Pod

Yes, this is on purpose. Pod is built for agents, so the default page is plain HTML.
Human? View the normal website.
Agent? You probably prefer /mcp/token-optimizer-mcp/issues.md or /mcp/token-optimizer-mcp/issues.json, or Pod over MCP.

Reported issues for token-optimizer-mcp

Pod holds 19 of 22 problems reported by people outside the maintainer team. Issues filed by the project's own owners, members and collaborators are excluded entirely — a maintainer's release checklist is not a warning to a prospective user.

Back to token-optimizer-mcp.

Most discussed

CRITICAL: 3-6x Performance Degradation - Add Server-Level Timing

CRITICAL Performance Degradation

Tasks that used to take 20 minutes are now taking 1-2 hours (3-6x slowdown)

Root Cause Analysis (from Gemini CLI)

Gemini identified THREE major bottlenecks in the PowerShell hooks:

1. Excessive External Process Spawns (10-50ms EACH)

Problem: Each call to invoke-mcp.ps1 creates a NEW PowerShell process Impact: Happens on EVERY hook invocation Fix: Dot-source invoke-mcp.ps1 as functions instead of spawning processes

Read the thread · 2025-11-01 · closed · 3 comments

Living wiki graph: Derived-Knowledge Retrieval (retrieve verdicts, not evidence)

Summary

Build a per-project living wiki graph the agent writes as a side effect of working and reads instead of re-deriving. Full design: docs/WIKI_GRAPH.md (#203).

Why this is the differentiator

Every token optimizer on the market — including ours today — reduces the cost of an operation. Compress the output, cache the read, diff the re-read. That is a bounded win: you

Read the thread · 2026-07-30 · closed · 2 comments

feat: add token savings data to current-session.txt for easy real-time monitoring

Problem

Currently, current-session.txt only tracks operation counts, NOT actual token savings:

{
    "lastOptimized": 0,
    "sessionId": "aad1f532-43d3-44af-ba65-6f89c409b122",
    "sessionStart": "20251031-082211",
    "totalOperations": 2201,
    "totalTokens": 0
}

Users must manually call get_session_stats() MCP tool to see their savings. This creates friction for:

Read the thread · 2025-10-31 · closed · 2 comments

smart_glob/smart_grep hang >120s (unbounded traversal) on large repo trees; native dir/rg return instantly

Summary

smart_glob and smart_grep (token-optimizer MCP v5.5.0) hang past the 120s tool timeout and get backgrounded/killed when run against a large/deep repository tree, whereas the native equivalents (cmd /c dir, ripgrep) return in well under a second on the same paths. The tools are advertised as bounded/paginated replacements for the built-ins, so timing out where the built-in is instant is a correctness/perf defect — and it forces a fallback to the very built-ins the routing p

Read the thread · 2026-08-28 · closed · 1 comment

contradictedAt is stored provenance with no reader

Found during #315 (wiki-graph gap closure, part 1 of 3).

hooks-core/curate.mjs writes two fields when a contradiction is recorded: contradictionReason and contradictedAt. The reason now has a reader — it is surfaced in the dispute disclosure when a finding is served. contradictedAt is read by nothing.

It is stored provenance with no consumer, which is the defect class #315 exists to close, in its quietest form.

Two honest options

  1. Give it a reader. The plausible one: age

Read the thread · 2026-08-26 · closed · 1 comment

[BUG]

Bug: Invalid schema for cache_warmup function

Description

The mcp__token-optimizer__cache_warmup tool fails with a schema validation error when called through Claude Code MCP client.

Error Message

{ "error": { "message": "Invalid schema for function 'mcp__token-optimizer__cache_warmup': In context=('properties', 'accessHistory'), array schema missing items.", "type": "invalid_request_error", "param": "tools[96].parameters", "code": "invalid_func

Read the thread · 2026-04-15 · closed · external user · 1 comment

feat: implement granular per-hook/per-action/per-MCP-server token analytics

Problem

Currently, token-optimizer-mcp tracks aggregate token savings at the session level, but lacks granular visibility into:

This makes it impossible to identify optimization opportunities and pinpoint which hooks/actions/servers are consuming the most tokens.

Proposed Solution

Implement server-

Read the thread · 2025-10-31 · closed · 1 comment

Claude Code runs no semantic harvest at all: #300 unregistered stop-harvest.mjs

Measured while executing Plan 2 (#320) of the #204 work.

What happened

#300 unregistered plugin/hooks/stop-harvest.mjs. Four capabilities went with it, silently:

  1. The transcript archive — restored in Plan 2 Task 5, because the local extractors read it.
  2. The harvest-worker spawn — still dead.
  3. Refusal detection — still dead.
  4. The harvest off-notice — still dead.

The consequence

On Claude Code, with a valid credential present, no semantic harvest runs. `h

Read the thread · 2026-08-26 · closed · 0 comments

Most recent

Plan 3: close the defect class, not its instances (#204, part 3 of 3)

Tracking issue for Plan 3 of the #204 wiki-graph work. Part 3 of 3; Plan 1 shipped as #315, Plan 2 is #320.

The plan is committed and self-sufficient: docs/superpowers/plans/2026-08-25-wiki-graph-plan-3-detector-and-allowlist.md on branch feat/close-wiki-graph-gaps-plan2. Read the "Execution state and corrections" section at the bottom first.

Runs last. Its allowlist-emptying task is incomplete by construction until Plans 1 and 2 remove their entries.

Why this is the load-beari

Read the thread · 2026-08-26 · closed · 0 comments

Plan 2: make the graph produce findings, and measure whether they pay (#204, part 2 of 3)

Tracking issue for Plan 2 of the #204 wiki-graph work. Part 2 of 3; Plan 1 shipped as #315.

The plan is committed and self-sufficient: docs/superpowers/plans/2026-08-25-wiki-graph-plan-2-production-and-measurement.md on branch feat/close-wiki-graph-gaps-plan2. Its "Execution state and corrections" section at the bottom is authoritative over the task text above it ? the tasks were written before Plan 1 ran and are wrong in five specific places.

What Plan 2 is for

Plan 1 made the

Read the thread · 2026-08-26 · closed · 0 comments

Content-addressed anchors: a finding about a vendored file should reach every repo holding it

Extracted from #315 (wiki-graph gap closure, part 1 of 3), where contentAnchor was deleted rather than left dormant on the reachability allowlist.

The idea is worth keeping. From its docblock before deletion:

A vendored library file is the same file in every repository that holds it, whatever path each gives it. Anchoring to content as well as path means a finding about it appears in all of them with no promotion step and no path mapping — reach a per-session checkpoint cannot have even in

Read the thread · 2026-08-26 · closed · 0 comments

A drainer killed mid-claim strands its pending-invalidation records

Found during #315 (wiki-graph gap closure, part 1 of 3), in code that PR added.

drainInvalidations in hooks-core/pending.mjs claims the pending-invalidation queue by renaming it to a per-pid claim file, then reads and deletes the claim. That closes the original defect — a concurrent append being deleted unread — and two drainers cannot lose a record, because the loser's renameSync throws and its recovery block is guarded on its own pid.

The residual

A drainer **killed between the ren

Read the thread · 2026-08-26 · closed · 0 comments

Injected text is unescaped: a stored claim or dispute reason can forge lines in the model's context

Found during #315 (wiki-graph gap closure, part 1 of 3). Filed rather than fixed there, because it is a pre-existing class that PR widened rather than introduced.

The problem

Stored text reaches a model's context without escaping. Two surfaces:

Read the thread · 2026-08-26 · closed · 0 comments

Dashboard: /wiki route 404s + "active session" widget reads unwired path

Found two bugs in the shipped dashboard (npm run dashboard, dist/server/web-server.js), tested on v5.7.0 installed via npm install -g @ooples/token-optimizer-mcp@latest.

Bug 1: /wiki route 404s (res.sendFile fails despite file existing)

GET /wiki and clicking the "What it knows" nav tab both 404 with:

NotFoundError: Not Found
    at createHttpError (node_modules/send/index.js:861:12)
    at SendStream.error (node_modules/send/index.js:168:31)
    at SendStream.pipe (node_mo

[Read the thread](https://github.com/ooples/token-optimizer-mcp/issues/311) · 2026-08-20 · closed · external user · 0 comments

### Codex plugin installs but MCP server exits before registering tools on Windows

## Summary

The token-optimizer Codex plugin is installed and enabled, but its MCP server exits before registering any tools on Windows. The packaged doctor also fails its server-response check. In addition, the doctor appears to inspect a stale Claude plugin installation (5.5.0) while diagnosing the current Codex/npm package (5.7.0).

## Environment

- Windows 11 (`Microsoft Windows NT 10.0.22631.0`)
- Windows PowerShell `5.1.22621.4249`
- Codex CLI `0.147.0`
- Node.js `v22.2.0`
- npm `10.7.0`

[Read the thread](https://github.com/ooples/token-optimizer-mcp/issues/307) · 2026-08-13 · closed · 0 comments

### feat: implement compression for stored optimization results

## Summary
Implement gzip compression for stored optimization results, session files, and cache data to reduce disk I/O and storage costs.

## Background
Currently, the token optimizer stores large amounts of data uncompressed:
- **Session files**: `current-session.txt`, operation CSV files (grow to several MB)
- **Cache data**: File content cache, optimization results (can be 10-100MB)
- **Operation logs**: CSV logs with full tool input/output (grow rapidly)

Gemini CLI uses gzip compr

[Read the thread](https://github.com/ooples/token-optimizer-mcp/issues/126) · 2025-11-03 · closed · 0 comments

### feat: implement LRU cache for expensive operations

## Summary
Implement a Least Recently Used (LRU) cache to store results of expensive operations (file search, edit correction, token counting, MCP API calls) and prevent redundant computation.

## Background
Currently, the token optimizer performs expensive operations repeatedly:
- **File search**: Scanning repository for patterns (takes 100-500ms per search)
- **Edit correction**: Running MCP `correct_edits` API (takes 500-2000ms per file)
- **Token counting**: Calling Google AI API or p

[Read the thread](https://github.com/ooples/token-optimizer-mcp/issues/125) · 2025-11-03 · closed · 0 comments

### feat: implement sophisticated token counting (beyond character/4)

### Problem
The current token counting mechanism is a mix of a simple estimation (length / 4) and a call to the Google AI countTokens API. This is not a robust solution for a production environment for several reasons:

*   **Inaccurate Fallback:** The length / 4 estimation is highly inaccurate for different types of content (code, JSON, etc.) and different languages.
*   **Vendor Lock-in:** The countTokens API is specific to Google AI models. The system should be able to handle models from othe

[Read the thread](https://github.com/ooples/token-optimizer-mcp/issues/124) · 2025-11-03 · closed · 0 comments

### feat: implement context delta tracking (only send changes)

### Problem
The current system sends the entire context with every request. This is inefficient and can lead to higher costs and slower response times, especially for long conversations.

### Desired State
A system that can:

*   Track the conversation history and file state for each user session.
*   Calculate the delta between the previous and current context.
*   Send only the delta to the model.
*   Persist session information across restarts.

### Revised Implementation Plan
This revised pl

[Read the thread](https://github.com/ooples/token-optimizer-mcp/issues/122) · 2025-11-03 · closed · 0 comments

The remaining reports are on [the project's issue tracker](https://github.com/ooples/token-optimizer-mcp/issues).