cli/mcp/hooks: consolidated gaps from agent dogfooding (edit state, indexing, skill docs)

[swapnilpaliwal-sd]

Summary

Dogfood agents in all three languages hit the same problems in the shared surface: the CLI verbs, the MCP tools, the post-edit hook, indexing, and the skill docs. This issue collects them. Language-specific gaps are tracked in the per-language consolidated issues.

Root cause

• changed, test-impact and the post-edit hook answer from the state the graph was built from, not from the working tree against the base revision. After the graph is rebuilt from an edited tree the edit is part of the baseline, so changed reports no change and test-impact selects nothing (the full repro was changed / test-impact: after the graph is rebuilt from an edited tree, an edit is reported as no change and no test is selected #1566). After a complete rename, the hook still says the old name has callers because it does not re-read the tree.
• The hook's graph provenance check misreads a graph the agent just built as "built by an older axiomcode" and repeats the line on nearly every call.
• The skill docs and the tool schemas drift: SKILL.md documents an MCP context argument that the tool rejects.

$ git diff --stat
 src/util.ts | 7 +++++++
$ axiomcode changed
no change to a declaration the graph knows (working tree against <base sha>)
$ axiomcode test-impact
no changed declaration the graph can name, so no test can be selected from it.

In progress

• fix/index-skip-unneeded: skip vendored code, report the main language first, kill the process tree.

Checklist

• After the graph is rebuilt from an edited tree, changed and test-impact still report the edit and name the revision they compare against (from changed / test-impact: after the graph is rebuilt from an edited tree, an edit is reported as no change and no test is selected #1566; a Java round hit the same, "part of the baseline" and 0 tests selected).
• changed and test-impact on 61 edited files take 32 s on a Java project; make them proportional to the edit.
• After a complete rename, the post-edit hook and changed do not say "its callers still name X" when 0 uses remain. Seen in all three languages; re-check after PR changed: classify edits by place and parameters, not by name #1809.
• The hook does not call a graph the agent just built "built by an older axiomcode", and after an install move it says that line once per session.
• A facts view (a verb or an impact section) prints every graph fact attached to a declaration or a file:line: decorations and annotations with their arguments, ancestors, bound config keys, the call sites inside a span, and why a dispatch pair was dropped. It replaces the hand sqlite queries fix agents keep running.
• SKILL.md and the MCP context tool agree on refresh=false.
• changed says tests/cases files are read by tests/case_runner.py; it should name tests/run.py.
• The test-impact dispatch rung ranks tests by count; rank the override's own package first.

Related open issues

• impact: --in does not choose among same-named declarations, and next: points outside the scope #1572: impact --in does not choose among same-named declarations. Distinct, left open.
• skill: impact and path show no code under rows the graph could not resolve, so agents open a file to settle each one #1794: show code under rows the graph could not resolve. Distinct, left open.

[swapnilpaliwal-sd]

Folded in from #1572 and #1794:

• impact <name> --in <path> answers for the one same-named declaration inside the scope, and next: names only declarations inside it (from impact: --in does not choose among same-named declarations, and next: points outside the scope #1572). Repro: a name declared in three language extractors, --in one of them, still answers for all three.
• impact and path print a 3-line code block under rows that are not [resolved], and only under those (from skill: impact and path show no code under rows the graph could not resolve, so agents open a file to settle each one #1794). Measured over three runs: about 10% fewer tokens at the same accuracy.

[swapnilpaliwal-sd]

Front-door gaps from a dogfood round on the new surface (find, impact, path, tests, index). Fixes are on the branch fix/front-door-answers.

Documented entry

• SKILL.md documents <skill>/scripts/axiomcode, the dispatcher, which gives raw answers. The documented entry should be a front door.

Bare impact (your uncommitted edits)

• A parameter type change crashes it: file:9(user): user is not a type in the graph after (User user) becomes (String userId). The targets are asked together, so the other edits get no answer either.
• A git mv class rename shows up as added files, then "nothing edited is a declaration other code depends on".
• After methods are added, untouched methods whose lines only moved are reported as signature changes, and 3 of 6 added methods are missed.
• On C#, a bare impact through the dispatcher prints its usage page and exits 0. Re-check it at the front door.
• A method under a route attribute ([HttpPatch("{id}/cancel")]) is misparsed. The attribute's arguments are read as parameters ("-cancel", "return type / modifiers: [HttpPatch"), and impact stops with "cancel is not a type".
• A parameter reorder is reported only as "quantity: default 1 -> (none)".
• After a class is deleted, rows for an unrelated base read "must change · extends / implements it". After a rename, overrides of a same-named method on another class are marked "must change".
• For a removed or renamed declaration, the answer should lead with the references to the old name that remain (0 means done), then the places that referenced it.

tests

• The run line says mvn in a Gradle-only repository.
• The run line keeps 1 of the 6 test classes the answer lists.
• Rows show "hop None", and a row sits at line 1 with no code.
• An empty tests or impact answer is printed as 30-60 lines of JSON.
• An edit in a web handler gives no warning that tests driven over HTTP (Client.PostAsync("/orders/index")) are not traced.

find and impact <name>

• find refuses with "--in is required". The front door has no option to give, so it should pick a scope or answer across the repository.
• find matches "post" to "postgres". It should match whole words or identifier parts.
• find ranks receiver calls (file.read()) first as "declared nowhere".
• impact <name> lists the call sites but not the import lines that name the target, which a rename or a delete must edit too.
• A 219 KB line from a generated HTML report is printed whole. Printed source lines should be capped, and generated or minified text should not be shown as code.

Notes around answers

• The refresh line and other hints name removed options and verbs (--fresh, --no-refresh, --in, --page, test-impact, context). At the front door an answer should name only the four questions and index.
• A "graph refresh ... files edited" line kept showing after the edits were reverted.