Slice 4: posture-aware context injection — the Active mode block

[aarontrowbridge]

Important

Problem — The context splice is posture-blind: nothing tells a mid-campaign or post-compaction agent which director posture it is in. The plugin can resolve the session's agent but does not, so an autodev session that compacted re-enters as a generic helper — the exact "agent doesn't know it's in autodev mode" failure this campaign exists to fix.
Approach — Resolve the session's agent (primary: the session-info agent id, availability inside the transform hook verified by a fixture FIRST) and emit an ## Active mode block on director-posture sessions: posture, phase/gate summary read from the registry bundle, ledger path convention, resolved-agent + registry-digest stamp. Emitted per request — it survives compaction by construction.
Scope — in: agent resolution + the mode block + the posture-binding map + honest degradation · out: any fork change (if the session-API fixture shows the hook cannot resolve the agent, NO widening executes — the contingency is amendment-gated, see Key Decisions).

Acceptance Criteria

• A director-posture session emits the stamped mode block; a copilot session (every agent id outside the registry's mode agents — plan, build, role agents, custom) emits nothing
• A simulated compaction (fresh request, history dropped) still emits the block from session state alone (primary path)
• A staged bundle with an unresolvable agent emits the explicit posture: unresolvable — re-bind from the ledger line
• A resolved director posture with missing bundle parts emits a degraded line naming them
• A PLUGIN version gap emits the unresolvable block (the loud failure), never silence
• The fallback fires only for non-compaction resolution failures and DECLINES when the agent-switched stream shows a switch newer on a monotonic key from the same store (never wall-clock across stores, never prose matching) — the unresolvable line is emitted, never a wrong posture
• The session-API availability fixture runs before anything consumes it; its outcome is recorded whichever way it decides

Testing Decisions

New plugin fixtures (the injection cases); the session-API availability fixture lands FIRST and gates the rest.

Key Decisions

Posture binding is a MAP, not a heuristic: registry mode agents bind their mode; every other agent id binds the copilot posture and is silent; dispatched role-agent sessions re-bind the parent posture from the ledger (casts may not receive the transform at all). The compactor runs outside the splice and is NOT posture-aware — post-compaction summaries are audited against the ledger per the discovery rule, no new mechanism. THE CONTINGENCY: if the fixture shows the call is insufficient, NO widening executes — the widening is proposed to the harness campaign at its next pin and requires an explicit amendment of the spec's fork invariant before execution; until amended, the slice runs fallback-only with the unresolvable semantics preserved, and the mode-block criteria are evaluated against the amended-or-fallback contract, honestly recorded either way.

Constraints & Invariants

The plugin stays dependency-free (no new IPC without the fixture proving it); the block never guesses — named outcomes only.

Prior Art

Spec decision D4; the registry slice's bundle (the phase/gate summary source); the fork's session-info agent field and agent-switched event stream (existing, preserved by the harness contract).

Source

Spec spec-20260905-063000-modes-first-class · Campaign ledger session-20260905-modes-first-class · Blocked by the mode-registry slice (bundle). Published with Aaron's standing approval to execute (autodev, 2026-09-05).