Add a derived 'recommended_action' property on InspectedResult that
collapses the orchestrator's go / wait / skip decision into one value:
- 'skip' → no main_video, or media_type == 'other'
- 'ask_user' → media_type == 'unknown', or road == 'path_of_pain'
- 'process' → confident parse with a main video on disk
The ordering is part of the contract (skip > ask_user > process) —
documented in the property docstring.
Until now every consumer (workflows, the agent, the orchestrator
sketch) had to re-derive this from the road / media_type / main_video
triple, with subtle drift between sites. One place, one rule.
Exposed through the analyze_release tool so the LLM can route on it.
Spec YAML updated to describe the new field.
Suite: 1083 passed (+6 new tests in tests/application/test_inspect.py
covering the four branches and the precedence rules).
New application-layer entry point that composes the four inspection
layers in one call:
1. parse_release(name, kb) -> (ParsedRelease, ParseReport)
2. detect_media_type(parsed, path, kb) -> patch parsed.media_type
3. find_main_video(path, kb) -> Path | None (top-level scan)
4. prober.probe(video) + enrich -> when video exists and
media_type not in
{unknown, other}
Returns a frozen InspectedResult(parsed, report, source_path,
main_video, media_info, probe_used). kb and prober are injected — no
module-level singletons in inspect.py.
analyze_release tool now delegates to inspect_release; its output
gains two fields, confidence (0-100) and road (easy/shitty/path_of_pain),
surfaced from ParseReport so the LLM can route by confidence. Spec
updated to document them.
12 new tests covering happy paths, probe gating (no video, media_type
'other', probe failure), mutation contract (detect refining
parsed.media_type, enrich filling None fields), resilience
(nonexistent path), and frozen contract. Suite: 1058 passing.
Adds two STM components and a transparent cache hook in the agent loop so
read-only tools don't re-do work the agent already did in this session.
New STM components:
- ToolResultsCache — {tool_name: {key: result}}, session-scoped.
to_dict() exposes only the key inventory (not payloads) to keep the
prompt cheap.
- ReleaseFocus — current_release_path + working_set list, updated
automatically when a path-keyed inspector runs.
YAML spec layer:
- New optional 'cache: { key: <param_name> }' block in ToolSpec.
- Validated at load time: cache.key must be a declared parameter.
- Surfaced on Tool dataclass as cache_key: str | None.
Agent._execute_tool_call:
- Pre-exec cache lookup; hit short-circuits and adds _from_cache=true.
- Post-exec: stores successful results, updates release_focus for
path-keyed tools, refreshes episodic.last_search_results when
find_torrent's hit served the response (so get_torrent_by_index
keeps pointing at the right list).
Cacheable tools (5): analyze_release, probe_media, list_folder,
find_media_imdb_id, find_torrent.
Adds YAML specs for the 14 tools that were still description-from-docstring:
filesystem:
- set_path_for_folder, list_folder, analyze_release, probe_media,
move_media, manage_subtitles, create_seed_links, learn
api:
- find_media_imdb_id, find_torrent, get_torrent_by_index,
add_torrent_to_qbittorrent, add_torrent_by_index
language:
- set_language
Each spec follows the established shape (summary / description /
when_to_use / when_not_to_use / next_steps / parameters with
why_needed + example / returns) and the Python function docstring is
slimmed to a one-line pointer.
Registry now reports: 21 tools, 21 with YAML spec, 0 doc-only.
francwa