TLS or expose the required API.
+- Previous observation from 2026-08-28: The operator host was ARM64, while the published Nitro binary was x86_64.
+- Previous observation from 2026-08-28: The host had no Nitro process, Nitro routing, or AWS deployment credentials.
+- Previous observation from 2026-08-28: Production Sandbox had no Nitro driver allowlist or Nitro target configuration.
+- Previous observation from 2026-08-28: Four temporary host-agent sandboxes were deleted and excluded from confidential proof.
+- Release decision: block publication until LIVE-10 passes against the release record above.
+
+## Remaining Sequence
+
+1. Finish and verify warm-start performance against the published dependency cohort.
+2. Regenerate stable visual captures and inspect every corrected flow.
+3. Run the complete local, package, property, performance, and live-provider checks.
+4. Run independent architecture, security, UI, and release reviews and fix each valid finding.
+5. Merge one Braid pull request after current checks and reviews pass.
+6. Publish the candidate package recorded above only after every required live row, including LIVE-10, passes.
+7. Download the public Braid package and compare it with the approved candidate.
+
+## Decision
+
+- Keep npm `latest` on the last verified release while any required live check is unavailable.
+- Merge reusable product and proof improvements when their exact checks pass.
+- Do not replace provider proof with fixtures, local Sidecar runs, or host-agent sandboxes.
+- Do not describe an unavailable confidential path as shipped.
+
+## Timeline
+
+- 2026-08-28: Proved retained Tangle reconnect after forced Braid process termination.
+- 2026-08-28: Implemented multi-run focus, continuation, forks, interactions, analysis, and supervisor controls.
+- 2026-08-28: Published Agent Interface `1.7.1`, Tangle provider `0.14.1`, Sandbox `0.34.0`, and Runtime `0.177.0`.
+- 2026-08-28: Proved the public Sandbox Sidecar input, detach, reconnect, replay, and interaction flow.
+- 2026-08-28: Proved two retained Tangle multirun cohorts with exact cleanup.
+- 2026-08-28: Confirmed that Base Sepolia has zero blueprint-12 services and no reachable Nitro operator.