Files
gitlore/docs/plan-expose-discussion-ids.feedback-4.md
teernisse 5f9ae29993 feat(cli): implement 'lore trace' command (bd-2n4, bd-9dd)
Gate 5 Code Trace - Tier 1 (API-only, no git blame).
Answers 'Why was this code introduced?' by building
file -> MR -> issue -> discussion chains.

New files:
- src/core/trace.rs: run_trace() query logic with rename-aware
  path resolution, entity_reference-based issue linking, and
  DiffNote discussion extraction
- src/core/trace_tests.rs: 7 unit tests for query logic
- src/cli/commands/trace.rs: CLI command with human output,
  robot JSON output, and :line suffix parsing (5 tests)

Human output shows full content (no truncation).
Robot JSON truncates discussion bodies to 500 chars for token efficiency.

Wiring:
- TraceArgs + Commands::Trace in cli/mod.rs
- handle_trace in main.rs
- VALID_COMMANDS + robot-docs manifest entry
- COMMAND_FLAGS autocorrect registry entry

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-02-17 14:53:54 -05:00

207 lines
7.8 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
Below are the highest-impact revisions Id make to this plan. I excluded everything listed in your `## Rejected Recommendations` section.
**1. Fix a correctness bug in the ambiguity guardrail (must run before `LIMIT`)**
The current post-query ambiguity check can silently fail when `--limit` truncates results to one project even though multiple projects match the same `gitlab_discussion_id`. That creates non-deterministic write targeting risk.
```diff
@@ ## Ambiguity Guardrail
-**Implementation**: After the main query, if `gitlab_discussion_id` is set and no `--project`
-was provided, check if the result set spans multiple `project_path` values.
+**Implementation**: Run a preflight distinct-project check when `gitlab_discussion_id` is set
+and `--project` was not provided, before the main list query applies `LIMIT`.
+Use:
+```sql
+SELECT DISTINCT p.path_with_namespace
+FROM discussions d
+JOIN projects p ON p.id = d.project_id
+WHERE d.gitlab_discussion_id = ?
+LIMIT 3
+```
+If more than one project is found, return `LoreError::Ambiguous` (exit code 18) with project
+paths and suggestion to retry with `--project <path>`.
```
---
**2. Add `gitlab_project_id` to the Bridge Contract**
`project_path` is human-friendly but mutable (renames/transfers). `gitlab_project_id` gives a stable write target and avoids path re-resolution failures.
```diff
@@ ## Bridge Contract (Cross-Cutting)
Every read payload that surfaces notes or discussions **MUST** include:
- `project_path`
+- `gitlab_project_id`
- `noteable_type`
- `parent_iid`
- `gitlab_discussion_id`
- `gitlab_note_id`
@@
const BRIDGE_FIELDS_NOTES: &[&str] = &[
- "project_path", "noteable_type", "parent_iid",
+ "project_path", "gitlab_project_id", "noteable_type", "parent_iid",
"gitlab_discussion_id", "gitlab_note_id",
];
const BRIDGE_FIELDS_DISCUSSIONS: &[&str] = &[
- "project_path", "noteable_type", "parent_iid",
+ "project_path", "gitlab_project_id", "noteable_type", "parent_iid",
"gitlab_discussion_id",
];
```
---
**3. Replace stringly-typed filter/sort fields with enums end-to-end**
Right now `sort`, `order`, `resolution`, `noteable_type` are mostly `String`. This is fragile and risks unsafe SQL interpolation drift over time. Typed enums make invalid states unrepresentable.
```diff
@@ ## 3a. CLI Args
- pub resolution: Option<String>,
+ pub resolution: Option<ResolutionFilter>,
@@
- pub noteable_type: Option<String>,
+ pub noteable_type: Option<NoteableTypeFilter>,
@@
- pub sort: String,
+ pub sort: DiscussionSortField,
@@
- pub asc: bool,
+ pub order: SortDirection,
@@ ## 3d. Filters struct
- pub resolution: Option<String>,
- pub noteable_type: Option<String>,
- pub sort: String,
- pub order: String,
+ pub resolution: Option<ResolutionFilter>,
+ pub noteable_type: Option<NoteableTypeFilter>,
+ pub sort: DiscussionSortField,
+ pub order: SortDirection,
@@
+Map enum -> SQL fragment via `match` in query builder; never interpolate raw strings.
```
---
**4. Enforce snapshot consistency for multi-query commands**
`discussions` with `--include-notes` does multiple reads. Without a single read transaction, concurrent ingest can produce mismatched `total_count`, row set, and expanded notes.
```diff
@@ ## 3c. SQL Query
-pub fn query_discussions(...)
+pub fn query_discussions(...)
{
+ // Run count query + page query + note expansion under one deferred read transaction
+ // so output is a single consistent snapshot.
+ let tx = conn.transaction_with_behavior(rusqlite::TransactionBehavior::Deferred)?;
...
+ tx.commit()?;
}
@@ ## 1. Add `gitlab_discussion_id` to Notes Output
+Apply the same snapshot rule to `query_notes` when returning `total_count` + paged rows.
```
---
**5. Correct first-note rollup semantics (current CTE can return null/incorrect `first_author`)**
In the proposed SQL, `rn=1` is computed over all notes but then filtered with `is_system=0`, so threads with a leading system note may incorrectly lose `first_author`/snippet. Also path rollup uses non-deterministic `MAX(...)`.
```diff
@@ ## 3c. SQL Query
-ranked_notes AS (
+ranked_notes AS (
SELECT
n.discussion_id,
n.author_username,
n.body,
n.is_system,
n.position_new_path,
n.position_new_line,
- ROW_NUMBER() OVER (
- PARTITION BY n.discussion_id
- ORDER BY n.position, n.id
- ) AS rn
+ ROW_NUMBER() OVER (
+ PARTITION BY n.discussion_id
+ ORDER BY CASE WHEN n.is_system = 0 THEN 0 ELSE 1 END, n.created_at, n.id
+ ) AS rn_first_note,
+ ROW_NUMBER() OVER (
+ PARTITION BY n.discussion_id
+ ORDER BY CASE WHEN n.position_new_path IS NULL THEN 1 ELSE 0 END, n.created_at, n.id
+ ) AS rn_first_position
@@
- MAX(CASE WHEN rn = 1 AND is_system = 0 THEN author_username END) AS first_author,
- MAX(CASE WHEN rn = 1 AND is_system = 0 THEN body END) AS first_note_body,
- MAX(CASE WHEN position_new_path IS NOT NULL THEN position_new_path END) AS position_new_path,
- MAX(CASE WHEN position_new_line IS NOT NULL THEN position_new_line END) AS position_new_line
+ MAX(CASE WHEN rn_first_note = 1 AND is_system = 0 THEN author_username END) AS first_author,
+ MAX(CASE WHEN rn_first_note = 1 AND is_system = 0 THEN body END) AS first_note_body,
+ MAX(CASE WHEN rn_first_position = 1 THEN position_new_path END) AS position_new_path,
+ MAX(CASE WHEN rn_first_position = 1 THEN position_new_line END) AS position_new_line
```
---
**6. Add per-discussion truncation signals for `--include-notes`**
Top-level `has_more` is useful, but agents also need to know if an individual threads notes were truncated. Otherwise they cant tell if a thread is complete.
```diff
@@ ## Response Schema
{
"gitlab_discussion_id": "...",
...
- "notes": []
+ "included_note_count": 0,
+ "has_more_notes": false,
+ "notes": []
}
@@ ## 3b. Domain Structs
pub struct DiscussionListRowJson {
@@
+ pub included_note_count: usize,
+ pub has_more_notes: bool,
#[serde(skip_serializing_if = "Vec::is_empty")]
pub notes: Vec<NoteListRowJson>,
}
@@ ## 3c-ii. Note expansion query (--include-notes)
-Group by `discussion_id` in Rust and attach notes arrays...
+Group by `discussion_id` in Rust, attach notes arrays, and set:
+`included_note_count = notes.len()`,
+`has_more_notes = note_count > included_note_count`.
```
---
**7. Add explicit query-plan gate and targeted index workstream (measured, not speculative)**
This plan introduces heavy discussion-centric reads. You should bake in deterministic performance validation with `EXPLAIN QUERY PLAN` and only then add indexes if missing.
```diff
@@ ## Scope: Four workstreams, delivered in order:
-4. Fix robot-docs to list actual field names instead of opaque type references
+4. Add query-plan validation + targeted index updates for new discussion queries
+5. Fix robot-docs to list actual field names instead of opaque type references
@@
+## 4. Query-Plan Validation and Targeted Indexes
+
+Before and after implementing `query_discussions`, capture `EXPLAIN QUERY PLAN` for:
+- `--for-mr <iid> --resolution unresolved`
+- `--project <path> --since 7d --sort last_note`
+- `--gitlab-discussion-id <id>`
+
+If plans show table scans on `notes`/`discussions`, add indexes in `MIGRATIONS` array:
+- `discussions(project_id, gitlab_discussion_id)`
+- `discussions(merge_request_id, last_note_at, id)`
+- `notes(discussion_id, created_at DESC, id DESC)`
+- `notes(discussion_id, position, id)`
+
+Tests: assert the new query paths return expected rows under indexed schema and no regressions.
```
---
If you want, I can produce a single consolidated “iteration 4” version of the plan text with all seven revisions merged in place.