Map-close verification: full demo green on main #113

Activity

1. tonibergholm

   added

   ready-for-agentScoped and unambiguous enough for an agent to pick upScoped and unambiguous enough for an agent to pick up

   wayfinder:taskWayfinder ticket: work that unblocks a decision or lands a fixWayfinder ticket: work that unblocks a decision or lands a fix

   on Jul 9, 2026
2. tonibergholm

   added a parent issue on Jul 9, 2026

   • Wayfinder map: 2026-07-09 security-audit remediation #103
3. tonibergholm

   marked this as blocked by 7 issues on Jul 9, 2026

   • Branch-name validation in Refs (path traversal fix) #105
   • Decide socket DoS budgets: fix in-map or defer #109
   • MCP response caps: streaming byte limit, JSON depth, manifest size #107
   • Verify content hashes on every get_raw() read #110
   • Implement ADR-0014: destructive-rollback approval gate #111
   • Implement the decided MCP URL policy #112
   • Read-idle timeout on the agenticd connection loop #119
4. tonibergholm commented on Jul 10, 2026

   

   tonibergholm

   on Jul 10, 2026

   MemberAuthor

   More actions

   Scope addition (surfaced resolving #111): verify each of the six audit fixes is actually present in `main` HEAD, not merely that its PR merged. PR #115's merge-conflict resolution silently deleted PR #114's interim gate — a merged critical-security fix that vanished from `main`. So map-close must grep `main` for each fix's fingerprint, not trust the merge log:

   • Bump Rust toolchain to 1.95.0 #1 `validate_branch_name` in refs.rs
   • Chunk C (part 1): MCP fingerprinter + per-dimension diff #2 the ADR-0014 gate (`ApprovalError` / `approval_gate`) in rollback/mod.rs — the one that regressed
   • Chunk C (part 2): rollback orchestrator (memory + prompts + forward-record) #3 `check_integrity` + `fetch_map` in store.rs / gcs_store.rs
   • Chunk D (part 1): trigger-based streamer + delta-sized snapshots #4 `McpLimits` in mcp.rs
   • chore: ignore agent-harness runtime state, keep main clean #6 `postgres integration (pgvector)` job in ci.yml

   As of 2026-07-10, #1/#3/#4/#6 confirmed present in main HEAD; #2 lands with PR #121.
5. tonibergholm

   self-assigned this

   on Jul 10, 2026
6. tonibergholm commented on Jul 10, 2026

   

   tonibergholm

   on Jul 10, 2026

   MemberAuthor

   More actions

   Verification result — mixed, and the map is NOT cleanly done against its literal destination

   Security destination — MET and verified ✅

   • All six fixes present in main HEAD (grepped every fingerprint, not just "PR merged"):
     • Bump Rust toolchain to 1.95.0 #1 validate_branch_name (refs.rs) · Chunk C (part 1): MCP fingerprinter + per-dimension diff #2 approval_gate + ApprovalError (rollback/mod.rs — the one Validate branch names before building ref paths under refs/heads #115 had regressed) · Chunk C (part 2): rollback orchestrator (memory + prompts + forward-record) #3 check_integrity (store.rs) + fetch_map (gcs_store.rs) · #4a McpLimits (mcp.rs) · #4b validate_mcp_url (mcp.rs) · Chunk D (part 2): pg_advisory_lock serialises snapshots across processes #5 READ_IDLE_TIMEOUT (server.rs) · chore: ignore agent-harness runtime state, keep main clean #6 postgres integration (pgvector) job (ci.yml).
   • Regression tests exist and are green in CI on every PR, and the Postgres integration suites are green on main (the authoritative clean environment).
   • §9 perf: the get_raw integrity fix measured ~0.5 ms/MiB (recorded in benchmarks.md), negligible vs the 5 s rollback budget; Postgres restore suites green in CI. (A live rollback timing wasn't captured — see below.)
   • Daemon side of the demo works: release binaries build, Postgres comes up, seeding runs, agenticd binds the socket, attaches the memory backend (logical decoding), and listens — healthy.

   Demo-green gate — NOT met ❌

   scripts/run-demo.sh does not run green end-to-end on the dev box. It hangs at the first "baseline ask" — the Python LangGraph agent → AgenticCheckpointer → daemon — before any rollback. Killed at 30 s, no output, no traceback; the daemon logs no request handling and no error.

   The map's real question — did the six fixes break the demo? NO. ✅ (proven, not argued)

   • The Python SDK + demo agent are byte-identical between pre-map d78c805 and current main: git diff d78c805..main -- sdk/python examples/langgraph-rollback is empty. The six fixes touched only Rust daemon/core code.
   • The read-idle timeout (the only fix touching every connection, and a 30 s suspect against the 30 s hang) provably did not fire — full daemon-log grep for idle/read deadline/error came back clean.
   • The daemon's commit-dispatch path is unchanged by any of the six fixes; the other five don't touch the baseline-commit path (branch validation only rejects bad names; approval gate only gates accept_data_loss; get_raw only affects restore; MCP only with configured servers — the demo configures none; chore: ignore agent-harness runtime state, keep main clean #6 is CI infra).

   So the hang is a pre-existing property of the demo/SDK stack, not a regression from the security remediation. Full gap list + repro filed as #124.

   Decision needed (Toni's call — genuine destination reinterpretation)

   The destination says "…with scripts/run-demo.sh green end-to-end at close." That literal gate isn't met, for reasons outside this map's security work. CLAUDE.md already calls the fresh-clone demo run "aspirational, only measured on dev machines with warm caches." Two ways to close:

   • A (recommended): Close the map on the security destination (six findings fixed, tested, merged, verified present in main HEAD) and treat demo-runnability as separate follow-up (Demo (run-demo.sh) does not run green from a clean state — 5 gaps + baseline-ask hang #124). Rationale: the demo failure is proven independent of the security work, and CLAUDE.md's "aspirational" framing supports it.
   • B: Keep this ticket open until the demo actually runs green — i.e. block map-close on fixing Demo (run-demo.sh) does not run green from a clean state — 5 gaps + baseline-ask hang #124 (the SDK hang + portability gaps), which is a distinct, non-security effort.

   I've left this ticket open rather than mark "demo green" (which would be false). Reassign/close per your call.
7. tonibergholm commented on Jul 10, 2026

   

   tonibergholm

   on Jul 10, 2026

   MemberAuthor

   More actions

   Resolved — map closed on the security destination (Toni's call, 2026-07-10)

   Decision: close on the security destination. All six audit findings are fixed, tested, merged, and verified present in main HEAD; regression tests green in CI; Postgres integration suites green on main. The literal "run-demo.sh green end-to-end" criterion is reinterpreted per CLAUDE.md's standing "aspirational, only measured on dev machines with warm caches" framing — the demo does not run green from clean on the dev box, but that failure is proven independent of the security work (SDK+agent byte-identical to pre-map d78c805; the read-idle timeout provably did not fire; the daemon commit-dispatch path is unchanged).

   Demo-runnability — the baseline-ask hang plus four portability gaps (release strip on macOS, socket SUN_LEN, python vs python3, uninstalled agent deps) — is carried forward as regular issue #124, outside this map's destination.

   Verification evidence is in the prior comment. Closing this ticket closes the map.
8. tonibergholm

   closed this as completedon Jul 10, 2026