feat(report): add multi-backend summarizer and LLM task grouping - #633
Merged
Conversation
Contributor
|
The latest updates on your projects. Learn more about Vercel for GitHub. 1 Skipped Deployment
|
![cubic-dev-ai[bot]](https://avatars.githubusercontent.com/in/1082092?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
10 issues found across 8 files
Prompt for AI agents (unresolved issues)
Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.
<file name="crates/tokscale-core/src/content_extractor.rs">
<violation number="1" location="crates/tokscale-core/src/content_extractor.rs:262">
P2: The truncation guard uses byte length instead of character count, which incorrectly appends ellipses for non-ASCII input.</violation>
</file>
<file name="crates/tokscale-cli/src/commands/report.rs">
<violation number="1" location="crates/tokscale-cli/src/commands/report.rs:208">
P2: Unknown `--summarizer` values return `Ok(())` instead of an error, causing silent misconfiguration and skipped summarization.</violation>
<violation number="2" location="crates/tokscale-cli/src/commands/report.rs:250">
P1: DB errors are swallowed when writing summaries, which can silently lose summarization results while reporting success.</violation>
<violation number="3" location="crates/tokscale-cli/src/commands/report.rs:327">
P2: Task grouping is silently skipped for unsupported backends (including the default `apple-fm`), so the feature can appear to run successfully while never producing task groups.</violation>
<violation number="4" location="crates/tokscale-cli/src/commands/report.rs:347">
P2: DB errors are swallowed when saving `task_group`, so grouping can silently fail while the command reports success.</violation>
<violation number="5" location="crates/tokscale-cli/src/commands/report.rs:727">
P2: End-of-day filtering truncates the last 999ms, so some sessions at the end of the `--until` day are incorrectly excluded.</violation>
</file>
<file name="crates/tokscale-core/src/wiki.rs">
<violation number="1" location="crates/tokscale-core/src/wiki.rs:116">
P2: Fallback config path uses an unexpanded `~`, which can write the wiki DB to an unintended location.</violation>
<violation number="2" location="crates/tokscale-core/src/wiki.rs:375">
P2: The `until` filter is inclusive in `query_entries` but exclusive in other range methods, causing inconsistent date-scoped behavior.</violation>
</file>
<file name="scripts/wiki-summarizer.py">
<violation number="1" location="scripts/wiki-summarizer.py:129">
P2: Validate `task_category`/`complexity` against allowed values before storing. Right now any model output string is accepted, even when it violates the declared schema.</violation>
<violation number="2" location="scripts/wiki-summarizer.py:136">
P2: The recovery path can crash because `session['session_id']` is dereferenced inside the exception handler. A malformed session then raises a second `KeyError` and aborts summarization.</violation>
</file>
Reply with feedback, questions, or to request a fix.
Re-trigger cubic
Contributor
There was a problem hiding this comment.
6 issues found across 5 files (changes from recent commits).
Prompt for AI agents (unresolved issues)
Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.
<file name="crates/tokscale-cli/src/commands/optimize.rs">
<violation number="1" location="crates/tokscale-cli/src/commands/optimize.rs:195">
P2: `partial_cmp(...).unwrap()` on `f64` can panic if `total_cost` is NaN. Use the safer pattern already established in the codebase.</violation>
<violation number="2" location="crates/tokscale-cli/src/commands/optimize.rs:455">
P2: Truncating `String` with `[..27]` can panic on non-ASCII model names due to invalid UTF-8 boundary slicing.</violation>
</file>
<file name="crates/tokscale-cli/src/main.rs">
<violation number="1" location="crates/tokscale-cli/src/main.rs:763">
P1: `--optimize` currently runs during `--json` report output, appending human-formatted text and corrupting JSON output for automation.</violation>
</file>
<file name="README.md">
<violation number="1" location="README.md:683">
P2: The new optimize example claims it defaults to today, but the command actually uses all sessions unless a date flag is provided.</violation>
</file>
Reply with feedback, questions, or to request a fix.
Re-trigger cubic
| week, | ||
| month, | ||
| }); | ||
| if optimize && result.is_ok() { |
Contributor
There was a problem hiding this comment.
P1: --optimize currently runs during --json report output, appending human-formatted text and corrupting JSON output for automation.
Prompt for AI agents
Check if this issue is valid — if so, understand the root cause and fix it. At crates/tokscale-cli/src/main.rs, line 763:
<comment>`--optimize` currently runs during `--json` report output, appending human-formatted text and corrupting JSON output for automation.</comment>
<file context>
@@ -745,6 +759,38 @@ fn main() -> Result<()> {
week,
month,
+ });
+ if optimize && result.is_ok() {
+ let _ = commands::optimize::run_optimize(commands::optimize::OptimizeOptions {
+ json: false,
</file context>
Suggested change
| if optimize && result.is_ok() { | |
| if optimize && !json && result.is_ok() { |
| println!(" {}", "─".repeat(68)); | ||
| for m in report.model_insights.iter().take(8) { | ||
| let model_display: String = if m.model.len() > 28 { | ||
| format!("{}…", &m.model[..27]) |
Contributor
There was a problem hiding this comment.
P2: Truncating String with [..27] can panic on non-ASCII model names due to invalid UTF-8 boundary slicing.
Prompt for AI agents
Check if this issue is valid — if so, understand the root cause and fix it. At crates/tokscale-cli/src/commands/optimize.rs, line 455:
<comment>Truncating `String` with `[..27]` can panic on non-ASCII model names due to invalid UTF-8 boundary slicing.</comment>
<file context>
@@ -0,0 +1,554 @@
+ println!(" {}", "─".repeat(68));
+ for m in report.model_insights.iter().take(8) {
+ let model_display: String = if m.model.len() > 28 {
+ format!("{}…", &m.model[..27])
+ } else {
+ m.model.clone()
</file context>
Suggested change
| format!("{}…", &m.model[..27]) | |
| format!("{}…", m.model.chars().take(27).collect::<String>()) |
| Tokscale can analyze your usage patterns and generate actionable recommendations to reduce costs and improve productivity. | ||
|
|
||
| ```bash | ||
| # Standalone optimization analysis (defaults to today) |
Contributor
There was a problem hiding this comment.
P2: The new optimize example claims it defaults to today, but the command actually uses all sessions unless a date flag is provided.
Prompt for AI agents
Check if this issue is valid — if so, understand the root cause and fix it. At README.md, line 683:
<comment>The new optimize example claims it defaults to today, but the command actually uses all sessions unless a date flag is provided.</comment>
<file context>
@@ -674,6 +675,65 @@ tokscale report --workspace my-project --client opencode
+Tokscale can analyze your usage patterns and generate actionable recommendations to reduce costs and improve productivity.
+
+```bash
+# Standalone optimization analysis (defaults to today)
+tokscale optimize
+
</file context>
Suggested change
| # Standalone optimization analysis (defaults to today) | |
| # Standalone optimization analysis (uses all cached sessions by default) |
leecoder
force-pushed
the
feat/report-task-grouping
branch
3 times, most recently
from
May 29, 2026 09:47
2cf54a9 to
36c4ce5
Compare
- Group sessions by model and task title in summary tables - Show daily breakdown for --week/--month, session list for --today - Integrate Apple FM summarizer for session classification - Add wiki DB for caching session summaries
- Support claude, codex, gemini, kiro as summarizer backends (in addition to apple-fm) - Add 2nd LLM pass to cluster sessions into high-level task groups - Add --summarizer flag to select backend, --rebuild to reset cached summaries - Scope summarization to date range (--week, --since/--until) - Add task_group column to wiki DB with migration - Update README with Task-Attributed Report documentation
…agate DB errors - Return error instead of Ok(()) for unknown --summarizer values - Propagate DB errors from update_summary and update_task_group - Clarify skip message when task grouping backend is unsupported (apple-fm)
- Fix end-of-day filtering: use next_day_00:00 - 1ms instead of 23:59:59
- Fix wiki.rs fallback path: use dirs::home_dir() instead of literal '~/.config'
- Fix until filter inconsistency: query_entries now uses '<' (exclusive) matching other range methods
- Validate task_category/complexity against allowed values in wiki-summarizer.py
- Fix recovery path crash: use session.get('session_id') instead of session['session_id'] in exception handler
- content_extractor: use chars().count() instead of byte len() for truncation guard - wiki.rs: rename from_str to parse to avoid clippy::should_implement_trait - report.rs: use div_ceil() instead of manual ceiling division - usage/copilot.rs: use strip_prefix() instead of manual prefix stripping - usage/minimax.rs: remove redundant closure - usage/zai.rs: fix reference-to-reference pattern - usage/mod.rs: extract type alias for complex type, use .ok() instead of manual match
leecoder
force-pushed
the
feat/report-task-grouping
branch
from
June 1, 2026 03:00
36c4ce5 to
d6c5432
Compare
Contributor
There was a problem hiding this comment.
1 issue found across 5 files (changes from recent commits).
Prompt for AI agents (unresolved issues)
Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.
<file name="crates/tokscale-cli/src/main.rs">
<violation number="1" location="crates/tokscale-cli/src/main.rs:763">
P1: `--optimize` currently runs during `--json` report output, appending human-formatted text and corrupting JSON output for automation.</violation>
</file>
<file name="crates/tokscale-cli/src/commands/optimize.rs">
<violation number="1" location="crates/tokscale-cli/src/commands/optimize.rs:455">
P2: Truncating `String` with `[..27]` can panic on non-ASCII model names due to invalid UTF-8 boundary slicing.</violation>
</file>
<file name="README.md">
<violation number="1" location="README.md:683">
P2: The new optimize example claims it defaults to today, but the command actually uses all sessions unless a date flag is provided.</violation>
</file>
<file name=".opencode/skill/deploy.md">
<violation number="1" location=".opencode/skill/deploy.md:49">
P2: `docker compose exec` allocates a TTY by default; in non-interactive environments like AWS SSM this can fail with "the input device is not a TTY". Use `-T` to disable pseudo-TTY allocation.</violation>
</file>
Tip: Review your code locally with the cubic CLI to iterate faster.
Re-trigger cubic
| --instance-ids "i-078fe82953c3047b5" \ | ||
| --document-name "AWS-RunShellScript" \ | ||
| --region ap-northeast-2 \ | ||
| --parameters 'commands=["export HOME=/home/ubuntu && cd /home/ubuntu/tokscale/self-host && docker compose exec app npx drizzle-kit push --force"]' \ |
Contributor
There was a problem hiding this comment.
P2: docker compose exec allocates a TTY by default; in non-interactive environments like AWS SSM this can fail with "the input device is not a TTY". Use -T to disable pseudo-TTY allocation.
Prompt for AI agents
Check if this issue is valid — if so, understand the root cause and fix it. At .opencode/skill/deploy.md, line 49:
<comment>`docker compose exec` allocates a TTY by default; in non-interactive environments like AWS SSM this can fail with "the input device is not a TTY". Use `-T` to disable pseudo-TTY allocation.</comment>
<file context>
@@ -0,0 +1,61 @@
+ --instance-ids "i-078fe82953c3047b5" \
+ --document-name "AWS-RunShellScript" \
+ --region ap-northeast-2 \
+ --parameters 'commands=["export HOME=/home/ubuntu && cd /home/ubuntu/tokscale/self-host && docker compose exec app npx drizzle-kit push --force"]' \
+ --timeout-seconds 120 \
+ --output json
</file context>
leecoder
force-pushed
the
feat/report-task-grouping
branch
2 times, most recently
from
June 1, 2026 09:01
7657640 to
f00d277
Compare
…pair build after rebase After merging main, the report arm called build_date_filter with the old 5-arg form; the DateRangeFlags refactor changed it to take &DateRangeFlags. compute_msg_cost hardcoded flat per-token rates that only matched Sonnet and ignored cache_write/reasoning; it now prices through the canonical PricingService (calculate_cost_with_provider) exactly like the aggregator, loading a fresh dataset with a cached fallback for offline use. Display truncation used byte slicing on LLM-generated strings (UTF-8 boundary panic), per-model attribution divided by models_used.len() with no zero guard, and cost sorts used partial_cmp().unwrap() (NaN panic). Constraint: pricing must come from canonical PricingService, never hardcoded constants or fuzzy matching Confidence: medium Scope-risk: moderate Not-tested: live LLM summarizer backends (apple-fm / claude/codex/gemini CLIs)
Contributor
There was a problem hiding this comment.
1 issue found across 3 files (changes from recent commits).
Prompt for AI agents (unresolved issues)
Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.
<file name="crates/tokscale-cli/src/commands/report.rs">
<violation number="1" location="crates/tokscale-cli/src/commands/report.rs:114">
P1: Cost attribution can be permanently zeroed for newly ingested sessions when pricing data is temporarily unavailable</violation>
</file>
Tip: Review your code locally with the cubic CLI to iterate faster.
Re-trigger cubic
| }) | ||
| .map_err(|e| anyhow::anyhow!("{}", e))?; | ||
|
|
||
| let pricing = load_pricing_service(); |
Contributor
There was a problem hiding this comment.
P1: Cost attribution can be permanently zeroed for newly ingested sessions when pricing data is temporarily unavailable
Prompt for AI agents
Check if this issue is valid — if so, understand the root cause and fix it. At crates/tokscale-cli/src/commands/report.rs, line 114:
<comment>Cost attribution can be permanently zeroed for newly ingested sessions when pricing data is temporarily unavailable</comment>
<file context>
@@ -112,6 +111,8 @@ fn populate_wiki_from_sessions(db: &WikiDb, opts: &ReportOptions) -> Result<()>
})
.map_err(|e| anyhow::anyhow!("{}", e))?;
+ let pricing = load_pricing_service();
+
let mut session_map: HashMap<String, SessionAgg> = HashMap::new();
</file context>
Every LLM backend invocation in run_task_grouping, run_cli_summarizer, and the apple-fm path spawned a child and blocked forever with no timeout, so an auth prompt or network stall could hang `tokscale report` indefinitely. Add a pure-std run_command_with_timeout helper (spawn + reader threads draining stdout/stderr + try_wait deadline loop + kill on timeout, mirroring run_capture_command in main.rs) bounded by a generous 300s BACKEND_TIMEOUT. A timeout surfaces as an io::ErrorKind::TimedOut that each call site now degrades gracefully (log + skip the backend), matching the existing non-zero-exit behavior. The apple-fm path keeps its stdin-write behavior via the helper's optional stdin_bytes argument. Constraint: pure-std timeout (spawn + try_wait deadline + kill), mirroring main.rs — no new dependency Confidence: high Scope-risk: narrow
This was referenced Jun 17, 2026
junhoyeo
added a commit
that referenced
this pull request
Jun 22, 2026
… keying Addresses automated review feedback on the #633 report-summarizer-content fix (PR #746). The summarizer still surfaced (none) for normal Codex/Gemini sessions and could mis-route cross-client session_id collisions. - content_extractor: parse the current on-disk Codex format (event_msg with payload.type == "user_message", text in payload.message) and skip harness-injected context blocks (<environment_context>/<system-reminder>/ <user_instructions>), mirroring sessions::codex. - content_extractor: Gemini extractor now handles chat-recording JSON (messages[].type == "user" / content) and falls back to scanning line-delimited JSONL; empty/whitespace user text is treated as not-found. - extract_session_content: an empty/whitespace first_user_message no longer counts as success, so scanning continues to a later candidate with real text. - report: SessionPathIndex is keyed by (client, session_id) to prevent cross-client collisions, and Gemini files are keyed by their in-file sessionId (via gemini_session_id_for_file) rather than the filename stem, since the wiki entry's session_id is derived from inside the file. - Added fixture-based regression tests for all of the above. Constraint: wiki session_id for Gemini comes from the in-file sessionId, not the path stem Rejected: match any leading '<' for Codex injected blocks | drops legit prompts starting with markup Confidence: high Scope-risk: narrow
junhoyeo
added a commit
that referenced
this pull request
Jun 22, 2026
…746) * docs: sync EN/ja/ko/zh-cn for unreleased clients + breaking flag removal Pre-v3.2.0 documentation sync for the unreleased range (v3.1.3..main): - #465: replace the now-false legacy per-client flag notices in ja/ko/zh-cn (they claimed the removed flags still work) with a v3.2.0 breaking-change migration note; add the same note to README.md (which had none). - #728: document the MiniMax Token Plan subscription source (distinct from the MINIMAX_API_KEY row) in all locales; port the entire Subscription Usage section into ja/ko/zh-cn (was English-only). - #718: add the Jcode table row + detail section to the locales missing them. - #726: document the TOKSCALE_FM_DEBUG env var in all locales. - #633: add the missing task-attributed report bullet to README.ja Key Features. - drift: add Junie to the frontend Source-filtering list (all locales). - #710: fix the MiMo Code repo link (XiaomiMiMo/MiMo -> XiaomiMiMo/MiMo-Code). - #717: disclose Command Code token usage is estimated (~4 chars/token). Confidence: medium Scope-risk: narrow Directive: ja/ko/zh-cn translations of the ported Subscription Usage section are machine-generated and should get a native-speaker review pass Not-tested: #713 Antigravity CLI detail section was not added — no English source section exists to port from * fix(report): feed real session content to the summarizer (#633) extract_content_for_session unconditionally returned metadata_only_content() (first_user_message hardcoded None), so the report summarizer never saw any conversation content and the four real per-client extractors were dead code. Add content_extractor::extract_session_content, which dispatches to the correct per-client extractor (opencode/claude/codex/gemini) and falls back to metadata-only — never erroring or panicking — for unknown clients, missing candidates, or unreadable/unparseable files. report.rs builds a SessionPathIndex once (session_id -> transcript file, plus opencode DBs) and threads it through run_summarizer so each payload carries the real first user message. Confidence: high Scope-risk: moderate Rejected: thread file paths through core's scanner/WikiEntry | too invasive; indexed at the report layer instead Not-tested: end-to-end opencode/codex/gemini extraction in report.rs (core dispatcher covers claude + all fallback paths; per-client extractors are pre-existing) * docs: name the breaking release v4.0.0 (was v3.2.0) The per-client flag removal (#465) is a breaking change, so the next release is v4.0.0, not v3.2.0. Update the migration notes in all four README locales and the main.rs doc comment accordingly. * fix(report): real Codex/Gemini extraction + (client,session_id) index keying Addresses automated review feedback on the #633 report-summarizer-content fix (PR #746). The summarizer still surfaced (none) for normal Codex/Gemini sessions and could mis-route cross-client session_id collisions. - content_extractor: parse the current on-disk Codex format (event_msg with payload.type == "user_message", text in payload.message) and skip harness-injected context blocks (<environment_context>/<system-reminder>/ <user_instructions>), mirroring sessions::codex. - content_extractor: Gemini extractor now handles chat-recording JSON (messages[].type == "user" / content) and falls back to scanning line-delimited JSONL; empty/whitespace user text is treated as not-found. - extract_session_content: an empty/whitespace first_user_message no longer counts as success, so scanning continues to a later candidate with real text. - report: SessionPathIndex is keyed by (client, session_id) to prevent cross-client collisions, and Gemini files are keyed by their in-file sessionId (via gemini_session_id_for_file) rather than the filename stem, since the wiki entry's session_id is derived from inside the file. - Added fixture-based regression tests for all of the above. Constraint: wiki session_id for Gemini comes from the in-file sessionId, not the path stem Rejected: match any leading '<' for Codex injected blocks | drops legit prompts starting with markup Confidence: high Scope-risk: narrow
pinion05
added a commit
to pinion05/tokscale
that referenced
this pull request
Jun 23, 2026
…hoyeo#633) * feat(report): add report command with model/task/daily breakdown - Group sessions by model and task title in summary tables - Show daily breakdown for --week/--month, session list for --today - Integrate Apple FM summarizer for session classification - Add wiki DB for caching session summaries * feat(report): add multi-backend summarizer and LLM task grouping - Support claude, codex, gemini, kiro as summarizer backends (in addition to apple-fm) - Add 2nd LLM pass to cluster sessions into high-level task groups - Add --summarizer flag to select backend, --rebuild to reset cached summaries - Scope summarization to date range (--week, --since/--until) - Add task_group column to wiki DB with migration - Update README with Task-Attributed Report documentation * fix(report): address cubic review - error on unknown summarizer, propagate DB errors - Return error instead of Ok(()) for unknown --summarizer values - Propagate DB errors from update_summary and update_task_group - Clarify skip message when task grouping backend is unsupported (apple-fm) * fix(report): address remaining cubic review issues - Fix end-of-day filtering: use next_day_00:00 - 1ms instead of 23:59:59 - Fix wiki.rs fallback path: use dirs::home_dir() instead of literal '~/.config' - Fix until filter inconsistency: query_entries now uses '<' (exclusive) matching other range methods - Validate task_category/complexity against allowed values in wiki-summarizer.py - Fix recovery path crash: use session.get('session_id') instead of session['session_id'] in exception handler * fix: resolve clippy lint errors - content_extractor: use chars().count() instead of byte len() for truncation guard - wiki.rs: rename from_str to parse to avoid clippy::should_implement_trait - report.rs: use div_ceil() instead of manual ceiling division - usage/copilot.rs: use strip_prefix() instead of manual prefix stripping - usage/minimax.rs: remove redundant closure - usage/zai.rs: fix reference-to-reference pattern - usage/mod.rs: extract type alias for complex type, use .ok() instead of manual match * fix(report): use canonical pricing, fix UTF-8/NaN/div-zero panics, repair build after rebase After merging main, the report arm called build_date_filter with the old 5-arg form; the DateRangeFlags refactor changed it to take &DateRangeFlags. compute_msg_cost hardcoded flat per-token rates that only matched Sonnet and ignored cache_write/reasoning; it now prices through the canonical PricingService (calculate_cost_with_provider) exactly like the aggregator, loading a fresh dataset with a cached fallback for offline use. Display truncation used byte slicing on LLM-generated strings (UTF-8 boundary panic), per-model attribution divided by models_used.len() with no zero guard, and cost sorts used partial_cmp().unwrap() (NaN panic). Constraint: pricing must come from canonical PricingService, never hardcoded constants or fuzzy matching Confidence: medium Scope-risk: moderate Not-tested: live LLM summarizer backends (apple-fm / claude/codex/gemini CLIs) * style: apply rustfmt * fix(report): bound LLM summarizer subprocesses with a timeout Every LLM backend invocation in run_task_grouping, run_cli_summarizer, and the apple-fm path spawned a child and blocked forever with no timeout, so an auth prompt or network stall could hang `tokscale report` indefinitely. Add a pure-std run_command_with_timeout helper (spawn + reader threads draining stdout/stderr + try_wait deadline loop + kill on timeout, mirroring run_capture_command in main.rs) bounded by a generous 300s BACKEND_TIMEOUT. A timeout surfaces as an io::ErrorKind::TimedOut that each call site now degrades gracefully (log + skip the backend), matching the existing non-zero-exit behavior. The apple-fm path keeps its stdin-write behavior via the helper's optional stdin_bytes argument. Constraint: pure-std timeout (spawn + try_wait deadline + kill), mirroring main.rs — no new dependency Confidence: high Scope-risk: narrow --------- Co-authored-by: Junho Yeo <i@junho.io>
t1000040
pushed a commit
to tmobi-internal/tokscale
that referenced
this pull request
Jun 30, 2026
…hoyeo#633) * feat(report): add report command with model/task/daily breakdown - Group sessions by model and task title in summary tables - Show daily breakdown for --week/--month, session list for --today - Integrate Apple FM summarizer for session classification - Add wiki DB for caching session summaries * feat(report): add multi-backend summarizer and LLM task grouping - Support claude, codex, gemini, kiro as summarizer backends (in addition to apple-fm) - Add 2nd LLM pass to cluster sessions into high-level task groups - Add --summarizer flag to select backend, --rebuild to reset cached summaries - Scope summarization to date range (--week, --since/--until) - Add task_group column to wiki DB with migration - Update README with Task-Attributed Report documentation * fix(report): address cubic review - error on unknown summarizer, propagate DB errors - Return error instead of Ok(()) for unknown --summarizer values - Propagate DB errors from update_summary and update_task_group - Clarify skip message when task grouping backend is unsupported (apple-fm) * fix(report): address remaining cubic review issues - Fix end-of-day filtering: use next_day_00:00 - 1ms instead of 23:59:59 - Fix wiki.rs fallback path: use dirs::home_dir() instead of literal '~/.config' - Fix until filter inconsistency: query_entries now uses '<' (exclusive) matching other range methods - Validate task_category/complexity against allowed values in wiki-summarizer.py - Fix recovery path crash: use session.get('session_id') instead of session['session_id'] in exception handler * fix: resolve clippy lint errors - content_extractor: use chars().count() instead of byte len() for truncation guard - wiki.rs: rename from_str to parse to avoid clippy::should_implement_trait - report.rs: use div_ceil() instead of manual ceiling division - usage/copilot.rs: use strip_prefix() instead of manual prefix stripping - usage/minimax.rs: remove redundant closure - usage/zai.rs: fix reference-to-reference pattern - usage/mod.rs: extract type alias for complex type, use .ok() instead of manual match * fix(report): use canonical pricing, fix UTF-8/NaN/div-zero panics, repair build after rebase After merging main, the report arm called build_date_filter with the old 5-arg form; the DateRangeFlags refactor changed it to take &DateRangeFlags. compute_msg_cost hardcoded flat per-token rates that only matched Sonnet and ignored cache_write/reasoning; it now prices through the canonical PricingService (calculate_cost_with_provider) exactly like the aggregator, loading a fresh dataset with a cached fallback for offline use. Display truncation used byte slicing on LLM-generated strings (UTF-8 boundary panic), per-model attribution divided by models_used.len() with no zero guard, and cost sorts used partial_cmp().unwrap() (NaN panic). Constraint: pricing must come from canonical PricingService, never hardcoded constants or fuzzy matching Confidence: medium Scope-risk: moderate Not-tested: live LLM summarizer backends (apple-fm / claude/codex/gemini CLIs) * style: apply rustfmt * fix(report): bound LLM summarizer subprocesses with a timeout Every LLM backend invocation in run_task_grouping, run_cli_summarizer, and the apple-fm path spawned a child and blocked forever with no timeout, so an auth prompt or network stall could hang `tokscale report` indefinitely. Add a pure-std run_command_with_timeout helper (spawn + reader threads draining stdout/stderr + try_wait deadline loop + kill on timeout, mirroring run_capture_command in main.rs) bounded by a generous 300s BACKEND_TIMEOUT. A timeout surfaces as an io::ErrorKind::TimedOut that each call site now degrades gracefully (log + skip the backend), matching the existing non-zero-exit behavior. The apple-fm path keeps its stdin-write behavior via the helper's optional stdin_bytes argument. Constraint: pure-std timeout (spawn + try_wait deadline + kill), mirroring main.rs — no new dependency Confidence: high Scope-risk: narrow --------- Co-authored-by: Junho Yeo <i@junho.io>
t1000040
pushed a commit
to tmobi-internal/tokscale
that referenced
this pull request
Jun 30, 2026
…unhoyeo#746) * docs: sync EN/ja/ko/zh-cn for unreleased clients + breaking flag removal Pre-v3.2.0 documentation sync for the unreleased range (v3.1.3..main): - junhoyeo#465: replace the now-false legacy per-client flag notices in ja/ko/zh-cn (they claimed the removed flags still work) with a v3.2.0 breaking-change migration note; add the same note to README.md (which had none). - junhoyeo#728: document the MiniMax Token Plan subscription source (distinct from the MINIMAX_API_KEY row) in all locales; port the entire Subscription Usage section into ja/ko/zh-cn (was English-only). - junhoyeo#718: add the Jcode table row + detail section to the locales missing them. - junhoyeo#726: document the TOKSCALE_FM_DEBUG env var in all locales. - junhoyeo#633: add the missing task-attributed report bullet to README.ja Key Features. - drift: add Junie to the frontend Source-filtering list (all locales). - junhoyeo#710: fix the MiMo Code repo link (XiaomiMiMo/MiMo -> XiaomiMiMo/MiMo-Code). - junhoyeo#717: disclose Command Code token usage is estimated (~4 chars/token). Confidence: medium Scope-risk: narrow Directive: ja/ko/zh-cn translations of the ported Subscription Usage section are machine-generated and should get a native-speaker review pass Not-tested: junhoyeo#713 Antigravity CLI detail section was not added — no English source section exists to port from * fix(report): feed real session content to the summarizer (junhoyeo#633) extract_content_for_session unconditionally returned metadata_only_content() (first_user_message hardcoded None), so the report summarizer never saw any conversation content and the four real per-client extractors were dead code. Add content_extractor::extract_session_content, which dispatches to the correct per-client extractor (opencode/claude/codex/gemini) and falls back to metadata-only — never erroring or panicking — for unknown clients, missing candidates, or unreadable/unparseable files. report.rs builds a SessionPathIndex once (session_id -> transcript file, plus opencode DBs) and threads it through run_summarizer so each payload carries the real first user message. Confidence: high Scope-risk: moderate Rejected: thread file paths through core's scanner/WikiEntry | too invasive; indexed at the report layer instead Not-tested: end-to-end opencode/codex/gemini extraction in report.rs (core dispatcher covers claude + all fallback paths; per-client extractors are pre-existing) * docs: name the breaking release v4.0.0 (was v3.2.0) The per-client flag removal (junhoyeo#465) is a breaking change, so the next release is v4.0.0, not v3.2.0. Update the migration notes in all four README locales and the main.rs doc comment accordingly. * fix(report): real Codex/Gemini extraction + (client,session_id) index keying Addresses automated review feedback on the junhoyeo#633 report-summarizer-content fix (PR junhoyeo#746). The summarizer still surfaced (none) for normal Codex/Gemini sessions and could mis-route cross-client session_id collisions. - content_extractor: parse the current on-disk Codex format (event_msg with payload.type == "user_message", text in payload.message) and skip harness-injected context blocks (<environment_context>/<system-reminder>/ <user_instructions>), mirroring sessions::codex. - content_extractor: Gemini extractor now handles chat-recording JSON (messages[].type == "user" / content) and falls back to scanning line-delimited JSONL; empty/whitespace user text is treated as not-found. - extract_session_content: an empty/whitespace first_user_message no longer counts as success, so scanning continues to a later candidate with real text. - report: SessionPathIndex is keyed by (client, session_id) to prevent cross-client collisions, and Gemini files are keyed by their in-file sessionId (via gemini_session_id_for_file) rather than the filename stem, since the wiki entry's session_id is derived from inside the file. - Added fixture-based regression tests for all of the above. Constraint: wiki session_id for Gemini comes from the in-file sessionId, not the path stem Rejected: match any leading '<' for Codex injected blocks | drops legit prompts starting with markup Confidence: high Scope-risk: narrow
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Add multi-backend LLM summarizer support and automatic task grouping to the
tokscale reportcommand.Changes
Multi-backend summarizer
claude,codex,gemini,kiroas summarizer backends in addition to the defaultapple-fm--summarizerflag to select backend (default:apple-fm)LLM task grouping (2nd pass)
Date-scoped operations
--week,--since/--untildate filters--rebuildflag resets cached summaries within the date range and re-summarizes from scratchSchema changes
task_groupcolumn towiki_entriestable with auto-migrationreset_summaries_in_rangeandget_unsummarized_session_ids_in_rangeDB methodsDocumentation
Usage
Summary by cubic
Adds a task-attributed usage report with multi-backend LLM summarization and optional task grouping. Pricing now uses the canonical
PricingService, and the CLI is resilient across empty or malformed inputs.New Features
tokscale reportshows model and task-group breakdowns; daily view for--week/--month, session list for today. Results are cached in a local wiki DB;--rebuildre-summarizes in range. Auto-migration addstask_group.--summarizer:apple-fm(default),claude,codex,gemini,kiro; batch mode for CLI backends. A second LLM pass clusters titled sessions into 3–8 task groups (requires a CLI backend);apple-fmskips grouping with a clear message.--json) and filters for--workspaceand--client;--no-summarizeshows raw data without LLM calls.Bug Fixes
PricingService(no hardcoded rates), including cache write/read and reasoning tokens; uses a fresh dataset with cached fallback for offline runs.untilis exclusive (next-day 00:00 minus 1ms) and DB queries use<to match; unknown--summarizernow errors and DB errors from summary/group updates propagate to the CLI.--no-summarize --jsonon an empty home emits valid JSON.wiki-summarizer.pyvalidatestask_category/complexity, handles missing IDs, and resolves the fallback path correctly.Written for commit c1352d2. Summary will update on new commits.