[kshitijk4poor]

Internal gateway events (background-process notifications, watch/tick events) no longer re-render the session-context and channel prompt pins. The ephemeral prompt stays byte-stable across human → internal → human turns.

Salvage of #122048 by @Kyzcreig (opened 2026-09-25). Their commit is kept as authored, with teknium1's inline review suggestion from that PR applied on top.

What was wrong

The gateway pins the rendered session-context prompt and the channel prompt per session, so the ephemeral part of the prompt stays byte-stable. Internal events carry a synthetic source and no channel prompt. When one arrived, it re-keyed both pins: the context prompt shrank (1001 → 971 bytes in the probe), and the next human turn re-rendered it again. Each flip broke the cached prefix twice.

Change

• gateway/run_turn.py, gateway/run_agent_cache.py: an internal turn reuses the pinned session-context render and the pinned channel inputs instead of re-keying them.
• If a session has never pinned (its first turn is internal), it still pins once (teknium1's suggestion).
• The pinned source is restored with the existing session_identity.replace_source, so bot identity survives when several bots share a gateway.
• _pinned_channel_inputs takes explicit arguments; the three call sites pass the event's or pending event's channel_prompt/internal.
• Tests: two invariants driven through the real handler and the real prompt combiner.
  • human → internal → human keeps the prompt byte-identical;
  • an internal-first session pins once.

Verification

• End-to-end probe:
  • head: 1001 / 1001 / 1001 bytes;
  • base: 1001 / 971 / 1001;
  • controls identical on both: the pin still re-keys on a real rename, and internal-only sessions are handled.
• Mutation: reverting the headline commit turns 2 tests red; reverting the no-pin suggestion turns 1 red. The replace_source fix is covered by the probe only, because the stack is capped at two tests.
• 85 tests across 12 related gateway files pass.
• Review: hermes-pr-review 2a/2b/2c plus the reuse, quality and efficiency passes, over three rounds.

Known and unchanged (same as base, for a follow-up):

• interrupted or /steer plain-text follow-ups still drop the channel prompt for one turn;
• the first internal event after a gateway restart pins a reduced render.

Supersedes #122048.

[teknium1]

PR Review — #124179 (head a80ae6dcf1 after my push)

Verdict: correct; the CI red on 66b38cb6a9 was a test-stub gap, fixed in place. Merge decision stays with the maintainer; #122048 (@Kyzcreig's original) and this PR must land once.

Premise (origin/main 1c535d9689): confirmed. _pinned_session_context_prompt keys the pin on the event's source, and internal events (kanban wakes, delegation completions, watch ticks) carry a source without chat_name/user_name and channel_prompt=None, so human → internal → human re-renders the ephemeral system prompt twice (A→B→A).

Delta vs #122048: (1) an internal-first session now renders AND pins (the original rendered without pinning) — the inline suggestion on #122048, applied; (2) dataclasses.replace → session_identity.replace_source, which keeps the wire-invisible provenance attrs a plain replace drops; (3) getattr(source, "parent_chat_id", None) → source.parent_chat_id. (3) is what went red: tests/gateway/test_queued_final_ledger.py::_source builds a SimpleNamespace without the field, and two chained-followup tests reach _pinned_channel_inputs through _run_agent_queued_followup. The real SessionSource always has the field, so I gave the stub parent_chat_id=None (a80ae6dcf1) rather than re-adding a guard production cannot hit. Every test that reaches _run_agent_queued_followup / the two pin helpers (39 across 5 files) passes on the new head.

Tests: 2 new invariants pass on head; swapping the three prod files to origin/main turns both red.

Live test: handler-level with the real prompt combiner (the PR's own tests); no gateway process run here.

Suggestions

• _evict_agent (gateway/run_agent_cache.py ~L735) resets ephemeral_pin and vc_last ("prompt-stability state rides the agent-cache lifecycle") but not the new channel_pin. Harmless today (channel prompt and parent_chat_id are per-chat, not per-agent), but resetting it there keeps the three pins on one lifecycle. One line; non-blocking.
• The two "known and unchanged" items in the body (interrupted//steer follow-ups dropping the channel prompt for one turn; the first internal event after a restart pinning a reduced render) are real and pre-existing; a follow-up, not this PR.

Looks Good

• Explicit internal: bool / channel_prompt parameters instead of getattr on the event; the three call sites pass the event's fields.
• parent_chat_id is restored only when the internal source lacks one, so a legitimately different destination is never overwritten.

[kshitijk4poor]

Thanks @teknium1 for the stub fix in a80ae6dcf1. That's the right fix; production SessionSource always has the field. The _evict_agent reset of channel_pin is a one-line lifecycle tidy-up with no behaviour change today, so I'll send it as a follow-up rather than re-gate this head. The two known pre-existing items stay follow-ups too.