injected-context¶

Provenance¶

Source: .spec/spexcode/spec-cli/sessions/injected-context/spec.md
Source SHA-256: ce97eedf12ad97e4b228cc598ac5f42be0d8b103b2489a2762329b7397b172ba

injected-context¶

raw source¶

A dispatched session should begin already knowing which spec is its ground truth, and should be caught the moment it forgets — but the launcher must never inline a spec body: a pasted snapshot bloats the launch prompt toward truncation and freezes a copy that goes stale the instant the agent edits the file. The harness injects only pointers and reminders, so the agent always reads the live contract itself.

expanded spec¶

Four thin injections, all deliberately non-enforcing (the Stop gate is the enforcer):

[[spec-pointer]] — when a dispatch names an existing node, append one line: the absolute path to that node's live spec.md inside the new worktree. Never the body. Fail-quiet by absence — an unknown id or a node-agnostic prompt (no [[id]]) appends nothing.
[[spawner-pointer]] — when the launch has a parent, append one line naming that session's worktree and branch, and why the child needs it: the child's tree is branched off the base, so the spawner's in-flight work — including a spec node the pointer above therefore cannot resolve — is not in it. A pointer at a live tree, read-only, and fail-quiet when there is no spawner.
[[inject-spec-first]] — a one-shot PreToolUse read gate that advances only on code with a real governor, telling the agent to read that governing spec and its relevant neighbors before retrying. Irrelevant and ungoverned reads leave it armed, so uncovered exploration cannot mute a later governed boundary.
[[inject-spec-of-file]] — a per-edit PostToolUse annotation that, once per file, names the spec governing the file just edited (and flags a shared-hub file with many owners). Non-blocking — the contract kept in view at the edit, not just at the start.

Together they make spec-awareness the session's starting AND running condition without ever duplicating spec text into a prompt: point at the live file, ground before the first governed read, name actionable ownership problems at each edit, enforce elsewhere.

All three are passive and assume the agent already knows which node is its ground truth. A fourth, active injection is a retired experiment — an injected spec-consult sub-agent (the spec analog of code search) for behaviour questions not bound to one node — so a session can find its governing spec, not only be pointed at one it already knows. Pending; the three above are what ships today.