docs: add status-values.md and update README status input (#16)

* docs: add status-values.md, update README status input description

Made-with: Cursor

* fix: include cache writes in cache hit rate denominator

cache_writes are tokens processed fresh to populate the cache and
belong in the denominator alongside input_tokens. Previous formula
ignored them, inflating the hit rate (e.g. 100% instead of ~69%).

New formula: reads / (reads + writes + input)

Made-with: Cursor

* docs: add product screenshots to README

Made-with: Cursor

* docs: add product screenshot images

Made-with: Cursor

* chore: remove author

Author: adamhenson

README.md

@@ -20,6 +20,20 @@ Add this action after your AI agent step to:
 
 The action **never fails your workflow** — all API calls and comment posts use `core.warning()` for errors, not `core.setFailed()`.
 
+<p align="center">
+  <img src="images/screenshot-agentmeter-github-comment.png" alt="AgentMeter GitHub PR comment showing cost summary" width="700" />
+  <br/><sub>AgentMeter posts a cost summary directly on the PR.</sub>
+</p>
+
+<p align="center">
+  <img src="images/screenshot-agentmeter-run-detail.png" alt="Run detail with token breakdown" width="48%" />
+  &nbsp;
+  <img src="images/screenshot-agentmeter-runs.png" alt="Runs feed dashboard" width="48%" />
+</p>
+<p align="center">
+  <sub>Full token breakdown per run &nbsp;·&nbsp; All runs in one dashboard</sub>
+</p>
+
 ---
 
 ## Quickstart
@@ -281,7 +295,7 @@ Replace `$INPUT_TOKENS` etc. with however your agent exposes token counts (step
 | `api_key` | ✅ | — | Your AgentMeter API key (`am_sk_…`). Get it from [agentmeter.app/dashboard/settings](https://agentmeter.app/dashboard/settings). |
 | `model` | ❌ | `''` | The AI model used (e.g. `claude-sonnet-4-5`). Used for per-token cost display. |
 | `engine` | ❌ | `claude` | The AI engine (`claude`, `codex`). |
-| `status` | ❌ | `success` | Run status: `success`, `failed`, `timed_out`, `cancelled`, `needs_human`. |
+| `status` | ❌ | `success` | Run outcome. In companion `workflow_run` mode this is resolved automatically from the triggering workflow's conclusion. In inline mode pass `${{ steps.agent.outcome }}` or a custom value like `needs_human`. See [docs/status-values.md](docs/status-values.md). |
 | `agent_output` | ❌ | `''` | Raw stdout from the agent step. Used to auto-extract token counts from JSON. |
 | `input_tokens` | ❌ | `''` | Explicit input token count. Overrides extraction from `agent_output`. |
 | `output_tokens` | ❌ | `''` | Explicit output token count. |

__tests__/comment.test.ts

@@ -163,6 +163,43 @@ describe('buildCommentBody', () => {
     expect(body).toContain('approximate');
   });
 
+  it('calculates cache hit rate using reads / (reads + writes + input)', () => {
+    const body = buildCommentBody({
+      apiPricing: testPricing,
+      existingBody: null,
+      runData: {
+        ...baseRun,
+        tokens: {
+          inputTokens: 50,
+          outputTokens: 2172,
+          cacheWriteTokens: 55569,
+          cacheReadTokens: 124794,
+          isApproximate: false,
+        },
+      },
+    });
+    // 124794 / (50 + 55569 + 124794) = 124794 / 180413 ≈ 69%
+    expect(body).toContain('69% cache hit rate');
+  });
+
+  it('does not show cache hit rate when cacheReadTokens is 0', () => {
+    const body = buildCommentBody({
+      apiPricing: testPricing,
+      existingBody: null,
+      runData: {
+        ...baseRun,
+        tokens: {
+          inputTokens: 1000,
+          outputTokens: 500,
+          cacheWriteTokens: 0,
+          cacheReadTokens: 0,
+          isApproximate: false,
+        },
+      },
+    });
+    expect(body).not.toContain('cache hit rate');
+  });
+
   it('skips token details when tokens are not provided', () => {
     const body = buildCommentBody({
       apiPricing: testPricing,

action.yml

@@ -1,6 +1,5 @@
 name: 'AgentMeter'
 description: 'Track token usage and cost for AI agent runs in GitHub Actions'
-author: 'Foo.software'
 branding:
   icon: 'zap'
   color: 'green'

dist/index.js

@@ -0,0 +1,89 @@
+# Status Values
+
+Documents how the `status` input works in the AgentMeter Action and what values are valid.
+
+---
+
+## How status is determined
+
+### Companion workflow mode (`workflow_run` trigger)
+
+When `workflow_run_id` is set, status is resolved **automatically** from the GitHub Actions conclusion of the triggering workflow. The user does not need to pass `status` — the action reads it from the `workflow_run` event payload and normalizes it internally via `normalizeConclusion()`.
+
+```yaml
+- uses: foo-software/agentmeter-action@main
+  with:
+    api_key: ${{ secrets.AGENTMETER_API_KEY }}
+    workflow_run_id: ${{ github.event.workflow_run.id }}
+    # status is resolved automatically — no need to set it
+```
+
+### Inline mode (direct / same-workflow)
+
+When running inline (no `workflow_run_id`), the user passes `status` explicitly. It defaults to `'success'` if omitted.
+
+```yaml
+- uses: foo-software/agentmeter-action@main
+  if: always()
+  with:
+    api_key: ${{ secrets.AGENTMETER_API_KEY }}
+    status: ${{ steps.agent.outcome }}
+```
+
+---
+
+## Built-in GitHub conclusion mappings
+
+`normalizeConclusion()` maps GitHub's standard conclusion strings to AgentMeter's internal status enum:
+
+| GitHub conclusion | AgentMeter status | Notes |
+|---|---|---|
+| `success` | `success` | |
+| `failure` | `failed` | |
+| `timed_out` | `timed_out` | |
+| `cancelled` | `cancelled` | |
+| `skipped` | *(not ingested)* | Run is skipped entirely — nothing is sent to the API |
+
+**Source of truth:** `normalizeConclusion()` in `src/workflow-run.ts`.
+
+---
+
+## Custom statuses
+
+Any value **not** in the mapping table above is passed through to the API unchanged. This is intentional — unrecognized values are preserved so custom statuses are not silently replaced with `failed`.
+
+### `needs_human`
+
+The primary custom status. Use it when an agent run completes but requires human review before the result can be acted on (e.g. low-confidence output, a tool call was blocked, or the agent explicitly flagged escalation).
+
+**Example — conditionally set based on an agent output flag:**
+
+```yaml
+- uses: anthropics/claude-code-action@v1
+  id: agent
+  with:
+    anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
+    prompt: "..."
+
+- uses: foo-software/agentmeter-action@main
+  if: always()
+  with:
+    api_key: ${{ secrets.AGENTMETER_API_KEY }}
+    model: claude-sonnet-4-5
+    status: ${{ steps.agent.outputs.needs_human == 'true' && 'needs_human' || job.status }}
+```
+
+When `steps.agent.outputs.needs_human` is `'true'`, the run is recorded as `needs_human`. Otherwise it falls back to the job's actual status (`success` or `failure`).
+
+---
+
+## All valid AgentMeter status values
+
+| Value | Set via action | Description |
+|---|---|---|
+| `success` | ✅ | Agent run completed successfully |
+| `failed` | ✅ | Agent run failed |
+| `timed_out` | ✅ | Agent run exceeded its time limit |
+| `cancelled` | ✅ | Agent run was cancelled |
+| `needs_human` | ✅ | Run completed but requires human review |
+| `running` | ❌ internal only | Run is currently in progress — not settable via action |

docs/status-values.md

@@ -155,10 +155,11 @@ function buildTokenDetails({
   const { tokens, model, turns } = run;
   if (!tokens) return null;
 
+  // Cache hit rate = reads / (reads + writes + input). Cache writes are tokens processed
+  // fresh to populate the cache — they belong in the denominator alongside input tokens.
+  const totalPromptTokens = tokens.cacheReadTokens + tokens.cacheWriteTokens + tokens.inputTokens;
   const cacheHitRate =
-    tokens.cacheReadTokens + tokens.inputTokens > 0
-      ? Math.round((tokens.cacheReadTokens / (tokens.cacheReadTokens + tokens.inputTokens)) * 100)
-      : 0;
+    totalPromptTokens > 0 ? Math.round((tokens.cacheReadTokens / totalPromptTokens) * 100) : 0;
 
   const pricing = getPricing({ apiPricing, model });
   const perM = (count: number, pricePerM: number | null | undefined): string => {