claudeglass 0.9.0__py3-none-any.whl

claudeglass/__init__.py

@@ -0,0 +1,192 @@
+"""claudeglass: config-aware token and prompt-cache analytics for
+Claude Code transcripts.
+"""
+
+from __future__ import annotations
+
+__version__ = "0.9.0"
+
+#: Bump when transcript-parsing logic changes in a way that could change
+#: results computed from a previously cached file.
+#:
+#: Bumped to 2 by the wp11-cache/wp12a-fixtures merge: parse.py's
+#: ttl_split_unknown/pre_split_turns fix (main's 709239a) and
+#: workstyle.py's chat-only-before-single-model reorder (main's 0f892e3)
+#: both landed on main after wp11-cache branched from 25c68c1 (before
+#: PARSER_VERSION existed), so a cache entry written under the old
+#: parsing behaviour must not be treated as still valid.
+#:
+#: Bumped to 3 by f99901f (fix(privacy): redact relative Windows paths
+#: and @-tokens at parse time): ``_redact_paths`` in parse.py now catches
+#: relative ``Users\``/``home/`` paths and any ``@``-bearing token that
+#: previously survived redaction into ``cmd_prefix``, so a digest cached
+#: from before that fix reflects a leakier ``cmd_prefix`` and must be
+#: treated as stale.
+#:
+#: Bumped to 4 by the capture-improvements batch: parse.py now derives
+#: seven new additive ``Turn`` fields from a transcript's raw lines
+#: (``tool_wait_s``/``model_latency_s``/``tool_result_chars_by_tool``,
+#: ``agent_brief_chars``/``tool_input_chars_by_tool``,
+#: ``read_target_hashes``, ``human_prompt_chars``/``human_prompt_has_paste``
+#: -- see model.py's module docstring), none of which a pre-batch digest
+#: cache entry ever computed, so it must not be treated as still valid.
+#:
+#: Bumped to 5 by the v3-limits batch: parse.py now detects usage-limit
+#: pauses (LIMIT_HIT/LIMIT_RESUME/AGENT_TERMINATED events,
+#: Turn.synthetic_kind, Turn.gap_cause -- see model.py's module
+#: docstring), none of which a pre-batch digest cache entry ever
+#: computed, so it must not be treated as still valid.
+#:
+#: Bumped to 6 by the v4-wasted-turns batch: parse.py now derives two new
+#: additive ``Turn`` fields, ``tool_error_count``/``tool_error_chars``,
+#: from each turn's own tool_result blocks that carry ``is_error: true``
+#: (see model.py's module docstring), which no pre-batch digest cache
+#: entry ever computed, so it must not be treated as still valid.
+#:
+#: Bumped to 7 by the readability batch: events.py now measures an
+#: attachment's ``Event.size_chars`` from the real ``rendered`` shape (a
+#: list of ``{"content": str}`` blocks) or its content fields, where it
+#: previously read only a bare string and so recorded ``None``; it also
+#: adds per-source ``instructions`` sizes and ``prompt_snapshot``
+#: system/tool sizes to ``Event.detail``.
+#:
+#: Bumped to 8 by the context-files batch: events.py now keeps one record
+#: per instruction file (salted path hash, type, path-scoped, size) for
+#: ``instructions`` and ``nested_memory``, and one per skill (name, size)
+#: for ``skill_listing`` and ``invoked_skills``; parse.py adds
+#: ``Turn.skills_invoked``.
+#:
+#: Bumped to 9 by the quality-signals batch: parse.py adds
+#: ``Turn.stop_reason``, ``tool_calls_by_tool``, ``tool_errors_by_tool``, ``edit_target_hashes``
+#: and ``human_correction``; events.py records task-notification and
+#: agent-result outcomes in ``Event.detail``.
+#:
+#: Bumped to 10 by the same batch: discovery.py records a workflow
+#: agent's end state on ``TranscriptMeta.workflow_agent_state``, which a
+#: version-9 digest (already written by a development build) lacks.
+#:
+#: Bumped to 11: parse.py took a turn's usage from the first line of a
+#: streamed reply, whose output_tokens is a partial count and which has no
+#: thinking_tokens; it now keeps the most complete snapshot among the
+#: reply's lines, so older digests undercount output and thinking tokens.
+#:
+#: Bumped to 12: parse.py records why each failed tool call failed
+#: (``Turn.tool_errors_by_kind``), which waste.py needs to tell a failing
+#: test or a hook block from a call that couldn't run.
+#:
+#: Bumped to 13 by the three-pricing-fixes batch: parse.py now records
+#: ``Turn.speed`` from each turn's own ``usage.speed`` (see model.py's
+#: module docstring), which pricing.py needs to apply a model's fast-mode
+#: rate multiplier. A pre-13 digest has no ``speed`` recorded, so every
+#: transcript is re-parsed once to pick it up.
+#:
+#: Bumped to 14 by the edit-capture batch: ``Turn.edit_target_hashes`` now
+#: covers MultiEdit and the files a Bash or PowerShell command writes
+#: (``shell_writes.py``), drops edits whose tool call failed, and paths
+#: are normalised further before hashing (Git Bash's ``/c/`` form, ``.``
+#: and ``..``), so older digests undercount edits and can't match a file
+#: written by a shell command to the same file edited with Edit. The same
+#: bump adds ``Turn.retry_marker``/``result_marker`` (the quality markers
+#: Claude can be asked to write; see ``quality.MARKER_LINES``).
+#:
+#: Bumped to 15 by the metrics-capture batch: parse.py reads capture tags
+#: (``Turn.cap``, ``spawn_marker``, ``retry_marker`` gains ``scope``) and
+#: capture notes (``Turn.cap_note_chars``, ``TranscriptMeta.cap_*``), and
+#: records ``agent_result_chars``, ``prompt_flags``, ``plan_stats`` and
+#: ``commands_run``. Three miscounts are fixed at the same time:
+#: ``read_target_hashes`` covered edits as well as reads, so every edited
+#: file looked re-read; ``hook_system_message`` lines (shown to you, never
+#: to the model) were counted as hook context; and task notifications
+#: weren't sized, so a background agent's report had no size.
+#:
+#: Bumped to 16 by the feedback batch: parse.py reads your /tl-feedback
+#: answers (``Turn.feedback``) from the skill's ``[tl-fb: ...]`` line or,
+#: failing that, from the AskUserQuestion result itself. ``commands_run``
+#: now names skills you ran with a slash too: they are written
+#: ``<command-message>`` first, as your message, and were missed.
+#:
+#: Bumped to 17 by the P2 capture-integrity batch: capture_tags.py now
+#: strips the exact reminder sentence before the tail match (a tag glued
+#: to a reminder no longer swallows it); a ``[tl-fb: ...]`` line only
+#: counts when the cycle's first turn actually ran ``/tl-feedback``, a
+#: forged one is ignored; a captured tag key is kept only when the
+#: session's note said to ask for it (``cap_injections > 0`` and the key
+#: is a requested metric); ``reported_task`` counts once per cycle, not
+#: once per line; ``Turn.skills_invoked`` now runs skill names through
+#: ``SKILL_NAME_PATTERN`` and drops one whose ``Skill`` call errored, so a
+#: skill can no longer self-authorise a capture note by name alone; and
+#: capture.py's coverage denominator drops feedback/interrupted/max_tokens
+#: cycles, merges a cycle's tags key by key instead of replacing the
+#: whole set, sizes ``_big_output`` per call instead of per turn, and
+#: ``_WRAP`` carries a ``:Tool`` suffix. None of this is recoverable from
+#: an older digest, so every transcript is re-parsed once to pick it up.
+#:
+#: Bumped to 18 by the P3 hook-lifecycle batch: a HOOK_OUTPUT event's
+#: ``Event.detail`` now carries the closed hook-event bucket it ran under
+#: (``hookName``, e.g. ``"PreToolUse"``, never the matcher/tool-name
+#: suffix -- SURV-HE, G7), its real ``durationMs`` when Claude Code
+#: recorded one (CAP-9/F10: this used to be dropped), and, only when
+#: ``True``, whether the call ran ClaudeGlass's own capture hook script
+#: (``capture``, never the command string itself). ``hook_health.py``'s
+#: new ``count_hook_errors``/``measure_deep_wait`` both read these
+#: straight off already-parsed events; a pre-18 digest has none of them,
+#: so every transcript is re-parsed once to pick them up.
+#: Bumped to 19 by the parser-signals batch (SURV-4/5/6/7): ``thinking_drop``
+#: joins the CACHE_SIGNAL family; ``task_status``/``structured_output``
+#: attachments get their own ``EventKind``s instead of falling into the
+#: generic ATTACHMENT catch-all; ``cost-state`` lines are read for their
+#: own ``totalCostUSD``/``hasUnknownModelCost`` (``TranscriptMeta.
+#: cc_cost_usd``/``cc_cost_has_unknown_model``) instead of falling
+#: through to UNKNOWN; a tool_result's or human prompt's own image/
+#: document content blocks are sized by the documented image-token rule
+#: instead of silently counting as 0 chars; and a line type no detection
+#: rule recognises at all is now counted separately
+#: (``TranscriptResult.parser_notes["unknown_line_types"]``) from one
+#: this parser knows about and deliberately ignores. None of this is
+#: recoverable from an older digest, so every transcript is re-parsed
+#: once to pick it up.
+#:
+#: Bumped to 20 by the docs-and-privacy sweep: ``Diagnostics.
+#: ignored_line_types`` now keys on the sanitised line type (or
+#: ``"other"``), like ``unknown_line_types`` already did, instead of the
+#: raw ``type`` straight off the wire. A cached pre-20 digest still holds
+#: the raw keys, so every transcript is re-parsed once to drop them.
+#:
+#: Bumped to 21: ``capture_tags.filter_tag`` now drops tag keys only the
+#: other scope is asked for, and a subagent's ``[tl: ...]`` no longer
+#: counts as a tag (it could set a whole prompt cycle's task or level);
+#: a subagent's ``out=`` is kept when a large-output note asked for it.
+#: A cached pre-21 digest still holds the unfiltered tags.
+#:
+#: Bumped to 22: hook events now carry the hook's label (its script's file
+#: name), why a failing hook failed and whether its script path is
+#: relative; turns carry the context your hooks added and the calls they
+#: blocked (and any unchanged re-send). None of this is in a pre-22
+#: digest.
+#:
+#: Bumped to 23: a hook event also says whether its command uses a
+#: Windows ``%VAR%`` variable, and a quoted script path with a tab in it
+#: is no longer read as relative. A pre-23 digest has neither.
+#:
+#: Bumped to 24: discovery.py gives an agent under a workflow run's folder
+#: ``TranscriptMeta.kind == "workflow-agent"``; it was ``"subagent"``, so
+#: every table split by kind folded workflow agents into subagents. A
+#: pre-24 digest (the parse cache's and the service store's alike) still
+#: holds the old kind.
+#:
+#: Bumped to 25: a coaching note (``tl-coach v``, the capture hook's live
+#: hints) is its own ``coaching_note`` event, and counts to
+#: ``Turn.cap_note_chars`` but not to ``cap_injections``; a capture note
+#: sharing an attachment with one keeps the coaching part apart
+#: (``detail["coach"]``/``["coach_chars"]``). Notes on a message you send
+#: (``UserPromptSubmit``) keep that hook name. A pre-25 digest counted
+#: neither.
+PARSER_VERSION = 25
+
+#: Bump when the model.py contract changes in a way that invalidates the
+#: on-disk digest cache (see model.py's module docstring for the contract
+#: rules: fields may be added with defaults, never renamed or removed
+#: without a SCHEMA_VERSION bump).
+SCHEMA_VERSION = 1
+
+__all__ = ["__version__", "PARSER_VERSION", "SCHEMA_VERSION"]

claudeglass/__main__.py

@@ -0,0 +1,17 @@
+"""``python -m claudeglass`` entry point.
+
+Also the entry point a ``python -m zipapp``-built ``claudeglass.pyz``
+uses (see the plan's "Distribution without pip" section): zipapp runs
+``__main__.py`` at the archive root, and without ``sys.exit(main())``
+here the process would always exit 0 regardless of what ``main()``
+returned, silently losing every non-zero exit code (no-data, bad-input,
+...) a caller or CI script depends on.
+"""
+
+from __future__ import annotations
+
+import sys
+
+from .cli import main
+
+sys.exit(main())