fix(resume): re-register agent system prompt on sub-session resume — resumed sub-agents ran with no system prompt

[Brian Krabach (bkrabach)]

Summary

Resumed delegate sub-sessions ran with no system prompt at all, silently.

Root cause

1. spawn_sub_session() registers the agent's system instruction as an in-memory factory (session_spawner.py:766-802, context.set_system_prompt_factory) rather than a persisted message. context-simple's SimpleContextManager builds the system message into a per-request copy inside get_messages_for_request() and never writes it into self.messages — so it is never present in transcript.jsonl. SessionStore._save_transcript also explicitly skips system/developer role messages when persisting, so no persisted sub-session transcript ever contains a system message, in either mode.
2. resume_sub_session() (session_spawner.py:923-1406) constructs AmplifierSession directly and restores the transcript (was :1296-1300) — but never re-derived or re-registered the system instruction anywhere in that 483-line function. Every subsequent request on a resumed sub-session ran with system_msgs == [].
3. A live API probe confirmed omitting the system prompt on a chained request clears the provider's server-held prompt rather than preserving it (provider-openai omits instructions; provider-anthropic omits system, when system_msgs is empty).

Evidence: 8 of 92 captured sessions were delegate-resumed sub-sessions. All 8 lost their system instructions starting from the first post-resume request and never recovered — 111 of 2,460 requests (4.5%) ran with no system prompt. Perfect correlation, zero counter-examples. The instructions/system key was absent from the wire request, not present as an empty string.

Fix

In resume_sub_session, inserted before the transcript-restore block:

• Recover the agent's system instruction from persisted metadata: metadata["agent_overlay"] (the exact agent_config dict spawn_sub_session saved at session_spawner.py:878), falling back to metadata["config"]["agents"][<agent_name>] for sessions saved before agent_overlay existed or with an empty inherit-as-is overlay.
• Re-expand @-mentions using the just-restored mention_resolver / mention_deduplicator / session.working_dir capabilities, via the same amplifier_foundation.mentions.expand_mentions_in_instruction() helper the spawn path and the resumed-instruction path already use — mirrors spawn's behavior exactly instead of injecting the raw, unexpanded body.
• Register via context.set_system_prompt_factory() when supported (same closure shape as spawn), with an add_message() fallback for context modules without factory support.
• If no instruction is recoverable at all, log a loud warning naming the session and agent — previously this failure mode was completely silent. Resume still proceeds (never raises).
• The nested/grandchild resume capability (child_resume_capability, registered inside resume_sub_session) calls resume_sub_session() directly, so the fix applies transitively with no separate change needed.

Also corrected a stale/misleading comment on the root-session resume path (session_runner.py ~line 238). That "preserve fresh system prompt / re-inject if transcript lacks one" guard is dead code today: in factory mode context.get_messages() only ever returns self.messages, which never contains a system-role message (the factory's output is injected ephemerally by get_messages_for_request(), never persisted) — so system_msgs is always empty and the guard's re-injection branch never fires. The comment previously claimed this guard was the safety mechanism; it is not. Root-session resume is actually safe because create_session() (PreparedBundle, called earlier in the same flow) already re-registers a fresh set_system_prompt_factory() before this code runs, and context.set_messages() never touches that registered factory. Behavior unchanged — comment only.

Cache note: resume re-expands @-mentions from disk at resume time. If a mentioned file changed since the original spawn, the reassembled instructions will legitimately differ from the spawn-time version — a one-time cache miss, correct behavior, not a bug.

Tests

New file tests/test_resume_system_prompt.py (fixture style mirrors TestCapabilityRegistrationIntegration in test_session_spawner.py):

• test_resume_reregisters_system_prompt_via_factory — factory-capable context receives a working factory whose output contains the persisted instruction.
• test_resume_adds_system_message_when_no_factory_support — context without factory support gets a system-role add_message() call instead, ordered before the restored transcript history.
• test_resume_falls_back_to_merged_config_agents_map — covers metadata saved before agent_overlay existed.
• test_resume_with_no_recoverable_instruction_warns_but_succeeds — no instruction anywhere in metadata → loud warning fires, resume still succeeds.

Fail-before proof: all 4 new tests run against the pre-fix code and FAIL (4 failed in 0.05s). After the fix, all 4 PASS.

Full suite: uv run pytest -q → 1497 passed, 1 skipped, 13 deselected, 1 xfailed (baseline was ~1493 passed; +4 for the new tests, zero regressions).

Lint: uv run ruff check on all three changed/added files → All checks passed!

Disclosure

This PR is self-authored by an Amplifier agent acting on explicit maintainer direction (repo owner Brian Krabach (@bkrabach)), including the merge step (--admin only if a required review check is what's blocking, never over a failing check). Investigation, fix, and tests were produced and verified end-to-end before opening this PR; see the evidence above.

🤖 Generated with Amplifier

[Brian Krabach (bkrabach)]

Maintainer note: this PR is authored by an Amplifier agent acting on explicit direction from repo owner Brian Krabach (@bkrabach) (maintainer). All 9 required checks are green (license/cla + 6 pytest matrix jobs + 2 integration jobs). Merging with `--admin` solely to bypass the required-review gate (no reviewer available in this workflow) — never to bypass a failing check. Squash-merging via `gh pr merge --squash --delete-branch=false`.