@inneranimalmedia/agentsam-sdk 2.6.2 → 2.6.3

package/docs/plans/MULTI-AGENT-SPRINTS-2026-09-19.md

@@ -0,0 +1,373 @@
+# Multi-agent sprint plan — remaining AgentSam work
+
+**Purpose:** Finish the open post-PR51 backlog end-to-end across bounded
+AgentSam / Cursor Cloud Agent sessions, and produce a comparable grade for
+Cursor Cloud Agent API value on real AgentSam work (not toy prompts).
+
+**Status authority:** [`RECONCILIATION-2026-09-19.md`](./RECONCILIATION-2026-09-19.md)  
+**Canonical contract:** `AGENTSAM.md` (via `AGENTS.md` shim)  
+**Repo:** `SamPrimeaux/agentsam-sdk` · start every session from current `origin/main`
+
+**Baseline after push (2026-09-19):**
+```text
+origin/main: 67f4a75541855d16c5baecf29007cab05512b532
+  f6dafed  feat(instructions): AGENTSAM + AGENTS/CLAUDE shims
+  67f4a75  docs: post-PR51 reconciliation status authority
+```
+
+**Global hard rules (every session)**
+
+1. Read `./AGENTSAM.md` first. Do not redesign storage / invent Durable Object defaults.
+2. `git fetch --all --prune` · `git worktree list` · `git status -sb` before edits.
+3. No `npm publish`, Worker deploy, worktree delete, force-push, or branch rewrite without Sam approval.
+4. Prefer `git worktree add` / feature branch + PR over direct `main` commits for implementation sessions.
+5. Inspect named parallel branches before rebuilding equivalent work.
+6. Evidence required: commands run, test output, SHAs. No “should work.”
+7. Stop at the session’s Done definition — do not expand into the next session’s scope.
+
+**Recommended models (Cursor Cloud Agent)**
+
+| Session type | Model | Effort |
+| --- | --- | --- |
+| Default AgentSam implementation | Gemini 3.8 Flash | High |
+| Long tool-chain / multi-branch recon | Muse Spark 1.3 1M | Medium → High if stuck |
+| Already-approved mechanical apply | GPT-5.6 Terra | Fast |
+| Architecture stuck / merge conflict hell | GPT-5.6 Sol | default |
+
+**A/B grade track (optional, after S1–S2 have a golden brief):** run the *same*
+S1 brief on Gemini 3.8 Flash High vs Muse Spark 1.3 1M Medium and score both
+with the rubric in §Grade.
+
+---
+
+## Cloud Agent API evaluation design
+
+Each session below is a **Cloud Agent job unit**: fixed prompt, fixed repo,
+fixed Done checklist, fixed grade dimensions. Run them as separate API
+invocations (or sequential AgentSam CLI sessions) so you can score:
+
+| Dimension | What to score (1–5) |
+| --- | --- |
+| D1 Inspect-before-invent | Opens existing files/branches before writing parallel systems |
+| D2 Architecture fidelity | Honors AGENTSAM.md, storage law, session boundaries |
+| D3 Tool-chain survival | Completes 15–40+ tool steps without abandoning |
+| D4 Finish | Hits Done definition; opens PR or leaves reviewable branch |
+| D5 Verify | Runs named tests / live CLI checks; pastes evidence |
+| D6 No rebuild | Does not recreate studio-auth / CMS / connections / help-all |
+| D7 Prose junk | Low fluff; high signal in final report |
+| D8 Cost / latency | Tokens + wall time for the same brief (A/B only) |
+
+**Scorecard file (create once, append per job):**
+`docs/plans/CLOUD-AGENT-SCORECARD-2026-09.md` (session id, model, SHA in/out,
+D1–D8, notes, PR URL).
+
+---
+
+## Dependency graph
+
+```text
+S0  bootstrap (track UI draft + scorecard stub)
+ │
+ ├─► S1  interactive presence (UI req §§1–5)
+ │     └─► S2  status chrome + /model catalog surface
+ │
+ ├─► S3  studio-auth branch decision (parallel with S1)
+ ├─► S4  connections-binding branch decision (parallel with S1)
+ ├─► S5  CMS pattern docs + scaffold design (parallel)
+ │         └─► S6  scaffold-from-real-app implement
+ │
+ └─► S7  terminal_remote enrollment (platform; may touch IAM)
+         ⚠ not blocked by S1–S6; do NOT treat as OAuth
+
+S8  optional Muse vs Gemini A/B on frozen S1 brief
+S9  backlog only if capacity (errors / command manifest) — out of critical path
+```
+
+**Critical path to “remaining sprints done”:** S0 → S1 → S2 → (S3∨S4 as needed) → S5 → S6, with S7 on a platform track.
+
+---
+
+## S0 — Bootstrap · ~15–30 min · Terra Fast or Flash Medium
+
+**Goal:** Make the interactive UI brief git-tracked and give Cloud Agents a
+scorecard stub + updated recon tip SHA.
+
+**Branch:** `chore/sprint-bootstrap-20260919`
+
+**Prompt (paste into Cloud Agent):**
+```text
+Repo: agentsam-sdk. Start from origin/main. Read AGENTSAM.md.
+
+1. git fetch --all --prune; git worktree list; do not delete worktrees.
+2. Add and commit the currently untracked file
+   docs/plans/AGENTSAM-INTERACTIVE-UI-REQUIREMENTS.md if present locally;
+   if missing, stop and report.
+3. Create docs/plans/CLOUD-AGENT-SCORECARD-2026-09.md with a table template
+   (session, model, effort, sha_in, sha_out, D1–D8, pr_url, notes).
+4. Patch docs/plans/RECONCILIATION-2026-09-19.md baseline SHA to current
+   origin/main and note UI draft is now tracked.
+5. Open a PR. No npm publish. No deploy. No other features.
+```
+
+**Done when:** PR merges (or is merge-ready); UI requirements tracked; scorecard exists.
+
+**Grade focus:** D4 Finish, D5 Verify (file presence), D6 No rebuild.
+
+---
+
+## S1 — Interactive shell presence · ~1–2 h · Gemini 3.8 Flash High
+
+**Goal:** Implement UI requirements §§1,3,5 then §2, then scope §4 — presence,
+not architecture.
+
+**Depends on:** S0 (or attach the untracked UI file in the job context).  
+**Branch:** `feat/cli-interactive-presence`  
+**Inspect first:** `src/ui/cli/activity.js`, `footer.js`, `src/commands/shell.js`,
+existing diff helpers under `src/ui/` / `src/lib/`.
+
+**Prompt:**
+```text
+Repo: agentsam-sdk. origin/main. Read AGENTSAM.md and
+docs/plans/AGENTSAM-INTERACTIVE-UI-REQUIREMENTS.md completely.
+
+Implement in order: §1 interrupt hint, §3 footer project+action, §5 paste-collapse,
+then §2 shimmer (with truecolor/256 fallback), then §4 only after searching for an
+existing diff formatter (reuse; do not reinvent).
+
+Rules: extend existing helpers; no storage changes; no DO; no publish/deploy;
+run git worktree list first; open a PR with screenshots or CLI transcript evidence.
+
+Verify: node --check on edited files; focused tests if present; manual
+`node src/cli.js` interactive smoke notes in the PR.
+```
+
+**Done when:** All five sections addressed (§4 may be “reuse X + wire call sites”
+if a renderer already exists); PR + evidence.
+
+**Grade focus:** D1–D7 (primary Cloud Agent value test for TUI work).
+
+---
+
+## S2 — Status chrome + provider-agnostic `/model` · ~2–3 h · Flash High / Muse Medium
+
+**Goal:** Product surface that combines Codex run chrome + Cursor model menu:
+
+```text
+project · git · contract AGENTSAM.md ✓ · runtime local SQLite ·
+provider · model · effort · context · usage
+```
+
+`/model` opens scrollable provider-agnostic catalog (reuse existing models
+command / account-visible catalog — do not invent a second registry).
+
+**Depends on:** S1 footer foundation.  
+**Branch:** `feat/cli-status-model-catalog`  
+**Inspect first:** `src/commands/models.js`, shell slash commands, footer from S1,
+`feat/studio-auth-and-ai-wiring` (1 commit) before touching studio auth UI.
+
+**Prompt:**
+```text
+Build the persistent AgentSam session status strip and /model catalog UX
+described in docs/plans/MULTI-AGENT-SPRINTS-2026-09-19.md S2.
+Reuse account-visible models and existing provider credential surfaces.
+Do not rebuild OAuth. Do not change storage authority.
+Before any local-studio auth UI: git log / diff
+origin/feat/studio-auth-and-ai-wiring...main and report whether to cherry-pick.
+PR with before/after footer transcript.
+```
+
+**Done when:** Footer/status shows contract + runtime + model/effort/context/usage;
+`/model` lists real catalog entries; tests or transcript evidence; PR.
+
+**Grade focus:** D1, D2, D6 (highest risk of reinventing models/auth).
+
+---
+
+## S3 — Studio auth branch triage · ~45–90 min · Flash High
+
+**Goal:** Decide keep/merge/abandon for `feat/studio-auth-and-ai-wiring`
+(1 commit ahead / 7 behind) — no blind rewrite.
+
+**Branch:** `integrate/studio-auth-ai-wiring` or PR from rebased tip.
+
+**Prompt:**
+```text
+Inspect origin/feat/studio-auth-and-ai-wiring vs origin/main.
+Summarize the single unique commit. Rebase or cherry-pick onto a fresh branch
+from main if still valuable. Fix conflicts. Run relevant local-studio / auth tests.
+Open PR or document “abandon because already on main” with evidence.
+Do not expand into full identity redesign.
+```
+
+**Done when:** Merged PR **or** written abandon receipt with SHA evidence.
+
+---
+
+## S4 — Connections / binding evidence triage · ~1–2 h · Flash High / Muse Medium
+
+**Goal:** Land or extract useful bits from
+`origin/fix/connections-binding-evidence` (11 ahead) without breaking
+terminal bindings; feed evidence into later terminal work.
+
+**Prompt:**
+```text
+Diff origin/fix/connections-binding-evidence...origin/main.
+Classify commits: merge as-is / cherry-pick subset / abandon.
+If mergeable, rebase onto main, fix tests, open PR.
+Explicitly state how this relates (or does not) to terminal_remote
+platform_identity_missing — do NOT claim it fixes enrollment.
+```
+
+**Done when:** PR or abandon receipt; note for S7 updated.
+
+---
+
+## S5 — CMS pattern docs + scaffold design · ~1–2 h · Flash High
+
+**Goal:** Merge `docs/cms-scaffolding-pattern-reference` (4 commits) and write
+a short design for scaffold-from-real-app (`apps/client-cms-editor`,
+`apps/cad-creator`) — design only, no generator rewrite yet.
+
+**Prompt:**
+```text
+1. Rebase/merge origin/docs/cms-scaffolding-pattern-reference onto main → PR.
+2. Add docs/plans/SCAFFOLD-FROM-REAL-APP-DESIGN.md answering:
+   template vs fixed platform files; D1/R2/Worker per project vs shared;
+   subtree vs copy-parameterize; how IMPORT_PROVENANCE.json is used.
+3. Do not implement scaffold generators in this session.
+```
+
+**Done when:** Docs PR + design merged or merge-ready.
+
+---
+
+## S6 — Scaffold-from-real-app implement · ~2–4 h · Muse Medium/High or Flash High
+
+**Depends on:** S5 design approval (Sam ack in PR or chat).  
+**Branch:** `feat/scaffold-from-real-apps`
+
+**Prompt:**
+```text
+Implement agentsam scaffold cms (and cad-creator if design says so) per
+docs/plans/SCAFFOLD-FROM-REAL-APP-DESIGN.md.
+Stand up parameterized instances of apps/client-cms-editor / apps/cad-creator —
+do not regenerate the thin toy templates as the primary path.
+Keep thin templates only if design marks them deprecated/compat.
+Tests: scaffold dry-run or integration assertions; node --check; PR.
+```
+
+**Done when:** Scaffold produces real-app-shaped tree; tests green; PR.
+
+**Grade focus:** D1, D6 (reinvention risk extreme).
+
+---
+
+## S7 — terminal_remote enrollment (platform track) · separate job · Sol or Flash High
+
+**Repo note:** Likely spans `agentsam-sdk` **and** IAM / terminal host config.
+Treat as Cloud Agent job against the repo(s) Sam designates; SDK-only agents
+must stop at diagnosis + SDK-side receipts if host access is missing.
+
+**Prompt:**
+```text
+terminal_remote fails with platform_identity_missing under platform_bridge.
+Prior live evidence:
+  connection_key_set: false
+  platform_bridge_key_set: true
+  machine_key_set: true
+  account_id / connection_id / instance_id: null
+
+This is NOT the CLI OAuth bug. Do not reopen PR #450 / auth-closure as the fix.
+
+Mission: reconfirm fingerprint on remote exec host; implement enrollment/
+pairing/provisioning so connection identity is non-null; prove
+agentsam_terminal_remote (or equivalent) succeeds with evidence.
+If this session lacks host credentials, produce an exact ops runbook +
+SDK instrumentation PR only — do not invent OAuth token reshaping.
+```
+
+**Done when:** Live exec succeeds **or** blocked with precise missing secret/host
+access listed for Sam.
+
+**Grade focus:** D2 (correct diagnosis), D6 (no OAuth rebuild), D5 (live proof).
+
+---
+
+## S8 — Model A/B (optional) · same brief as S1 frozen
+
+Run identical S1 prompt on:
+
+1. Gemini 3.8 Flash High  
+2. Muse Spark 1.3 1M Medium (escalate High only if stalled)
+
+Fill scorecard D1–D8 + cost. Prefer the winner for S6/S7.
+
+---
+
+## S9 — Explicit non-goals / later backlog
+
+Do **not** assign as Cloud Agent “finish remaining sprints” work:
+
+- Storage / DO architecture redesign  
+- Errors-v1 divergent branches (56/40 ahead, 200+ behind) without a dedicated
+  errors mission  
+- Full command-manifest unification / lineage object / vocabulary CI  
+- `npm publish` / production Worker deploy  
+- Deleting worktrees
+
+---
+
+## Suggested Cloud Agent schedule (calendar)
+
+| Day | Sessions | Parallelism |
+| --- | --- | --- |
+| Day 1 | S0 → S1 | S3 and S4 can start after S0 in parallel |
+| Day 2 | S2 · finish S3/S4 | S5 in parallel |
+| Day 3 | S6 after Sam approves S5 | S7 platform track anytime |
+| Day 4 | S8 A/B if desired · merge lag · smoke | — |
+
+**Human gates (Sam only):**
+
+1. Approve S5 design before S6.  
+2. Approve any publish/deploy.  
+3. Provide remote host / IAM access for S7 if agents are blocked.  
+4. Pick A/B winner after S8.
+
+---
+
+## Session handoff packet (attach to every Cloud Agent job)
+
+```text
+AGENTSAM.md
+docs/plans/RECONCILIATION-2026-09-19.md
+docs/plans/MULTI-AGENT-SPRINTS-2026-09-19.md  (this file — your session id only)
+docs/plans/AGENTSAM-INTERACTIVE-UI-REQUIREMENTS.md  (S0–S2)
+docs/plans/CLOUD-AGENT-SCORECARD-2026-09.md  (append your row)
+```
+
+**Return format required from every agent:**
+
+```text
+## Session <id>
+Model / effort:
+SHA in / SHA out:
+Branch / PR:
+Done checklist: [ ] ...
+Grade self-score D1–D7:
+Evidence (commands + outcomes):
+Stopped because (if incomplete):
+Next session should start at:
+```
+
+---
+
+## Definition of “remaining sprints done”
+
+All of the following are true:
+
+- [ ] S0–S2 merged (interactive presence + status/`/model` chrome)
+- [ ] S3–S4 decided (merged or abandon receipts)
+- [ ] S5–S6 merged (CMS docs + real-app scaffold path)
+- [ ] S7 live green **or** blocked only on Sam-held host secrets (runbook filed)
+- [ ] Scorecard has ≥3 scored Cloud Agent jobs (enough to grade API value)
+- [ ] Still no surprise npm publish / Worker deploy / storage redesign

package/docs/plans/NEXT-AGENT-HANDOFF.md

@@ -0,0 +1,167 @@
+# Next-agent handoff — stack on inventory PR
+
+**Read first:** `AGENTSAM.md` · this file · open PR branch below  
+**Do not** start from stale Astra notes. Do not redesign storage. Do not invent DOs.
+
+## Where we are (2026-09-19)
+
+```text
+Branch (pushed): feat/per-user-model-inventory-dist @ dc24e62
+Open PR:         https://github.com/SamPrimeaux/agentsam-sdk/pull/new/feat/per-user-model-inventory-dist
+                 (gh auth was 401 last push — create/merge PR if not open yet)
+main tip before this work: 5affb1f (recon + AGENTS shims + sprint plan)
+verify:release:  PASS on this branch
+```
+
+Commits on the branch (do not redo):
+
+1. Cursor in `API_PROVIDERS` + live discovery  
+2. Shared `src/models/inventory-core.js`  
+3. Studio inventory picker + fail-closed `{provider, model_id}` chat  
+4. `/usage` panel + activity chrome  
+5. `agentsam app` + CMS/Studio manifests/bins  
+6. `scripts/install.sh` npm-bootstrap installer  
+
+Untracked (leave unless Sam says incorporate):  
+`docs/plans/AGENTSAM-INTERACTIVE-UI-REQUIREMENTS.md`
+
+---
+
+## Critical correction — “platform” is not a stub
+
+Local Studio **already ships with real production bindings and secrets**.
+`credential_plane: platform` means **desk/org Worker secrets + bindings**, not
+“we haven’t set anything up.”
+
+### Documented production surface (`agentsam-sdk` Worker)
+
+**Bindings (honor these):**
+
+| Binding | Type | Value |
+| --- | --- | --- |
+| `AGENTSAM_WAI` | Workers AI | catalog |
+| `DB` | D1 | `inneranimalmedia-business` |
+| `EXECOS` | Service | `execos` |
+| `HYPERDRIVE` | Hyperdrive | supabase hyperdrive |
+| `PTY_SERVICE` | VPC Service | `iam-vpc` |
+| `WEBSITE_ASSETS` | R2 | `agentsam-os-blueprint-content` |
+| `ASSETS` | Assets | `.output/public` |
+
+**Runtime secrets/vars already present (do not print values):**
+
+- `OPENAI_API_KEY`, `GEMINI_API_KEY`, `CURSOR_API_KEY`
+- `CLOUDFLARE_API_TOKEN`, `CLOUDFLARE_OAUTH_*`, break-glass admin token
+- `AGENTSAM_BRIDGE_KEY`
+- `IAM_CLIENT_ID` / `IAM_CLIENT_SECRET` / `IAM_OAUTH_ISSUER`
+
+So Studio inventory/chat **can and should** discover models from these
+platform credentials today. That is intentional desk authority.
+
+### What is still missing (your job)
+
+**Layer vault BYOK on top of platform — prefer vault when the logged-in user has it.**
+
+```text
+authenticated session
+        │
+        ▼
+  user_id / account_id
+        │
+        ├─► D1 user_secrets unwrap (DB + VAULT_MASTER_KEY)  → source: user_vault
+        │         preferred per provider when present
+        │
+        └─► else platform Worker secret / AGENTSAM_WAI      → source: platform
+```
+
+Hard rules (unchanged):
+
+- Never read `~/.agentsam` from Studio.  
+- Never return secrets to the browser.  
+- Never silently swap providers (no XAI→OpenAI).  
+- Selection stays `{ provider, model_id }`.  
+- Sam’s machine `env.d` ≠ Connor’s Studio options.
+
+Worker already has vault routes + `GET /api/llm/inventory` provenance.  
+TanStack `/api/llm/inventory` + `/api/chat` currently use **platform** via
+`apps/local-studio/shared/agentsam/src/studio-inventory.ts`.
+
+**Next implementation slice:** bind IAM/browser session → `user_id`, unwrap
+that user’s `user_secrets` server-side, merge vault-over-platform into
+`buildStudioInventory` / chat credential resolve. Reuse
+`apps/local-studio/backend/worker/index.js` vault decrypt path — do not invent
+a second vault.
+
+Inspect before rebuild:
+
+- `feat/studio-auth-and-ai-wiring` (worktree) — **auth portal / session gate only**;
+  **reject** any XAI→OpenAI chat fallback hunk.
+
+---
+
+## Exact next batches (pick one, finish it, PR)
+
+### A — Session→vault inventory (highest leverage) · Flash High
+
+1. Checkout `feat/per-user-model-inventory-dist` (or merge it to main first if Sam approves).  
+2. Wire Studio auth session so inventory/chat get a real `user_id` (not `studio-local`).  
+3. Server-side: vault unwrap for that user → merge over platform credentials.  
+4. Prove: user A inventory ≠ user B; secrets never in JSON; mismatch still 409.  
+5. Route Cloudflare curated models through `env.AGENTSAM_WAI` when provider is
+   `cloudflare` / Workers AI lane (binding already exists).
+
+### B — Auth portal live · Flash High
+
+1. Inspect `feat/studio-auth-and-ai-wiring` vs main.  
+2. Cherry-pick **portal + Worker identity routes only** (no AI fallback).  
+3. Prove `/auth/login` `/auth/signup` `/auth/reset` on Studio host.  
+4. That unblocks A’s session→user_id.
+
+### C — Host `/install` · Terra Fast / Flash Medium
+
+SDK already has `scripts/install.sh`. IAM/host must serve:
+
+`GET https://agentsam.inneranimalmedia.com/install` → that script body.
+
+App selectors: `/install/cad` `/install/cms` `/install/studio` (same engine, different `--app`).
+
+### D — Interactive UI polish · Flash High
+
+Use untracked `docs/plans/AGENTSAM-INTERACTIVE-UI-REQUIREMENTS.md` against
+activity/footer/shell (partially done by `/usage` commit). Finish interrupt
+hint / paste-collapse / shimmer if still open — extend, don’t rebuild.
+
+---
+
+## Do not touch / already done
+
+| Item | Status |
+| --- | --- |
+| Project SQLite storage / no default DOs | Done (PR #51) |
+| `agentsam help all` | Works on main |
+| Cursor machine discovery | Done on this branch |
+| Static Studio Grok-only menu | Replaced on this branch |
+| `agentsam app` manifests | Done on this branch |
+| npm publish / surprise deploy | Forbidden without Sam |
+
+---
+
+## Operating rules for this handoff
+
+1. `git fetch --all --prune` · `git worktree list` before create/delete.  
+2. One focused PR; keep stacking on `feat/per-user-model-inventory-dist` unless merged.  
+3. Run `npm run verify:release` before claiming done.  
+4. End with: SHA, tests, remaining blockers with evidence, next slice letter (A/B/C/D).  
+5. If blocked on IAM session cookies / vault master key / deploy, stop with exact missing access — don’t invent OAuth reshapes.
+
+## Return format
+
+```text
+## Session
+Slice: A|B|C|D
+SHA in / out:
+PR:
+Done:
+Evidence:
+Blocked (if any):
+Next slice:
+```