This proposal is beyond reproach

This seems like a great idea

Adapter-level live evidence for #19958, head 6f5852b190c5, glm-acp-agent 1.14.0.

Scope, stated plainly: this run drives makeAcpRegistryAdapterV2 on the PR branch against the real installed, authenticated agent (Z.ai), capturing the raw PromptResponse.usage off the wire (via the adapter's protocol-logging hook, allowlisted to the six counters) side by side with the turnTokenUsage the adapter records on each provider turn. It does not go through orchestration persistence or any client surface — the harness iterated over ~10 runs while being built; this is the final one, and no earlier run failed a usage invariant (failures were harness wiring only).

GLM per-turn usage evidence (PR #19958)

Provenance

• head: 6f5852b
• gitStatus: clean
• agentId: glm-acp-agent
• agentVersionDir: 1.14.0
• agentResolvedVersion: 1.14.0
• agentDist: index.js
• agentEntrypointSha256: d0d9237bffcbd5017a0085aa3c4625c29e2b750f87c358fa9c455d3cd4cdbd98
• agentDistTreeSha256: 8e5a7b2ca91e0f583cf26102cbc3693fcc2a2576155c2e147a0f1cdc1519e386
• authSource: agent credentials file present; not read by the harness
• harness: AcpLiveUsageEvidence.test.ts (untracked) via makeAcpRegistryAdapterV2
• isolation: T3 home isolated (temp baseDir, network disabled); agent state redirected via XDG_STATE_HOME
• capturedAt: 2026-10-20T17:34:35.063Z
• tempToolsManifestUnchanged: yes
• liveAgentManifestUnchanged: yes

Wire samples (raw PromptResponse.usage, allowlisted counters)

Recorded turnTokenUsage

Scenario log

[
  {
    "scenario": "S1",
    "prompts": 4,
    "cachedReadPositive": true,
    "thoughtPositive": true,
    "inputSeries": [
      2857,
      3839,
      4196,
      4553
    ],
    "outputSeries": [
      629,
      4,
      4,
      4
    ],
    "cachedReadSeries": [
      1280,
      2816,
      3776,
      4160
    ],
    "thoughtSeries": [
      337,
      0,
      0,
      0
    ]
  },
  {
    "scenario": "S2",
    "delayMs": 40,
    "interruptAccepted": true,
    "wireSample": null,
    "stopReason": "cancelled",
    "recordedStatus": "unavailable",
    "recordedStatusDetail": "interrupted",
    "recordedCounters": null
  },
  {
    "scenario": "S2",
    "delayMs": 150,
    "interruptAccepted": true,
    "wireSample": null,
    "stopReason": "cancelled",
    "recordedStatus": "unavailable",
    "recordedStatusDetail": "interrupted",
    "recordedCounters": null
  },
  {
    "scenario": "S2",
    "delayMs": 600,
    "interruptAccepted": true,
    "wireSample": null,
    "stopReason": "cancelled",
    "recordedStatus": "unavailable",
    "recordedStatusDetail": "interrupted",
    "recordedCounters": null
  },
  {
    "scenario": "S2",
    "delayMs": 2500,
    "interruptAccepted": true,
    "wireSample": null,
    "stopReason": "cancelled",
    "recordedStatus": "unavailable",
    "recordedStatusDetail": "interrupted",
    "recordedCounters": null
  },
  {
    "scenario": "S2",
    "delayMs": "response-triggered",
    "interruptAccepted": true,
    "wireSample": {
      "inputTokens": 2847,
      "outputTokens": 1109,
      "totalTokens": 3956,
      "cachedReadTokens": 2432,
      "thoughtTokens": 394
    },
    "stopReason": "end_turn",
    "recordedStatus": "partial",
    "recordedStatusDetail": "interrupted",
    "recordedCounters": {
      "inputTokens": 2847,
      "outputTokens": 1109,
      "cachedInputTokens": 2432,
      "reasoningTokens": 394
    }
  },
  {
    "scenario": "S2",
    "delayMs": "mid-stream",
    "interruptAccepted": true,
    "wireSample": null,
    "stopReason": "cancelled",
    "recordedStatus": "unavailable",
    "recordedStatusDetail": "interrupted",
    "recordedCounters": null
  },
  {
    "scenario": "S3",
    "postResumeSample": {
      "inputTokens": 3270,
      "outputTokens": 4,
      "totalTokens": 3274,
      "cachedReadTokens": 2816,
      "thoughtTokens": 0
    },
    "postResumeRecorded": {
      "inputTokens": 3270,
      "outputTokens": 4,
      "cachedInputTokens": 2816,
      "reasoningTokens": 0
    },
    "terminal": "completed",
    "resumeMethods": [
      "session/load"
    ]
  }
]

Readings

• Per-turn, not cumulative: output goes 629 → 4 → 4 → 4 on a long answer followed by one-word replies; a session-cumulative counter would read 629, 633, 637, 641. Each turn's input ≈ previous input + previous output + ~360 (the new prompt); a cumulative input would read 2857 / 6696 / 10892 / 15445.
• Input includes cache: total === input + output exactly on every usage-bearing sample, and warm-turn input stays ≥ cachedRead throughout (e.g. turn 4: input 4553 ≥ cachedRead 4160). If input excluded cached tokens, warm turns would read below turn 1's baseline.
• Recorded equals wire: across all samples and counters, cachedInput === cachedRead, reasoning === thought, inputs/outputs identical — i.e. the PR's defensive clamps never fired on real GLM data (a clamp firing would show as a mismatch). cachedWriteTokens is absent from every GLM 1.14.0 sample (marked N/A); the read+write ≤ input bound is exercised by the PR's unit tests with synthetic samples.
• Stop outcomes (live): four timed cancels and one mid-stream cancel landed before the agent returned usage on the wire — recorded unavailable, no counters. One response-triggered Stop (interrupted the moment the response landed) recorded partial with exactly the wire sample (2847/1109/2432/394). Cancelled-with-usage was not observed live: GLM 1.14.0 only attaches usage to a cancelled stopReason after a completed model call within a multi-call (tool) prompt, which these no-tool prompts never produce; the PR's tap is stopReason-agnostic, so the observed partial exercises the same capture path, and the pure-function tests cover the mapping.
• Resume: leg 1's process was closed (scope teardown) before a fresh adapter loaded the same agent session via session/load; the post-resume turn's recorded usage equals that turn's own wire sample (3270/4/2816/0), with restored context visible in the input (~3270 vs ~2840 fresh-session baseline).

Isolation and redaction: the run used a temp T3 home (network disabled, cached registry) with the agent's own state redirected via XDG_STATE_HOME; the live ~/.t3 agent tree is verified unchanged (name+size manifest) and the dist tree is content-hashed in the provenance above. The artifact contains only counter values, statuses, stop reasons, method names, ordinals, and timestamps — no conversation content, request params, credentials, or identifiers, verified by an independent audit of every string value in the JSON.

Independent review: the approach and the results (including a recomputation of every invariant from the artifact and a string-by-string redaction audit) were reviewed by a separate Opus 5.5 pass; its corrections are incorporated above. Remaining open item for the gate's "check the resulting usage in a real client": an end-to-end dev-server + client run with the persisted provider-turn records, to follow.

Tier A: end-to-end evidence for #19958 — real client -> dev server -> orchestration -> adapter -> glm-acp-agent 1.14.0 (Z.ai), head 32727b572a.

Setup: a git worktree of the PR branch with its own gitignored .t3 (dev-runner's worktree home) seeded with the agent and registry cache plus a minimal settings file holding only the Z.ai registry instance; pnpm dev on that worktree; the web client driven end to end (paired browser session, Z.ai -> GLM-5.3 Flash). The agent self-authenticates; no credentials were read or copied. Persisted records were read back read-only (mode=ro) from .t3/userdata/statev2.sqlite, orchestration_v2_projection_provider_turns.payload_json -> $.turnTokenUsage.

Scenarios and the persisted records

Readings

• Per-turn, not cumulative: cumulative input running sums would read 31474 / 63559 / 96008 / 128834; instead each turn adds the previous answer plus a near-constant ~360 tokens of framing (611−251=360, 364−4=360, 377−4=373). Output tracks each prompt (251 -> 4 -> 4 -> 1035).
• Stop while the response settles: Stop was accepted ~14s after send, ~4s into the visible answer (after ~6s of reasoning); the prompt RPC settled 35 ms after Stop; the turn persisted interrupted with usageStatus: unavailable and no counters — conservative, never overcounting. (Wire-side stopReason was not captured in this run; the adapter-level run below shows the settled-soft -> partial-with-sample outcome, and the PR's unit tests cover response-after-finalize.)
• Resume across a restart: the server process was killed and restarted between turns 5 and 6. The server trace shows the session/load span at 18:05:24Z re-attaching the same native session id; turn 6 answered the codeword set before the restart (word described, not quoted, per the no-contents rule) and recorded as its own per-turn sample — output 67 versus a cumulative >=1361, no subtraction of pre-restart counters.
• Cache: cachedInput <= input on every sample (27520/32826 = 83.8%, 34240/37031 = 92.5%), both multiples of 64 as block prefix caching would produce; reasoning <= output on every sample.
• Cross-checks: the durable event log's terminal provider-turn.updated events carry turnTokenUsage equal to the projection rows exactly, including turn 5's unavailable. The client-received projection was also verified: fetching the client's own thread API from the driven browser session returns the six turnTokenUsage records field-identical to the DB. Plainly stated: no client surface renders this field in this PR (the usage meter reads different fields); what is verified end-to-end is receipt by the real client and persistence. The provider.turn.completed analytics event is derived from these same records (ProviderEventIngestor.ts:103-127).
• Non-claims: a restart during a turn was not exercised; the cancelled-with-usage -> partial leg is covered at adapter level and by unit tests, not end-to-end.

Provenance: head 32727b572a67e4e51638f05fae08f7ddfc2ddbb9 (worktree clean, .t3 gitignored), agent 1.14.0 npm layout (entrypoint sha256 d0d9237b...cdbd98, dist-tree content hash 8e5a7b2c...19e386 — the same binary as the adapter-level run), read-only capture against the live WAL. This completes the pairing with the adapter-level run earlier in this thread: that one proved wire-sample equality per counter; this one proves the full pipeline persists and re-serves exactly those records. Both runs were reviewed by an independent Opus pass before posting (its corrections are incorporated).

The linked implementation records per-turn ACP usage for the verified GLM flavor. Other flavors report unavailable until their counters are validated, as the proposal requested. #19958 adapter.ts

Follow-up: #32173 proposes showing these GLM turn records on the Usage page. It's scoped to glm-acp-agent, so nothing with a native reader is counted twice.