Repository navigation
Release v3.0.0: Honest call answers
v3.0.0 is a major release. Several default outputs and one CLI behaviour
changed since v2.2.0, and the index format moved forward. Existing indexes are
rebuilt automatically on the next refresh. Read Breaking / behaviour changes
and migration below before upgrading an agent or script that parses TSA output.
Highlights:
- Honest call answers.
nav action=callersno longer answerscompletewhen a real caller exists.
The release gate (dogfood round 12b, on this release) found 0 wrong
completeacross 7 real repos (rich, pydantic, fastapi, axios, gson, zod,
hono) in both bare and qualified form: about 1,570 sweep answers plus every
ground-truth battery.- Bare and qualified method queries now pass the same gates (#1740, #1759).
- Unresolved receivers keep every same-named method possible (#1758),
including structuralProtocol/ duck-typed implementations in test
files (#1772).
- Faster warm queries. Warm
nav action=callersis 44–67% faster than
v2.2: pydantic 0.39 → 0.15 s, fastapi 0.46 → 0.22 s, gson 0.24 → 0.08 s,
rich 0.14 → 0.08 s. The call graph is built at most once per query and
reused while the index snapshot is unchanged (#1774). - TypeScript and JavaScript resolution. Caller recall on zod rose from
64/87 to 85/87 and on hono from 62/87 to 81/103, with hono false callers
16 → 0.- Calls bind through ES-module imports and re-exports (#1764).
- Calls also bind through imported default-export objects (#1741).
- Arrow-function constants are functions (#1763).
this.m()dispatches through the class'sextendschain (#1765).- String-literal members match their element-access calls (#1762).
- Java. Typed receivers bind, and overrides stay possible (#1743).
- Refresh cost. A one-file refresh is much faster on a large project:
- Index safety.
- Typed admission errors are shared by MCP, the CLI and the watcher
(RFC-0032 slice 4, #1554). - Custom
db_pathcaches are fenced against obsolete and concurrent writers
(RFC-0032 slice 5).
- Typed admission errors are shared by MCP, the CLI and the watcher
Breaking / behaviour changes and migration
edit action=impact/ CLI--change-impact --change-impact-fullsummarize
by default (#1617).- What changed: the full response used to return every list in full, 70–224 KB
for a multi-file diff. It now keeps the top entries of its bulk lists,
with honest totals (*_total,details_truncated,details_next_step),
within a 24 KB budget. - Migration: pass
include_details=true(CLI
--change-impact-include-details, which implies--change-impact-full)
to get the full lists and every command field, as before.
- What changed: the full response used to return every list in full, 70–224 KB
ast_cache/index action=cachelists only changed or errored files
(#1582).- What changed: the
filesdetail omitsstatus=cachedentries and lists
at most 50 files, errors first, withfiles_total/files_listed/
files_cap/files_truncated/unchanged_omitted. - Migration: pass
include_details=true(CLI
--ast-cache-include-details) for the old full list.
- What changed: the
- CLI
--watchholds until Ctrl-C (#1685).- What changed: before,
--watchstarted a watcher and returned at once, so the
process exited mid-sync. It now prints its answer, keeps watching until
interrupted, then stops the watcher. It exits 1 if the stop does not finish. - Migration: scripts that ran
--watchand expected it to return must
run it in the background or send an interrupt. Use--incremental-sync
for a one-shot refresh.
- What changed: before,
- Bare-name queries are as strict as qualified ones (#1740, #1759).
- What changed:
nav action=callerswith a bare method name plusfile_path
now runs the same name-only and declaring-file checks asClass.method.
A bare name that names exactly one method in the file is answered as that
method. Answers that used to saycompletemay now sayunknownor
incompletewithNAME_ONLY_INBOUNDorUNRESOLVED_IN_DECLARING_FILE, and
list the evidence sites (name_only_sites,unresolved_inbound_sites). - Migration: treat these as "read the named sites", not as errors.
- What changed:
- New and changed reason / error codes.
SYMBOL_IN_EXCLUDED_FILEreplacesSYMBOL_NOT_DEFINEDwhen the definition
sits in a file the index excludes (#1635, #1648). The reply names the file,
line and exclude pattern.- Typed admission errors across MCP, the CLI and the watcher (RFC-0032 slice
4, #1554):INDEX_WRITER_BUSY,INDEX_CANDIDATE_SUPERSEDED,
INDEX_RECOVERY_REQUIRED. Each reply carrieserror_code,cause_code,
retryable,retry_after_ms,next_actionandnext_step.
INDEX_GENERATION_SOURCE_CHANGEDmoves fromerror_codetocause_code.
The codes are documented indocs/agent-envelope-contract.md. INDEX_CANDIDATE_UNBOUND(not retryable,next_action: obtain_bound_candidate) reports acandidate_snapshotthat was not
captured bybuild_index_candidate_snapshot. It used to be reported as
INDEX_CANDIDATE_SUPERSEDED.- Migration: branch on
error_code/retryableinstead of matching
error text.
- Visualization node ids changed.
- What changed:
codegraph_visualize/viz action=graphids are now
<file>::<Owner>.<name>@<line>(sanitized), so same-named functions no
longer collapse into one node. Labels are
<basename>::<Owner>.<name>:<line>. - Migration: clients that stored or matched the old ids must re-derive them.
- What changed:
- Index format: manifest v3, schema 18, extractor version 53.
- What changed:
- Migration: nothing to do. The next refresh re-extracts files written by
an older extractor and upgrades the database in place. Older TSA versions
(v2.x) read a v3 manifest as not certified and fail closed. They answer
without certification, never with a wrong certified result. Downgrading
therefore costs a full re-index, not a corrupt answer.
Full details: CHANGELOG.md · Migration guide: docs/MIGRATION.md