fix(unit-only): Primarily a DX/discoverability fix (stateless behavior is... (#1591)

[aidandaly24]

Refs aws#1591

Issues

• bug: default agent has no session context aws/agentcore-cli#1591 — An agent created with the default memory option ('none') has no session continuity: every invoke behaves like a fresh conversation. The agent never gets a session_manager, so nothing is loaded from or written to a durable store, and the no-memory template uses a single process-global agent that is not keyed by session_id, so --session-id has no effect at all in that path. Meanwhile the CLI tells the user the opposite: --session-id is labeled "Use specific session ID for conversation continuity" and both the CLI stream output and the interactive TUI print "To resume this session, run: agentcore invoke --session-id ..." even for memory-less agents. New customers reasonably expect continuity and are confused when there is none.

Root cause

Reproduces exactly at HEAD (v0.20.2) and is by-design: session continuity is gated entirely on memory being configured. In src/assets/python/http/strands/base/main.py, the default (memory='none', no truncation, no config-bundle, no payment) renders the branch at lines 441-468: a single module-global _agent created once via get_or_create_agent() with conversation_manager=NullConversationManager() and NO session_manager. Two consequences: (1) with no session_manager, nothing is loaded/persisted to a durable store, so history is lost on container cold-start; (2) the global _agent is NOT keyed by session_id, so the runtime session_id (set as x-amz-bedrock-agentcore-session-id at src/cli/commands/invoke/action.ts:175) is functionally inert in this path — --session-id changes nothing. The session_id is only consumed by src/assets/python/http/strands/capabilities/memory/session.py get_memory_session_manager, which returns None when MEMORY_ID is unset; that manager (SDK AgentCoreMemorySessionManager extending Strands RepositorySessionManager, bedrock-agentcore-sdk-python v1.9.1, src/bedrock_agentcore/memory/integrations/strands/session_manager.py) is the sole mechanism that restores cross-invoke history, and it only exists when hasMemory is true. Memory defaults to 'none' (src/cli/commands/create/command.tsx:285,343,588,608; validate.ts:90; --defaults => "no memory" at command.tsx:394). The 'none' option description is just "No memory" (src/cli/tui/screens/generate/types.ts:253), --session-id is mislabeled "Use specific session ID for conversation continuity" (src/cli/commands/invoke/command.tsx:146), and the unconditional "To resume" notice fires for memory-less agents in both the CLI stream path (command.tsx:105-120) and the TUI (src/cli/tui/screens/invoke/InvokeScreen.tsx:216).

The fix

Primarily a DX/discoverability fix (stateless behavior is correct without a session store), plus tightening of actively misleading copy. (1) Reword MEMORY_OPTIONS so the tradeoff is explicit: 'none' = "No memory - each turn is a fresh, stateless session", 'shortTerm' = "Remembers prior turns within a session" (src/cli/tui/screens/generate/types.ts:252-256). (2) Make the continuity messaging memory-aware: soften the --session-id help text (command.tsx:146) away from "conversation continuity", and suppress or reword the "To resume --session-id" notice (command.tsx:105-120) and the TUI exit hint (InvokeScreen.tsx:216) when the resolved agent has no memory configured — in the no-memory path --session-id is genuinely inert, so promising resume is wrong, not just vague. (3) Add a post-create/first-invoke hint for memory-less agents pointing to agentcore add memory; AddFlow.tsx already renders a memory-aware success note (AddFlow.tsx:370-394, 411-417) so reuse that copy. Product decision required: document/warn (low effort, recommended) vs change the default to short-term memory (larger; provisioning/cost impact). Do NOT change the NullConversationManager branch itself. Optional follow-up: address the latent cross-session-bleed smell of the single non-keyed global _agent in the no-memory path.

Files touched: src/cli/tui/screens/generate/types.ts:252-256 (MEMORY_OPTIONS descriptions); src/cli/commands/invoke/command.tsx:146 (--session-id help) and 105-120 ("To resume" notice, make memory-conditional); src/cli/tui/screens/invoke/InvokeScreen.tsx:216 (TUI resume hint, make memory-conditional); optional post-create hint reusing src/cli/tui/screens/add/AddFlow.tsx:370-394. Template src/assets/python/http/strands/base/main.py NullConversationManager branch (lines 441-468) is correct and should NOT change.

Validation evidence

The fix was verified by reproducing the original symptom and re-running after the change:

BEFORE (base/HEAD): command.tsx:144 help text = 'Use specific session ID for conversation continuity'; printInvokeResult unconditionally emitted 'To resume: agentcore invoke --session-id ' in both stream (HEAD:105) and non-stream (HEAD:118) branches regardless of memory; InvokeScreen.tsx:213 unconditionally set 'To resume this session, run: ...' exit hint; MEMORY_OPTIONS descriptions = 'No memory' / 'Context within a session' (never named the stateless tradeoff). AFTER (fix/1591): printInvokeResult only prints the 'To resume:' line when result.hasMemory is true (Session: still always prints); hasMemory derived once in handleInvoke (action.ts:54 = (project.memories?.length ?? 0) > 0) and threaded through all 4 InvokeResult return paths; InvokeScreen exit hint gated on config?.hasMemory (InvokeScreen.tsx:213) with useInvokeFlow populating hasMemory on InvokeConfig; help text reworded to '...continuity requires memory — see agentcore add memory'; MEMORY_OPTIONS 'none' -> 'No memory - each turn is a fresh, stateless session', 'shortTerm' -> 'Remembers prior turns within a session'. ADVERSARIAL PROOF: temporarily reverted printInvokeResult to unconditional 'To resume' -> the new no-memory test FAILED (expected '\nSession: abc\nTo resume...' not to contain 'To resume'), confirming the test guards the symptom; restored fix and it passes. REGRESSION GUARD: 'shows the To resume hint when target has memory' test passes. EXPLICITLY-NOT-CHANGED verified: base/main.py untouched, cre

Test suite: green.

---

Staged on the fork as a draft for human review. Promote to aws/agentcore-cli after vetting.

[github-actions]

Coverage Report

Status	Category	Percentage	Covered / Total
🔵	Lines	37.19%	13604 / 36579
🔵	Statements	36.45%	14463 / 39670
🔵	Functions	31.81%	2334 / 7337
🔵	Branches	31.14%	9011 / 28937

Generated in workflow #122 for commit 459af3d by the Vitest Coverage Report Action