Historical evidence: any Serena references below describe a provider retired in contract 4.0.0. They are not operational setup instructions.

v1 readiness#

This page is the acceptance record for the published releases. The evidence below the “Latest 1.0.5 verification” heading was measured against that release and stays filed under it; what was checked for 1.1.0 has a section of its own at the end. It separates what is implemented and tested in this repository from capabilities that would require a new contract or an external provider decision. The normative policy is in v1.0 policy.

Code-backed and shipped#

Acceptance command#

Run from a clean checkout:

bash scripts/v1-readiness.sh

The gate checks repository hygiene, removal of references to the retired backend, formatting, module tidiness, vet, native build, the complete race suite, benchmark thresholds and shell entry points. It does not publish a release. The declared provider matrix is checked separately from live integrations by scripts/provider-matrix-check.sh. It proves that the shipped catalog still contains the required capability-to-provider edges and that the effective configuration parses. A live MCP probe is opt-in with ATENEA_MCP_CHECK=1; credentials, paid model calls and provider-side indexing are not fabricated by this local gate. The v1-readiness workflow exposes two opt-in jobs: ATENEA_LIVE_MCP starts an isolated service and performs the MCP bridge handshake, while ATENEA_OPENCODE_SMOKE runs the real provider/MCP matrix. Neither job runs on ordinary pull requests unless the repository variable is explicitly enabled. The same jobs can be launched manually with live_mcp=true and live_opencode=true; credentials and provider availability remain external inputs and a failed live job is recorded as an environment/provider result, not converted into a local readiness claim.

CI runs the race-enabled suite with a coverage profile and publishes the measurement and per-platform artifacts. Coverage is informational: no global percentage target, matrix floor or historical percentage regression can fail the workflow.

On 2026-08-22, Hugo 0.165.0 built docs/ locally without errors. atenea wrap opencode --version completed MCP handshakes for all 8 configured team servers: 2 were declared directly and 6 were retained as raw or without a direct surface; no external tool was invoked.

The wrapper-only handshake check repeated the handshake through atenea wrap for opencode, claude and codex: all 8 servers were reachable, with 2 declared directly and 6 held as raw or internal surfaces. The installed versions were Claude Code 2.1.228, Codex CLI 0.149.0, OpenCode 1.18.15, Serena 1.28.1, Context7 4.0.2, Semgrep 1.23.3, claude-mem 13.15.3, agent-device 0.20.10, Maestro 1.0.0, Headroom 1.29.0 and Chrome DevTools MCP 1.7.0. That handshake-only check performed no model turn, scan, write or device action; the later provider validation is recorded below.

Claude Code and Codex remain optional provider surfaces. The earlier minimal code.search for CapabilityIndex passed through each wrapper within the configured 0.25 USD ceiling: Claude reported 0.210193 USD and Codex did not report monetary usage. In the latest external-target recheck, Claude Code reached its spending ceiling with 0.310367 USD observed usage. Codex later completed an isolated external search in 81.8s; its effective timeout is now 120s. The MCP health display remains on-demand by design; a handshake proves reachability, not the behavior of every raw tool.

The same day, direct calls through the official MCP transport passed against Serena, Context7 4.0.2, Semgrep 1.23.3, claude-mem 13.15.3, agent-device 0.20.10, Maestro 1.0.0, Headroom 1.29.0 and Chrome DevTools MCP 1.7.0. Read-only and diagnostic tools were exercised; device, browser interaction, write and destructive actions were excluded. On the external TypeScript target, claude-mem answered but its structural search returned no symbols and its outline could not parse server.ts.

Maintenance recheck#

The maintenance recheck on 2026-08-22 replaced timing sleeps in the workflow resume tests with process-start markers. TMPDIR=/tmp go test -race -count=5 ./internal/workflow passed all five repetitions. The MCP safety harness also passed in normal and race mode across internal/core, internal/passthrough, internal/mcpstdio and internal/mcpprobe; destructive write, process and external fixtures were denied before the backend and receipts remained available for refused calls.

The configured claude-mem process (13.15.3) was queried read-only: its handshake and tools/list exposed 14 tools, list_corpora returned [], the memory search returned no observations, the backend TypeScript tree scanned 101 files but produced no structural symbols, and src/server.ts could not be parsed. No memory, corpus or session was written.

The opt-in OpenCode matrix passed all 6 cases with version 1.18.15: three free models without MCP and the same three with the isolated Atenea MCP bridge. All six reported cost_usd=0.000000; the MCP cases called atenea_catalog_repositories. Claude Code was not called again because its previous 0.310367 USD observation already demonstrated the provider overspend; repeating it would add cost without new evidence.

Kivgraph 0.3.4 was then configured with Atenea as a fifth repository and indexed successfully as generation 000009. Atenea’s real service detected the index as ready; graph.status, symbol.definition, symbol.references, symbol.overview, symbol.consumers, symbol.get and symbol.unresolved passed through kivgraph against the Atenea repository. The adapter also accepts the installed compact groups[].files[].at outline format. The local binary was rebuilt with LadybugDB v0.13.1 and now publishes get_unresolved_references. The Kivgraph registry excludes docs/, which is an independent Go module rather than part of the root graph.

The Kivgraph adapter also provides code.impact: it parses a Git baseline diff, resolves current declaration roots with get_file_outline, and walks incoming consumers with get_blast_radius. repository.index runs the official kivgraph index --full --json command only with explicit write and process permission, then verifies the published counters through graph_status for readiness while returning the authoritative counters from the index command’s final JSONL result. Both paths have unit coverage; impact intentionally omits foreign repositories because its contract returns repository-relative paths without a repository field.

The final workspace recheck also ran the official index command from the TaxiPrime workspace root. It completed as generation 000011 with 5 logical repositories, 656 files, 19,885 symbols, 72,034 edges and 5,080 unresolved references. taxiprime-root remains intentionally absent from the published graph: it is the aggregate workspace root, while Kivgraph registers the five logical repositories beneath it. Atenea therefore keeps indexed_by=[] for that aggregate target; marking it as ready would misrepresent the provider’s repository model. The five logical repositories remain ready and provider backed.

Tokensave 7.10.0 was installed from the official Homebrew tap, initialized for /Users/gtrave/Documents/atenea, and configured as an on-demand runner. Its live MCP server reported 11,586 nodes, 27,170 status edges and 311 files; code.context, symbol.overview and symbol.calls passed through Atenea. The anonymous upload counter was disabled. Large tokensave_entities answers can be truncated by the upstream server; Atenea detects that case and retries with the official kinds filter, then merges and deduplicates the bounded responses. A real large-file overview returned 45 symbols successfully.

Latest 1.0.5 verification#

On 2026-08-23 the complete current snapshot was applied to an isolated temporary clone and committed only inside that clone so the clean-tree gate could run. bash scripts/v1-readiness.sh passed all nine stages, including go test -race -count=1 ./..., policy anchors and the 1.0.5/3.3.0 build identity. The release was subsequently published by the release workflow.

The end-to-end test internal/agent/review_integration_test.go launched the real agent-exec reviewer through internal/agent.Runner; it accepted two content-checked citations, validated the strict result schema and confirmed the reviewer trace points to the work run.

The existing public release was also checked with:

bash scripts/release-smoke.sh 1.0.5

Checksum verification, install, same-version update, rollback and uninstall passed on macOS arm64. This validates the public 1.0.5 release lifecycle; the smoke test itself does not publish a release.

The current phase was rechecked on 2026-08-23. TMPDIR=/tmp go test ./..., TMPDIR=/tmp go test -race -count=1 ./..., go vet ./..., go build ./..., the policy gate and the 1.0.4 release smoke passed. The timing margins in the Claude/OpenCode fixtures now tolerate loaded machines and the race detector while remaining below the child-process lifetime.

The same recheck ran the complete 13-capability matrix against the active team overlay. All 13 returned ok, including code.impact and repository.index with explicit test/index permissions, plus symbol.unresolved through Kivgraph, a large-file symbol.overview through Tokensave, and symbol.calls through a real indexed symbol. Serena and Kivgraph alternative implementations also returned ok; atenea detect --repo atenea reached all 8 configured MCP servers. Hugo 0.165.0 built the documentation site successfully. The two minimal Claude/Codex searches used the configured budget; the OpenCode matrix used the configured free models. The release workflow and post-release smoke also passed.

The team service is now installed from /Users/gtrave/.local/bin/atenea rather than a Go build-cache path. Serena runs persistently for six repositories with --open-web-dashboard False; Headroom serves 127.0.0.1:8787/dashboard with memory disabled, and Maestro’s Viewer serves 127.0.0.1:8765 through a launchd wrapper using OpenJDK 21. None of these launch agents opens a browser automatically.

Latest real-provider smoke#

On 2026-08-21, scripts/opencode-smoke.sh passed against the installed OpenCode 1.18.15 CLI and the opencode/hy3-free model. The run completed a real --format json --pure turn in 80.03 seconds, produced text plus a structured { "ok": true } answer, and passed Atenea’s local schema check. It supplied no Atenea MCP tools and did not use --auto.

This is evidence for one real model and one local CLI version, not a guarantee for every OpenCode provider or future release. The smoke remains opt-in and may create the provider client’s normal local session state.

Provider, MCP and permission matrix#

On 2026-08-22, scripts/opencode-matrix.sh ran the installed OpenCode 1.18.15 twice per model: once without MCP and once with an ephemeral Atenea service. The free models below passed both turns, returned the requested local JSON object and reported zero provider cost across two OpenCode providers:

The MCP half used the actual atenea mcp bridge and an isolated state/socket directory; the service announced 13 capabilities. The runner recorded tool_use events instead of trusting the model’s prose. The matrix and the unit contract both verify the safe headless arguments --format json --pure and the absence of --auto; no paid model was used. This is evidence for three free models across two providers on one CLI version, not provider-wide compatibility.

On 2026-08-24, the manually dispatched readiness workflow passed the general gate, all four native builds and the live MCP bridge. Its first OpenCode run passed opencode/mimo-v2.5-free, while opencode/hy3-free and opencode-go/ox-alpha-free returned an external Unexpected server error. The targeted retry passed hy3-free and mimo-v2.5-free and reproduced the same error only for ox-alpha-free. This keeps the live job red as an honest provider-availability signal; it does not indicate a failure in Atenea’s MCP bridge or local readiness gate.

External target recheck — 2026-08-22#

The optional MCPs were rechecked against the TaxiPrime workspace and a public documentation target using read-only or diagnostic calls. Serena, Kivgraph, Context7, Semgrep, agent-device, Maestro, Headroom and Chrome DevTools answered real calls. Tokensave refused taxiprime-backend because its configured root is Atenea, which is the expected scope boundary. claude-mem answered but its structural search found no symbols in the TypeScript target and its outline could not parse server.ts.

The optional model providers were also exercised again against taxiprime-backend with the existing 0.25 USD budget. Claude Code was authenticated but stopped at its spending ceiling; the provider reported 0.310367 USD observed usage, including 0.060367 USD above the requested ceiling, and returned no search result. The adapter now also rejects a success envelope whose reported cost exceeds the permission, while retaining the observed charge. Codex completed the diagnostic search in 81.8s with the timeout override. Neither result changes the core readiness gate. The external calls did not index repositories, launch apps, open a persistent browser, or perform mutating device/MCP operations.

The readiness script now defaults its temporary test state to /tmp on macOS to stay below the kernel’s UNIX-socket path limit; an explicit ATENEA_TEST_TMPDIR may be supplied when a different temporary root is required.

The final local provider smoke used the installed Claude Code 2.1.228 and Codex CLI 0.149.0 against disposable repositories. Claude Code exposed two streaming cost events and was canceled by the observer at the test ceiling after reporting 0.055262 USD. Codex initially exposed an invalid nested JSON-schema error; the adapter now closes nested objects with additionalProperties: false, and the follow-up smoke completed with ok, one match, 40,558 tokens and one provider observation. This validates the real Codex completion path as well as the adapter’s streaming event path.

1.1.0 verification#

On 2026-08-25 the audit remediation was merged to main as fifteen commits and this is what was run against the merged tree, on this machine unless the run is named as a CI one:

• bash scripts/v1-readiness.sh passed all nine stages, including go test -race -count=1 ./..., the policy anchors and the 1.1.0/3.3.0 build identity.
• CI passed all seven jobs on main: four native builds, the lint job on both ubuntu and macOS, and the widget parser.
• Coverage was 81.3% on ubuntu and 81.4% on macOS. Those values were reported as measurements, not used as merge gates.
• The three fuzz targets ran in CI for thirty seconds each and found nothing.
• golangci-lint v2.12.2 reported zero issues on both platforms.

What was NOT re-run for this release: no live MCP probe, no paid provider call, no OpenCode matrix, and no provider-side indexing. Those are the opt-in jobs and they stay opt-in; the observations recorded above them in this page are still the ones from 1.0.5 and are dated as such.

Two changes in this release can stop an existing installation from starting, and both are stated in the changelog: an [[agent]] block that omits max_tokens is now refused rather than silently given no ceiling, and a service whose catalog declares a repository by a relative path is refused rather than resolving it against whatever directory the unit file left it in.

Current post-Kivgraph audit — 2026-08-29#

The tables and measurements above are retained as historical snapshots. This dated entry is the active status after the Kivgraph 0.9.2 integration:

Deliberately deferred#

These are not hidden failures in the current product. They are explicit v1.0 decisions or later contracts:

• an implicit interactive permission prompt: no adapter and no daemon path stops mid-run to ask. A service under systemd or launchd has no terminal to be asked on, so such a prompt would hang the run or wait for nobody; the security model stays explicit grant/refusal through policy and --allow. The explicit confirmation is not deferred and already ships: --confirm on task, ask, decide --run and agent prints the budget and the effects about to be granted and refuses outright without a TTY, agent demands it for any type declaring write or external effects, and backup discard takes no other word for an answer;
• exact OpenCode parity with Claude Code: the opt-in provider adapter is hardened, maps common provider errors and checks reported cost against the requested budget, but OpenCode still lacks a native schema flag, a provider hard cost cap and a permanently reliable terminal event guarantee;
• exact provider-side hard per-turn token enforcement: Atenea now carries limits.max_tokens into model turns, stops contained streams as soon as an observed over-limit event arrives and rejects the result locally, but supported external providers still do not expose one uniform mid-turn hard cap, so provider work already in flight can overshoot;
• default semantic verification of narrative claims: the opt-in semantic-reviewer can judge whether a conclusion follows from its evidence, but it is model-backed, costs the configured explore budget and may return indeterminate; deterministic citation verification remains the acceptance gate. Renamed paths with a different basename remain unresolved until the answer supplies the current path.

The historical design notes remain in What is not built yet. This page is the current acceptance record and should be updated when one of the deferred contracts is actually chosen and implemented.

Dashboard contract#

MCP declarations may include an optional dashboard HTTP(S) URL. Atenea validates the URL and derives a DNS-safe alias from the MCP id. Dashboards are reported by status and detect, but starting or probing an MCP never opens a browser. Use atenea dashboard <id> for an explicit accessibility check and browser launch, or atenea dashboard <id> --check to keep the operation headless. atenea dashboard hosts --dry-run previews the managed hosts block; the real /etc/hosts file is changed only by the explicit hosts command. Serena is discovered by the active project because each persistent repository instance receives a dynamic dashboard port; its instances are warmed sequentially to avoid dashboard port collisions.