Tighten what-happened lookup guardrails · Entire
Tighten what-happened lookup guardrails
0515631→main·
pfleidi·2mo ago·2 files·+46 added/-23 removed
Sessions
6393adf4d72cView transcript
?\ I have added a new skill to plugins/entire/skills/what-happenedCodex·GPT-5.5·2 steps
Changes
2
MREADME.md+4/-2
plugins/entire/skills/what-happened
MSKILL.md+42/-21
45 unmodified lines
46
47
48
49
50
49
50
51
52
53
54
55
45 unmodified lines
Current behavior:
- resolves file lines, ranges, or pasted snippets to exact line numbers
- groups matching blame ranges by unique commit before running `entire explain`
- summarizes `entire explain` output without dumping raw transcripts by default
- asks for a concrete file line, range, or snippet before running provenance commands
- deduplicates commit and checkpoint lookups before running expensive transcript commands
- asks before expanding broad ranges with many unique commits
- summarizes `entire explain` output without dumping raw transcripts by default; raw transcript expansion is opt-in
- falls back to clearly labeled current-code analysis when checkpoint-backed context is unavailable
Examples:
MREADME.md+4/-2
44 unmodified lines
45
46
47
48
49
50
51
52
53
9 unmodified lines
63
64
65
63
66
67
68
69
70
66
67
71
72
73
69
70
74
75
76
77
78
73
79
80
81
75
82
83
84
78
85
86
87
88
82
89
90
91
92
93
94
95
89
90
91
92
96
97
98
99
100
101
102
103
104
105
1 unmodified line
107
108
109
100
110
111
112
113
17 unmodified lines
131
132
133
124
134
135
136
137
18 unmodified lines
156
157
158
159
160
161
162
163
164
165
19 unmodified lines
185
186
187
174
188
189
190
191
192
193
194
195
196
197
180
198
199
200
201
202
203
17 unmodified lines
221
222
223
204
205
206
207
224
225
226
227
228
229
230
231
44 unmodified lines
- `path:start-end`
- `path` plus a pasted code snippet from that file
If the user asks a vague provenance question without a file path, line range, or pasted
snippet, ask for the target code and stop without running commands or using the header.
## Goal
Find the most recent change blocks matching the user's target lines, list the matching
9 unmodified lines
2. Use the installed `entire` binary from `PATH`, not `./entire` from the current repo.
3. Prefer `git blame` for provenance and `entire explain --commit` for transcript-backed context.
Do not use experimental `entire why` for this skill.
4. Do not manually hunt through `.git/entire-sessions/` or raw transcript files for commit
4. Use this skill for latest-change provenance on a specific block. For broad original intent
of a symbol, file, or feature, prefer the `explain` skill.
5. Do not manually hunt through `.git/entire-sessions/` or raw transcript files for commit
provenance. If `entire explain` cannot provide transcript context, report the exact
missing or unavailable state.
5. If the user provides a snippet, resolve it to exact line numbers before explaining anything.
6. If multiple blame blocks match, include all distinct ranges. Deduplicate commit hashes
6. If the user provides a snippet, resolve it to exact line numbers before explaining anything.
7. If multiple blame blocks match, include all distinct ranges. Deduplicate commit hashes
before running `entire explain`; run transcript lookups once per unique commit, not once
per range.
7. Distinguish these states explicitly:
per range. Also deduplicate checkpoint IDs before expanding checkpoint transcripts; run
checkpoint expansion once per unique checkpoint, not once per commit or range.
8. Distinguish these states explicitly:
- no checkpoint is referenced for the commit
- a checkpoint is referenced but is unavailable locally or remotely
- a checkpoint is available, but full transcript expansion failed
- a checkpoint is available, but full transcript expansion failed and raw transcript
expansion was not explicitly requested
- the code is untracked, uncommitted, or otherwise has no committed history
8. For every resolved code block, include either checkpoint-backed history or a fallback
9. For every resolved code block, include either checkpoint-backed history or a fallback
explanation of what the current code does. Label fallback explanations as "not
checkpoint-backed" and do not imply intent or historical rationale from checkpoints.
9. Treat `entire explain` command output as intermediate source material for summarization.
10. Treat `entire explain` command output as intermediate source material for summarization.
Do not paste raw command output or full transcripts into the user response unless the user
explicitly asks for raw output. Include only short error excerpts when they help
the user fix a failed lookup.
10. Keep the final explanation concise and block-focused. Do not summarize unrelated parts
11. Keep the final explanation concise and block-focused. Do not summarize unrelated parts
of the file.
## Workflow
### 1. Resolve the target block
If the user gave `path:line` or `path:start-end`, use that line or range directly and read only
that target from the file before explaining it. If the path does not exist, the file cannot be
read, or the line/range is outside the file, say so plainly and stop without using the
`Entire What Happened:` header.
If the user did not provide a file path plus exact line/range or pasted snippet, ask them for
the target code and stop. Do not run commands or use the `Entire What Happened:` header.
If the user gave `path:line` or `path:start-end`, use that line or range directly and read only
that target from the file before explaining it. If the path does not exist, the file cannot be
read, or the line/range is outside the file, say so plainly and stop without using the
`Entire What Happened:` header.
If the user gave a path and a snippet:
1 unmodified line
matching to find candidate locations:
```bash
grep -n -F -- "<distinctive snippet line>" <path>
grep -n -F -- "<distinctive snippet line>" "<path>"
```
- Read the small candidate windows around each hit, not the whole file unless the file is
17 unmodified lines
Run:
```bash
git blame --porcelain -L <start>,<end> -- <path>
git blame --porcelain -L <start>,<end> -- "<path>"
```
If the command fails because the file is untracked, mark the whole target range as an untracked
18 unmodified lines
commit SHA to all target ranges blamed to that commit. Do not run `entire explain` separately
for multiple ranges that share the same commit.
If the resolved target spans more than 5 unique real commits, stop before running `entire
explain`. Report the matched ranges and ask the user to narrow the range or confirm the deeper
lookup. Do not use the `Entire What Happened:` header for this confirmation response.
Keep the exact snippets from the target-resolution read so the final answer can show users
which code each provenance entry refers to. Only reread a matched block if the snippet for
that range was not already captured.
19 unmodified lines
authentication or missing remote configuration.
If the commit view reveals a checkpoint ID but is still not enough to answer the user's
question, run:
question, collect the checkpoint ID for expansion. Deduplicate checkpoint IDs across all
commit views before running checkpoint lookups; if several commits reference the same
checkpoint, expand that checkpoint once and map the result back to every relevant range.
For each unique checkpoint ID that needs more detail, run:
```bash
entire explain --checkpoint <checkpoint-id> --full --no-pager
```
If `--full` fails and raw transcript is necessary to answer the user's question, run:
Do not run raw transcript expansion automatically. If `--full` fails or is insufficient,
mark the affected ranges for current-code fallback analysis unless the user explicitly asked
for raw transcript detail. Only when explicitly requested, run:
```bash
entire explain --checkpoint <checkpoint-id> --raw-transcript --no-pager
```
not available locally or remotely." Include the command error only if it helps the user fix
the issue, such as authentication or missing remote configuration.
If the checkpoint loads but `--full` or `--raw-transcript` fails, say that checkpoint metadata
was available but transcript expansion failed. Answer checkpoint-backed facts from the
`entire explain --commit` output, and use current-code fallback analysis for anything that
output cannot support.
If the checkpoint loads but `--full` fails, say that checkpoint metadata was available but
full transcript expansion failed. If raw transcript detail was not explicitly requested, say
it was not expanded automatically. Answer checkpoint-backed facts from the `entire
explain --commit` output, and use current-code fallback analysis for anything that output
cannot support.
Map each unique commit explanation back to every target range blamed to that commit.