Skip to content

feat: AgentCore login attach and consent - #6600

Closed
kyleseaman wants to merge 1 commit into
cursor/agentcore-4-gateway-r2-1b02from
cursor/agentcore-5-login-1b02
Closed

kyleseaman wants to merge 1 commit into
cursor/agentcore-4-gateway-r2-1b02from
cursor/agentcore-5-login-1b02

Conversation

@kyleseaman

@kyleseaman kyleseaman commented Aug 28, 2026 •

Copy link
Copy Markdown
Contributor

Stacked PR 5 of 6. Trust boundary: login-posture inbound attach, 3LO consent, and unattended-session policy.

What this PR does

Login posture stages a per-session 0600 inbound sidecar (JWT or URL-only OAuth challenge). Dispatchers call prepare_turn_gateway / prepare_session_gateway to stage identity. SessionManager.get_or_create installs that sidecar under a per-key creation reservation that covers both warm-pool claimants and cold creators, before provider.start(), so session/new / session/load cannot read a leftover or concurrent bearer. The reservation is acquired before _drain_and_claim and released by one outer finally across install/start/register — a factory exception, a cancel while waiting for the lock or the start semaphore, or an install failure cannot wedge the key or orphan a claimed pool process (the finally puts an unregistered claim back). A warm-pool process already completed session/new at fill; when AgentCore is on (login JWT or workload SigV4) that process is discarded after install so a fresh start() injects Gateway. After the same-key winner holds the per-session semaphore it applies again and recycles if the fingerprint changed. Attach-before-acquire raced concurrent JWTs and killed in-flight streams; attach-after-return leaked a leftover bearer.

Unattended cron:, taskrunner:, subagent:, hook:, wf:, wf-pool:, and wf-author: prefixes and the exact keys _bg / _hb never bind. Login inject denial is SEL-audited even when attach was never called. clear_session_principal after get_or_create would recycle the provider this turn just acquired, so post-acquire retract is metadata-only. A failed unlink/recycle raises GatewayCredentialError rather than leaving the old bearer active.

Rebuild withholds non-managed MCP from the emitted --agent spec (kirocrew.json is the filtered runtime). Authored leftovers are stashed to owner-only agentcore-authored-mcp.json. Restore runs after current MCP sources merge (setdefault, current sources win) and the sidecar is unlinked only after the runtime write succeeds, so leaving login cannot resurrect a deleted server.

GET /api/agentcore/consent is owner-dashboard-cookie only. Consent URLs must pass security.allow_agentcore_consent_url. Catalog inspect and the consent GET both go through surface_consent_url, which SEL-audits grant and deny (host+path only, never token bytes).

Inert by default

No capability + no login posture = no sidecar, no withhold, no consent URL.

Two changes to the PR 4 workload path ride along and are intended, not incidental:

  • Workload user/OBO sessions without a vaulted owner token are denied Gateway (SEL reason user_without_vault), the same rule unattended keys already met. Denial is the absence of a session entry: the Gateway is only ever session-injected and rebuild pops any persisted managed leftover, so no placeholder element is needed and none is emitted. The inject tests were re-keyed from agent:main:main to dashboard:1 because the former is now a denied key under this rule.

  • Gateway boot runs the agent rebuild off the event loop with app_agent_refresh=False so a slow rebuild cannot stall boot readiness; app-agent files are refreshed by the normal rebuild path afterwards.

  • The withhold gate reads the EFFECTIVE posture (agentcore_aws.resolved_posture, strict mode: ceiling when loaded, else launch env / home authoring), the same read the session attach path uses -- so an env-only login instance (a CloudFormation launch configured through its unit environment, no ceiling document) withholds non-managed MCP at rebuild exactly where it attaches login Gateways; an unavailable ceiling still withholds.

  • Login withhold decides ownership by the FILE kiro-cli resolves (--agent by declared name, project dir first, declared name over stem): the app-agent ownership ledger (<config_dir>/app-agent-trust/ownership.json, on the sensitive-path floor and sandbox-masked, written before each spec lands, never replaced when unreadable) plus the manifest say which materialized specs are the app's; a bare declared name is no exemption, a user's <app>--mine.json is never touched, unknown ownership fails closed.

  • The authored-MCP stash never mistakes an unreadable source for a deleted one: _live_source_inventory reports completeness; on an incomplete inventory the stash writer omits sourceServers (the merge prunes nothing), keeps every source-owned stash entry and its refs as stashed instead of merging the partial extract's entry for it (the runtime spec was assembled from the readable sources only, so an alias the unreadable source owns resolves to a sibling there), passes no partial live specs to the merge (no alias is classified as taken over), and restore likewise skips takeover classification, drops no name and keeps the sidecar.

Out of this PR

Tests

  • test_agentcore_credential_lifecycle.py: factory / _start_sem cancel cannot leak the create lock; leftover bearer + concurrent pooled/cold creators; stale JWT replaced before both session/new and session/load; AgentCore-on pool hit discards the fill process and re-issues session/new (login and workload); identity-off still reuses the pool; cancel-while-waiting-for-lock and install failure return the claim; human first turn injects then synthetic unbind retracts
  • test_agentcore_login_withhold.py: authored sidecar stash/restore; second login rebuild does not overwrite a good stash; source mcp.json stays byte-identical
  • test_identity_topology.py: every human dispatcher names await prepare_turn_gateway immediately before get_or_create
  • test_agentcore_inspect.py: a refused catalog consent URL is dropped and SEL-audited
  • test_agentcore_unattended.py, test_agentcore_consent.py

Stacked on #6598. Part of stack #6601. Supersedes #6387.

Credential lifecycle: staged install runs under the creation reservation before provider.start(); unbind retracts the live Gateway/bearer (ACP recycle + sidecar drop). publish_turn_identity without bind is metadata-only so a post-acquire retract cannot kill the provider just created. Authored MCP is stashed to agentcore-authored-mcp.json and unlinked after restore. Unattended prefixes include hook: (not channel:).

@github-actions

github-actions Bot commented Aug 28, 2026 •

Copy link
Copy Markdown
Contributor

First Principles Review (Fable 5) — 🟡 CONCERNS

Premise-level review of 0f7bd89e1ceb7e632777222b61ffd88826feea38 — why this exists and whether the shipped surface is the smallest honest version. Updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

First-Principles-Verdict: CONCERNS

drain_expired_gateway_transport and UNATTENDED_SESSION_PREFIXES ship with zero callers anywhere — the RFC's expiry drain has no trigger wired in this change.

Not justified as shipped

  • Item 10 — zero consumers: grepped both names across src/ and test/ — 0 references outside their definitions (agentcore_gateway.py:472, :70; the constant's own comment says "kept for docs and tests" yet no test or doc names it). The RFC (line 602) plans the drain, but nothing invokes it; shipped expiry handling is the attach-path fingerprint recycle.
    Subtraction: delete UNATTENDED_SESSION_PREFIXES; defer drain_expired_gateway_transport to the PR that wires the expiry watch that calls it.
    Clears when: a production caller lands in this diff or the two symbols are removed.
  • Item 7 — undeclared: optional client_ids on oauth_endpoints.json entries (exfil.py, consumed only by allow_agentcore_consent_url) is spec'd in governance.md but absent from the description.
  • Item 8 — rides along: declared; base RFC already records the workload user/OBO vault rule.
  • Item 9 — rides along: declared; boot-readiness cause named.

What this change ships

Inventory (10 items) — 6 justified

Intent: let a human's channel/CLI session carry their own identity to the AgentCore Gateway under login posture, while automation and non-managed MCP are withheld — an ADDITION (RFC rfc-agentcore-identity-gateway is in-progress on base and names the sidecar, consent endpoint, withhold, and unattended denial).

  1. Human channel/CLI turns stage a per-session 0600 Gateway credential sidecar installed before session start — justified
  2. Cron/taskrunner/subagent/hook/workflow/unknown session keys never get a login Gateway; denial is SEL-audited — justified
  3. Under login the rendered agent file keeps only managed MCP; operator servers stash to an owner-only sidecar and restore on leaving login — justified
  4. App agents lose embedded MCP under login; a new hidden ownership ledger (app-agent-trust/ownership.json) records app-owned agent files — justified
  5. An app agent still carrying withheld MCP refuses to spawn under login, on all three harnesses — justified
  6. New owner-only GET /api/agentcore/consent plus allowlisted catalog authorization URLs (UI lands in stacked PR 6, per RFC) — justified
  7. oauth_endpoints.json entries accept optional client_ids; query-bearing consent URLs need a listed client and loopback redirect — undeclared (spec'd, not in description)
  8. Workload user/OBO sessions without a vaulted owner token are denied Gateway; inject tests re-keyed accordingly — rides along (declared, RFC-covered)
  9. Gateway boot runs the config rebuild off the event loop and skips the app-agent walk — rides along (declared, boot readiness)
  10. Two new public helpers (drain_expired_gateway_transport, UNATTENDED_SESSION_PREFIXES) have no caller — zero consumers

[FIRST-PRINCIPLES-REVIEWED] 0f7bd89

@github-actions

github-actions Bot commented Aug 28, 2026 •

Copy link
Copy Markdown
Contributor

Design Review (Fable 5) — ✅ PASS

Design-level review of 0f7bd89e1ceb7e632777222b61ffd88826feea38 — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

Design-Verdict: PASS

Verify the warm-pool claim→discard→replenish cycle under login/workload posture — every human session creation burns one pooled spawn plus a cold start.

Suggestions

  • _gateway_requires_fresh_session inputs (key, resolved claim crew) are available before _drain_and_claim; checking it pre-drain would skip the pool for identity-on keys instead of claiming, hard-killing, and refilling a process per session — and would stop recording a "hit" that delivered nothing.

[DESIGN-REVIEWED] 0f7bd89

@github-actions

github-actions Bot commented Aug 28, 2026 •

Copy link
Copy Markdown
Contributor

GPT 5.6 Review — 🔴 changes requested (blocking)

GPT 5.6 found at least one blocking issue that must be resolved before merging 0f7bd89e1ceb7e632777222b61ffd88826feea38. 1 of 1 blocking finding(s) are security-class and were withheld from adjudication, so the blocking verdict stands.

This comment is updated in place on each push.

BLOCKING -- src/kiro_crew/agent.py:1755 -- stale ownership records can delete a user-created agent
leftover = _prune_unkept_app_agent_files(name, set())
Disabled app update drops an agent -> deregistration retains its ledger entry -> user reuses filename -> login rebuild deletes it.
Anchor: residual/crash-data-loss-corruption
Fix: Remove successfully deregistered filenames from the ownership ledger.
[BLOCK-MERGE] 0f7bd89
[GPT-REVIEWED] 0f7bd89

Adjudication (Opus 4.8) — is blocking on each finding proportionate?

The adjudicable block is empty (0 findings). One fenced finding, F1.

F1 — src/kiro_crew/agent.py:1755, fenced (unbounded: irreversible loss of user data).

Conditions traced in this run:

  • _deregister_agents (bridges.py:1142–1148) unlinks app agent files by prefix but never calls forget_app_agent_ownership, so a deregistered/dropped name stays in the append-only ledger. Append-only is confirmed at owned_app_agent_names (bridges.py:3155–3160) and _record_app_agent_ownership "never removes" (bridges.py:3249).
  • owned_app_agent_names returns declared | recorded (bridges.py:3182); a formerly-owned name reused by a user resolves as owned.
  • The disabled-app fail_closed branch calls _prune_unkept_app_agent_files(name, set()) (agent.py:1755) → unlinks every owned file with empty keep (bridges.py:1376–1387).

Recovery path: none — the user's file is unlinked and there is no backup; loss is irreversible.

Rarity assessment: The triggering collision (a user naming a custom agent <app>--<dropped-name>.json) is precisely the scenario the entire ownership mechanism was built around — its own docstrings (bridges.py:3148, 1345) treat user app-prefixed custom agents as expected, not exotic. The conditions are neither mutually contradicting nor impossible, so I cannot complete the rarity record required to FLAG. Per the asymmetry rule, torn on the unrecoverable class → UPHOLD-FENCED.

[ADJUDICATION] 0f7bd89e1ceb7e632777222b61ffd88826feea38 total=0 uphold=0 downgrade=0
[GPT-ADJUDICATED] 0f7bd89e1ceb7e632777222b61ffd88826feea38
[ADJUDICATION-FENCED] 0f7bd89e1ceb7e632777222b61ffd88826feea38 fenced=1 flagged=0
UPHOLD-FENCED F1 src/kiro_crew/agent.py:1755 -- The append-only ledger retains deregistered names, and a user reusing an app-prefixed filename is the exact documented scenario the ownership system exists for, so the deletion of unrecoverable user data is plausible in real operation.
[GPT-ADJUDICATED-FENCED] 0f7bd89e1ceb7e632777222b61ffd88826feea38

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 0f7bd89e1ceb7e632777222b61ffd88826feea38: <one-sentence reason>

@github-actions github-actions Bot added readiness: action required A blocking check or review needs attention and removed readiness: checking Automated validation is still running labels Aug 28, 2026
@github-actions

github-actions Bot commented Aug 28, 2026 •

Copy link
Copy Markdown
Contributor

Opus 4.8 Review — ✅ no blocking findings

Reviewed 0f7bd89e1ceb7e632777222b61ffd88826feea38 — this comment is updated in place on each push.

Review details

No findings.

The single candidate (_recycle_live_session reporting a busy-skip reset as success) rests on a TOCTOU window that I could not establish as reachable. is_busy (session_allocation.py:1002) and reset's skip check (session_lifecycle.py:438) test the same session.semaphore.locked(), so there is no systematic mismatch. In _recycle_live_session there is no await between the synchronous is_busy pre-check and creation of the reset(...) coroutine; reset re-evaluates busyness atomically under owner._lock, and asyncio.Lock.acquire() on an uncontended lock returns without yielding — so in the common path the reset's busy check runs in the same tick as the pre-check and cannot diverge. Divergence requires owner._lock to be contended at that instant AND a foreign turn to acquire the same key's lease during reset's lock-wait suspension — a concurrent same-key operation the per-key creation reservation is built to serialize. Grounding the harm reduces to "if a foreign turn were to acquire the lease in that window," which is exactly the speculative call path the bar rejects. The discovery pass itself rated it low and could not confirm reachability.

[OPUS-REVIEWED] 0f7bd89

Verdict parsed from the review's SHA-scoped output markers for commit 0f7bd89e1ceb7e632777222b61ffd88826feea38.

False positive or not applicable? A repository writer can comment:
/ai-review override fable 0f7bd89e1ceb7e632777222b61ffd88826feea38: <one-sentence reason>

@kyleseaman
kyleseaman marked this pull request as ready for review August 28, 2026 18:56
@kyleseaman
kyleseaman requested review from a team and Zedmor August 28, 2026 18:56
@kyleseaman
kyleseaman force-pushed the cursor/agentcore-5-login-1b02 branch from e78c0fd to 7397d76 Compare August 28, 2026 19:31
@github-actions github-actions Bot added readiness: checking Automated validation is still running and removed readiness: action required A blocking check or review needs attention labels Aug 28, 2026
@github-actions

github-actions Bot commented Aug 28, 2026 •

Copy link
Copy Markdown
Contributor

UX Review (Fable 5) — ⏭️ skipped

Revision 0f7bd89e1ceb7e632777222b61ffd88826feea38 touches no user-facing surface (no changes under website/ or committed screenshots), so the UX review was skipped. Advisory — does not block merge.

@kyleseaman
kyleseaman force-pushed the cursor/agentcore-5-login-1b02 branch from 7397d76 to c6be17d Compare August 28, 2026 19:54
@kyleseaman
kyleseaman force-pushed the cursor/agentcore-5-login-1b02 branch from c6be17d to bf26519 Compare August 28, 2026 20:21
@github-actions github-actions Bot added readiness: action required A blocking check or review needs attention and removed readiness: checking Automated validation is still running labels Aug 28, 2026
Comment thread src/kiro_crew/agent.py Outdated
if not isinstance(servers, dict):
return
for name in [n for n in list(servers) if n not in managed_names]:
del servers[name]

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

[P1] Do not implement withholding by deleting the only durable copy of custom MCP configuration. For an agent-file-only server, enabling login posture removes the entry and writes the reduced config back to disk; switching to workload later has no source from which to restore it. Preserve the authored configuration separately and filter only the emitted session view, or leave these entries untouched until a reversible withholding path exists.

@kyleseaman
kyleseaman force-pushed the cursor/agentcore-5-login-1b02 branch from bf26519 to 625dfb5 Compare August 28, 2026 20:53
@github-actions github-actions Bot added readiness: checking Automated validation is still running and removed readiness: action required A blocking check or review needs attention labels Aug 28, 2026
@github-actions github-actions Bot added readiness: checking Automated validation is still running readiness: action required A blocking check or review needs attention and removed readiness: checking Automated validation is still running labels Aug 29, 2026
@kyleseaman
kyleseaman force-pushed the cursor/agentcore-5-login-1b02 branch from 701bdf4 to 65debdd Compare August 29, 2026 01:54
@github-actions github-actions Bot added readiness: checking Automated validation is still running and removed readiness: action required A blocking check or review needs attention labels Aug 29, 2026
@kyleseaman
kyleseaman force-pushed the cursor/agentcore-5-login-1b02 branch from 65debdd to 4e1a20f Compare August 29, 2026 02:00
@github-actions github-actions Bot added readiness: action required A blocking check or review needs attention and removed readiness: checking Automated validation is still running labels Aug 29, 2026
@kyleseaman
kyleseaman force-pushed the cursor/agentcore-5-login-1b02 branch from 4e1a20f to 706ba27 Compare August 29, 2026 02:18
@github-actions github-actions Bot removed the readiness: action required A blocking check or review needs attention label Aug 29, 2026
@kyleseaman

Copy link
Copy Markdown
Contributor Author

/ai-review override gpt 2b5f71a: Boot leftover-verify was already rejected because it delays readiness by profile size; Identity PUT and leave-login still rematerialize fail-closed.

@github-actions

Copy link
Copy Markdown
Contributor

Human judgment recorded

@kyleseaman marked the gpt AI finding as false positive, not applicable, or explicitly accepted for 2b5f71aec9ada72eef73a6db646b5c1b71cd29d7.

Boot leftover-verify was already rejected because it delays readiness by profile size; Identity PUT and leave-login still rematerialize fail-closed.

This decision applies only to this commit. A new push requires a new judgment.

@github-actions

github-actions Bot commented Sep 11, 2026 •

Copy link
Copy Markdown
Contributor

Security Scope Review -- 🟡 CONCERNS

Which legitimate operations does 0f7bd89e1ceb7e632777222b61ffd88826feea38 newly refuse? Updated in place on each push. A script-confirmed regression blocks; the model proposes candidates and never decides a refusal.

Verdict basis: the reviewer's own judgement; no confirmed regression.

Security scope review -- adjudicated candidates

  • macos-latest (platform posix): 9 classified, 0 newly refused, 9 still allowed.
  • ubuntu-latest (platform posix): 9 classified, 0 newly refused, 9 still allowed.
  • windows-latest (platform windows): 5 classified, 0 newly refused, 5 still allowed.

No candidate flipped from allowed to refused.

Scope-Verdict: CONCERNS

TIGHTENING: The read+write path fence (security/paths.py _CREW_SECRET_LEAVES, expanded under both ~/.kiro/crew and ~/.kirocrew) gains two directory leaves — APP_AGENT_TRUST_DIR_NAME ("app-agent-trust") and "agentcore-authored-mcp" — refusing the one threat the hunks name: "an agent that could write it would erase its own record and run a stale app spec's MCP under Login" (app-agent ownership ledger), and "an agent that could write it would restore arbitrary MCP commands when posture leaves login; a reader learns the operator's withheld servers" (authored-MCP stash). The same two names are also bind-masked in sandbox._CREW_HIDDEN_LEAVES so a spawned command's open() cannot route around the tool gate.

ADJUDICATED: 14 candidates, 0 confirmed newly-refused
No confirmed rows were handed to me for this run. The candidates work the new leaves' boundary: the crew-home parent listing (ls -la ~/.kiro/crew), the apps sibling sharing the app prefix with app-agent-trust, the legacy ~/.kirocrew prefix, the routine config.json read in the same directory, source searches carrying the new leaf tokens in argv (grep/findstr/Select-String), the sanctioned pytest invocation of this branch's own test_agentcore_login_withhold.py, and git reads/diffs naming the guard modules.

PLATFORM COVERAGE: posix and windows; no any rows, since every candidate carries a path, separator, or executable name. The diff declares both leaves with POSIX / separators (_CREW_HOME_PREFIXES = (".kiro/crew", ".kirocrew")), which is the platform assumption the windows rows probe: the same crew-home listing and config read in \-separated %USERPROFILE% (cmd) and $env:USERPROFILE (PowerShell) spellings, and the leaf tokens in argv under the Windows tokenizer (findstr, Select-String). Posix rows cover the spelling the fence was written in, including the legacy-home prefix. Per-leg verdicts are the deterministic fold's to report, not mine.

UNADJUDICATED: Four guards in this change take no shell command line, so no verdict is available for them and the honest verdict is CONCERNS:

  • allow_agentcore_consent_url (security/exfil.py): reached via surface_consent_url / consent_snapshot in platform/agentcore_gateway.py, which the new dashboard handler dashboard/handlers/agentcore_consent.py serves to the operator. Candidate legitimate inputs I could not get a verdict for: a builtin-allowlisted IdP authorize URL carrying its real query (?client_id=...&redirect_uri=https://app.example/cb&response_type=code&scope=...&state=...) — refused twice over, since builtin endpoints have no client_ids list and a non-loopback redirect_uri is refused; a consent endpoint on an explicit port (https://idp.corp:8443/authorize); an operator-extended endpoint whose oauth_endpoints.json entry predates the new optional client_ids key, whose query-bearing URLs are now all refused.
  • The login MCP withhold (_login_mcp_withhold() inside agent.py rebuild_agent_config, deciding by the app-agent-trust ledger): under Login posture it withholds authored non-managed MCP servers from the runtime spec. Candidate legitimate inputs: an operator's own hand-authored MCP server under Login posture; a materialized app spec whose ledger record is absent (written by a pre-ledger version, or the ledger lost to a restore), which the gate can no longer tell from a user's custom agent.
  • The sandbox bind-mask additions (sandbox._CREW_HIDDEN_LEAVES): an OS-level mount mask, not the tool-call composite. The hunks state nothing inside the sandbox reads or writes either directory, so I have no legitimate in-sandbox input to name — but no classifier verdict covers the mask itself.
  • prepare_turn_gateway (messaging/identity.py, called from every channel dispatcher and drive_turn before session/new): GatewayCredentialError deliberately propagates, so when AgentCore identity is on, a transiently failed credential recycle now refuses an ordinary chat turn on messaging, Slack, Discord, and Telegram. A chat turn has no command line, so no verdict is available.

[SCOPE-REVIEWED] 0f7bd89

@github-actions

Copy link
Copy Markdown
Contributor

GPT 5.6 Review — ⚠️ review incomplete

GPT 5.6 did not produce a complete verdict for 296fb9b9b077916724adbd015baa63ecd1b75c71; inspect the workflow logs and re-run it.

This comment is updated in place on each push.

See the GPT 5.6 Review job logs; this commit has no completed GPT verdict.

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 296fb9b9b077916724adbd015baa63ecd1b75c71: <one-sentence reason>

Stage Gateway bind before get_or_create; apply after the per-session
lease, immediately before session/new, so concurrent turns cannot
overwrite each other's JWT or recycle an in-flight stream. A failed
recycle or sidecar unlink raises so the turn cannot continue with a
leftover bearer. Authored MCP is merged into an owner-only sidecar
with sourceServers deletion tracking so leaving login cannot
resurrect a deleted server; --clean discards that sidecar.
Unattended is the default: only dashboard, cli_chat, and channel
session namespaces bind as a human. A same-named live source does
not discard a stashed agent override. A nonempty consent query
must bind client_id and redirect_uri.

The login leftover-verify catalog fails closed: an app directory
whose installed.json the listing could not represent aborts the
login emit instead of leaving that app's materialized agents, and
their non-managed MCP commands, unrefreshed and loadable.
@bolichen97

Copy link
Copy Markdown
Contributor

Maintainer note: PR #11780 (merged 2026-09-19, bc0d2b0cf, "simplify member memory to one SQLite store") removed the protected-binding / private-execution member memory model. Two PRs written against the old model merged after it with a textually clean merge and broke main at import time (#11947, #11960); both are now reverted. This PR carries the same hazard.

This PR is MERGEABLE, but on this head src/kiro_crew/session_allocation.py:1801 calls await cast(Any, provider).prepare_private_memory() under is_acp_provider. No provider on main defines prepare_private_memory any more, so every ACP session allocation would raise AttributeError. Please do not merge as-is: rebase onto current main and delete that call (the _private_memory provider flag it fed is gone too).

@bolichen97 bolichen97 left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Blocking merge until this is rebased onto post-#11780 main and the removed-symbol references listed in my comment above are dropped. Textually the merge is clean, so nothing else would stop it from breaking main at import time.

@iamwhatever

Copy link
Copy Markdown
Contributor

Thanks a lot for this careful, well-split AgentCore stack. We are closing #6595, #6793, #6828, #6600 and #6602 together for now: the branches are about 7,500 commits behind main, #6595 still has an open review on the identity contract, and the upper PRs reference symbols that #11780 removed, so a rebase would be close to a rewrite. The need is still tracked in docs/request-for-change/rfc-agentcore-identity-gateway.md (status: partial; phase 1 is on main via #6594). When AgentCore is picked up again, please open a fresh stack from current main, starting with the principal-propagation layer. Happy to reopen if you would rather continue on these branches.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants