feat(agent-runtimes): add github copilot sdk runtime

.env.example

@@ -0,0 +1,32 @@
+# Template for podman-compose.yaml. Copy to `.env` and replace every secret.
+# Generate strong values: `openssl rand -hex 32`
+# cp .env.example .env
+
+# --- Core infra secrets (override the # CHANGEME defaults in the compose) ---
+POSTGRES_PASSWORD=postgres
+CLICKHOUSE_PASSWORD=clickhouse
+MINIO_ROOT_PASSWORD=miniosecret
+REDIS_AUTH=myredissecret
+SALT=mysalt
+ENCRYPTION_KEY=0000000000000000000000000000000000000000000000000000000000000000
+NEXTAUTH_SECRET=mysecret
+NEXTAUTH_URL=http://localhost:3000
+
+# --- Headless initialization ---
+# On first boot langfuse-web auto-creates this org/project/user and provisions
+# the API keys below. No manual UI signup needed.
+# Docs: https://langfuse.com/self-hosting/administration/headless-initialization
+LANGFUSE_INIT_ORG_ID=case
+LANGFUSE_INIT_ORG_NAME=Case
+LANGFUSE_INIT_PROJECT_ID=case
+LANGFUSE_INIT_PROJECT_NAME=Case
+LANGFUSE_INIT_PROJECT_PUBLIC_KEY=pk-lf-00000000-0000-0000-0000-000000000000
+LANGFUSE_INIT_PROJECT_SECRET_KEY=sk-lf-00000000-0000-0000-0000-000000000000
+LANGFUSE_INIT_USER_EMAIL=admin@case.local
+LANGFUSE_INIT_USER_NAME=Case Admin
+LANGFUSE_INIT_USER_PASSWORD=changeme123
+
+# --- Client config (the `ca` app reads these; match the INIT keys above) ---
+LANGFUSE_HOST=http://localhost:3000
+LANGFUSE_PUBLIC_KEY=pk-lf-00000000-0000-0000-0000-000000000000
+LANGFUSE_SECRET_KEY=sk-lf-00000000-0000-0000-0000-000000000000

.gitignore

@@ -24,3 +24,14 @@ dist/
 
 # bun artifacts
 *.bun-build
+.todos/
+
+# Local secrets for podman-compose (use .env.example as the template)
+.env
+.sidecar/
+.sidecar-agent
+.sidecar-task
+.sidecar-pr
+.sidecar-start.sh
+.sidecar-base
+.td-root

AGENTS.md

@@ -8,7 +8,7 @@ Humans steer. Agents execute. When agents struggle, fix the harness.
 Run the session command to gather context before doing anything else:
 
 ```bash
-SESSION=$(ca session <target-repo-path> --task <task.json>)
+SESSION=$(ca session <target-repo-path> --task <td-id>)
 echo "$SESSION"
 ```
 
@@ -40,7 +40,7 @@ Full metadata (commands, remotes, evidence strategy): `~/.config/case/projects.j
 
 ## Task Dispatch
 
-Tasks are markdown files that agents execute. Runtime task files live in the target repo's ignored `.case/tasks/active/`.
+Tasks are `td` issues that agents execute. Each task is a `td` issue in the target repo's `.todos/` store, identified by a `td-…` issue handle.
 
 - **Format spec**: `tasks/README.md`
 - **Templates**: `tasks/templates/`
@@ -49,11 +49,11 @@ Pipeline: scout → implementer → verifier → reviewer → closer → (retros
 
 Onboarding agent (out of the pipeline): `interviewer` — invoked by `ca onboard --interview` to capture evidence strategy rationale, verification notes, and repo learnings.
 
-Lifecycle: `.case/tasks/active/` → PR opened/merged status in the task JSON
+Lifecycle: `td` issue created → PR opened/merged status tracked on the task record
 
 ## Working in a Target Repo
 
-0. Run `ca session {repo-path} --task {task-json}` to gather context
+0. Run `ca session {repo-path} --task {td-id}` to gather context
 1. Read the repo's `CLAUDE.md` (or `CLAUDE.local.md`) for project-specific instructions
 2. Run `ca bootstrap {repo-name}` to verify readiness
 3. Follow the repo's PR checklist before opening a PR

CLAUDE.md

@@ -50,14 +50,14 @@ AGENTS.md # Entry point for agents (routing map)
 CLAUDE.md # This file (meta-instructions for case itself)
 projects.schema.json # JSON Schema for the project manifest
 docs/
- architecture/ # Canonical patterns per repo type
+ architecture/ # Canonical patterns per repo type (+ pipeline.md for case itself)
  conventions/ # Shared rules (commits, testing, PRs)
  failure-matrix.md # Phase × outcome → action lookup (synced with code)
  golden-principles.md # Invariants enforced across all repos
  playbooks/ # Step-by-step guides for recurring operations
 tasks/
-active/  # Current task files for agent execution
-templates/# Reusable task templates
+templates/ # Spec scaffolds (a filled template becomes a td issue)
+ # Live task state is a td issue in the target repo's .todos/, not a file here
 src/commands/
  check.ts # Cross-repo convention enforcement
  bootstrap.ts # Per-repo readiness verification

README.md

@@ -43,10 +43,10 @@ ca --agent 1234
 
 `ca --agent` starts an interactive orchestrator session. It can inspect context, fetch issues, help shape the task, create the task file, and then run the pipeline. It should not implement directly. This is the primary interface for “humans steer.”
 
-For an existing task file:
+For an existing td task, pass its issue handle:
 
 ```bash
-ca run --task .case/tasks/active/cli-1-issue-53.task.json
+ca run --task td-a1b2c3 --repo-path ../cli/main
 ```
 
 To resume an interrupted issue run, re-run the same command:
@@ -109,15 +109,15 @@ ca --agent 1234 # steering session with issue context
 ca onboard <path> # add a repo to projects.json
 ca onboard <path> --interview # add a repo with an interactive interview
 ca onboard <repo> --re-interview # re-interview an already-onboarded repo
-ca run --task <file> # run an existing task JSON
+ca run --task <td-id> # run an existing td task by its issue handle
 ca watch <task-slug> # live-tail the event log
 ```
 
 Agent-facing commands:
 
 ```bash
-ca session <repo-path> --task <task.json>
-ca status <task.json> [field value...]
+ca session <repo-path> --task <td-id>
+ca status <td-id> [field value...]
 ca mark-tested
 ca mark-manual-tested
 ca mark-reviewed --critical 0
@@ -134,8 +134,8 @@ Common flags:
 
 ```bash
 ca --model claude-opus-4-5 1234
-ca run --task <file> --mode unattended
-ca run --task <file> --dry-run
+ca run --task <td-id> --mode unattended
+ca run --task <td-id> --dry-run
 ca run --fresh 1234
 ```
 
@@ -150,18 +150,15 @@ Package-level config lives under `~/.config/case/`. Per-repo runtime state lives
  agent-versions/
 
 <target-repo>/.case/
- active
  learnings.md
  amendments/
  run-log.jsonl
- tasks/
- active/
- <task-slug>.md
- <task-slug>.task.json
  <task-slug>/
  events/
  plan.json
  working-memory.json
+
+<target-repo>/.todos/ # td issue store (SQLite); a task is a td issue, not a .task.json file
 ```
 
 Override the config/cache directory with:
@@ -178,7 +175,7 @@ For portable binary installs, keep `projects.json` in `~/.config/case/` via `ca
 
 ## Pipeline
 
-The runtime uses a deterministic TypeScript pipeline executor for phase transitions. The LLMs do the work inside each phase; TypeScript decides which phase runs next.
+The runtime drives phase transitions with a deterministic [LangGraph](docs/architecture/pipeline.md) `StateGraph`: the LLMs do the work inside each phase, while TypeScript — the graph routers plus the failure matrix — decides which phase runs next. Runs carrying a SQLite checkpointer resume from the last phase after a crash. Each agent spawn is provider-routed — Claude models run on the Claude Agent SDK, everything else on LangChain (override with `CASE_AGENT_RUNTIME`). See [docs/architecture/pipeline.md](docs/architecture/pipeline.md) for the full picture.
 
 Profiles:
 
@@ -241,7 +238,7 @@ Configure models in `~/.config/case/config.json`:
 {
  "$schema": "https://raw.githubusercontent.com/workos/case/main/config.schema.json",
  "models": {
- "default": { "provider": "anthropic", "model": "claude-sonnet-4-20250514" },
+ "default": { "provider": "anthropic", "model": "claude-sonnet-4-6" },
  "reviewer": { "provider": "google", "model": "gemini-2.5-pro" },
  "verifier": null
  }
@@ -251,7 +248,7 @@ Configure models in `~/.config/case/config.json`:
 Priority:
 
 ```text
---model flag > explicit spawn options > config file > hardcoded default
+--model flag > CASE_MODEL_OVERRIDE env > per-agent config > config default > hardcoded default
 ```
 
 ## Repository Map
@@ -280,7 +277,7 @@ For case itself:
 
 ```bash
 bun run typecheck
-bun test ./src/__tests__/
+bun run test
 bun run lint
 bun run format:check
 ```

agents/closer.md

@@ -1,19 +1,18 @@
 ---
 name: closer
-description: PR creation agent for /case. Drafts thorough PR descriptions from task file and verification evidence. Verifies all evidence gates before PR creation. Never implements or tests.
+description: PR creation agent for /case. Drafts thorough PR descriptions from the task and verification evidence. Verifies all evidence gates before PR creation. Never implements or tests.
 tools: ['Read', 'Bash', 'Glob', 'Grep']
 ---
 
 # Closer — PR Creation Agent
 
-Create a pull request with a thorough description based on the task file, progress log, and verification evidence. You are the only agent that runs `gh pr create`. You must verify all evidence gates yourself before attempting to create the PR.
+Create a pull request with a thorough description based on the task, progress log, and verification evidence. You are the only agent that runs `gh pr create`. You must verify all evidence gates yourself before attempting to create the PR.
 
 ## Input
 
 You receive from the orchestrator:
 
-- **Task file path** — absolute path to the `.md` task file under the target repo's ignored `.case/tasks/active/`
-- **Task JSON path** — the `.task.json` companion
+- **td issue handle** — the `td-…` id for this task (shown as **td issue** in the Task Context block); pass it to `ca status`/`ca session`
 - **Target repo path** — absolute path to the repo
 - **Verifier AGENT_RESULT** — structured output from the verifier (screenshot URLs, evidence markers, pass/fail)
 
@@ -24,26 +23,26 @@ You receive from the orchestrator:
 Run the session command to orient yourself:
 
 ```bash
-SESSION=$(ca session <target-repo-path> --task <task.json>)
+SESSION=$(ca session <target-repo-path> --task <td-id>)
 echo "$SESSION"
 ```
 
-Read the output to understand: current branch, last commits, task status, which agents have run, and what evidence exists. This replaces manual git log / task file discovery.
+Read the output to understand: current branch, last commits, task status, which agents have run, and what evidence exists. This replaces manual git log / task discovery.
 
 ### 0.5. Record Start
 
 Mark yourself as running with a start timestamp immediately:
 
 ```bash
-ca status <task.json> agent closer status running
-ca status <task.json> agent closer started now
+ca status <td-id> agent closer status running
+ca status <td-id> agent closer started now
 ```
 
 ### 1. Gather Context
 
-1. Read the task file (`.md`) — full content including progress log entries from all agents
-2. Read the task JSON for issue reference, repo, branch
-3. Read verification evidence markers (get task slug from `.case/active`, markers are under `.case/<task-slug>/`):
+1. Read the task (`td show <td-id>`) — full content including progress log entries from all agents
+2. Read the task record for issue reference, repo, branch
+3. Read verification evidence markers (the task slug is the taskId — the **Task** id in the Task Context block, or `SLUG=$(ca status <td-id> id)`; markers are under `.case/<task-slug>/`):
  - `.case/<task-slug>/tested` — should have `output_hash` field
  - `.case/<task-slug>/manual-tested` — should have `evidence` field (if src/ files changed)
  - `.case/<task-slug>/reviewed` — should have `critical: 0` (review findings summary)
@@ -108,12 +107,12 @@ Closes #<number>
 
 Before running `gh pr create`, verify every requirement.
 
-**CRITICAL: Check the task JSON first.** Read the task JSON and confirm the reviewer agent phase shows `"status": "completed"`. If the reviewer never ran, STOP — do not attempt to create the PR. Report the missing reviewer phase in your error output so the orchestrator can dispatch the reviewer.
+**CRITICAL: Check the task record first.** Read the task and confirm the reviewer agent phase shows `"status": "completed"`. If the reviewer never ran, STOP — do not attempt to create the PR. Report the missing reviewer phase in your error output so the orchestrator can dispatch the reviewer.
 
-1. **Reviewer ran**: Read the task JSON and confirm `agents.reviewer.status` is `"completed"`
+1. **Reviewer ran**: Read the task and confirm `agents.reviewer.status` is `"completed"`
 
  ```bash
- test "$(ca status <task.json> agent reviewer status)" = "completed"
+ test "$(ca status <td-id> agent reviewer status)" = "completed"
  ```
 
 2. **Branch**: Verify not on main/master
@@ -128,7 +127,7 @@ Before running `gh pr create`, verify every requirement.
 3. **Test evidence**: Read `.case/<task-slug>/tested` — must exist with `output_hash` field
 
  ```bash
- SLUG=$(cat .case/active | tr -d '[:space:]')
+ SLUG=$(ca status <td-id> id)
  test -f ".case/${SLUG}/tested" && grep -q "output_hash:" ".case/${SLUG}/tested"
  ```
 
@@ -168,7 +167,7 @@ The body must contain verification keywords (any of: "verif", "tested", "test pl
 If the reviewer produced warnings or info findings (check `.case/<task-slug>/reviewed` for `warnings` and `info` counts), post them as a PR review comment:
 
 ```bash
-# Read findings from the reviewer's progress log entry in the task file
+# Read findings from the reviewer's progress log entry in the task
 # Format as a comment
 gh api repos/{owner}/{repo}/pulls/{pr_number}/reviews \
  --method POST \
@@ -188,18 +187,18 @@ Only post if there are actual findings to share. Skip this step if the reviewer
 
 ### 5. Record
 
-1. **Update task JSON** — set agent phase completed, then transition status and record PR URL:
+1. **Update the task** — set agent phase completed, then transition status and record PR URL:
 
  ```bash
- ca status <task.json> agent closer status completed
- ca status <task.json> agent closer completed now
- ca status <task.json> status pr-opened
- ca status <task.json> prUrl "<PR URL>"
+ ca status <td-id> agent closer status completed
+ ca status <td-id> agent closer completed now
+ ca status <td-id> status pr-opened
+ ca status <td-id> prUrl "<PR URL>"
  ```
 
  Extract the PR URL from the `gh pr create` output. A null `prUrl` makes the task record incomplete — this is not optional.
 
-2. **Append to the task file's Progress Log**:
+2. **Append to the task's Progress Log**:
 
  ```markdown
  ### Closer — <ISO timestamp>