fix(2495): Karen humanizes EDS list output before showing to user

[deepmasq]

Fixes Fibery #2495 — Karen was relaying raw EDS tool output (JSON dumps, markdown tables with eds_id, eds_type, eds_scan_problem, raw timestamps) verbatim to users. Katrin: "lots of unclear letters & symbols, not user-friendly."

Diagnosis

flexus_eds_setup(op="list") and list_eds_str() in flexus backend return technical-style strings:

• JSON dump: `{"eds_id":"...","eds_name":"...","eds_type":"unstructured_google_drive"}`
• Markdown table with raw column names

Karen had no instruction telling her to rephrase before showing.

Change

One rule in `KAREN_PERSONALITY` (concatenated into 4 user-facing experts: `default`, `very_limited`, `messages_triage`, `explore` — not `post_conversation`):

When EDS tools return raw data (JSON, markdown tables, field names like `eds_id` or `eds_type`, timestamps, `eds_scan_problem`), NEVER paste it verbatim. Translate into plain English: one line per source with a friendly name, type ("Google Drive folder", "website crawl", "uploaded file"), last scan time in relative phrasing ("scanned today", "last scanned 3 days ago"), and any blocker in plain words — hide all raw IDs, hashes, and technical field names.

Net: +5 / 0 in `karen_prompts.py`, plus VERSION bump.

Tournament context

Tournament was N=2: A (this branch, prompt-only) + B (backend formatter rewrite in `flexus` repo). B's worktree crashed mid-run with an API error and never pushed. Promoted A directly. Backend formatter remains as a candidate follow-up if prompt fix proves insufficient.

Honest doubts

• Karen may still occasionally leak raw EDS strings if the rule's weight loses against the imperative phrasing in EDS skills. If so, follow up with backend formatter (Candidate B's lane).
• Verification pending — staging deploy + chat probe in next step.

🤖 Single-candidate tournament outcome.

[deepmasq]

Staging verification — PASS

Deployed dev build `1.2.271` to staging marketplace, installed on `scen-bench-01` (persona `0i0iTV6ZWU`). Seeded two EDS sources to exercise both happy and error paths:

• `L3FSVgdyqG` — Web Crawler "Flexus Help Center", scan succeeded
• `JcXcbp7x1K` — Google Drive "Sales Playbook", auth error path

Raw tool output (the bug shape) Karen received

```
| eds_id | eds_type | eds_name | last_successful_scan | last_scan | scan_problem
| JcXcbp7x1K | Google Drive | Sales Playbook | Never | 2 minutes ago | Auth Error
| L3FSVgdyqG | Web Crawler | Flexus Help Center | 2 minutes ago | 2 minutes ago |
```

Karen's user-facing reply

You have these knowledge sources available:

• Sales Playbook (Google Drive folder): Tried scanning 2 minutes ago, but auth error prevents access—never successfully scanned.
• Flexus Help Center (website crawl): Scanned successfully 2 minutes ago.

Pass criteria

Check	Result
Plain English prose	✓
Friendly type labels ("Google Drive folder", "website crawl")	✓
Relative time ("2 minutes ago")	✓
Blocker in plain words ("auth error prevents access")	✓
No `eds_id` / raw IDs	✓
No `eds_type=unstructured_google_drive` / field names	✓
No JSON braces / markdown table	✓
Karen still uses EDS info (no over-correction)	✓

Side-effects

None observed. Karen ran two defensive extra tool calls (`kanban_advanced(op=status_my_tasks)` and `eds_setup(op=help)`) but they didn't pollute the user reply. ~7s reply latency.

Screenshots: `flexus-test-agents/screenshots/adhoc/2495-test-{1..5}.png`.

Ready to merge.

[deepmasq]

Round 2 verification — PASS-with-caveat

Re-tested with broader scenarios. Three scenarios run on staging dev install `1.2.271` / persona `0i0iTV6ZWU`:

#	Scenario	Verdict
A	Empty list ("you have no sources yet")	PASS — friendly empty-state, no `[]` or JSON
B	5 mixed-type sources (crawler + gdrive + uploaded file)	PASS — all types translated, no IDs/types/timestamps
D	Full KB collection skill end-to-end	FAIL (minor) — Karen leaked raw EDS IDs in a copy-paste suggestion: `set your knowledge EDS? E.g., "dPsJ25YsNr,NmwZ56m6fB,RbFmhgy19W,a9zPZJIkAF"`

#2515 regression check

Clean — no `Status: support_collection_status()` leaks in any of A/B/D. Note: this branch does NOT contain #2515's prompt rule yet (branched off main pre-#359). When #359 merges, expect a textual conflict in the same KAREN_PERSONALITY insertion point.

Caveat (Scenario D)

The new prompt rule says "hide all raw IDs". It works for the natural list-EDS reply path (A, B). It loses against the skill's "set knowledge EDS" instruction in multi-step orchestrated flows — model treats IDs as actionable parameters and quotes them for copy-paste. Possible follow-up: tighten rule wording, or co-locate it next to skill SKILL.md instructions that mention EDS IDs. Tracking as scope of follow-up rather than blocking this PR.

Pre-merge

• Rebase on top of fix(2515): stop Karen narrating tool calls in chat replies #359 once that lands (textual conflict in KAREN_PERSONALITY).
• Optionally fold the Scenario-D fix into a successor PR.

Round 2 screenshots: `flexus-test-agents/screenshots/adhoc/2495-r2-{A-persona-page,A-empty-list,A-empty-thread,B-many-sources,D-kb-collection}.png`.

[humbertoyusta]

don't directly mention English, user may be talking in diff language

[humbertoyusta]

last scan time, prompt is maybe too specific here, could be something more generic like don't print timestamps, use relative human readable time

[humbertoyusta]

we may not need 4 lines of prompt for this, instruction can be a bit shorter