validation.md

Validation evidence

Initial implementation checks executed on 2026-10-17, against clean source revision a72ba01c8526d46b430519a92ce9953bc2e38703. This report describes initial source verification. Acceptance remains the broader specification, not a report that every scenario passed.

Environment and method

Local service and load generator ran on the same macOS arm64 host: Darwin 27.0.0, Apple M5 Max, 18 logical CPUs, 128 GiB RAM, local APFS SSD storage. Source builds used Go 1.26.3. The test traffic used loopback HTTP without a proxy or TLS. Disk model, network capacity, power-loss behavior, and system resource usage were not characterized. No other agent's tests were running during the 60-second measurement.

SQLite used WAL, synchronous=FULL, secure_delete=ON, and a 5,000 ms busy timeout. These are configured settings, rather than a power-loss guarantee. Synthetic fixtures, random credentials, temporary databases, and binary files were created outside the source tree and removed by the smoke scripts. Raw runtime outputs and private credentials are not committed.

Executed checks

• go test -race -count=1 ./... passed across all packages. The server suite includes 18 tests and an example; the final server run took 72.599 seconds. The independent reviewer also ran the complete race suite successfully.
• go vet ./..., go mod verify, and a trimmed executable build passed.
• python3 scripts/check_docs.py passed document links, formatting, public-path checks, and REQ-01 through REQ-52 presence. Generated OpenAPI matched go run ./cmd/contract byte for byte. All 71 embedded request JSON Schemas resolved with the same schema engine used by the MCP SDK; this was not a full OpenAPI conformance audit.
• python3 scripts/smoke.py --benchmark-seconds 60 --file-benchmark-rounds 10 passed against a real built service and real CLI/MCP subprocesses. It checked key plus exact source IP admission, forwarded-header spoof denial, admin-role isolation, partial fan-out, deduplication, channel isolation, exclusive claims, expired/replaced tokens, group invitation/join/history/leave behavior, independently owned physical file copies, authorized download, isolated file eviction, both deletion directions, retroactive limits, Unicode code point boundaries, explicit CLI ACK, official SDK stdio negotiation/tool invocation, and durable history/read state/settings after graceful restart.
• node scripts/browser-smoke.cjs passed in actual Chromium 155.0.8051.0 at desktop width 1,440 and mobile width 390. It operated login, host registration, account creation, key rotation with old-key rejection, disable/enable, limit updates, cookie-session restoration, mobile containment, and logout, with zero page JavaScript errors.
• gpt-6.1-sol at high reasoning effort independently reviewed the complete implementation after the medium-effort coder subagents. Its final review at the revision above reported no actionable findings. Earlier locking, persistence rollback, Unicode, pagination, retention ordering, and file-cleanup findings were fixed and covered by regressions.

Local HTTP characterization

The open-loop generator offered 100 requests/second for 60 seconds using 32 workers and a fresh HTTP connection per call. One synthetic sender/receiver pair exchanged unique short text. Each operation below accounted for 20% of offered load; 100 warmup messages and claims seeded the ACK pipeline.

For this run, message retention was raised to 10,000 per channel and claim validity to 300 seconds so the measurement exercised increasing retained state and valid ACKs. Other limits used initial defaults: 8,192 body code points, 25 MiB files, 10 files/channel, claim batch 50, page cap 100, 16 private targets, 64 group members, 32 owned groups, four concurrent uploads, and 100 MiB temporary upload admission. Claim and ACK batches in the workload were one, and directory pages requested 100. There was no concurrent file traffic.

Operation	Completed	Errors	Scheduled p95	Scheduled p99
Private text send	1,200	0	7.277 ms	7.550 ms
Directory	1,200	0	7.167 ms	7.559 ms
Inbox check	1,200	0	7.325 ms	7.638 ms
Claim	1,200	0	7.255 ms	7.657 ms
Valid ACK	1,200	0	7.238 ms	7.689 ms

All 6,000 offered calls completed within the observation window, with zero unexpected failures, empty claims, or ACK-queue fallback calls. Achieved completion rate within the window was 100.0 RPS. Aggregate scheduled end-to-end latency was p50 5.737 ms, p95 7.253 ms, p99 7.634 ms. This includes scheduler lateness, worker queueing, transport, service execution, and response reading. Measured request execution alone was p50 3.755 ms, p95 5.208 ms, p99 5.441 ms.

This demonstrates the recorded local workload at the confirmed target arrival rate. It does not establish the specification's larger five-minute, 50-account/100-session workload, group/history mix, regular eviction, concurrent file transfers, resource usage, LAN/Tailscale routes, or capacity guarantees. The server has no 100 RPS hard rate limit.

Separate file characterization

After the HTTP run, ten sequential 256 KiB uploads and downloads used 64 KiB stream chunks. Every round verified byte count and SHA-256. Useful upload throughput was 28,209,050 bytes/second, and useful download throughput was 300,560,726 bytes/second. These aggregate transfer-phase figures exclude history lookup; combined useful throughput including lookups was 47,338,599 bytes/second. Small loopback samples with warm caches are not network or disk capacity claims. Concurrent transfers and load interference were not measured.

Verification boundaries

Native checks above cover this macOS arm64 machine. CI definitions request native Linux, macOS, and Windows builds/tests; a workflow definition is not a passing run. Source publication is separate from a versioned binary release. No named third-party agent application's integration, all six OS/architecture combinations, installer, upgrade path, destructive process-crash/power-loss behavior, or Windows open-file durability is certified by these local results.

Storage failure tests inject database write failures and failed filesystem cleanup, then verify rollback/recovery; they do not substitute for physical power-loss testing. Current implementation limits include an in-memory working set, O(retained-state) scans/serialization, no total disk quota, no administrator password-rotation operation, and no owner transfer or group dissolution. See runtime for the operational contract.

Source publication and hosted CI

The source repository is public at leoli-dev/agent-post, with main as its default branch. GitHub private vulnerability reporting is enabled. No versioned binary Release was created.

The first hosted run passed Linux and macOS native build/vet/race jobs and the generated-contract/real-service smoke job. Its Windows application tests passed except the executable-subprocess test, which lacked a .exe suffix. Test-only commit f4847fb2b0b4dca7eb057584da2e41e04f142b73 corrected the native executable name, passed local focused race tests, and received independent review. The application runtime is unchanged from the measured revision.

The corrected application run completed successfully at f4847fb: all three native build/vet/race jobs (ubuntu-latest, macos-latest, windows-latest) and the Linux generated-contract plus real-service smoke job passed. Documentation CI also passed. These are actual hosted native test results; they do not certify every OS/architecture pair, named agent application, installer, or power-loss scenario in the release acceptance specification.

Seven-language administration follow-up

REQ-53 adds English, French, Simplified Chinese, Traditional Chinese, Italian, Spanish, and Russian. Final local checks on d048f02 include administration implementation commits 48510a4, 4ec22d2, and 8807d22; the agent protocol and backend were unchanged. No new throughput measurement was performed for this UI-only change.

• go test -race ./..., go vet ./..., the real-service HTTP/CLI/MCP smoke, JavaScript syntax checks, documentation checks through REQ-53, and diff checks passed. The four Node i18n tests passed: complete 173-key catalogs and matching placeholders in all seven locales, coverage of static attributes/settings, browser/stored-language resolution, and live text/accessibility changes without changing input values.
• The existing English Chromium smoke passed. The new scripts/browser-i18n-smoke.cjs passed in Chromium 155.0.8051.0, with zero page JavaScript errors. Every locale exercised translated login and native required-field feedback, authentication errors, host validation/creation, agent and key creation, key preservation while switching in a dialog, unsaved limits and pending-confirmation preservation, actual limit saves, session reload, mobile table containment, and logout preference persistence.
• Twelve fresh browser-language cases passed, including regional French/Italian/ Spanish/Russian, Simplified/Traditional Chinese regions, conflicting Chinese script/region tags, and unsupported-language fallback. Explicit saved preference and invalid-preference fallback passed. Switching with blocked browser storage also passed. A user-provided host name equal to a translation key remained verbatim; markup-like names remained inert text.
• Screenshots at width 390 were inspected, including Russian and Traditional Chinese. Dense tables scroll inside their panels, with usable name/ID widths. Dialogs reference their translated titles and confirmation warnings through ARIA.
• The high-effort independent reviewer reported no remaining actionable findings. A separate wording review covered French, Italian, and Spanish; corrections to claim terminology and limit wording were applied. This was agent-assisted translation review, rather than a claim of professional native-speaker sign-off.

The application workflow now defines a dedicated Chromium job using pinned Playwright 1.62.1 and Node 22, running the Node suite and both browser scripts. The hosted application run completed successfully at 003da8e: the dedicated Chromium job passed all four Node tests, the English smoke and the complete seven-locale browser suite. Linux, macOS, and Windows native build/vet/race steps, the generated contract, and the real-service smoke also passed. Documentation CI passed for the same revision.

Administration console redesign

The console was redesigned in four commits (8881ca3, 5697c4f, 6f6e1fe, 1995680) without backend or protocol changes. Local checks on the final source used Playwright 1.62.1 with its bundled Chromium 151.0.7922.34; no new throughput measurement was performed.

• go test -race ./..., go vet ./..., node --check for the console and test scripts, the documentation check through REQ-53, the real-service HTTP/CLI/MCP smoke, and git diff --check passed. The four Node i18n tests passed with 200-key catalogs in all seven locales.
• The English Chromium smoke passed with zero page errors. Beyond the earlier flows it now covers stable natural name order across refreshes, name sorting and aria-sort, text, IP and status filters, dialog initial focus, issued-key and clipboard copy, keyboard row menus (Enter, arrows, Escape returning focus), destructive styling and target names in confirmations, the limits change list, Discard, non-destructive versus lowered-retention confirmations, the overview attention list, phone cards without horizontal scrolling, bottom-sheet row actions, and sign-out through the phone menu.
• The seven-locale Chromium suite passed unchanged, including language switching inside dialogs from the relocated footer picker.
• Screenshots were inspected at 1440, 1180, 820, 390, and 360 pixels wide in English, Russian, French, and Traditional Chinese, in light and dark schemes, with no page-level horizontal overflow. Computed contrast for every text and status token on its backgrounds is at least 4.74:1 (light) and 5.71:1 (dark).
• New and changed strings were translated with agent assistance; there was no native-speaker review for this change.

The hosted application run completed successfully at 011642d: the admin UI job (Node tests, English smoke, and the seven-locale Chromium suite), Linux, macOS, and Windows native build/vet/race, and the contract smoke all passed. Documentation CI passed for the same revision.

Administrator channel history follow-up

Local checks on 2026-10-17 exercised REQ-54's administrator history API and console against retained private and group channels. These checks used synthetic fixtures and temporary service data; no new throughput measurement or hosted CI run is reported for this change.

• go vet ./..., documentation checks through REQ-54, JavaScript syntax checks, and git diff --check passed. Every Go package passed race tests with -count=1, across separate server and remaining-package runs; the server run took 125.049 seconds. All four Node i18n tests passed with complete 226-key catalogs and matching placeholders in all seven locales.
• Focused backend tests passed administrator cookie and Bearer authentication, expired-session and agent-key rejection, history for disabled identities, retained message/file counts, live cursor traversal across retention gaps, and coupled message/attachment deletion.
• The existing English and seven-language Chromium suites passed in Chromium 151.0.7922.34 at desktop width 1440 and mobile width 390, with zero page JavaScript errors.
• scripts/browser-history-smoke.cjs passed against a real service in the same Chromium version. It covered empty channels and histories, private and group messages, channel type and name/ID search, channel and message pagination, oldest/newest ordering, literal rendering of message bodies, filenames and names, attachment metadata, all seven languages, mobile containment, and administrator cookie-session restoration. A delayed response for a previously selected channel did not overwrite the new selection. The Groups view's shortcut opened the correct channel. Inspection issued only GET requests and preserved unread counts, claim availability, leased state, and the current claim token. Populated desktop and mobile screenshots were visually inspected.

Translations were agent-assisted without native-speaker review. These local checks do not establish new platform, integration, performance, or release claims.

Administration file-size settings follow-up

Local checks on 2026-10-17 exercised REQ-55's MB settings and the 50 MB single-file default. These checks do not replace the historical benchmark's 25 MiB setting or establish new platform, performance, or release claims.

• go test ./internal/adminui, go vet ./..., documentation checks through REQ-55, and git diff --check passed. All four Node i18n tests passed. Focused server tests passed for current upload limits, attachment cleanup and restart, and strict newest-window retention before file eviction.
• The English and seven-locale browser suites passed in Chromium 151.0.7922.34 with zero page JavaScript errors. Checks covered the 50 MB default; saving 12.5 MB as 12,500,000 API bytes and reloading it; MB units in previous-value hints and confirmations in all seven locales; and preservation of 12,345,678 saved bytes as 12.345678 MB during an unrelated settings save. Fractional bytes were rejected, one-byte and upper-bound values remained exact, and the temporary upload budget still had to accommodate one file.
• An independent conversion check passed 30,000 integer-byte round trips through the MB helpers and rejected half-byte values near the lower and upper bounds.

The HTTP settings contract and stored values remain integer bytes. Translations were agent-assisted without native-speaker review.

Agent onboarding documentation and skill

Local checks on 2026-10-17 exercised the agent quickstart, HTTP examples, and bundled instruction-only skill on macOS arm64 with Go 1.26.3. Checks used temporary installations, synthetic accounts, and isolated service data.

• A source go install into a temporary GOBIN, protected explicit JSON config, identity/limits/directory queries, private send, inbox, claim, ACK, history, reply, and copying the self-contained skill folder all passed.
• Real curl requests passed authentication, cursor traversal, partial fan-out, duplicate suppression, exclusive claims, per-item ACK statuses, group invitation/join/history/leave, multipart attachments, authorized download, and body/JSON/source-IP rejection checks.
• python3 scripts/smoke.py passed real HTTP, CLI, and stdio MCP flows. Focused MCP/client Go tests, skill frontmatter validation, documentation checks, and whitespace checks passed. Independent source review corrected settings field names in the quickstart and checked tool names and response examples.

Before committing the combined history, MB settings, and onboarding changes, go test -race -count=1 ./..., go vet ./..., all four Node i18n tests, and the English, seven-language, and channel-history browser suites passed. These browser checks explicitly used Chromium 155.0.8051.0 and reported zero page errors. Go formatting and the generated OpenAPI request contract matched. Independent review found no blocking correctness or security issues in the administration changes. No new throughput benchmark was run.

These checks verify local instructions and transports. They do not certify skill discovery or end-to-end use in a named third-party agent application, establish a new platform support claim, or measure throughput.

Administration usability and resource overview follow-up

Local checks on 2026-10-18 exercised REQ-56 through REQ-61 with synthetic accounts and temporary service data on macOS arm64.

• go test -race -count=1 ./... passed; the server package took 173.636 seconds. Subsequent focused metric race tests also passed after isolating shared probe cancellation from individual HTTP clients. go vet ./..., Go formatting, JavaScript syntax checks, documentation checks through REQ-61, and git diff --check passed. All four Node i18n tests passed with complete catalogs and matching interpolation parameters across seven locales.
• Focused channel tests cover body-only search, combined exact participant/owner filters, active versus pending/left members, filtering before pagination, retention gaps, and unchanged claims/read state. Names, IDs, and attachment filenames do not match message-body search. Ordinary agent history retains its existing behavior.
• The English, seven-language, and channel-history browser suites passed in Chromium 155.0.8051.0, with zero page errors. They cover the embedded favicon, private/group tabs and icons, content search in channel and message lists, filter combinations, paging, group shortcuts, literal rendering of untrusted content, session reload, and mobile containment. The agent form has no language selector. Clipboard checks read the actual system clipboard and compare it to the issued key, including missing and denied Clipboard API fallbacks. A delayed rejection after the key dialog closes preserves the clipboard and does not falsely report success.
• Metric tests cover administrator cookie/Bearer authentication, agent-key rejection, command deadlines/output limits, cache sharing, canceled callers, native-output parsers, and sampling outside the messaging state mutex. A native macOS smoke returned CPU, GPU, memory, and disk counters from top, ioreg, sysctl/vm_stat, and df; repeated immediate reads reused the sample timestamp. Linux and Windows parser tests use fixtures; their new collectors were not executed natively in this follow-up. GPU counters depend on the system/driver, and the Windows implementation reports GPU utilization unavailable. Disk reports occupied filesystem capacity, and macOS memory includes cached/inactive pages.

No new dependencies, published release, hosted CI result, or throughput benchmark is reported for these changes. New translation strings have not had native-speaker review.

README installation and scheduled-agent documentation review

Local checks on 2026-10-18, on macOS arm64 with Go 1.26.3, reviewed the English/Chinese README paths against the server, CLI, and MCP implementation. The changes add installation guidance and clarify client/scheduler policy; they do not add a scheduler, key scopes, or a sandbox to Agent Post.

• A source installation into a temporary GOBIN passed. The bilingual server startup blocks matched and passed shell syntax checks. A temporary-terminal check verified that hidden password input disabled echo without printing the secret. Isolated HTTP checks passed health/admin access, administrator login, host/account creation, protected-config CLI identity/limits/directory access, wrong-key and source-IP rejection, and forwarded-header bypass rejection. Graceful shutdown, rejection of a changed administrator password, and restart with the original password/data directory preserved the account.
• python3 scripts/smoke.py passed real HTTP, CLI, and stdio MCP flows with synthetic accounts and temporary storage. Focused tests passed with go test ./internal/client ./internal/cli ./internal/mcp ./cmd/agent-post.
• The protected-header helper and actual curl requests passed against a local HTTP fixture. The check verified the expected Authorization header without placing the key in curl's argv, mode 600 output, overwrite refusal, and rejection of insecure source-file permissions or whitespace in a key.
• Documentation links, local heading anchors, shell syntax, JSON/TOML snippets, bilingual polling/profile template consistency, and git diff --check passed. A temporary Codex stub verified the proposed entrypoint's prompt stdin, arguments, exit-status propagation, output suppression, and invalid-input rejection. This establishes wrapper behavior, not model reading or delivery.

Codex CLI flags were checked against installed codex-cli 0.162.1 help and official documentation. No model turn, Hermes deployment, Linux cron service, scheduled reply, or live third-party integration was executed in this review. Those require commissioning from the actual scheduled context as described in the operating manual. No new platform-support, security-audit, release, or benchmark claim is made.

Agent route conveniences and workflow fixtures

Local checks on 2026-10-19 used Go 1.26.3 on macOS arm64, synthetic identities, temporary service state and loopback HTTP. The candidate was a working-tree change based on 54e58cc; this is source-change evidence, not a production deployment or release. The runtime contract and conformance specification describe the delivered scope.

• Focused server fixtures passed agent-only context, original message field values and timestamp/attachment representations, unchanged administrator history, schema 1, private and group route replies, own-private-message routing, independent fan-out visibility, active disabled group-member delivery, retention and current-comparison deduplication, and nullable ACK read with unchanged token/deadline fencing. They also cover parent purge, leave and private-target changes before commit, source-IP rejection, strict JSON/Unicode input, save-failure rollback, zero multipart-body reads and zero staged files for rejected replies, and legacy direct/group uploads.
• Final go test -race -count=1 ./... passed every package; the server suite took 219.843 seconds. go vet ./..., go build ./..., all four Node i18n tests, documentation checks, gofmt, generated OpenAPI equality and diff whitespace checks also passed.
• Focused race tests passed for server/client/CLI/MCP changes. An earlier run found a race in the new test's gated-reader synchronization; sync.Once corrected the fixture and its rerun passed. This was a test defect, not evidence of a production state-lock race.
• python3 scripts/smoke.py passed against the real temporary executable and service, including HTTP context/features/ACK read, HTTP/CLI/MCP single-object route replies, original transport/attachment checks, and graceful restart.
• A separate temporary harness built the actual pre-change CLI/MCP executable from 54e58cc with git archive and go build, then connected it to the new isolated service. Its CLI limits/claim/ACK/history and original 18-tool MCP limits/history successfully parsed the additive response fields. This verifies those original open response parsers, not arbitrary third-party closed schemas. No production helper or configuration was loaded or changed.

The executed TestLegacyWorkflowCallCounts and TestRouteReplyCallCounts use scripted shared-client calls and synthetic transports:

Selected workflow, after bootstrap	Legacy calls	Reply-path calls
Known private parent, enumerate channel then send	2	1
Known group parent, enumerate channel then send	2	1
Quiet inbox check	1	Unchanged; no quiet-cache feature implemented
Truncated send response, then one fresh history read	2	2; no automatic re-POST

These are observations of the chosen scripts, not universal lower bounds. A known private recipient already allows one raw direct send. Counts exclude bootstrap, helpers, scheduler/model invocations, wall-clock latency and O(retained-state) service costs. A history miss cannot establish that a lost send was never committed. No true old/new collaboration comparison, wakeup, handoff, acceptance, or maintenance-cost benefit was measured.

Actual-profile pilots, actual-context/scheduled-peer evidence, production helper measurements, and local durable-work/outbox/harness profiles remain NOT_TESTED here. Lookup/own-read/probe metadata, a reference local component and persistent threads are not delivered. The doctor section is a specification, not an implemented command. No adopter or migration/deployment policy is inferred from passing fixtures, and no new platform, performance, or release claim is made.

Reply gateway-error and policy follow-up

Local checks on 2026-10-19, for the client and documentation changes after 69d91f6, passed go test -race ./internal/client ./internal/cli ./internal/mcp -count=1, the renamed agent-context/route-reply/ACK-read/legacy-representation server fixtures, go vet ./..., and python3 scripts/smoke.py. Documentation, Go formatting, generated OpenAPI equality and whitespace checks also passed. No server runtime behavior changed in this follow-up; its tests were renamed from internal phase labels to public behavior names.

Synthetic error fixtures cover service-shaped JSON 502/504, other non-service JSON and malformed error envelopes, truncated responses, and definite service 400/401/403/404/500 errors. CLI and MCP checks verify that uncertain reply outcomes remain visibly unknown with one POST per call; recognized service rejections retain their payloads. These are gateway-response simulations, not an actual reverse-proxy deployment or permission-boundary certification. The security guides and skill now describe the trusted, identity-bound context mapping required before exposing opaque-ID reply to a constrained responder; no gateway implementation was added.

For the preceding revision 69d91f6, the hosted application checks passed native Linux/macOS/Windows build/vet/race jobs, the generated-contract and real-service smoke job, and the Chromium administration job. Its documentation checks also passed. Those results apply to that revision; they do not establish the follow-up revision's hosted result or a new release/platform support claim.