MCP: add update test results mcps

rust/crates/sift_cli/assets/docs/src/agents/mcp.md

@@ -51,6 +51,10 @@ not run interactively.
 | `count_test_measurements` | Count test measurements matching a filter, without fetching them.    |
 | `create_test_report` | Create a test report with its steps and measurements from a JSON document.  |
 | `append_test_measurements` | Append measurements to an existing test step.                          |
+| `export_test_report` | Snapshot a test report's full tree to a JSON file for audit/diff and dry-runs. |
+| `update_test_report` | Update a test report's fields (status, name, times, run link, archive).     |
+| `update_test_step` | Update a test step's fields (name, status, times, error info, metadata).      |
+| `update_test_measurement` | Update a test measurement's value, bounds, verdict, unit, or metadata. |
 | `get_data`       | Download channel data for an asset/run to a Parquet file, with optional decimation. |
 | `sql`            | Run SQL over one or more Parquet files; chain after `get_data` for analysis.  |
 | `upload_dataset` | Stream a Parquet dataset into Sift.                                           |

rust/crates/sift_cli/assets/docs/src/agents/prompts.md

@@ -33,6 +33,7 @@ then in Claude Code each prompt is available as a slash command of the form
 - `/mcp__sift__explore_asset`
 - `/mcp__sift__analyze_run`
 - `/mcp__sift__derive_and_upload`
+- `/mcp__sift__audit_test_report`
 
 The `sift` in the command is your registered server name, so a different name
 in `.mcp.json` changes the prefix accordingly. Arguments are passed positionally
@@ -116,4 +117,35 @@ Specify the destination explicitly:
 The agent extracts the source data, applies the transform with `sql` (keeping
 `timestamp_unix_nanos` as the first column, as Sift requires), confirms the
 target asset, run, and any tags with you, then uploads the result with
-`upload_dataset`.
+`upload_dataset`.
+
+## `audit_test_report`
+
+Snapshots a test report (the test-results feature: a report owns steps, which
+own measurements) to a JSON file for audit and diffing. Given a description of
+intended edits, it dry-runs them first: it projects the changes, diffs them
+against the baseline for you to confirm, and only then applies the writes and
+verifies the result. Omit the edits for an audit snapshot alone.
+
+| Argument      | Required | Description                                                        |
+| ------------- | -------- | ----------------------------------------------------------------- |
+| `test_report` | yes      | Report to audit, by id or name.                                   |
+| `changes`     | no       | Plain-language description of the batch edits to dry-run.         |
+| `output_dir`  | no       | Directory for the snapshot files. Defaults to the working directory. |
+
+Snapshot a report for the record:
+
+```
+/mcp__sift__audit_test_report "nightly-regression 2024-05-01"
+```
+
+Dry-run a batch of edits before applying:
+
+```
+/mcp__sift__audit_test_report tr_abc123 "mark every step under power-on as FAILED and set its error_info"
+```
+
+The agent resolves the report, exports a baseline snapshot, projects the edits
+to a second file, and diffs the two so you can confirm. Once you approve, it
+applies the changes with the `update_test_*` tools, re-exports, and diffs
+against the projection to confirm the server recorded what you intended.

rust/crates/sift_cli/assets/skills/agents-md/AGENTS.md

@@ -25,6 +25,14 @@ to combine them when working with Sift.
    - `create_test_report`, `append_test_measurements`: create a test report with
      its step/measurement tree, or add measurements to an existing step (writes —
      confirm with the user first).
+   - `export_test_report`: snapshot a report's full tree to a JSON file (reads
+     only; omits server-managed fields, deterministically ordered). Use it to
+     record a baseline for audit/diff, or to dry-run bulk edits — export, apply
+     changes to a copy, diff, then run the update tools and re-export to confirm.
+   - `update_test_report`, `update_test_step`, `update_test_measurement`: update
+     fields of an existing report, step, or measurement (writes — only the fields
+     you set change; metadata/channel_names use replace semantics; confirm current
+     versus proposed values with the user first).
    - `get_data`: download channel data for an asset/run to a Parquet file.
    - `sql`: run SQL over one or more Parquet files (chain after `get_data`).
    - `upload_dataset`: stream a Parquet dataset into Sift. Returns an

rust/crates/sift_cli/assets/skills/claude-code/SKILL.md

@@ -41,6 +41,14 @@ to combine them when working with Sift.
    - `create_test_report`, `append_test_measurements`: create a test report with
      its step/measurement tree, or add measurements to an existing step (writes —
      confirm with the user first).
+   - `export_test_report`: snapshot a report's full tree to a JSON file (reads
+     only; omits server-managed fields, deterministically ordered). Use it to
+     record a baseline for audit/diff, or to dry-run bulk edits — export, apply
+     changes to a copy, diff, then run the update tools and re-export to confirm.
+   - `update_test_report`, `update_test_step`, `update_test_measurement`: update
+     fields of an existing report, step, or measurement (writes — only the fields
+     you set change; metadata/channel_names use replace semantics; confirm current
+     versus proposed values with the user first).
    - `get_data`: download channel data for an asset/run to a Parquet file.
    - `sql`: run SQL over one or more Parquet files (chain after `get_data`).
    - `upload_dataset`: stream a Parquet dataset into Sift. Returns an

rust/crates/sift_mcp/src/prompt/mod.rs

@@ -30,6 +30,13 @@ pub struct DeriveAndUploadArgs {
     target_run: Option<String>,
 }
 
+#[derive(Debug, Deserialize, JsonSchema)]
+pub struct AuditTestReportArgs {
+    test_report: String,
+    changes: Option<String>,
+    output_dir: Option<String>,
+}
+
 #[prompt_router(vis = "pub(crate)")]
 impl SiftMcpServer {
     #[prompt(
@@ -157,4 +164,61 @@ impl SiftMcpServer {
 
         vec![PromptMessage::new_text(PromptMessageRole::User, body)]
     }
+
+    #[prompt(
+        name = "audit_test_report",
+        description = "Snapshot a test report for audit, and optionally dry-run a batch of edits: diff the projected changes before applying, then apply and verify. Writes only after you confirm the diff."
+    )]
+    pub async fn audit_test_report(
+        &self,
+        params: Parameters<AuditTestReportArgs>,
+    ) -> Vec<PromptMessage> {
+        let Parameters(AuditTestReportArgs {
+            test_report,
+            changes,
+            output_dir,
+        }) = params;
+
+        let dir = output_dir.unwrap_or_else(|| "the current working directory".to_string());
+
+        let tail = match changes {
+            Some(changes) => format!(
+                "The user wants to dry-run this batch of edits before applying them: \"{changes}\".\n\n\
+                 4. DRY RUN: apply the intended edits to a COPY of the baseline file (write \
+                 `<dir>/<test_report_id>.projected.json`). Do NOT call any write tool yet. Field names in \
+                 the snapshot mirror the update-tool inputs, so map each edit to the step or measurement \
+                 by its id. Remember `metadata` and `channel_names` are REPLACE, and changing a \
+                 measurement value does not recompute `passed` — set it explicitly if the verdict should \
+                 change.\n\
+                 5. Diff baseline vs projected and show the user exactly what would change. Get explicit \
+                 confirmation before writing.\n\
+                 6. On approval, apply the edits with `update_test_report`, `update_test_step`, and \
+                 `update_test_measurement`, using the ids from the snapshot.\n\
+                 7. Re-export to `<dir>/<test_report_id>.after.json` and diff it against the projected \
+                 file. Report any discrepancy between what was intended and what the server recorded."
+            ),
+            None => {
+                "This is an audit snapshot only. Report the baseline file path as the record and \
+                     make no changes."
+                    .to_string()
+            }
+        };
+
+        let body = format!(
+            "Help the user audit a Sift test report (the test-results feature: a report owns steps, which \
+             own measurements). The user referred to the report as: \"{test_report}\".\n\n\
+             Write snapshot files under: {dir}.\n\n\
+             Steps:\n\
+             1. Resolve the report with `list_test_reports`. If \"{test_report}\" looks like an id, filter \
+             `test_report_id == \"{test_report}\"`; otherwise filter `name == \"{test_report}\"`. If \
+             several match, ask the user which one before continuing.\n\
+             2. Snapshot the baseline with `export_test_report` to \
+             `<dir>/<test_report_id>.before.json`. This is the audit record; it omits server-managed \
+             fields and is deterministically ordered, so diffs show only user-editable changes.\n\
+             3. Confirm with the user what the baseline captured (step and measurement counts).\n\
+             {tail}"
+        );
+
+        vec![PromptMessage::new_text(PromptMessageRole::User, body)]
+    }
 }