|
| 1 | +# Qualify a locally configured ACP agent |
| 2 | + |
| 3 | +Use one local JSON configuration for qualification and ordinary orchestration. |
| 4 | +The command reaches XMD's production ACPX provider as a literal command; no agent |
| 5 | +registry edit is needed. Another agent changes configuration rather than copying |
| 6 | +this workflow. |
| 7 | + |
| 8 | +Prepare this checkout with `deno task setup`. Install and authenticate your agent |
| 9 | +separately. Copy `kimi.example.json` outside this checkout and edit its command |
| 10 | +and expectations. Keep credentials in the agent's existing local authentication; |
| 11 | +the configuration accepts no credentials or environment map. ACPX's raw command |
| 12 | +resolution has platform limits on Windows; this example adds no launcher or |
| 13 | +registry to change them. |
| 14 | + |
| 15 | +## Authorize one bounded run |
| 16 | + |
| 17 | +Selecting a scenario authorizes its model turns and cancellation of the suite's |
| 18 | +own test turn. Live execution is opt-in and may cost money. Use a fresh private |
| 19 | +working directory. The runner generates unique local A/B/C Session names. The maximum is six Prompts, |
| 20 | +one attempt each: seed A, recall A, independent B, reconnect A, cancel C, then |
| 21 | +recall A from another OS invocation. Turns have 60 seconds; the cancellation |
| 22 | +request targets the still-active C turn after five seconds. Teardown observation |
| 23 | +has 30 seconds and the outer work deadline has 600 seconds. Expiry ends admission; |
| 24 | +the actual owner still joins. No uncertain turn is retried. |
| 25 | + |
| 26 | +The runner computes setup dependencies before provider work. Continuity, |
| 27 | +separation, reconnect and resume require one completed A seed. Only selected |
| 28 | +continuity adds A's second turn; separation adds independent B. Cancellation |
| 29 | +needs no marker scenario. Selecting reconnect or resume authorizes their seed |
| 30 | +setup, even when the setup capability rows are unselected. |
| 31 | + |
| 32 | +The checked-in initial document keeps one provider and A Session alive for the |
| 33 | +seed and recall, then tests B. Reconnect opens a new same-process invocation. |
| 34 | +Cancellation owns a separate document; its five-second interval begins at actual |
| 35 | +transport-confirmed start and structured cancellation joins that document. Resume uses one |
| 36 | +separate CLI invocation and ordinary retained A history. Its private comparison |
| 37 | +file is neither provider history nor continuation authority. |
| 38 | + |
| 39 | +Ordinary production validation, including its cached availability probe, remains |
| 40 | +in force. After successful owner settlement independent stages can proceed even |
| 41 | +though exact native closing is unverified. Pending joins, actual teardown failure |
| 42 | +or production ownership refusal stop admission. A marker failure preserves earlier |
| 43 | +basic/settings evidence and permits independent cancellation. No capability passes |
| 44 | +from configuration, availability, process exit or matching display names. |
| 45 | + |
| 46 | +From the private working directory, use absolute paths for the checked-in document, |
| 47 | +component directory, local configuration and report: |
| 48 | + |
| 49 | +```sh |
| 50 | +xmd run /checkout/examples/agent-certification/certify.md --include /checkout/examples/agent-certification/components --props-config /private/local-agent.json --props-report /private/capabilities.json --deny-all --timeout 600s --raw |
| 51 | +``` |
| 52 | + |
| 53 | +The same command through this source checkout is `deno task xmd run ...`. |
| 54 | +Ordinary CI exercises controlled agents without credentials or model calls. |
| 55 | + |
| 56 | +## Read the report |
| 57 | + |
| 58 | +The current JSON format is closed and unversioned. Each of thirteen capability rows |
| 59 | +records selected and expected flags, an outcome, a fixed reason and bounded facts. |
| 60 | +`pass`, `failure`, `unsupported` and `not_run` are distinct. An absent selector is |
| 61 | +unsupported; a selection the target Session does not apply is a failure. Omitted |
| 62 | +model/effort settings supply no explicit-selection evidence. |
| 63 | + |
| 64 | +`complete=false` includes missing provenance and selected evidence blocked by |
| 65 | +prerequisites. A measured unsupported capability is a non-pass. Overall success |
| 66 | +requires complete evidence and all selected/expected capabilities passing. |
| 67 | +Acceptance and native identity are separate rows: a standard command can pass |
| 68 | +basic authentication, streaming and completion while acceptance is |
| 69 | +`unsupported/acceptance-unavailable` and identity is unsupported. An actual native |
| 70 | +assertion can pass identity independently. Only a strong-declared adapter's gated |
| 71 | +started event proves acceptance; marker-like output from a literal command does |
| 72 | +not. Durable certification additionally requires acceptance, identity, continuity, separation, reconnect, |
| 73 | +separate-invocation resume and verified cleanup. Native cleanup remains |
| 74 | +`not_run/cleanup-unverified` after a successful owner join; actual observed cleanup |
| 75 | +failure is failure. Correct stop attribution alone cannot pass cancellation. |
| 76 | + |
| 77 | +Only categorical values, booleans, counts, strict version tokens, fixed platform |
| 78 | +facts and fingerprints enter this report. Missing executable/adapter/runtime |
| 79 | +provenance remains null and prevents complete evidence. Configuration and source |
| 80 | +fingerprints identify exact noncredential inputs; no digest is invented. The |
| 81 | +report contains no prompts, replies, private marker, credential, environment |
| 82 | +value, path, command, raw identifier, exception or provider output. |
| 83 | + |
| 84 | +Your agent can retain private native history under its normal policy. XMD's |
| 85 | +optional diagnostic journal is also private and is not this report or continuation |
| 86 | +input. Neither certification nor a report grants durable identity or tools. |
| 87 | + |
| 88 | +## Use the same local configuration |
| 89 | + |
| 90 | +Ordinary orchestration uses the same command and exact optional model/effort IDs: |
| 91 | + |
| 92 | +```sh |
| 93 | +xmd run /checkout/examples/agent-certification/orchestrate.md --include /checkout/examples/agent-certification/components --props-config /private/local-agent.json --props-prompt 'Summarize this task without tools.' --deny-all --raw |
| 94 | +``` |
| 95 | + |
| 96 | +This is a separate explicitly requested model call, outside the qualification |
| 97 | +suite. Ordinary name/cwd retention is not workflow Session-site identity. |
| 98 | +A literal command uses standard ACP without the private acceptance barrier. |
| 99 | +Sequential prompts share the live Session; later reopening loads the exact ACP |
| 100 | +identity if supported and refuses otherwise. Interrupted requests are never |
| 101 | +automatically resent. Strong-declared adapters retain their acceptance barrier. |
| 102 | + |
| 103 | +Kimi's selected live scope uses `kimi acp`, provider-default model and effort, |
| 104 | +and no retries. A later useful live run requires changed-source review, exact fingerprints and |
| 105 | +a fresh explicit selected-scenario grant. The earlier zero-Prompt attempt is |
| 106 | +retained invocation evidence only; it establishes no capability battery. Gemini was absent on PATH in prepared observations; its sample command is a placeholder for the installed adapter's documented command and is |
| 107 | +configuration syntax, not availability or certification evidence. This delivery |
| 108 | +authorizes zero Gemini live calls and installs no agent. |
| 109 | + |
| 110 | +The terminal matrix always lists all thirteen capabilities with selected/expected flags, |
| 111 | +outcomes and fixed reasons, plus observed Prompt/cancellation counts and overall |
| 112 | +flags. Failure preserves earlier independent rows and a meaningful nonzero status. |
| 113 | +Private markers, replies, Session identities and child fragments remain local. |
| 114 | + |
| 115 | +Prompt counts describe turns observed through the production stream. A refusal |
| 116 | +before stream disclosure can leave native submissions unobserved; these counts |
| 117 | +are not billing evidence. Resume follows its own CLI work deadline and the parent |
| 118 | +waits for actual process settlement even if outer work is cancelled. |
| 119 | + |
| 120 | +The 30-second teardown observation begins when an element's own work exits, |
| 121 | +before its resources close. Expiry permanently ends later admission even if |
| 122 | +cleanup subsequently succeeds; the runner still waits for the real join. The |
| 123 | +same observation runs inside the resume child. Demonstrated invocation teardown |
| 124 | +failure remains failure/cleanup-failed after later successful joins. |
0 commit comments