release(v1.6.16): agent observability + state_cache bake fix#2
Conversation
… td_self_update + state_cache bake fix
- Added: `_read_journal` response hint on every dispatched MCP tool
(`src/td_mcp/read_journal.py`, wired in `_forward()`). Each response
carries {call_count, first_seen_at, last_seen_at, result_unchanged}
keyed by (tool_name, args_fingerprint). Claude can see across MCP
request boundaries which reads have moved and which haven't.
- Added: 200-entry server-side activity_log ring buffer + new
`td_get_activity_log` MCP tool (`src/td_mcp/activity_log.py`,
`src/td_mcp/registry/tools_meta.py`).
- Added: in-TD `activity_log` Table DAT mirror at /local/mcp_server/
activity_log — rolling 200-row window so users can wire agent
activity into their visual patch. `td_component/event_emitter.py`
gains `append_activity_row`; `mcp_webserver_callbacks.py`
`_record_request_safe` appends on every HTTP request.
- Added: `td_self_update` MCP tool + `td_mcp/self_updater.py`. Hits
GitHub releases API, compares semver, optionally writes the latest
.tox to all three install paths (repo, plugin cache, ~/.tdpilot/)
with md5 sync reporting. Closes the seven-layer staleness saga.
Pure stdlib so it also runs inside TD's Python via Textport.
- Fixed: `td_component/state_cache.py` source was never baked into
the textDAT — listed in `_TOX_SOURCE_FILES` since v1.6.7 for hash
tracking, but `_populate_component` callers never passed
`state_cache_code`. The DAT shipped empty, `.module` raised
tdError, and hasattr propagated through (only catches AttributeError),
killing the state_cache writer. Latent for 9 releases. Fixed in
`td_component/build_export_mcp_tox.py:build_and_export`.
- Fixed: `_record_request_safe` dual-writer separation — state_cache
and activity_log writers now in independent try blocks. A
state_cache compile error can no longer take down the activity_log
mirror. Also switched `parent()` to `me.parent()` (instance method,
always defined) to be resilient across TD module-load contexts.
- Bumped: tool count 104 → 106 (EXPECTED_MIN_TOOL_COUNT). Updated
README, npm/README, plugin_README, docs/{API_REFERENCE,USER_GUIDE,
GETTING_STARTED,MANUAL}, skills/{tdpilot-core,tdpilot-production},
marketplace+plugin descriptions, npm description, mcp/manifest.json,
td_component/renderer.py, CHANGELOG.
- Bumped: version 1.6.15 → 1.6.16 across all 7 fields (pyproject,
__init__, .claude-plugin/plugin.json, marketplace.json, npm/package,
mcp/manifest, API_VERSION in mcp_webserver_callbacks.py).
- Added: `scripts/audit_v1_6_16_live.py` — end-to-end live audit
script that exercises all v1.6.16 surfaces against a running TD
(29 checks, all PASS). Joins patch_session_smoke.py + full_td_mcp_e2e.py
as a permanent regression tool in scripts/.
Verification: 970 pytest pass, 29-check live audit pass, check_versions
+ check_tox_freshness + smoke_mcp_registry all green. .tox rebuilt via
Textport (hash 796a3d488aa48052..., 18:34 UTC).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three trivial auto-fixes flagged by CI lint on the v1.6.16 release commit: - F541 in scripts/audit_v1_6_16_live.py: drop extraneous `f` prefix on f-string with no placeholder. - UP035 in src/td_mcp/self_updater.py: import `Callable` from `collections.abc` instead of `typing` (PEP 585 deprecation). - I001 in tests/test_self_updater.py: import-block ordering. All other CI jobs on v1.6.16 already pass (test 3.11, test 3.12, npm publish, install parse-checks macos+windows, Build & upload release assets v1.6.16 with tdpilot.plugin + tdpilot.mcpb attached). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
CI lint step runs BOTH `ruff check` and `ruff format --check`. The previous fix only ran `ruff check --fix`, missing two files the formatter wanted to reflow: - scripts/audit_v1_6_16_live.py - tests/test_read_journal.py No behavior change — pure whitespace/style. 970 tests still pass; `ruff check` + `ruff format --check` both clean locally. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 23294bd75d
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| try: | ||
| from td_mcp import self_updater | ||
|
|
||
| result = self_updater.run(check_only=check_only) |
There was a problem hiding this comment.
td_self_update is declared async but calls self_updater.run() synchronously, and that function performs blocking network (urllib) and filesystem writes. In practice this can block the MCP server’s event loop for up to the HTTP timeout (30s+) during an update/check, stalling unrelated tool calls in the same process. Wrapping the call in asyncio.to_thread(...) (or making the updater async) avoids request-wide head-of-line blocking.
Useful? React with 👍 / 👎.
| text = json.dumps(value, default=str, separators=(",", ":")) | ||
| except Exception: | ||
| text = repr(value) |
There was a problem hiding this comment.
The summary path serializes the entire result with json.dumps(...) and only then truncates to 120 chars. Since _forward now logs every tool result, large responses (for example image/base64-heavy or large graph payloads) incur an extra full serialization/allocation per call, which can add significant latency and memory pressure under load. The summary should be size-bounded before full serialization or use a cheap structural preview.
Useful? React with 👍 / 👎.
1 participant
Summary
_read_journalhint (call_count,result_unchanged,first_seen_at,last_seen_at) so Claude can see across MCP request boundaries which reads have moved.td_get_activity_logMCP tool + in-TDactivity_logTable DAT mirror for wiring agent activity into visuals.td_self_update: new MCP tool that hits GitHub releases API, optionally writes the latest .tox to all three install paths with md5 sync reporting — closes the seven-layer staleness saga.state_cache.pysource was never baked into its textDAT since v1.6.7 — silently broke the state_cache writer. Now properly threaded through_populate_component._record_request_safenow has independent try blocks for state_cache + activity_log, plusme.parent()for resilience across TD module-load contexts.Verification
scripts/audit_v1_6_16_live.py)check_versions,check_tox_freshness,smoke_mcp_registry,check_no_personal_pathsall green.toxrebuilt inside TD; activity_log mirror + state_cache writer verified live with real MCP trafficTest plan
🤖 Generated with Claude Code