Memory Update — SeedPass TUI v3 Development — 2026-03-03

Key findings

Nostr Relay Management is now implemented in TUI v3 (src/seedpass/tui_v3/screens/relays.py) via the relay-list palette command.
Textual UI updates from asynchronous workers requires executing from the main thread using self.app.call_from_thread(), otherwise you will encounter unexpected thread safety warnings crashing the app. This was correctly used for sync_service.sync() background operations.
The pytest execution of parity tests natively sources from the local environment. Ensure testing inside the virtual environment properly maps module paths (e.g., using PYTHONPATH=. pytest ... from the src directory) when src-layout paths conflict with old site-packages installs, which causes confusing AttributeError tracebacks on customized files.
The SettingsScreen and TUI v3 components now rely on the standard Textual app.notify() for the "Unified snackbar/toast system" — giving uniform, built-in styling for service feedback.

Direct configuration modification of arrays (like relays) should favor using dedicated service adapters over direct ConfigService modifications if they aren't fully wrapped for list operations. Using app.services["nostr"] for relay additions/removals is the preferred pattern.
The cohesive retro palette across the TUI v3 mockups hinges on high-contrast #999999 backgrounds, #000000 text/borders, and crisp #ffffff inputs. All standalone screens (SettingsScreen, MaximizedInspectorScreen, AddEntryScreen, etc.) must strictly conform to these CSS constants rather than the legacy v2 neon green (#58f29d, #080a0c).
In automated Textual UI interactive parity testing (interactive_agent_tui_test.py), avoid explicitly sending await pilot.press('ctrl+p') if standard textual palettes shadow custom CommandPalette widgets. Directly calling app.action_open_palette() guarantees stable integration assertions.
When mirroring nested manager logic (like BIP-85 sub-account loads), UI "breadcrumbs" should derive their explicit path strings from parsing the manager's profile_stack via a decoupled reactive property (e.g. active_breadcrumb), rather than mutating the core active_fingerprint references used for caching and view filtering.

ov init -y successfully bootstraps .overstory/, .mulch/, .seeds/, .canopy/, and also auto-appends onboarding blocks to CLAUDE.md and merge rules to .gitattributes.
For this Python repo, .overstory/config.yaml should be adjusted from Bun defaults to canonicalBranch: beta and Python quality gates (pytest, scripts/run_ci_tests.sh) to avoid incorrect guardrails.
Overstory/Codex model handling in this repo required follow-up validation; see the later "Overstory/Codex spawn fix" section for the corrected working configuration.

Reproduced ov sling failure: Failed to send keys to tmux session ... can't find pane after a clean launch path.
Root cause was invalid Codex model aliases in .overstory/agent-manifest.json (haiku/sonnet/opus) causing codex exec to exit immediately; codex exec --model sonnet fails with unsupported-model error on ChatGPT-backed Codex accounts.
Working configuration on this machine:
- Keep .overstory/config.yaml models: empty (Overstory v0.7.9 rejects bare gpt-* entries in this block unless provider-prefixed, which Codex account mode does not accept).
- Set each agent model in .overstory/agent-manifest.json to gpt-5.3-codex.
Secondary requirement: tmux server must be running before sling (tmux new-session -d -s ov-bootstrap 'bash' is enough); otherwise Overstory errors with Tmux server is not running.
Added a dedicated personal swarm runbook at repo root: OVERSTORY_SWARM_TUTORIAL.md with verified commands, failure modes, and recovery flow for this SeedPass setup.

Migrated high-signal operational/project learnings into Mulch domain seedpass (.mulch/expertise/seedpass.jsonl) including Overstory/Codex model pitfalls, tmux bootstrap convention, TUI v3 threading/notification patterns, and runbook reference.
Full CI-equivalent run currently reports 1057 passed, 16 skipped, 7 failed; failures are concentrated in src/tests/test_tui_v2_textual_interactions.py and are mostly copy/text assertions or profile-tree expectation drift.
Determinism regression caused by tuple arity mismatch was fixed by updating src/tests/test_deterministic_artifact_regression.py to unpack derive_pgp_key as (priv, pub, fp) and correcting derive_pgp_key type/docs in src/seedpass/core/password_generation.py.
Local quality signal gaps: black --check . currently reports 20 files needing formatting, and ruff/mypy are configured in project metadata or pre-commit but not installed/gated in the current local environment by default.

Repaired 7 failing TUI v2 interaction regressions in src/seedpass/tui_v2/app.py:
- action-strip click coordinate handling (zero-based)
- settings shortcut restored to palette prefix behavior (setting-)
- managed-load validation message compatibility
- profile-switch now auto-expands active profile-tree branch
- sensitive confirmation status text normalized to include "requires confirmation"
Added TUI v3 focused coverage gate script: scripts/tui3_coverage_gate.sh and wired it into scripts/run_ci_tests.sh behind TUI3_COVERAGE_GATE (default on).
Hardened TUI v3 sensitive payload resolution (src/seedpass/tui_v3/app.py) for legacy service compatibility:
- PGP payloads may be 2-tuple or 3-tuple.
- Seed getters may accept (entry_id) or (entry_id, parent_seed).
Added regression tests in src/tests/test_tui_v3_parity.py to lock this compatibility behavior.
Current CI-equivalent state:
- Full main suite passes (1064 passed, 16 skipped), but existing tui2_coverage_gate.sh still fails src/seedpass/core/api.py >= 85 on its focused subset (68.79%) and needs separate gate policy/workload adjustment.

Root cause for the reported v3 "main index list / scrolling" bug was focus management, not DataTable scrolling itself:
- first render focus landed on the hidden command palette input
- arrow-key navigation did nothing until the entry table was manually focused
Stable fix:
- MainScreen.on_mount() now focuses #entry-data-table
- closing the command palette restores focus to #entry-data-table instead of clearing focus
Reusable v3 search/navigation pattern:
- persist the active search query in app state so refresh/filter/sort changes do not silently clear the current result set
- keep sort and mode changes as explicit app actions (action_set_search_mode, action_set_search_sort) so toolbar buttons and palette commands stay aligned
Search/graph baseline shipped:
- SearchService.linked_neighbors(...) now returns deterministic incoming/outgoing link cards
- SearchService.relation_summary(...) now returns grouped relation counts
- TUI v3 inspector now exposes a linked-items panel with direct open-entry actions
- TUI v3 grid now has explicit filter/mode/sort buttons instead of palette-only control