Kiro: forward per-turn token counts + session model for the pricing hedge

[realtonyyoung]

KiroUsage now reads Kiro's per-turn input_token_count/output_token_count (carried only when > 0) and the session model_id (rts_model_state.model_info) from {id}.json, and injects them into the _kcap_usage object on the turn's final assistant line.

Kiro persists these counts as 0 today — the Amazon-Q CLI discards the streaming API's TokenUsage (upstream aws/amazon-q-developer-cli#2397). So this is a dormant hedge: nothing changes now, but the server stamps a canonical $usage (and pricing/cost lights up) automatically the moment a future kiro-cli release starts recording tokens — no further code change. Credits/context% behaviour is unchanged.

Server side (normalizer $usage stamping + the credits chip) is in the paired kcap-server PR.

Tests: KiroUsageTests cover token+model capture, injection, and the zero-guard (dormant) case. Full CLI unit suite green (2240/2240); AOT publish clean.

Closes #275
Linear: AI-1196

🤖 Generated with Claude Code

[linear-code]

AI-1196

[qodo-code-review]

PR Summary by Qodo

Forward Kiro per-turn token counts and session model into _kcap_usage (dormant hedge)

✨ Enhancement 🧪 Tests 🕐 20-40 Minutes

AI Description

• Read per-turn input/output token counts and session model_id from Kiro {id}.json metadata.
• Inject token/model fields into _kcap_usage only when token counts are non-zero.
• Add unit tests covering capture, injection, and zero-count dormant behavior.

Diagram

graph TD
  A["Kiro session {id}.json"] --> B["KiroUsage.AnchorMap"] --> C[("Anchor→usage map")]
  D["Transcript line (.jsonl)"] --> E["KiroUsage.EnrichLine"] --> F["Transcript line + _kcap_usage"] --> G["Server normalizer"]
  C --> E
  subgraph Legend
    direction LR
    _file["File/Input"] ~~~ _fn["Function"] ~~~ _data[("Data")]
  end

Loading

High-Level Assessment

The following are alternative approaches to this PR:

1. Always include token/model fields (even when 0) and let server ignore

• ➕ Simpler client logic (no non-zero guard).
• ➕ Server can uniformly read fields without null checks.
• ➖ Risks treating current persisted zeros as real usage if server logic changes.
• ➖ Violates stated goal of staying fully dormant until upstream starts recording tokens.

2. Have CLI stamp canonical $usage directly (instead of _kcap_usage)

• ➕ Makes usage payload explicit and potentially reduces server coupling.
• ➕ Easier to validate end-to-end behavior in CLI outputs alone.
• ➖ Duplicates server-side normalization logic and pricing semantics in the CLI.
• ➖ Higher risk of divergence across clients; harder to roll out model/pricing changes centrally.

Recommendation: Current approach (injecting optional token/model into _kcap_usage only when counts are >0) is the safest hedge: it prevents today’s persisted zeros from being misinterpreted, while enabling the server to light up canonical $usage automatically when upstream starts populating token counts. The alternatives either increase divergence (client-side $usage stamping) or introduce risk around zero values.

Files changed (2) +112 / -17

Enhancement (1) +54 / -17

KiroUsage.csCapture token/model from metadata and conditionally inject into _kcap_usage +54/-17

Capture token/model from metadata and conditionally inject into _kcap_usage

• Extends TurnUsage to include optional input/output token counts and session model_id. Parses session_state.rts_model_state.model_info.model_id and per-turn input_token_count/output_token_count, carrying token/model only when counts are > 0. Updates JSONL enrichment to inject these fields onto the anchor assistant line alongside credits/context%.

src/Capacitor.Cli.Core/Kiro/KiroUsage.cs

Tests (1) +58 / -0

KiroUsageTests.csAdd tests for token/model forwarding hedge and zero-count guard +58/-0

Add tests for token/model forwarding hedge and zero-count guard

• Adds unit coverage ensuring non-zero token counts and session model are captured in AnchorMap and injected by EnrichLine. Verifies that zero token counts remain dormant (no token fields injected).

test/Capacitor.Cli.Tests.Unit/KiroUsageTests.cs

[qodo-code-review]

Code Review by Qodo

🐞 Bugs (0) 📘 Rule violations (0) 📎 Requirement gaps (0) 📜 Skill insights (0)

Context used

✅ Tickets: 🎫 Kiro: forward per-turn token counts + model so server can price them when available

1. _kcap_usage uses model key ✗ Dismissed 📎 Requirement gap ≡ Correctness

Description

When forwarding the session model identifier, the code writes it to _kcap_usage.model instead of
_kcap_usage.model_id. This can prevent the server from recognizing the model identifier required
to stamp canonical $usage when non-zero token counts are present.

Code

src/Capacitor.Cli.Core/Kiro/KiroUsage.cs[R115-117]

+            if (u.InputTokens is { } inTok)       usage["input_token_count"]        = inTok;
+            if (u.OutputTokens is { } outTok)     usage["output_token_count"]       = outTok;
+            if (u.Model is { Length: > 0 } model) usage["model"]                    = model;

Evidence

PR Compliance ID 1 explicitly requires forwarding the session model_id in _kcap_usage alongside
non-zero token counts. The added enrichment code injects the model under the key model, not
model_id, so the required field name is not present on _kcap_usage.

Forward Kiro per-turn token usage counts and session model_id in _kcap_usage (only when counts > 0)
src/Capacitor.Cli.Core/Kiro/KiroUsage.cs[115-117]

Agent prompt

The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
The compliance requirement expects the session `model_id` to be included in `_kcap_usage` when forwarding non-zero token counts, but the implementation writes the model identifier under the key `model`.

## Issue Context
PR Compliance ID 1 requires forwarding Kiro per-turn `input_token_count`/`output_token_count` (only when > 0) and the session `model_id` in `_kcap_usage`. The current code injects `usage["model"] = model` rather than `usage["model_id"] = modelId`.

## Fix Focus Areas
- src/Capacitor.Cli.Core/Kiro/KiroUsage.cs[115-117]
- src/Capacitor.Cli.Core/Kiro/KiroUsage.cs[24-29]
- test/Capacitor.Cli.Tests.Unit/KiroUsageTests.cs[78-100]

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools

---

2. Model parse aborts usage ✓ Resolved 🐞 Bug ☼ Reliability

Description

KiroUsage.AnchorMap reads session_state.rts_model_state.model_info.model_id via GetValue<string>(),
which can throw if model_id is present but not a JSON string; the outer catch then aborts parsing
and returns an empty/partial anchor map. This can silently drop credits/context% enrichment during
Kiro import for malformed/variant metadata blobs.

Code

src/Capacitor.Cli.Core/Kiro/KiroUsage.cs[R46-50]

+            // Session-level model id (e.g. "minimax-m2.5", "auto"). Rides alongside the
+            // token counts so the server can stamp a canonical $usage with a model —
+            // AccumulateTokens drops entries with no model (HasModel guard).
+            var sessionModel = root?["session_state"]?["rts_model_state"]?["model_info"]?["model_id"]
+                ?.GetValue<string>();

Evidence

The session model is extracted with GetValue<string>() inside the same broad try/catch that
guards the entire AnchorMap parse; an exception at this point short-circuits the rest of the
per-turn usage loop. Kiro import depends on AnchorMap to enrich transcript lines with
credits/context%, so aborting AnchorMap drops that enrichment.

src/Capacitor.Cli.Core/Kiro/KiroUsage.cs[40-86]
src/Capacitor.Cli/Commands/KiroImportSource.cs[253-268]

Agent prompt

The issue below was found during a code review. Follow the provided context and guidance below and implement a solution

## Issue description
`KiroUsage.AnchorMap` uses `GetValue<string>()` for `model_id`. If the JSON contains `model_id` but not as a string (e.g., `null`, number, object), `GetValue<string>()` can throw, and because the method wraps the entire extraction in one broad `try/catch`, this can abort anchor-map construction and drop otherwise usable turn usage data.

## Issue Context
Usage enrichment is explicitly best-effort and should not be disabled by a single malformed optional field. Today the new `model_id` extraction is session-level, but an exception there prevents credits/context% mapping for the whole session metadata.

## Fix Focus Areas
- src/Capacitor.Cli.Core/Kiro/KiroUsage.cs[40-52]
- src/Capacitor.Cli.Core/Kiro/KiroUsage.cs[84-93]

## Suggested change
- Replace `?.GetValue<string>()` with a defensive pattern:
 - check `is JsonValue` and use `TryGetValue<string>(out var s)`
 - optionally apply `string.IsNullOrWhiteSpace` normalization
- Alternatively, isolate model parsing in its own `try/catch` so a bad `model_id` cannot abort turn parsing.

ⓘ Copy this prompt and use it to remediate the issue with your preferred AI generation tools

---