sieve-core 0.0.0-stage → 0.1.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +26 -3
- package/config/buckets.json +188 -0
- package/config/hooks.json +89 -0
- package/config/instructions/bucket-actions.md +6 -0
- package/config/instructions/bucket-commitments_tasks.md +7 -0
- package/config/instructions/bucket-current_state.md +7 -0
- package/config/instructions/bucket-decisions.md +9 -0
- package/config/instructions/bucket-goals_requirements.md +7 -0
- package/config/instructions/bucket-knowledge_facts.md +7 -0
- package/config/instructions/bucket-problems_corrections.md +8 -0
- package/config/instructions/bucket-questions.md +8 -0
- package/config/instructions/bucket-results_evidence.md +7 -0
- package/config/instructions/bucket-rules.md +11 -0
- package/config/instructions/judge-readback.md +22 -0
- package/config/instructions/readback.md +6 -0
- package/config/instructions/reading-pass.md +26 -0
- package/config/rules.json +576 -0
- package/config/settings.json +19 -0
- package/config/systems.json +62 -0
- package/dist/atlas/identity.d.ts +65 -0
- package/dist/atlas/identity.js +148 -0
- package/dist/atlas/identity.js.map +1 -0
- package/dist/atlas/map.d.ts +205 -0
- package/dist/atlas/map.js +710 -0
- package/dist/atlas/map.js.map +1 -0
- package/dist/atlas/units.d.ts +32 -0
- package/dist/atlas/units.js +256 -0
- package/dist/atlas/units.js.map +1 -0
- package/dist/brief.d.ts +25 -0
- package/dist/brief.js +75 -0
- package/dist/brief.js.map +1 -0
- package/dist/check.d.ts +14 -0
- package/dist/check.js +198 -0
- package/dist/check.js.map +1 -0
- package/dist/config.d.ts +130 -0
- package/dist/config.js +117 -0
- package/dist/config.js.map +1 -0
- package/dist/examples.d.ts +15 -0
- package/dist/examples.js +72 -0
- package/dist/examples.js.map +1 -0
- package/dist/hook.d.ts +31 -0
- package/dist/hook.js +305 -0
- package/dist/hook.js.map +1 -0
- package/dist/index.d.ts +16 -0
- package/dist/index.js +17 -0
- package/dist/index.js.map +1 -0
- package/dist/install.d.ts +21 -0
- package/dist/install.js +92 -0
- package/dist/install.js.map +1 -0
- package/dist/judge.d.ts +20 -0
- package/dist/judge.js +38 -0
- package/dist/judge.js.map +1 -0
- package/dist/models.d.ts +44 -0
- package/dist/models.js +110 -0
- package/dist/models.js.map +1 -0
- package/dist/parts/atlas.d.ts +40 -0
- package/dist/parts/atlas.js +211 -0
- package/dist/parts/atlas.js.map +1 -0
- package/dist/parts/changes.d.ts +86 -0
- package/dist/parts/changes.js +337 -0
- package/dist/parts/changes.js.map +1 -0
- package/dist/parts/readback.d.ts +36 -0
- package/dist/parts/readback.js +249 -0
- package/dist/parts/readback.js.map +1 -0
- package/dist/parts/record.d.ts +9 -0
- package/dist/parts/record.js +72 -0
- package/dist/parts/record.js.map +1 -0
- package/dist/parts/rulecheck.d.ts +20 -0
- package/dist/parts/rulecheck.js +69 -0
- package/dist/parts/rulecheck.js.map +1 -0
- package/dist/parts/stopcheck.d.ts +3 -0
- package/dist/parts/stopcheck.js +88 -0
- package/dist/parts/stopcheck.js.map +1 -0
- package/dist/reading.d.ts +45 -0
- package/dist/reading.js +204 -0
- package/dist/reading.js.map +1 -0
- package/dist/repair.d.ts +14 -0
- package/dist/repair.js +64 -0
- package/dist/repair.js.map +1 -0
- package/dist/repo.d.ts +12 -0
- package/dist/repo.js +41 -0
- package/dist/repo.js.map +1 -0
- package/dist/rules.d.ts +28 -0
- package/dist/rules.js +169 -0
- package/dist/rules.js.map +1 -0
- package/dist/setup.d.ts +39 -0
- package/dist/setup.js +541 -0
- package/dist/setup.js.map +1 -0
- package/dist/state.d.ts +4 -0
- package/dist/state.js +27 -0
- package/dist/state.js.map +1 -0
- package/dist/systems.d.ts +114 -0
- package/dist/systems.js +221 -0
- package/dist/systems.js.map +1 -0
- package/dist/transcript.d.ts +4 -0
- package/dist/transcript.js +64 -0
- package/dist/transcript.js.map +1 -0
- package/package.json +14 -6
package/README.md
CHANGED
|
@@ -1,3 +1,26 @@
|
|
|
1
|
-
#
|
|
2
|
-
|
|
3
|
-
|
|
1
|
+
# sieve-core
|
|
2
|
+
|
|
3
|
+
Sieve keeps the right things in a Claude Code agent's context and enforces the owner's rules, through Claude Code
|
|
4
|
+
hooks. Install it with the `sieve` command from `sieve-cli`:
|
|
5
|
+
|
|
6
|
+
```
|
|
7
|
+
npm install --save-dev sieve-cli
|
|
8
|
+
npx sieve install # writes Sieve's hooks into .claude/settings.json
|
|
9
|
+
npx sieve check # validates Sieve's configuration
|
|
10
|
+
npx sieve app # a local page that shows and configures Sieve
|
|
11
|
+
```
|
|
12
|
+
|
|
13
|
+
What it does, at Claude Code's hook moments:
|
|
14
|
+
|
|
15
|
+
- **Before execution:** a brief of what is still live in the project (goals, open questions, decisions, outstanding
|
|
16
|
+
work, what went wrong), and the approval read-back: when the owner answers or approves, the agent reads back what
|
|
17
|
+
it will do and changes nothing until the owner says yes.
|
|
18
|
+
- **Before a tool:** the hold on changes while a read-back waits, and the core rules checked by pattern (no secrets
|
|
19
|
+
written into files; files edited with the edit tools, not shell commands).
|
|
20
|
+
- **After a tool:** a record of every call and of exactly what it changed, tied to the approved item it was for, with
|
|
21
|
+
lasting ids for the code it touched.
|
|
22
|
+
- **On stop and compaction:** what was said is filed into buckets by a model through the local cliproxy; changes made
|
|
23
|
+
without approval and promises left undone are reported; what a compaction summary dropped is brought back.
|
|
24
|
+
|
|
25
|
+
Configuration is in three layers (the core's files in `config/`, the user's `~/.sieve/config/`, the repository's
|
|
26
|
+
`.sieve/config/`). Models are called only through a cliproxy on the same machine, never with a provider API key.
|
|
@@ -0,0 +1,188 @@
|
|
|
1
|
+
{
|
|
2
|
+
"schema_version": 1,
|
|
3
|
+
"what": "The ten buckets (docs/design/buckets.md). Each: what it holds, its reader's instructions file, what the brief shows from it, the hook moments that use it, its status values and where its entries may come from. A layer adds a bucket with an entry here and an instructions file. Status values and what the brief shows are approved (owner, 2026-11-17); the definitions and moments are a draft.",
|
|
4
|
+
"sources": {
|
|
5
|
+
"rule_files": "CLAUDE.md and the files it loads (fills Rules only)",
|
|
6
|
+
"owner": "the owner's messages",
|
|
7
|
+
"assistant": "the agent's replies",
|
|
8
|
+
"thinking": "the agent's thinking, marked as such",
|
|
9
|
+
"hook_output": "hook output and injected text",
|
|
10
|
+
"tool_calls": "tool calls, filed mechanically",
|
|
11
|
+
"tool_results": "tool output, filed mechanically"
|
|
12
|
+
},
|
|
13
|
+
"buckets": {
|
|
14
|
+
"rules": {
|
|
15
|
+
"number": 1,
|
|
16
|
+
"name": "Rules",
|
|
17
|
+
"contains": "Instructions, restrictions, permissions, preferences, exceptions and conditions",
|
|
18
|
+
"instructions": "bucket-rules",
|
|
19
|
+
"brief": "What the agent must, must not or should do, with when each applies",
|
|
20
|
+
"moments": [
|
|
21
|
+
"before_execution",
|
|
22
|
+
"before_tool",
|
|
23
|
+
"after_tool",
|
|
24
|
+
"on_stop"
|
|
25
|
+
],
|
|
26
|
+
"statuses": [],
|
|
27
|
+
"sources": [
|
|
28
|
+
"rule_files"
|
|
29
|
+
]
|
|
30
|
+
},
|
|
31
|
+
"goals_requirements": {
|
|
32
|
+
"number": 2,
|
|
33
|
+
"name": "Goals & Requirements",
|
|
34
|
+
"contains": "What we're trying to achieve, what the system must do and what success looks like",
|
|
35
|
+
"instructions": "bucket-goals_requirements",
|
|
36
|
+
"brief": "What we're trying to achieve",
|
|
37
|
+
"moments": [
|
|
38
|
+
"before_execution",
|
|
39
|
+
"on_stop"
|
|
40
|
+
],
|
|
41
|
+
"statuses": [],
|
|
42
|
+
"sources": [
|
|
43
|
+
"owner"
|
|
44
|
+
]
|
|
45
|
+
},
|
|
46
|
+
"questions": {
|
|
47
|
+
"number": 3,
|
|
48
|
+
"name": "Questions",
|
|
49
|
+
"contains": "Questions asked, questions answered, unresolved questions and questions that need revisiting",
|
|
50
|
+
"instructions": "bucket-questions",
|
|
51
|
+
"brief": "Unanswered questions (open or needing revisiting), and the questions each current decision answered",
|
|
52
|
+
"moments": [
|
|
53
|
+
"before_execution",
|
|
54
|
+
"on_stop"
|
|
55
|
+
],
|
|
56
|
+
"statuses": [
|
|
57
|
+
"open",
|
|
58
|
+
"answered",
|
|
59
|
+
"needs_revisiting"
|
|
60
|
+
],
|
|
61
|
+
"sources": [
|
|
62
|
+
"owner",
|
|
63
|
+
"assistant"
|
|
64
|
+
]
|
|
65
|
+
},
|
|
66
|
+
"decisions": {
|
|
67
|
+
"number": 4,
|
|
68
|
+
"name": "Decisions",
|
|
69
|
+
"contains": "Historical decisions, current decisions, reasoning, rejected alternatives and superseded decisions",
|
|
70
|
+
"instructions": "bucket-decisions",
|
|
71
|
+
"brief": "Current decisions with their reasons; superseded and rejected ones with why, linked to the choice made instead",
|
|
72
|
+
"moments": [
|
|
73
|
+
"before_execution",
|
|
74
|
+
"before_tool"
|
|
75
|
+
],
|
|
76
|
+
"statuses": [
|
|
77
|
+
"current",
|
|
78
|
+
"superseded",
|
|
79
|
+
"rejected"
|
|
80
|
+
],
|
|
81
|
+
"sources": [
|
|
82
|
+
"owner",
|
|
83
|
+
"assistant",
|
|
84
|
+
"thinking"
|
|
85
|
+
]
|
|
86
|
+
},
|
|
87
|
+
"commitments_tasks": {
|
|
88
|
+
"number": 5,
|
|
89
|
+
"name": "Commitments & Tasks",
|
|
90
|
+
"contains": "What someone agreed to do, what's assigned, what's outstanding and what's completed",
|
|
91
|
+
"instructions": "bucket-commitments_tasks",
|
|
92
|
+
"brief": "Outstanding work with who agreed; completed work linked to its evidence",
|
|
93
|
+
"moments": [
|
|
94
|
+
"before_execution",
|
|
95
|
+
"on_stop"
|
|
96
|
+
],
|
|
97
|
+
"statuses": [
|
|
98
|
+
"outstanding",
|
|
99
|
+
"completed",
|
|
100
|
+
"dropped"
|
|
101
|
+
],
|
|
102
|
+
"sources": [
|
|
103
|
+
"owner",
|
|
104
|
+
"assistant"
|
|
105
|
+
]
|
|
106
|
+
},
|
|
107
|
+
"actions": {
|
|
108
|
+
"number": 6,
|
|
109
|
+
"name": "Actions",
|
|
110
|
+
"contains": "What the agent actually did: tool calls, commands, file edits, searches, executions",
|
|
111
|
+
"instructions": "bucket-actions",
|
|
112
|
+
"brief": "Not shown for its own sake: the evidence behind completed work and corrections",
|
|
113
|
+
"moments": [
|
|
114
|
+
"after_tool"
|
|
115
|
+
],
|
|
116
|
+
"statuses": [],
|
|
117
|
+
"sources": [
|
|
118
|
+
"tool_calls"
|
|
119
|
+
]
|
|
120
|
+
},
|
|
121
|
+
"results_evidence": {
|
|
122
|
+
"number": 7,
|
|
123
|
+
"name": "Results & Evidence",
|
|
124
|
+
"contains": "What happened as a result of those actions, including test results, failures and verification",
|
|
125
|
+
"instructions": "bucket-results_evidence",
|
|
126
|
+
"brief": "Not shown for its own sake: the evidence linked from completed work",
|
|
127
|
+
"moments": [
|
|
128
|
+
"after_tool",
|
|
129
|
+
"on_stop"
|
|
130
|
+
],
|
|
131
|
+
"statuses": [],
|
|
132
|
+
"sources": [
|
|
133
|
+
"tool_results"
|
|
134
|
+
]
|
|
135
|
+
},
|
|
136
|
+
"knowledge_facts": {
|
|
137
|
+
"number": 8,
|
|
138
|
+
"name": "Knowledge & Facts",
|
|
139
|
+
"contains": "Established information, discoveries, technical details, relationships and assumptions",
|
|
140
|
+
"instructions": "bucket-knowledge_facts",
|
|
141
|
+
"brief": "Not shown for its own sake: the evidence behind decisions and the current state; assumptions marked as such",
|
|
142
|
+
"moments": [
|
|
143
|
+
"before_execution"
|
|
144
|
+
],
|
|
145
|
+
"statuses": [
|
|
146
|
+
"established",
|
|
147
|
+
"assumption"
|
|
148
|
+
],
|
|
149
|
+
"sources": [
|
|
150
|
+
"owner",
|
|
151
|
+
"assistant",
|
|
152
|
+
"thinking"
|
|
153
|
+
]
|
|
154
|
+
},
|
|
155
|
+
"problems_corrections": {
|
|
156
|
+
"number": 9,
|
|
157
|
+
"name": "Problems & Corrections",
|
|
158
|
+
"contains": "Mistakes, failures, contradictions, user corrections, attempted fixes and lessons learned",
|
|
159
|
+
"instructions": "bucket-problems_corrections",
|
|
160
|
+
"brief": "What went wrong and what was learned",
|
|
161
|
+
"moments": [
|
|
162
|
+
"before_execution",
|
|
163
|
+
"after_tool"
|
|
164
|
+
],
|
|
165
|
+
"statuses": [],
|
|
166
|
+
"sources": [
|
|
167
|
+
"owner",
|
|
168
|
+
"assistant",
|
|
169
|
+
"hook_output",
|
|
170
|
+
"tool_results"
|
|
171
|
+
]
|
|
172
|
+
},
|
|
173
|
+
"current_state": {
|
|
174
|
+
"number": 10,
|
|
175
|
+
"name": "Current State",
|
|
176
|
+
"contains": "Where the project, task, environment or investigation currently stands",
|
|
177
|
+
"instructions": "bucket-current_state",
|
|
178
|
+
"brief": "Where everything stands right now: the latest statement about each thing",
|
|
179
|
+
"moments": [
|
|
180
|
+
"before_execution",
|
|
181
|
+
"after_tool",
|
|
182
|
+
"on_stop"
|
|
183
|
+
],
|
|
184
|
+
"statuses": [],
|
|
185
|
+
"sources": []
|
|
186
|
+
}
|
|
187
|
+
}
|
|
188
|
+
}
|
|
@@ -0,0 +1,89 @@
|
|
|
1
|
+
{
|
|
2
|
+
"schema_version": 1,
|
|
3
|
+
"what": "The four hook moments, which parts run at which Claude Code event, and the hook command. `sieve install` writes Claude Code's hooks from this file. command: null runs the installed core's dist/hook.js; a layer may set its own. events gives an event its own settings: async runs it in the background (its output arrives on the next turn). background names parts that run in a second, async hook for an event (the same command with --background), so the session never waits for them: SemAtlas's first map at SessionStart. Sieve sets no time limit of its own (owner's rule: no cap without proof it is needed); a layer may set timeout_seconds, for all events or one. claude_code records the limits Claude Code itself imposes, with where they are documented. SemAtlas (part atlas, switched by settings atlas.enabled) also takes each change's code ids and moves its map after changes and commits; that rides on the part changes, which has the trees.",
|
|
4
|
+
"command": null,
|
|
5
|
+
"events": {
|
|
6
|
+
"Stop": {
|
|
7
|
+
"async": true
|
|
8
|
+
}
|
|
9
|
+
},
|
|
10
|
+
"claude_code": {
|
|
11
|
+
"what": "Limits Claude Code imposes on command hooks, from its hooks reference. Sieve works within them; they are not Sieve's own.",
|
|
12
|
+
"source": "https://code.claude.com/docs/en/hooks.md (checked 2026-11-21): 'Hook handler fields' (timeout) and 'JSON output' (output size)",
|
|
13
|
+
"timeout_seconds": {
|
|
14
|
+
"default": 600,
|
|
15
|
+
"UserPromptSubmit": 30
|
|
16
|
+
},
|
|
17
|
+
"async_timeout": null,
|
|
18
|
+
"output_chars": 10000
|
|
19
|
+
},
|
|
20
|
+
"moments": {
|
|
21
|
+
"before_execution": {
|
|
22
|
+
"events": [
|
|
23
|
+
"SessionStart",
|
|
24
|
+
"UserPromptSubmit"
|
|
25
|
+
]
|
|
26
|
+
},
|
|
27
|
+
"before_tool": {
|
|
28
|
+
"events": [
|
|
29
|
+
"PreToolUse",
|
|
30
|
+
"PermissionRequest"
|
|
31
|
+
]
|
|
32
|
+
},
|
|
33
|
+
"after_tool": {
|
|
34
|
+
"events": [
|
|
35
|
+
"PostToolUse",
|
|
36
|
+
"PostToolUseFailure",
|
|
37
|
+
"PreCompact",
|
|
38
|
+
"PostCompact"
|
|
39
|
+
]
|
|
40
|
+
},
|
|
41
|
+
"on_stop": {
|
|
42
|
+
"events": [
|
|
43
|
+
"Stop",
|
|
44
|
+
"SubagentStop",
|
|
45
|
+
"SessionEnd"
|
|
46
|
+
]
|
|
47
|
+
}
|
|
48
|
+
},
|
|
49
|
+
"parts": {
|
|
50
|
+
"SessionStart": [
|
|
51
|
+
"brief",
|
|
52
|
+
"repair"
|
|
53
|
+
],
|
|
54
|
+
"UserPromptSubmit": [
|
|
55
|
+
"rules",
|
|
56
|
+
"brief",
|
|
57
|
+
"readback"
|
|
58
|
+
],
|
|
59
|
+
"PreToolUse": [
|
|
60
|
+
"readback",
|
|
61
|
+
"rulecheck",
|
|
62
|
+
"atlas",
|
|
63
|
+
"changes"
|
|
64
|
+
],
|
|
65
|
+
"PostToolUse": [
|
|
66
|
+
"record",
|
|
67
|
+
"changes"
|
|
68
|
+
],
|
|
69
|
+
"PostToolUseFailure": [
|
|
70
|
+
"record",
|
|
71
|
+
"changes"
|
|
72
|
+
],
|
|
73
|
+
"Stop": [
|
|
74
|
+
"reading",
|
|
75
|
+
"stopcheck"
|
|
76
|
+
],
|
|
77
|
+
"PreCompact": [
|
|
78
|
+
"repair"
|
|
79
|
+
],
|
|
80
|
+
"PostCompact": [
|
|
81
|
+
"repair"
|
|
82
|
+
]
|
|
83
|
+
},
|
|
84
|
+
"background": {
|
|
85
|
+
"SessionStart": [
|
|
86
|
+
"atlas"
|
|
87
|
+
]
|
|
88
|
+
}
|
|
89
|
+
}
|
|
@@ -0,0 +1,6 @@
|
|
|
1
|
+
Bucket 6, Actions: What the agent actually did: tool calls, commands, file edits, searches, executions.
|
|
2
|
+
(Definition from docs/design/buckets.md, a draft the owner has not yet approved.)
|
|
3
|
+
|
|
4
|
+
Belongs: what the agent actually did: every tool call. Filed by the mechanical pass.
|
|
5
|
+
Does not belong: what the agent said it did or would do.
|
|
6
|
+
Example: `Bash: List ssh keys and git credential settings`
|
|
@@ -0,0 +1,7 @@
|
|
|
1
|
+
Bucket 5, Commitments & Tasks: What someone agreed to do, what's assigned, what's outstanding and what's completed.
|
|
2
|
+
(Definition from docs/design/buckets.md, a draft the owner has not yet approved.)
|
|
3
|
+
|
|
4
|
+
Belongs: someone agreeing to do something, or being told to: who, what, and whether it is done.
|
|
5
|
+
Does not belong: the doing of it (Actions), or the outcome wanted (Goals & Requirements).
|
|
6
|
+
Examples: "I'll push the branch, open the PR, set the metric's start time, merge and update this PC." (agent) ·
|
|
7
|
+
"yes do that then report the results in a summary to me" (you, assigning)
|
|
@@ -0,0 +1,7 @@
|
|
|
1
|
+
Bucket 10, Current State: Where the project, task, environment or investigation currently stands.
|
|
2
|
+
(Definition from docs/design/buckets.md, a draft the owner has not yet approved.)
|
|
3
|
+
|
|
4
|
+
Belongs: where things stand at the end of the session. Nothing is filed here directly: it is worked out from the
|
|
5
|
+
other nine, and only the latest statement about each thing holds.
|
|
6
|
+
Does not belong: how things stood earlier in the session. That history stays in the other buckets.
|
|
7
|
+
Example of the evidence it draws on: "PR 200 is merged and installed on Ubuntu."
|
|
@@ -0,0 +1,9 @@
|
|
|
1
|
+
Bucket 4, Decisions: Historical decisions, current decisions, reasoning, rejected alternatives and superseded decisions.
|
|
2
|
+
(Definition from docs/design/buckets.md, a draft the owner has not yet approved.)
|
|
3
|
+
|
|
4
|
+
Belongs: one choice being settled, by you or by the agent, with the reasoning (the entry's `reason`) and
|
|
5
|
+
whatever was turned down (each its own entry, status `rejected`, with why, linked to the choice made instead).
|
|
6
|
+
Does not belong: a standing instruction (Rules). A decision to add or change a rule is filed here; the rule itself
|
|
7
|
+
arrives in Rules only through the rule file.
|
|
8
|
+
Examples: "Approve rule 10" · "use what i said --claude-review or --codex-review ignore review-model that was a
|
|
9
|
+
mistake" (a decision, a rejected alternative and a correction in one message)
|
|
@@ -0,0 +1,7 @@
|
|
|
1
|
+
Bucket 2, Goals & Requirements: What we're trying to achieve, what the system must do and what success looks like.
|
|
2
|
+
(Definition from docs/design/buckets.md, a draft the owner has not yet approved.)
|
|
3
|
+
|
|
4
|
+
Belongs: what must be true at the end: an aim, something the system must do, or what success looks like.
|
|
5
|
+
Does not belong: who is going to do it (Commitments & Tasks), or how the agent should behave (Rules).
|
|
6
|
+
Example: "the orchestrator should be able to decide exactly what its lineage needs in the VM from the CUA's full
|
|
7
|
+
catalogue"
|
|
@@ -0,0 +1,7 @@
|
|
|
1
|
+
Bucket 8, Knowledge & Facts: Established information, discoveries, technical details, relationships and assumptions.
|
|
2
|
+
(Definition from docs/design/buckets.md, a draft the owner has not yet approved.)
|
|
3
|
+
|
|
4
|
+
Belongs: something that stays true beyond the action that revealed it: a discovery, a technical detail, how two
|
|
5
|
+
things relate. An unverified one is kept and marked as an assumption.
|
|
6
|
+
Does not belong: raw output (Results & Evidence), or where things stand right now (Current State).
|
|
7
|
+
Example: "T3 lists it under Servers because the agent opened the port."
|
|
@@ -0,0 +1,8 @@
|
|
|
1
|
+
Bucket 9, Problems & Corrections: Mistakes, failures, contradictions, user corrections, attempted fixes and lessons learned.
|
|
2
|
+
(Definition from docs/design/buckets.md, a draft the owner has not yet approved.)
|
|
3
|
+
|
|
4
|
+
Belongs: something that went wrong and what was done about it: a mistake, a failure, a contradiction, you
|
|
5
|
+
correcting the agent, an attempted fix, a lesson. Labelled failures are filed by the mechanical pass.
|
|
6
|
+
Does not belong: a rule. A correction stays here even when it says how to behave from now on; it becomes a rule
|
|
7
|
+
only when it is written into the rule file.
|
|
8
|
+
Examples: "stop sending me voice please" (you) · "I was wrong to leave that waiting on a session there." (agent)
|
|
@@ -0,0 +1,8 @@
|
|
|
1
|
+
Bucket 3, Questions: Questions asked, questions answered, unresolved questions and questions that need revisiting.
|
|
2
|
+
(Definition from docs/design/buckets.md, a draft the owner has not yet approved.)
|
|
3
|
+
|
|
4
|
+
Belongs: a question asked by you or by the agent. The entry holds the question; its status says whether it was
|
|
5
|
+
answered, and it links to where the answer landed.
|
|
6
|
+
Does not belong: the answer itself (usually Knowledge & Facts or Decisions), or a request phrased as a question
|
|
7
|
+
("can you fix that?" is a task).
|
|
8
|
+
Examples: "whats that server thats running?" (you) · "Is the one you're looking at on the Windows PC?" (agent)
|
|
@@ -0,0 +1,7 @@
|
|
|
1
|
+
Bucket 7, Results & Evidence: What happened as a result of those actions, including test results, failures and verification.
|
|
2
|
+
(Definition from docs/design/buckets.md, a draft the owner has not yet approved.)
|
|
3
|
+
|
|
4
|
+
Belongs: what one action returned: tool output, test results, errors, background task outcomes. Each stays tied to
|
|
5
|
+
the action that produced it. Filed by the mechanical pass.
|
|
6
|
+
Does not belong: a conclusion drawn from results (Knowledge & Facts).
|
|
7
|
+
Example: `Exit code 1`, linked to the action above.
|
|
@@ -0,0 +1,11 @@
|
|
|
1
|
+
Bucket 1, Rules: Instructions, restrictions, permissions, preferences, exceptions and conditions.
|
|
2
|
+
(Definition from docs/design/buckets.md, a draft the owner has not yet approved.)
|
|
3
|
+
|
|
4
|
+
Belongs: what the written rule files say: `CLAUDE.md` and the files it loads. One rule is one line of such a
|
|
5
|
+
file. A rule is fixed by the file, not by the conversation, so it changes only when the file changes. The transcript
|
|
6
|
+
shows which version of the file was in force at each point.
|
|
7
|
+
Does not belong: anything said in the conversation. You correcting the agent is a Problem & Correction, and you
|
|
8
|
+
choosing to add or change a rule is a Decision. Neither is a rule until it is written into the file.
|
|
9
|
+
Example: "No force-push or history rewrite." (from the global `CLAUDE.md` loaded in the sample session)
|
|
10
|
+
Also recorded for each rule (decided by the owner): when it applies, and which of the four hook moments checks it
|
|
11
|
+
(before execution, before tool, after tool, on stop).
|
|
@@ -0,0 +1,22 @@
|
|
|
1
|
+
You read one message from the owner of a software project to the AI agent working for them, together with the end of
|
|
2
|
+
the agent's last reply. Decide what the owner's message does. Reply with exactly one of these four words and nothing
|
|
3
|
+
else: confirm, answer, question, other.
|
|
4
|
+
|
|
5
|
+
confirm: the agent's last reply is a read-back (it lists what it takes the owner to have approved and asks for a yes),
|
|
6
|
+
and the owner confirms it as it stands, with nothing added or changed. "yes", "all yes", "a yes b yes",
|
|
7
|
+
"confirmed, go", "yep do it" all confirm. A yes that adds, removes or changes anything is answer, not confirm.
|
|
8
|
+
answer: the owner responds to something the agent's last reply asked, proposed, recommended or read back: a yes or no,
|
|
9
|
+
a choice between options (even a bare option name with a question mark, like "Postgres?"), a correction, a counter-
|
|
10
|
+
proposal ("why not just use the built-in one?", "can we leave out the second part?"), a pushback that
|
|
11
|
+
overrules the agent ("why would it need that? just do whatever works"), or a new requirement added to what was
|
|
12
|
+
read back. A message that asks a question and also decides or overrules something is answer. A yes, a go-ahead or
|
|
13
|
+
an approval is always answer (or confirm), even when the agent's last reply asked nothing: it may approve something
|
|
14
|
+
said earlier. A short reply that names one of the agent's numbered or listed options ("just the first", "3?") picks
|
|
15
|
+
it, so it is answer, with or without a question mark.
|
|
16
|
+
question: the owner only asks something (how, why, what, whether, can you explain, where is it) and decides,
|
|
17
|
+
approves, proposes or overrules nothing. Spelling and punctuation don't matter: "isnt that done already" is a
|
|
18
|
+
question even without a question mark, and so is a remark that the owner can't find or see something.
|
|
19
|
+
other: anything else: a new instruction that doesn't respond to anything the agent asked or proposed ("now write
|
|
20
|
+
tests for the parser" after the agent reports it is done), thanks, a greeting, a status remark, pasted output.
|
|
21
|
+
|
|
22
|
+
Judge the owner's words against the agent's reply. Reply with the one word only.
|
|
@@ -0,0 +1,6 @@
|
|
|
1
|
+
The owner's message answers or approves something you asked or proposed. Their rule: before you act on any of it, read back what they are approving, then wait.
|
|
2
|
+
- Start your reply with a line "Read-back:".
|
|
3
|
+
- List every question or proposal from your earlier replies that still waits for the owner, numbered, each with the answer you take them to have given, in their words; or "still open" when their message doesn't answer it. A partial answer approves only what it names.
|
|
4
|
+
- Where the owner handed the choice to you ("you decide", or a workflow that lets you decide), write "decided by me, because you said: <their exact words>" and give your choice.
|
|
5
|
+
[impact] - Under each answered one, write "Approving this will:" and say what it changes (files, commands, behaviour), and what already on record it replaces or overrides (an earlier decision, a rule, an open task), or "replaces nothing on record".
|
|
6
|
+
- End by asking the owner to confirm. Change no files until they say yes.
|
|
@@ -0,0 +1,26 @@
|
|
|
1
|
+
You read part of a conversation between the owner of a software project and the AI agent working for them, and file
|
|
2
|
+
what matters into buckets. Each message is a numbered segment: [O1], [O2] … are the owner's, [A1], [A2] … the agent's.
|
|
3
|
+
|
|
4
|
+
Reply with one JSON object and nothing else:
|
|
5
|
+
|
|
6
|
+
{"items": [{"bucket": "<key>", "text": "<one short line>", "status": "<status or empty>", "reason": "<optional why>",
|
|
7
|
+
"source": "<segment, e.g. O2>", "quote": "<the exact words from that segment>"}],
|
|
8
|
+
"updates": [{"id": "<an open item's id>", "status": "<its new status>", "source": "<segment>", "quote": "<exact words>"}],
|
|
9
|
+
"current_state": "<two or three sentences: where the work stands now, or empty if nothing changed>"}
|
|
10
|
+
|
|
11
|
+
Rules for every item and update:
|
|
12
|
+
- "quote" must be copied word for word from the segment named in "source" (a short phrase is enough). An item whose
|
|
13
|
+
quote isn't in that segment is thrown away, so never paraphrase it.
|
|
14
|
+
- "text" says the item in plain words, one line, so it reads on its own later.
|
|
15
|
+
- File only what the segments say. Nothing from before them, nothing guessed.
|
|
16
|
+
- Don't add conditions, exceptions or reasons the words don't state. If the owner says "yes" to something, file what
|
|
17
|
+
was proposed, and no "unless" or "because" that neither side said.
|
|
18
|
+
- When it's unclear what a short remark refers to, don't guess: file nothing for it. Don't tie it to the nearest
|
|
19
|
+
topic.
|
|
20
|
+
- One thing per item. A message can hold several (a decision, a rejected alternative and a correction).
|
|
21
|
+
- Use a status only from the bucket's list; buckets without statuses take "".
|
|
22
|
+
- Use "updates" when the segments settle an open item listed below: a question answered, a commitment completed or
|
|
23
|
+
dropped, a decision superseded. Don't file a new item for it as well.
|
|
24
|
+
- File nothing for small talk, thanks or tool chatter. An empty list is a fine answer.
|
|
25
|
+
|
|
26
|
+
The buckets you file into, each with what belongs in it:
|