tsa-pr-review

name	tsa-pr-review
description	| Use when this capability is needed.

tsa-pr-review — AST-grounded PR review with merge-gate verdict

Replaces "LLM reads the diff and guesses" with a deterministic pipeline that grounds the review in (a) the AST diff, (b) the persisted call graph, and (c) architectural-constraints.yml. Output is a structured verdict an agent can act on — including the exact pytest command to run before merge.

When to use

• Local diff review: unstaged, staged, or branch-vs-main
• GitHub PR review by URL (uses gh under the hood)
• CI / agent pre-merge gate
• Any time a generic "read the diff" review would be too shallow

Don't use when:

• The diff is a one-line typo in a comment or doc
• The change is in a file with no AST coverage (e.g. binary, image, .lock)
• You only need to describe the diff, not judge it → use plain git diff

Why this beats a generic LLM diff-read

Concern	LLM-only diff read	tsa-pr-review
Classify signature vs body change	Heuristic	edit action=pr AST diff
Exact tests to run	Invented / hallucinated	edit action=impact verification_command
Caller blast radius	Grep guess	nav action=callers from persisted graph
Callee regression surface	Read N files	nav action=callees from persisted graph
Architecture rule violations	None	edit action=constraints vs YAML rules
Per-file pre-merge verdict	Subjective	edit action=safe SAFE/REVIEW/CAUTION/UNSAFE
Behaviour on Python 3.14	n/a	Stable (grep-ast crashes, cgc silent-0)

(Source: docs/internal/COMPETITOR_HEAD_TO_HEAD_2026-05-23.md — the 4 fault-tolerance advantages: oracle-consistent AST, indexed-not-silent, Python 3.14-safe, no external DB dependency.)

Procedure

Step 1 — Diff-level fan-out (parallel, 3 calls)

In one message, call all three:

1. edit action=pr mode="staged" — AST diff classification, per-symbol change category (signature_change / body_change / new_symbol / deletion), and per-file changed-symbol list.
   • For PR URL: edit action=pr mode="pr" pr_url="<url>"
   • For unstaged: mode="diff"; for branch-vs-main: mode="branch"
2. edit action=impact mode="staged" — returns verification_command, queue_ledger (in-scope vs out-of-scope dirty files), pytest_required, and per-file impact rows.
3. edit action=constraints — full repo audit against architectural-constraints.yml, filtered to files the diff touches if you want noise-free output: edit action=constraints path_filter="<glob covering changed files>".

Step 2 — Per-file fan-out (parallel, N calls)

For each changed file from Step 1, call:

edit action=safe file_path="<abs>" edit_type="<refactor|add_feature|fix_bug|rename>"

Pick edit_type from the dominant change category in Step 1's edit action=pr output. Returns per-file verdict + risk_factors.

Step 3 — Per-symbol fan-out (parallel, ≤2N calls)

For each changed signature (not body-only) symbol from Step 1, fan out two calls:

• nav action=callers function_name="<sym>" language="<lang>" limit=20 — who depends on this; the blast radius if its signature moved.
• nav action=callees function_name="<sym>" language="<lang>" limit=20 — what it calls; the regression surface if its body changed.

Filter callers/callees where callee_resolution == "stdlib" or "unknown" — they're noise for blast-radius scoring (per tsa-graph guidance).

Step 4 — Fold into verdict

verdict_precedence (worst wins):
  - BLOCK   ← any edit action=constraints violation with severity=error
              OR any edit action=safe verdict == UNSAFE
  - REVIEW  ← any edit action=safe verdict in {REVIEW, CAUTION}
              OR pytest_required=true with no nearby tests
              OR signature_change with ≥5 external callers
  - APPROVE ← all per-file verdicts == SAFE,
              all constraints pass,
              pytest_required=false OR verification_command is tightly scoped

Surface the precise verification_command verbatim — never paraphrase it.

Worked example (this repo, 2-file diff)

Suppose git diff HEAD~2 HEAD shows changes in two files. To walk through without committing anything new, simulate the workflow against a real diff:

git -C /Users/aisheng.yu/git-private/tree-sitter-analyzer diff HEAD~2 HEAD --name-only
# → tree_sitter_analyzer/mcp/tools/safe_to_edit_tool.py
# → tree_sitter_analyzer/mcp/tools/change_impact_tool.py

Diff-level fan-out

# Call 1
edit action=pr mode="branch" include_call_graph=true
# returns:
#   changed_files:
#     - path: ".../safe_to_edit_tool.py"
#       changed_symbols:
#         - name: "SafeToEditTool.execute"
#           category: "body_change"
#         - name: "SafeToEditTool._score_risk"
#           category: "signature_change"  # ← signature changed
#     - path: ".../change_impact_tool.py"
#       changed_symbols:
#         - name: "ChangeImpactTool.execute"
#           category: "body_change"

# Call 2
edit action=impact mode="branch" agent_summary_only=true
# returns:
#   verification_command: "uv run pytest tests/unit/test_safe_to_edit_tool.py
#                          tests/unit/test_change_impact_tool.py -q"
#   pytest_required: true
#   queue_ledger:
#     in_scope: [".../safe_to_edit_tool.py", ".../change_impact_tool.py"]
#     out_of_scope_dirty: []

# Call 3
edit action=constraints path_filter="tree_sitter_analyzer/mcp/**"
# returns:
#   verdict: SAFE
#   violations: []

Per-file fan-out

edit action=safe file_path=".../safe_to_edit_tool.py" edit_type="refactor"
# → verdict: REVIEW, risk_factors: [{factor: "high_caller_count", ...}]
edit action=safe file_path=".../change_impact_tool.py" edit_type="fix_bug"
# → verdict: SAFE

Per-symbol fan-out (only the signature-change)

nav action=callers function_name="_score_risk" language="python" limit=20
# → 7 callers; 6 in tests/, 1 in cli/commands/mcp_commands.py
nav action=callees function_name="_score_risk" language="python" limit=20
# → 4 callees; all local (good — no stdlib noise)

Folded verdict

verdict: REVIEW
files:
  - path: ".../safe_to_edit_tool.py"
    verdict: REVIEW
    changed_symbols: 2
    signature_changes: 1
    blast_radius: 7         # callers of _score_risk
    health_grade: B
  - path: ".../change_impact_tool.py"
    verdict: SAFE
    changed_symbols: 1
    signature_changes: 0
    blast_radius: 0
constraint_violations: []
verification_command: "uv run pytest tests/unit/test_safe_to_edit_tool.py
                       tests/unit/test_change_impact_tool.py -q"
stop_condition: "tests exit successfully AND no NEW failures vs main"
notes:
  - "Signature change on _score_risk affects 1 non-test caller — verify mcp_commands.py still passes."

CLI equivalents (1-to-1 with MCP calls)

# Step 1 — Diff fan-out
uv run tree-sitter-analyzer --pr-review --change-impact-mode staged --output-format toon
uv run tree-sitter-analyzer --change-impact --change-impact-mode staged --agent-summary-only --output-format json
uv run tree-sitter-analyzer --check-constraints --constraint-path-filter "tree_sitter_analyzer/mcp/**" --output-format toon

# Step 2 — Per-file
uv run tree-sitter-analyzer <file> --safe-to-edit --edit-type refactor --output-format json

# Step 3 — Per-symbol
uv run tree-sitter-analyzer --callers <FUNC> --output-format toon
uv run tree-sitter-analyzer --callees <FUNC> --output-format toon

Exit codes mirror --change-impact / --check-constraints:

• 0 → APPROVE
• 2 → REVIEW (warn-only)
• 1 → BLOCK (error severity)

Anti-patterns

• DO NOT ask the LLM to invent the pytest command — quote verification_command verbatim. r36 / past incidents confirm hallucinated pytest invocations skip the regressions TSA already pinpointed.
• DO NOT skip edit action=constraints because "the LLM can spot architecture drift". It can't — the rules live in YAML and are evaluated against the persisted edge graph.
• DO NOT run callers/callees on body-only changes. Bodies don't change signatures; callers don't care. Reserve graph queries for signature changes, deletions, and new public symbols.
• DO NOT mix mode="staged" for pr_review with mode="branch" for change_impact — keep modes consistent so both tools see the same diff.
• DO NOT swallow queue_ledger.out_of_scope_dirty — it tells you the working tree has files outside the diff that may still break the build.
• DO NOT flip MCP output_format from toon to json — locked by user (see CLAUDE.md "Deliberate design decisions"). Tokens matter.

Decision surface (output schema)

verdict: APPROVE | REVIEW | BLOCK
files:
  - path: <abs>
    verdict: SAFE | REVIEW | CAUTION | UNSAFE
    changed_symbols: <int>
    signature_changes: <int>
    body_changes: <int>
    new_symbols: <int>
    deletions: <int>
    blast_radius: <int>            # external callers of changed signatures
    regression_surface: <int>      # non-stdlib callees touched
    health_grade: A | B | C | D | F
    risk_factors: [{factor, detail, severity}]
constraint_violations: [
  {rule_id, caller_file, caller_line, callee_file, severity, reason}
]
verification_command:       
 

   []
   []     
    
 [ ]

Verdict cascade (worst wins):

• BLOCK → constraint error OR per-file UNSAFE
• REVIEW → per-file REVIEW/CAUTION OR ≥5-caller signature change OR pytest_required and tests aren't co-located
• APPROVE → all per-file SAFE, no constraint violations, tight verify cmd

---

Source: aimasteracc/tree-sitter-analyzer — distributed by TomeVault.