# Reported issues for AIppocampus

Pod holds 23 of 77 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 [AIppocampus](/mcp/aippocampus).

## Most discussed

### Compact recall can still surface generic route labels that are not discriminative enough for foreground choice

## Summary

Dogfood recall still has a route-label quality gap: compact `agent recall` can return reopenable routes whose public labels are technically safe but too generic for a foreground agent to choose between without immediately falling back to manual search/deepen.

This looks like a recurrence of the route-label UX family previously addressed in issues such as #2080, #2340, #2501, #2526, #2555, #2562, and #2716. The current failure is narrower: low-specificity detection exists, but the co

[Read the thread](https://github.com/Sapientropic/AIppocampus/issues/2876) · 2026-06-28 · closed · external user · 4 comments

### Add a shared foreground action and recovery contract for CLI/MCP surfaces

## Problem

The recent foreground UX sweep is finding the same failure mode across many surfaces:

- successful or near-successful commands return placeholder actions (`<cue>`, `old cue`);
- parent commands with `--json` still fall through to raw argparse;
- recovery cards use different field names and precedence rules;
- MCP schemas and CLI cards express required selectors differently;
- some surfaces are operator-only but still look like foreground frontdoors;
- individual issues fix one comma

[Read the thread](https://github.com/Sapientropic/AIppocampus/issues/2057) · 2026-06-16 · closed · external user · 11 comments

### Make architecture-native designs discoverable as callable agent lanes

## Summary

Several discussion / research designs now have real runtime pieces, but the current safe projection posture can still make those capabilities product-invisible to a foreground agent.

This is not asking to promote Dream / avatar / Macro / topology / source-shape material into source truth. The current source boundary is correct. The gap is narrower: when a capability is present or requested, the foreground agent should be able to discover the safe capability lane and the next callabl

[Read the thread](https://github.com/Sapientropic/AIppocampus/issues/1693) · 2026-06-15 · closed · external user · 4 comments

### Add compact agent-facing JSON modes for health/update/MCP diagnostics

## Summary
Several diagnostic commands are either human-compact or full-audit JSON, but there is no compact machine-readable middle layer for foreground agents.

## Environment
- `main` at `037a642b8a0728c048f037fe04298bdf6173d248`
- local install refreshed to `v0.3.1`
- macOS / Codex Desktop dogfood run

## Reproduction
```bash
aippocampus health --cwd "$PWD" --compact-json
# error: unrecognized arguments: --compact-json

aippocampus update status --compact-json
# error: unrecognized arguments:

[Read the thread](https://github.com/Sapientropic/AIppocampus/issues/1492) · 2026-06-14 · closed · external user · 5 comments

### Distinguish current artifacts from host-exposed AIppocampus CLI/MCP/plugin readiness

## Summary

Second-user dogfood found a confusing readiness gap: AIppocampus can report the CLI/MCP/plugin surfaces as current while the current Codex foreground agent still cannot use them as first-class host tools.

This can make product dogfood look worse than the full intended UX. In our local testing, several agents had to fall back to `PYTHONPATH=... python3 -m ...`, direct clean-source/registry inspection, or broad shell searches. After checking the environment more carefully, that fricti

[Read the thread](https://github.com/Sapientropic/AIppocampus/issues/1279) · 2026-06-11 · closed · external user · 3 comments

### Define graded trust semantics for ambient source-backed packets

## Summary

AIppocampus should not teach agents to distrust its own source-backed packets.

Right now the safe default wording around ambient/hook output often collapses too many packet types into the same operational stance: “treat this as a hint, do not trust it, reopen source before use.” That is correct for semantic/cognitive hints, but it is too blunt for bounded source-backed evidence.

This is related to #201, but distinct enough to track separately. #201 is already a long recall-quality 

[Read the thread](https://github.com/Sapientropic/AIppocampus/issues/786) · 2026-06-06 · closed · external user · 7 comments

### Improve human-facing diagnostics for live-provider setup and progressive recall smokes

## Summary

A second-user macOS/Codex Desktop smoke on `main` at `6442c81` found that the new progressive recall and live-provider surfaces are working, but the operator/human path is still hard to reason about. This may be fine for agents in the steady state, but first-time users and maintainers still need a simpler diagnostic path when provider keys, smoke scripts, or MCP recall handles do not behave as expected.

Related umbrella/work items: #397, #470, #465.

## What worked

- Updated local 

[Read the thread](https://github.com/Sapientropic/AIppocampus/issues/478) · 2026-06-02 · closed · external user · 4 comments

### Foreground semantic cache warms exact prompts but paraphrases still cold-time out under default hook budget

## Summary

Follow-up to #53 and #99 on current `main`: the semantic cache works once a prompt has successfully completed, but in an external installed-hook run it only made the exact same prompt shape hot. Semantically close paraphrases still missed the cache and repeatedly hit foreground read timeouts under the default installed hook budget.

This issue is intentionally scoped to cache/warming and foreground budget behavior. It is not a provider-wide latency claim.

## Tested state

- Commit: 

[Read the thread](https://github.com/Sapientropic/AIppocampus/issues/157) · 2026-05-30 · closed · external user · 5 comments

## Most recent

### agent_recall can emit a route_note deepen action that fails with source_ref_not_found

## Summary

A real dogfood recall on `b375c0dd` emitted a primary `agent_deepen` action for a low-confidence `route_note` route, but following the emitted action failed with `source_ref_not_found`.

This is not just low confidence. The foreground contract gave an executable next step, but that next step could not reopen source.

## Environment

- Local checkout: `b375c0dd`
- Skill/plugin refreshed locally
- `aippocampus mcp status --json` reports key tools present
- Ordinary source search and fr

[Read the thread](https://github.com/Sapientropic/AIppocampus/issues/2969) · 2026-07-01 · closed · external user · 0 comments

### Current-thread MCP can expose stale search_memory schema even after install verifies

## Summary

After syncing to `b375c0dd`, refreshing the Codex plugin, and applying the skill update, a real foreground Codex MCP call still exposed a stale/mismatched `search_memory` tool path. The host/package checks looked healthy, but the current-thread MCP call failed at runtime.

This is not a claim that the packaged/fresh MCP server is broken. Fresh in-repo MCP probes passed. The gap appears to be current foreground MCP/runtime freshness detection and follow-through.

## Environment

- Loc

[Read the thread](https://github.com/Sapientropic/AIppocampus/issues/2968) · 2026-07-01 · closed · external user · 0 comments

### APW foreground labels should discount low-signal/meta cue anchors before passing source-anchor gates

## Summary

APW semi-default recovery can currently turn low-signal or meta-discussion cue terms into a foreground route label. In a dogfood recall-label UX cue, the compact primary action became:

```text
Open APW source route: too / Relationship / continuity
```

This is technically source-reopenable, but the selected anchors do not describe the user's real intent (`route labels are too generic`). They make the foreground action look specific while still being semantically weak.

## Reproducti

[Read the thread](https://github.com/Sapientropic/AIppocampus/issues/2877) · 2026-06-28 · closed · external user · 3 comments

### Define a product-level progressive disclosure contract for foreground surfaces

## Summary

AIppocampus has improved many individual foreground surfaces, but the same product problem keeps reappearing: compact/default output still often exposes too many fields, status layers, fallback actions, diagnostics, or internal state before the agent or user gets the smallest useful next step.

This issue asks for a product-level progressive disclosure contract across setup, status, plugin, MCP, hooks, recall, search, and frontdoor cards. The goal is not just smaller JSON; it is atte

[Read the thread](https://github.com/Sapientropic/AIppocampus/issues/2841) · 2026-06-27 · closed · external user · 1 comment

### Quick tier fails on macOS because Claude Code clean-source metadata keeps cwd under /private/var

## Problem

On current `main` / `45e42680`, the quick tier fails on this macOS host in the Claude Code provider normalization test. The failure is triggered by `provider_metadata.cwd` retaining the local temp cwd, which resolves under `/private/var/...` on macOS.

This looks like both a flaky test and a clean-source boundary question: clean-source visible messages are supposed to omit private/tool material, but the normalized message still carries a local cwd in `provider_metadata`.

## Environm

[Read the thread](https://github.com/Sapientropic/AIppocampus/issues/2840) · 2026-06-27 · closed · external user · 0 comments

### MCP agent_recall can fail with stale runtime after plugin update while CLI and host probe pass

## Problem

After syncing `main` to `7c45a3c3` and reinstalling/verifying the Codex plugin, the current Codex thread still has a foreground MCP `agent_recall` runtime that appears stale. The host probe and CLI path pass, but the actual foreground MCP tool call fails with a Python signature mismatch.

This is a UX/product issue because an agent sees AIppocampus tools, tries to use the recommended MCP path, and gets a hard tool failure even though the local install/cache look correct.

## Environm

[Read the thread](https://github.com/Sapientropic/AIppocampus/issues/2765) · 2026-06-27 · closed · external user · 2 comments

### Plugin verify can fail key-tool smoke with JSONDecodeError even when tools are visible and CLI recall works

## Problem

On current `main` (`ca25b426`) with AIppocampus `0.3.3`, plugin install/verify can report a failed host live probe even though:

- the plugin is installed and enabled;
- 17 MCP tools are visible;
- key tools are present (`agent_recall`, `agent_aippo`, `agent_background`, `agent_deepen`, `agent_explain`);
- the CLI `aippocampus agent recall ... --json` path returns valid JSON.

This makes setup look broken from the user/agent-facing readiness card, even when core CLI recall is usable.

[Read the thread](https://github.com/Sapientropic/AIppocampus/issues/2718) · 2026-06-25 · closed · external user · 2 comments

### Do not mark write commands read_only in update/plugin foreground actions

## Summary

Second-user UX probe on current `main` (`30b975bc`) found that setup/status foreground cards can mark write commands as `mutation_risk: read_only`.

This is a shared foreground-action contract problem, not a single-command typo.

## Repro

```bash
aippocampus plugin status --json
aippocampus update status --foreground-tools-visible --agent-json
aippocampus update plan --json
```

Observed primary action in `plugin status` / `update status`:

```json
{
  "id": "review_agent_callable",

[Read the thread](https://github.com/Sapientropic/AIppocampus/issues/2669) · 2026-06-24 · closed · external user · 1 comment

### Respect include_private_identifiers=false in MCP list_threads full detail

## Summary

Second-user UX/privacy probe on current `main` (`30b975bc`) found that MCP `list_threads` can still expose private identifiers and a very large raw session metadata block even when called with `include_private_identifiers=false` and `include_private_paths=false`.

## Repro

MCP call:

```text
list_threads(max=2, detail="full", include_private_paths=false, include_private_identifiers=false)
```

## Observed

The compact call is well behaved, but `detail=full` returned fields shaped li

[Read the thread](https://github.com/Sapientropic/AIppocampus/issues/2666) · 2026-06-24 · closed · external user · 1 comment

### Do not source-open the first registry hit when usefulness says query_anchor_missing

## Summary

Second-user UX probe on current `main` (`30b975bc`) found a foreground-action mismatch in registry-wide search.

The search result correctly says the result is not useful enough, but the primary foreground action still tells the agent to open the first hit.

## Repro

CLI:

```bash
aippocampus search --all "机械飞升 基因飞升" --json
```

MCP:

```text
search_memory(query="机械飞升 基因飞升", scope="all_registered_sources", include_private_paths=false)
```

## Observed

CLI summary fields:

```json
{

[Read the thread](https://github.com/Sapientropic/AIppocampus/issues/2662) · 2026-06-24 · closed · external user · 1 comment

### Align update status agent-callable readiness with MCP status visibility

## Problem

On a second-user macOS install after syncing and reinstalling the Codex plugin at `30b975bc`, the MCP status and update status disagree about foreground agent-callable readiness.

Setup/verification:

```text
aippocampus plugin install --codex --verify
=> plugin install: ok
=> agent callable: host probe checked
```

MCP status then reports tools are visible:

```text
aippocampus mcp status --json
=> ok=true
=> tool_count=17
=> foreground_action=inspect_current_thread_tool_discovery
`

[Read the thread](https://github.com/Sapientropic/AIppocampus/issues/2661) · 2026-06-24 · closed · external user · 1 comment

### Make live source-chain readiness prerequisite-aware when private dogfood sources are absent

## Problem

On a second-user macOS install after syncing `main` to `3d61cd95`, the normal PR/unit gates pass, but the live recall readiness gate fails because it uses maintainer-private dogfood source-chain cues whose expected source is not present in this local registry.

Verified first:

```text
python tools/aippocampus/docs/check_docs_health.py --json
=> ok=true

python tools/aippocampus/run_tests.py --tier pr
=> 311 tests OK, 1 skipped

python -m unittest \
  tests.aippocampus.test_mcp_defau

[Read the thread](https://github.com/Sapientropic/AIppocampus/issues/2652) · 2026-06-23 · closed · external user · 0 comments

### Normalize explicit clean-source paths across macOS /var symlink boundaries

## Problem

A focused clean-source resolution test fails on macOS because an explicit clean-source path under `/var/...` is returned as the realpath-normalized `/private/var/...` path.

This is a small cross-platform path identity issue, but it matters because several recent recall/MCP fixes now depend on source-chain identity and explicit clean-source directory handling. A user- or host-provided explicit path should not fail equality checks merely because macOS exposes `/var` as a symlink to `/

[Read the thread](https://github.com/Sapientropic/AIppocampus/issues/2610) · 2026-06-22 · closed · external user · 1 comment

### Do not mark foreground MCP verified from assertion-only status flags

## Summary

A real Codex foreground-thread retest found a mismatch between host/plugin verification and the current foreground MCP transport. `plugin install --codex --verify` can pass, and `update status` can be lifted to `full_continuity_path` via assertion flags, while a direct AIppocampus MCP tool call in the same foreground conversation still fails with `Transport closed`.

This is not a claim that the plugin install path is broken. The CLI fallback and the diagnostic downgrade path are use

[Read the thread](https://github.com/Sapientropic/AIppocampus/issues/2117) · 2026-06-17 · closed · external user · 1 comment

### Add a shared clean-source ingestion and projection sanitizer contract

## Problem

AIppocampus has many privacy/redaction/source-boundary protections, but the same class of issue keeps reappearing at different edges: a surface has to decide whether some text or metadata is clean source, private operator detail, public-safe projection, audit-only material, or a host/control envelope.

The latest concrete example is #2049: `agent deepen` reopened a clean-source window whose message text contained a Codex host-internal continuation/goal wrapper. That should not be ord

[Read the thread](https://github.com/Sapientropic/AIppocampus/issues/2059) · 2026-06-16 · closed · external user · 1 comment

The remaining reports are on [the project's issue tracker](https://github.com/Sapientropic/AIppocampus/issues).