start-session

New session. Before doing anything else, run your bootstrap cycle:

1. Bootstrap check: verify .opencode/PROJECT-PROFILE.md exists AND plan/ contains all four subfolders (draft, in-progress, qa, complete) with plan/README.md. If either is missing, delegate to profiler before any other routing. Skip only when both conditions are already satisfied. 0.5. Session memory load: read .context/progress.md, .context/decisions.md, .context/issues.md, and .opencode/PROJECT-PROFILE.md (if they exist) before planning any routing. If PROJECT-PROFILE.md reports Code Graph: present, note it for the Act step — it is informational only and must not delay or block delegations.
2. Observe: understand the request and read only what is needed for routing.
3. Orient: classify the request and estimate scope.
4. Decide: choose one agent, a sequence, or parallel subtasks.
5. Act: state the step in one line of plain text, then delegate via subagent — the dispatch tool's runtime name in OpenCode 2 (task is the V1 spelling and is not a tool in this runtime). todowrite MAY be used when the runtime exposes it, but never rely on it: if it is unavailable, the one-line plan in your reply IS the progress record. Do not start any new work until the user gives the next instruction.

After the cycle, summarize in 3-4 lines, in the user's language (the language the user is speaking): the detected stack (from PROJECT-PROFILE.md), the current status of the work (from progress.md), and the latest relevant decision/issue if present. Then wait for the next instruction.

The comprehension telemetry write (.context/comprehension-log.md) is NOT part of bootstrap. It is never loaded during bootstrap either: the loaded set stays exactly .context/progress.md, .context/decisions.md, .context/issues.md (the comprehension log is telemetry, not session memory). The manual /recall command may read the log (and the per-evaluation records under .context/comprehension/) only under explicit user invocation — never during bootstrap. Phases 1-4 above are strictly read-only with respect to telemetry: no append, no file creation, no header emission. Telemetry writes happen only AFTER a comprehension-coach verdict is final, in a later turn — never during bootstrap.

The research dataset (.context/research-dataset.jsonl) is also NOT part of bootstrap and is NOT in the loaded set: it is excluded by the same isolation rule as the comprehension log. The dataset is read and written only inside the research lifecycle when research mode is active — never during session start, never during bootstrap.

Runtime tool names (OpenCode 2)

Tool names changed between V1 and V2. Calling a name the runtime does not have costs a full turn and returns No tool named "<name>" is currently available; it never degrades silently. Names observed in this runtime:

• dispatch a subagent: subagent — never task.
• run shell commands: shell — never bash. A permission.bash / tools.bash entry may still gate shell, but the TOOL to call is shell.
• read / search / list: read, grep, glob.
• change a file: edit — there is NO write tool in 2.0.x (patch is the other accepted name). execute runs code without a shell.
• ask the user: question — primary session only. A subagent has no interlocutor: a question call from a dispatched agent is rejected or silently unanswered, so NEVER dispatch a step whose only path forward is asking the user.
• load a skill: skill; fetch a URL: webfetch (only for agents whose permissions allow it).

Routing consequence: a request carrying a URL, an issue link or any remote document MUST be routed to an agent that can fetch it (e.g. librarian) — never reformulated as a local search. When no agent has the tool a step needs, say so instead of inventing a name or a workaround.