lean-ctx-review
GitHub分析 lean-ctx 工具在会话中的表现,提取日志并审查失败原因。用于评估工具性能、排查回归问题及上报缺陷,辅助优化上下文管理逻辑。
Trigger Scenarios
Install
npx skills add yvgude/lean-ctx --skill lean-ctx-review -g -y
SKILL.md
Frontmatter
{
"name": "lean-ctx-review",
"description": "Review how the lean-ctx ctx_* MCP tools performed in the current session and file upstream issues for confirmed problems. Use when the user asks to review\/audit lean-ctx tool performance, says \"lean-ctx review\", asks why ctx_* tools failed or fell back to native tools, or wants to report a lean-ctx bug."
}
lean-ctx session review
Upstream repo: yvgude/lean-ctx (https://github.com/yvgude/lean-ctx).
Reads a Claude Code session transcript, so it applies to agents that write
~/.claude/projects/*/*.jsonl. Other harnesses need their own extractor; the
triage rules below are harness-independent.
1. Extract
Run from this skill's directory (needs python3):
python3 scripts/scan_session.py
Defaults to the newest transcript under ~/.claude/projects/*/ — the live
Claude Code session. Pass a path to review a different one. -d N dumps call
#N in full.
The script reports facts, not verdicts: every ctx_* call in order, which
results the harness flagged is_error, which calls repeated identical
arguments, which asked for raw/fresh output, and how many native
Read/Grep/Bash calls happened. It deliberately does not guess which of
those is a problem — that judgment is yours, below.
Transcripts run to megabytes, so read the table, not the raw JSONL.
2. Triage
Start from is_error. In practice every genuine tool failure carries it.
Then read the call list for the softer signals: identical retries, raw/fresh
re-reads, native calls clustered right after a ctx_* call, or a sequence that
only completed once ctx_* was abandoned.
Before absolving anything, do these two things
Skipping them is how this review returns a false "nothing to report". Both cost one command each.
Search the tracker for the guard or tool you are about to excuse. Do this during triage, not at step 4. A closed issue defines intended behaviour: if a shipped fix says a form is supported and you just watched it fail, that is a regression and one of the most valuable things this review finds. It also stops you re-filing a settled design decision.
gh issue list --repo yvgude/lean-ctx --search "<guard or tool keyword>" --state all
Probe the stated rationale — never accept a message's self-description. An error explains itself, and that explanation can be wrong. Three checks, each one command:
- Does the sanctioned alternative achieve the same thing? If the block is bypassable through a route the message itself recommends, it constrains syntax, not capability. Sometimes that is deliberate (structured input parses reliably where a shell string does not) — decide which, do not assume.
- Does the stated reason match the observed rule? Vary one axis at a time. A guard blaming payload size that actually keys on path, or blaming a pipe on a command that is piped, is misdiagnosing itself.
- Does the message contradict itself? A rejection that also states the rejected form is allowed is a finding on its own, whichever way the policy should go.
A wrong diagnostic is a real bug even when the block is correct: it sends the next caller down a route that cannot work.
Not a finding
- Words inside returned content. A file containing
http.StatusConflict, agit rebaseprintingCONFLICT (content):,ghreturning"mergeable":"CONFLICTING", a test run printingnot found. The tool returned exactly what was asked for. Judge the tool's own behaviour, never the payload's vocabulary. - Self-reference. Any result that quotes a previous scan, this skill, or a transcript will echo every error word in it.
- Guard blocks — only after they survive the probe above. Inline env
overrides (
GIT_EDITOR=), shell redirects (>/>>), non-allowlisted commands, paths outside the project root. "It offered an alternative and the alternative worked" is not enough to clear one: that is true of a guard whose reason is wrong, whose rule is different from its description, or that contradicts itself. Clear it only once the stated rationale holds up. - Malformed tool input.
InputValidationErroron__unparsedToolInputis the caller's own JSON serialization — commonly a raw tab or newline inside a string — rejected before lean-ctx ever ran. - My own wrong arguments, a genuinely missing file, a failing build.
Confirmable
The tool did something wrong or unhelpful given correct input:
- wrong, lossy, or truncated output where the mode promises otherwise
- content injected into output documented as verbatim, or anything that breaks a documented output format (e.g. corrupting batch-read separators)
- a crash, hang, or schema mismatch against the documented parameters
- a documented mode behaving differently than described
- a retry that only succeeded after dropping to native tools
- a wrong or self-contradicting diagnostic, even where the block itself is right — naming the wrong cause sends the next caller down a route that cannot work, and costs more than the block did
- a regression against a closed issue: a shipped fix says the form is supported, and it is not
3. Reproduce
No minimal repro, no issue. Re-run the exact tool and arguments. If it passes, try to isolate the trigger; if you cannot, do not file — report the observation to the user instead, saying plainly that it did not reproduce and what you tried. Guessing at internals is worse than silence.
lean-ctx --version for the version line.
Zero confirmed findings is a valid and common result. Say so and file nothing.
4. File
Only for confirmed findings, one issue per distinct problem. You searched the tracker during triage; widen it here if the finding shifted:
gh issue list --repo yvgude/lean-ctx --search "<keywords>" --state all
Skip if already reported — and read the close text before re-filing, since a closed issue may have already conceded the point you are about to raise. When a finding contradicts a closed fix, cite it and frame the issue as the gap in that fix. Otherwise:
gh issue create --repo yvgude/lean-ctx --title "..." --body-file <file>
Body: what happened, expected, minimal repro (exact tool + args), version, OS.
No speculation about internals. End every issue body with the Claude Code
attribution footer. Report the issue html_url back to the user.
Version History
- e710e85 Current 2026-09-09 03:29


