Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 4 additions & 2 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -421,6 +421,7 @@ Implemented today:
through trace evidence,
- subprocess MCP invalid client-parameter guards before empty identifiers can
reach downstream servers,
- compact denial summaries on blocked MCP tool, resource, and prompt responses,
- subprocess MCP response timeout handling for hung downstream servers,
- subprocess MCP transport-close handling for child exits and broken pipes,
- a runnable MCP killer demo that blocks poisoned-output exfiltration and
Expand All @@ -437,8 +438,9 @@ Implemented today:
transport-close checks, mixed interop, public interop transcripts, resource
subscription no-passthrough, pre-ready notification guards,
notification-burst/flood checks, config and metadata guards,
client-intent redaction, invalid client-parameter guards, no-passthrough
checks, the MCP shim eval, inspect, and MCP server smoke checks.
client-intent redaction, invalid client-parameter guards, denial summaries,
no-passthrough checks, the MCP shim eval, inspect, and MCP server smoke
checks.

Not implemented yet:

Expand Down
4 changes: 3 additions & 1 deletion docs/architecture.md
Original file line number Diff line number Diff line change
Expand Up @@ -137,13 +137,15 @@ Replay modes:

`agentk trace-inspect` is the human review path. It verifies the hash chain, summarizes signature status, groups blocked events by policy rule, groups boundary events by syscall and evidence-ref type, and prints one compact row per event. Known hash evidence refs such as `args_sha256`, `descriptor_sha256`, and `response_sha256` are preserved. Any raw input ref is replaced with a fresh `input_sha256` ref in the inspection report.

Blocked MCP tool, resource, and prompt responses also carry compact `denial` summaries at the response boundary. Those summaries surface verdict, policy rule, reason, syscall, target, and any missing capability without requiring reviewers to dig through the full nested event body.

`agentk replay` records deterministic `stub_output_sha256` evidence refs for allowed `model.call`, `tool.invoke`, and `network.send` events. Blocked side effects stay blocked, do not get stub outputs, and are summarized by policy rule.

`agentk fork-replay` compares the recorded log against another policy and reports both per-event changes and transition counts such as `deny:rule->allow:rule`. This makes policy drift reviewable without manually counting every changed event.

`agentk fork-replay-behavior` accepts a JSON array of changed hashed output refs and emits a divergence report. Overrides are bound to the recorded step, syscall, and target, and raw output strings are rejected.

`agentk release-audit` packages the local release ritual into one report. It runs readiness, git hygiene checks, formatting, tests, clippy, a fresh demo trace, signature verification with signer summaries, signer-pinning and trusted-signer manifest smoke coverage, brokered secret-handle, secret-reference validation, and secret-store availability smoke tests, MCP taint-flow, subprocess MCP boundary, lifecycle-redaction, initialize-guard, tool/resource/prompt shape guards, bad-response redaction, response-timeout, transport-close, mixed-interop, public interop transcript, resource subscription no-passthrough, pre-ready notification no-passthrough, notification-burst/flood, no-passthrough, config-guard, AgentK metadata-redaction, client-intent hashing, and invalid-client-param smoke tests, redacted inspect, replay blocked-rule summaries, fork replay decision summaries, behavior fork replay, and an MCP server smoke test. It does not configure remotes or push.
`agentk release-audit` packages the local release ritual into one report. It runs readiness, git hygiene checks, formatting, tests, clippy, a fresh demo trace, signature verification with signer summaries, signer-pinning and trusted-signer manifest smoke coverage, brokered secret-handle, secret-reference validation, and secret-store availability smoke tests, MCP taint-flow, subprocess MCP boundary, lifecycle-redaction, initialize-guard, tool/resource/prompt shape guards, bad-response redaction, response-timeout, transport-close, mixed-interop, public interop transcript, resource subscription no-passthrough, pre-ready notification no-passthrough, notification-burst/flood, no-passthrough, config-guard, AgentK metadata-redaction, client-intent hashing, invalid-client-param smoke tests, and denial-summary smoke tests, redacted inspect, replay blocked-rule summaries, fork replay decision summaries, behavior fork replay, and an MCP server smoke test. It does not configure remotes or push.

### MCP Proxy MVP

Expand Down
3 changes: 3 additions & 0 deletions docs/roadmap.md
Original file line number Diff line number Diff line change
Expand Up @@ -99,6 +99,8 @@ Status: in progress.
- [x] Hash client-provided AgentK intent metadata in subprocess MCP evidence.
- [x] Reject empty subprocess MCP tool, resource, and prompt identifiers before
forwarding.
- [x] Surface compact denial summaries in blocked MCP tool/resource/prompt
responses.
- [x] Add an operator contract for subprocess MCP proxy boundaries.
- [x] Default-deny unsupported subprocess MCP request methods instead of generic passthrough.
- [x] Add release-audit smoke coverage for unsupported subprocess MCP no-passthrough.
Expand Down Expand Up @@ -133,6 +135,7 @@ Status: in progress.
- [x] Record stub outputs for model/tool/network syscalls.
- [x] Summarize blocked policy rules in deterministic replay output.
- [x] Summarize decision transitions in fork replay output.
- [x] Surface blocked MCP denial details directly at the response boundary.
- [x] Fork replay with changed model/tool behavior.
- [x] Emit divergence reports.

Expand Down
Loading