Added compact for SHELL.md and removed HEARTBEAT OK.

Author: gtapps

.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-code-hermit",
-  "version": "0.1.0",
+  "version": "0.1.1",
   "description": "A personal assistant that lives in your project — memory-driven learning, daily rhythm, idle agency, and operational hygiene for Claude Code",
   "author": {
     "name": "gtapps"

CHANGELOG.md

@@ -1,5 +1,17 @@
 # Changelog
 
+## [0.1.1] - 2026-03-30
+
+### Fixed
+
+- **Cost tracker now captures real token data** — The Stop hook payload never included token fields; `cost-tracker.js` was silently writing zero-cost entries (or on newer installs, not writing at all). It now reads the last assistant turn's `usage` from `transcript_path` in the hook payload, which is where Claude Code actually exposes token counts.
+- **Cache token pricing** — Added separate pricing rates for `cache_creation_input_tokens` (write) and `cache_read_input_tokens` (read) for all models. Previously cache tokens were ignored entirely, understating cost for heavily-cached sessions.
+- **Efficient transcript tail-read** — Instead of loading the full transcript file, reads only the last 128KB using a positioned `fs.readSync` call. Avoids unnecessary memory use on long sessions.
+
+### No manual upgrade steps required.
+
+---
+
 ## [0.1.0] - 2026-03-30
 
 ### Added

agents/session-mgr.md

@@ -86,6 +86,9 @@ When the main session requests an idle transition (not a full close):
    - Clear `## Plan` table (reset to the 3-row placeholder from the template)
    - Clear `## Progress Log`, `## Blockers`, `## Findings`, `## Changed` (all task-specific — already preserved in the archived report)
    - Preserve `## Monitoring`, `## Cost`, `## Session Summary` (session-scoped, accumulates across tasks)
+   - **Compact preserved sections if over threshold.** Read `compact` settings from `.claude-code-hermit/config.json`:
+     - **Monitoring:** Count non-empty, non-comment lines. If count exceeds `monitoring_threshold`: summarize all entries except the most recent `monitoring_keep` into a single `[Earlier]` line — e.g., `[Earlier] 14 alerts, 3 self-evals (03-28 08:00 — 03-30 18:00)`. If an `[Earlier]` line already exists, merge the counts and extend the time range. Keep the most recent `monitoring_keep` entries intact.
+     - **Session Summary:** Count non-empty, non-comment lines. If count exceeds `summary_threshold`: summarize all entries except the most recent `summary_keep` into a single `[Earlier]` line — e.g., `[Earlier] 15 tasks archived (S-001 — S-015)`. If an `[Earlier]` line already exists, merge counts and extend the range. Keep the most recent `summary_keep` entries intact.
    - Append a summary line to `## Session Summary`:
      `**S-NNN** (YYYY-MM-DD): [one-line task summary] — [status] ($X.XX)`
 5. If cost data is available, preserve the cumulative total in the Cost section
@@ -100,7 +103,8 @@ When the main session requests an idle transition (not a full close):
 ## Rules
 
 - Session IDs are sequential and never reused
-- Always preserve the full content of SHELL.md — never truncate progress logs
+- Never truncate progress logs during a task — only clear them on idle transition (after archiving to report)
+- Monitoring and Session Summary may be compacted on idle transition per the `compact` config thresholds
 - If SHELL.md exists but has Status `completed` or `blocked`, treat it as needing a new session
 - If SHELL.md exists with Status `idle`, treat it as ready for a new task (not a new session) — do not create a new SHELL.md
 - Keep session reports factual and concise — no filler text

docs/ALWAYS-ON-OPS.md

@@ -81,7 +81,7 @@ hermit-start -> [in_progress] -> task done -> [idle] -> new task -> [in_progress
 | **Reflection runs** | Yes                             | Yes                              |
 | **Heartbeat**       | Keeps running (or starts)       | Stopped                          |
 | **Channels**        | Keep running (always-on only)   | Stopped                          |
-| **SHELL.md**        | Reset in-place to `idle`        | Replaced with fresh template     |
+| **SHELL.md**        | Reset in-place to `idle`, Monitoring & Summary compacted if over threshold | Replaced with fresh template     |
 | **Applies to**      | Both interactive and always-on  | Both interactive and always-on   |
 
 Default: idle transition when work finishes. Full shutdown only via explicit `/session-close` or `hermit-stop`.

docs/ARCHITECTURE.md

@@ -73,7 +73,7 @@ SHELL.md plan,   status,   S-NNN-REPORT.md,
 
 **Close:** Defaults to idle transition at every task boundary — your hermit says "What's next?" and waits. Reflection fires. Full shutdown only via `/session-close`. See [Always-On Lifecycle](ALWAYS-ON-OPS.md#2-always-on-lifecycle).
 
-**Archive:** SHELL.md -> `S-NNN-REPORT.md`. Fresh template with carry-forward items. Any session can pick up where the last one left off.
+**Archive:** SHELL.md -> `S-NNN-REPORT.md`. Fresh template with carry-forward items. Monitoring and Session Summary sections are compacted if over threshold (configurable via `compact` in config.json). Any session can pick up where the last one left off.
 
 ---
 

docs/CONFIG-REFERENCE.md

@@ -74,6 +74,21 @@ Modify with `/hermit-settings routines`, `/hermit-settings idle`.
 
 ---
 
+## `compact`
+
+Controls automatic compaction of SHELL.md sections during idle transitions. When a section exceeds its threshold, older entries are summarized into a single `[Earlier]` line and only the most recent entries are kept.
+
+| Key | Type | Default | Description |
+|-----|------|---------|-------------|
+| `monitoring_threshold` | integer | `30` | Compact `## Monitoring` when it exceeds this many lines. |
+| `monitoring_keep` | integer | `20` | Keep this many recent entries after compacting. |
+| `summary_threshold` | integer | `30` | Compact `## Session Summary` when it exceeds this many lines. |
+| `summary_keep` | integer | `15` | Keep this many recent entries after compacting. |
+
+Setting `keep` equal to `threshold` effectively disables compaction for that section. Modify with `/hermit-settings compact`.
+
+---
+
 ## `docker`
 
 | Key | Type | Default | Description |

docs/OBSIDIAN-SETUP.md

@@ -6,11 +6,13 @@ An optional read-mostly companion dashboard. Your hermit works without it.
 
 1. Open Obsidian -> "Open folder as vault" -> select the repo root
 2. Add `.obsidian/` to `.gitignore`
-3. Install the **Dataview** plugin (Community plugins -> Browse -> "Dataview")
-4. Create `dashboard.md` at the repo root (add to `.gitignore`) and paste:
+3. **Install Folder Bridge and add .claude and .claude-code-hermit folder:** plugin (Community plugins -> Browse -> "FolderBridge")(https://github.com/tescolopio/Obsidian_FolderBridge)
+4. Install the **Dataview** plugin (Community plugins -> Browse -> "Dataview")
+5. Create `dashboard.md` at the repo root (add to `.gitignore`) and paste:
 
 ````markdown
 ## Sessions
+
 ```dataview
 TABLE status, date, duration, cost_usd AS "Cost", tags
 FROM ".claude-code-hermit/sessions"
@@ -19,6 +21,7 @@ SORT date DESC
 ```
 
 ## Proposals
+
 ```dataview
 TABLE status, source, category, session, created
 FROM ".claude-code-hermit/proposals"
@@ -27,7 +30,7 @@ SORT created DESC
 ```
 ````
 
-5. Pin SHELL.md in the right pane for live updates (right-click tab -> "Pin")
+7. Pin SHELL.md in the right pane for live updates (right-click tab -> "Pin")
 
 That's it. You now have a queryable dashboard of all sessions and proposals.
 
@@ -148,19 +151,20 @@ This is a manual, operator-curated view. The hermit does not generate or modify
 
 ## Advanced: Multi-Hermit Setup
 
-If you run multiple hermit instances (e.g., dev hermit + Glam), you can create a cross-hermit observatory:
+If you run multiple hermit instances, you can create a cross-hermit observatory:
 
 ```
 observatory/
-  dev-hermit/     -> symlink to project-a/.claude-code-hermit/
-  glam/           -> symlink to project-b/.claude-code-hermit/
-  dashboard.md    -> cross-hermit Dataview queries
+  dev-hermit/        -> symlink to project-a/.claude-code-hermit/
+  accountant-hermit/ -> symlink to project-b/.claude-code-hermit/
+  dashboard.md       -> cross-hermit Dataview queries
 ```
 
 Open the `observatory/` directory as an Obsidian vault. Dataview queries work across symlinked directories:
 
 ````markdown
 ## All Hermits — Recent Sessions
+
 ```dataview
 TABLE id, status, date, cost_usd AS "Cost"
 FROM ""

docs/SKILLS.md

@@ -70,7 +70,7 @@ This section is only active when `idle_behavior` is set to `"discover"` (default
 
 | Skill             | What it does                                                                                                                                                                     | Auto-triggers |
 | ----------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------- |
-| `hermit-settings` | View or change project config. Subcommands: `name`, `language`, `timezone`, `escalation`, `sign-off`, `channels`, `remote`, `model`, `budget`, `brief`, `permissions`, `heartbeat`, `routines`, `idle-agency`, `env`, `docker`. | --             |
+| `hermit-settings` | View or change project config. Subcommands: `name`, `language`, `timezone`, `escalation`, `sign-off`, `channels`, `remote`, `model`, `budget`, `brief`, `permissions`, `heartbeat`, `routines`, `idle-agency`, `env`, `compact`, `docker`. | --             |
 | `init`            | One-time project setup. Creates state directory, runs the wizard, scans your project, and writes OPERATOR.md.                                                                    | --            |
 | `upgrade`         | Run after updating the plugin. Detects version gaps, refreshes templates, prompts for new settings.                                                                              | --            |
 

scripts/cost-tracker.js

@@ -11,9 +11,9 @@ const path = require('path');
 
 // Per-1M-token pricing (USD)
 const PRICING = {
-  haiku: { input: 0.8, output: 4.0 },
-  sonnet: { input: 3.0, output: 15.0 },
-  opus: { input: 15.0, output: 75.0 },
+  haiku:  { input: 0.80, cacheWrite: 1.00,  cacheRead: 0.08, output: 4.0  },
+  sonnet: { input: 3.00, cacheWrite: 3.75,  cacheRead: 0.30, output: 15.0 },
+  opus:   { input: 15.0, cacheWrite: 18.75, cacheRead: 1.50, output: 75.0 },
 };
 
 const MAX_STDIN = 1024 * 1024; // 1MB safety limit
@@ -30,11 +30,45 @@ function detectModel(modelStr) {
   return 'sonnet';
 }
 
-function calculateCost(model, inputTokens, outputTokens) {
+function calculateCost(model, inputTokens, cacheWriteTokens, cacheReadTokens, outputTokens) {
   const pricing = PRICING[model] || PRICING.sonnet;
-  const inputCost = (inputTokens / 1_000_000) * pricing.input;
-  const outputCost = (outputTokens / 1_000_000) * pricing.output;
-  return inputCost + outputCost;
+  return (inputTokens      / 1_000_000) * pricing.input
+       + (cacheWriteTokens / 1_000_000) * pricing.cacheWrite
+       + (cacheReadTokens  / 1_000_000) * pricing.cacheRead
+       + (outputTokens     / 1_000_000) * pricing.output;
+}
+
+function readLastTurnUsage(transcriptPath) {
+  const TAIL_BYTES = 131072; // 128KB — read from end, avoid loading full transcript
+  try {
+    const stat = fs.statSync(transcriptPath);
+    const readFrom = Math.max(0, stat.size - TAIL_BYTES);
+    const fd = fs.openSync(transcriptPath, 'r');
+    const buf = Buffer.alloc(Math.min(TAIL_BYTES, stat.size));
+    fs.readSync(fd, buf, 0, buf.length, readFrom);
+    fs.closeSync(fd);
+
+    const lines = buf.toString('utf-8').split('\n');
+    // Drop the first line when mid-file (it's a partial line)
+    if (readFrom > 0) lines.shift();
+
+    for (let i = lines.length - 1; i >= 0; i--) {
+      try {
+        const entry = JSON.parse(lines[i]);
+        if (entry.type === 'assistant' && entry.message?.usage) {
+          const u = entry.message.usage;
+          return {
+            inputTokens:      u.input_tokens || 0,
+            cacheWriteTokens: u.cache_creation_input_tokens || 0,
+            cacheReadTokens:  u.cache_read_input_tokens || 0,
+            outputTokens:     u.output_tokens || 0,
+            model:            entry.message.model || '',
+          };
+        }
+      } catch {}
+    }
+  } catch {}
+  return null;
 }
 
 function parseLogEntries() {
@@ -266,18 +300,27 @@ async function main() {
 
     const data = JSON.parse(raw);
 
-    // Extract token counts — handle various property names
-    const inputTokens = data.input_tokens || data.inputTokens || data.usage?.input_tokens || 0;
-    const outputTokens = data.output_tokens || data.outputTokens || data.usage?.output_tokens || 0;
-    const model = detectModel(data.model || data.model_name || '');
-    const sessionId = data.session_id || data.sessionId || 'unknown';
+    const sessionId = data.session_id || 'unknown';
+    const transcriptPath = data.transcript_path;
+
+    if (!transcriptPath) {
+      process.exit(0);
+    }
+
+    const turn = readLastTurnUsage(transcriptPath);
+    if (!turn) {
+      process.exit(0);
+    }
+
+    const { inputTokens, cacheWriteTokens, cacheReadTokens, outputTokens, model: rawModel } = turn;
+    const model = detectModel(rawModel);
 
-    const totalTokens = inputTokens + outputTokens;
+    const totalTokens = inputTokens + cacheWriteTokens + cacheReadTokens + outputTokens;
     if (totalTokens === 0) {
       process.exit(0);
     }
 
-    const cost = calculateCost(model, inputTokens, outputTokens);
+    const cost = calculateCost(model, inputTokens, cacheWriteTokens, cacheReadTokens, outputTokens);
     const roundedCost = Math.round(cost * 10000) / 10000;
 
     // Log to JSONL
@@ -286,6 +329,8 @@ async function main() {
       session_id: sessionId,
       model,
       input_tokens: inputTokens,
+      cache_write_tokens: cacheWriteTokens,
+      cache_read_tokens: cacheReadTokens,
       output_tokens: outputTokens,
       total_tokens: totalTokens,
       estimated_cost_usd: roundedCost,
@@ -314,7 +359,7 @@ async function main() {
     writeCostSummary();
 
     // Output brief summary
-    console.log(`[cost-tracker] ${model}: ${Math.round(totalTokens / 1000)}K tokens, $${cost.toFixed(4)} (cumulative: ${costStr})`);
+    console.log(`[cost-tracker] ${model}: ${Math.round(totalTokens / 1000)}K tokens (${Math.round(cacheReadTokens / 1000)}K cached), $${cost.toFixed(4)} (cumulative: ${costStr})`);
   } catch (err) {
     // Non-fatal — never block on cost tracking failure
     console.error(`[cost-tracker] Error: ${err.message}`);

skills/heartbeat/SKILL.md

@@ -49,7 +49,7 @@ Execute one heartbeat tick immediately. Useful for testing the checklist.
 10. Determine: does anything need operator attention?
 
 **If nothing actionable:**
-- Append to SHELL.md `## Monitoring` section: `[HH:MM] Heartbeat: OK`
+- Do NOT append to SHELL.md (the tick is already recorded via `total_ticks` in config.json)
 - Read config `heartbeat.show_ok`:
   - If `true` AND channels are configured: send "Heartbeat OK" to channel
   - If `false` (default): no channel message — silent acknowledgment
@@ -79,21 +79,22 @@ Triggered every N ticks (configured by `heartbeat.self_eval_interval`, default 2
 
 **Steps:**
 
-1. Read recent heartbeat entries from SHELL.md `## Monitoring` section — collect all `[HH:MM] Heartbeat:` lines in the current session
-2. For each checklist item in HEARTBEAT.md: check how many recent ticks flagged it vs how many were OK
-3. **Stale check detection:** If a specific checklist item was OK for all ticks in the evaluation window, suggest to the operator:
+1. Read `heartbeat.total_ticks` and `heartbeat.self_eval_interval` from config.json. The evaluation window covers the last `self_eval_interval` ticks (i.e., ticks `total_ticks - self_eval_interval` through `total_ticks`).
+2. Read SHELL.md `## Monitoring` — collect all `[HH:MM] Heartbeat:` alert/warning lines. These are the only entries present (OK ticks are not logged).
+3. For each checklist item: count how many alert lines reference that item. Compare against the window size (`self_eval_interval`). If zero alerts across the window, the item has been clean for all ticks in the evaluation period.
+4. **Stale check detection:** If a specific checklist item had zero alerts across the evaluation window, suggest to the operator:
    > "Heartbeat self-check: the item '[check description]' has been OK for [N] consecutive ticks. Consider removing it to save tokens, or reducing heartbeat frequency."
-4. **Checklist weight check:** Count items in HEARTBEAT.md. If count > 10, suggest:
+5. **Checklist weight check:** Count items in HEARTBEAT.md. If count > 10, suggest:
    > "HEARTBEAT.md has {count} items (recommended: ≤10). Consider moving periodic checks to routines (`/hermit-settings routines`) to reduce per-tick token cost. Items that only need daily/weekly checking are good routine candidates."
-5. **Missing check suggestion:** Scan `.claude-code-hermit/proposals/` for recent auto-detected proposals about recurring issues. If a proposal describes a problem that could be caught by a heartbeat check, suggest:
+6. **Missing check suggestion:** Scan `.claude-code-hermit/proposals/` for recent auto-detected proposals about recurring issues. If a proposal describes a problem that could be caught by a heartbeat check, suggest:
    > "PROP-NNN identified a recurring [issue]. Consider adding a [relevant check] to HEARTBEAT.md."
-6. Append self-evaluation findings to SHELL.md `## Monitoring` with timestamp:
+7. Append self-evaluation findings to SHELL.md `## Monitoring` with timestamp:
    ```
    [HH:MM] Heartbeat self-eval: [summary of suggestions]
    ```
-7. If channels are configured: send self-eval findings as a notification
-8. Do NOT auto-modify HEARTBEAT.md — suggest only. The operator decides what to check. Offer to make the edit if the operator agrees.
-9. After the operator acts on a suggestion (adds or removes a check): reset `heartbeat.total_ticks` to 0 in config.json
+8. If channels are configured: send self-eval findings as a notification
+9. Do NOT auto-modify HEARTBEAT.md — suggest only. The operator decides what to check. Offer to make the edit if the operator agrees.
+10. After the operator acts on a suggestion (adds or removes a check): reset `heartbeat.total_ticks` to 0 in config.json
 
 ### start