extract-agent-sessions

Name: extract-agent-sessions
Author: shareAI-lab

304stars52forks

How to install extract-agent-sessions?

shareai-lab/shareai-skills/extract-agent-sessions

$npx -y skills add shareAI-lab/shareAI-skills --skill extract-agent-sessions

Installs into the current project.

Summary

Recover, export, or summarize recent human and assistant conversations from local agent session stores while excluding tool calls, reasoning, system injection, copied parent history, and subagent noise. Use for recent-session inventories, clean conversation recovery, or topic sum

Files · 1

View on GitHub

SKILL.md

1	# Extract Agent Sessions
2
3	Use this skill as a storage map and decision guide, not as a stable parser. Treat
4	all local session layouts as unversioned implementation details. Inspect the files
5	that exist now, adapt shell commands to their observed shape, and fail closed when
6	the shape is ambiguous.
7
8	Do not create or install a persistent extraction script. Compose one-off commands
9	from the current filesystem and shell capabilities.
10
11	## Safety defaults
12
13	- Treat every session log as sensitive. It may contain source code, personal data,
14	secrets typed by the user, absolute paths, or credentials echoed by tools.
15	- Keep discovery local. Never upload a log or extracted artifact unless the user
16	explicitly asks for that destination.
17	- List candidates as temporary aliases such as `S1` and `S2`. Do not display prompt
18	previews, project names, working directories, complete IDs, or absolute paths by
19	default. Truncation is not redaction.
20	- Read message bodies only after the requested scope identifies the relevant
21	session or the user selects an alias. If content would be sent to another service
22	or person, disclose that boundary and obtain confirmation.
23	- Write private intermediate artifacts only under an owner-restricted temporary
24	directory. Remove them on normal exit and interruption unless the user explicitly
25	requests a private export.
26	- Never inspect authentication files, credential stores, environment snapshots,
27	shell history, debug logs, telemetry, crash dumps, clipboard data, or browser
28	storage for this task.
29
30	## Output modes
31
32	Choose one mode before reading conversation bodies:
33
34	- Inventory: aliases, store family, UTC activity time, transcript availability,
35	state, and warnings. This is the default discovery mode.
36	- Private full: every persisted human message and visible assistant text block
37	in order. Keep this owner-only and do not claim to redact it.
38	- Shareable summary: themes and outcomes with paths, stable IDs, attachment
39	references, secrets, and unnecessary verbatim text removed.
40
41	“Lossless” means lossless relative to persisted visible text. Preserve a stored
42	placeholder such as `[Pasted Content ...]` or `[Image #1]` in private-full mode, but
43	state that the externalized body is unavailable.
44
45	## Select the shell guidance
46
47	Resolve the current user's home directory without guessing another user's account.
48	This version supports Linux and macOS hosts:
49
50	- Read [references/posix-shell.md](references/posix-shell.md) and compose commands
51	for the shell and utilities actually present.
52	- Under WSL, treat the Linux distribution as a separate environment and inspect it
53	only when the agent client ran there. Native Windows session stores are outside
54	this version's supported scope.
55
56	Read [references/session-layouts.md](references/session-layouts.md) for the relevant
57	store family before selecting identity or message fields.
58
59	## Workflow
60
61	### 1. Freeze a UTC window
62
63	Interpret “recent two days” as a rolling 48-hour interval unless the user asks for
64	calendar days. Capture one `[start, end]` snapshot and apply both bounds throughout
65	the scan so new writes do not move the reported window.
66
67	Use history or visible-message timestamps for discussion recency. Use file size and
68	mtime only to detect active writes or prioritize fallback inspection; never treat
69	mtime as the conversation time.
70
71	### 2. Discover from small indexes
72
73	Inspect keys and timestamp units before filtering the small history indexes. Keep
74	only the latest in-window entry per session ID in a private candidate map. Display
75	anonymous aliases and metadata only.
76
77	History is a discovery index, not canonical conversation text. It can contain slash
78	commands, local shell commands, or abbreviated pasted content.
79
80	### 3. Resolve and prove root identity
81
82	Resolve transcripts only for selected candidates. Constrain every path to an
83	approved data root, reject symlinks and unsafe IDs, and compare basenames exactly.
84
85	For tree-shaped stores, accept only root-level project transcripts and exclude
86	every `subagents/` descendant. For rollout-shaped stores, inspect the first
87	projected `session_meta` before reading later records; reject explicit subagents
88	before they can contribute copied parent history.
89
90	### 4. Fingerprint the schema without content
91
92	Project only key names, value types, role/type enums, content-block type names, and
93	counts. Do not print IDs, paths, timestamps, or text values. Collapse unrecognized
94	enum strings to an `unknown-string` marker, then compare the fingerprint to the
95	observed patterns in the reference.
96
97	If require

People who liked this also viewed

6 picks

Type

extract-agent-sessions

304stars52forks

View on GitHub

How to install extract-agent-sessions?

shareai-lab/shareai-skills/extract-agent-sessions

$npx -y skills add shareAI-lab/shareAI-skills --skill extract-agent-sessions

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Extract Agent Sessions
2
3	Use this skill as a storage map and decision guide, not as a stable parser. Treat
4	all local session layouts as unversioned implementation details. Inspect the files
5	that exist now, adapt shell commands to their observed shape, and fail closed when
6	the shape is ambiguous.
7
8	Do not create or install a persistent extraction script. Compose one-off commands
9	from the current filesystem and shell capabilities.
10
11	## Safety defaults
12
13	- Treat every session log as sensitive. It may contain source code, personal data,
14	secrets typed by the user, absolute paths, or credentials echoed by tools.
15	- Keep discovery local. Never upload a log or extracted artifact unless the user
16	explicitly asks for that destination.
17	- List candidates as temporary aliases such as `S1` and `S2`. Do not display prompt
18	previews, project names, working directories, complete IDs, or absolute paths by
19	default. Truncation is not redaction.
20	- Read message bodies only after the requested scope identifies the relevant
21	session or the user selects an alias. If content would be sent to another service
22	or person, disclose that boundary and obtain confirmation.
23	- Write private intermediate artifacts only under an owner-restricted temporary
24	directory. Remove them on normal exit and interruption unless the user explicitly
25	requests a private export.
26	- Never inspect authentication files, credential stores, environment snapshots,
27	shell history, debug logs, telemetry, crash dumps, clipboard data, or browser
28	storage for this task.
29
30	## Output modes
31
32	Choose one mode before reading conversation bodies:
33
34	- Inventory: aliases, store family, UTC activity time, transcript availability,
35	state, and warnings. This is the default discovery mode.
36	- Private full: every persisted human message and visible assistant text block
37	in order. Keep this owner-only and do not claim to redact it.
38	- Shareable summary: themes and outcomes with paths, stable IDs, attachment
39	references, secrets, and unnecessary verbatim text removed.
40
41	“Lossless” means lossless relative to persisted visible text. Preserve a stored
42	placeholder such as `[Pasted Content ...]` or `[Image #1]` in private-full mode, but
43	state that the externalized body is unavailable.
44
45	## Select the shell guidance
46
47	Resolve the current user's home directory without guessing another user's account.
48	This version supports Linux and macOS hosts:
49
50	- Read [references/posix-shell.md](references/posix-shell.md) and compose commands
51	for the shell and utilities actually present.
52	- Under WSL, treat the Linux distribution as a separate environment and inspect it
53	only when the agent client ran there. Native Windows session stores are outside
54	this version's supported scope.
55
56	Read [references/session-layouts.md](references/session-layouts.md) for the relevant
57	store family before selecting identity or message fields.
58
59	## Workflow
60
61	### 1. Freeze a UTC window
62
63	Interpret “recent two days” as a rolling 48-hour interval unless the user asks for
64	calendar days. Capture one `[start, end]` snapshot and apply both bounds throughout
65	the scan so new writes do not move the reported window.
66
67	Use history or visible-message timestamps for discussion recency. Use file size and
68	mtime only to detect active writes or prioritize fallback inspection; never treat
69	mtime as the conversation time.
70
71	### 2. Discover from small indexes
72
73	Inspect keys and timestamp units before filtering the small history indexes. Keep
74	only the latest in-window entry per session ID in a private candidate map. Display
75	anonymous aliases and metadata only.
76
77	History is a discovery index, not canonical conversation text. It can contain slash
78	commands, local shell commands, or abbreviated pasted content.
79
80	### 3. Resolve and prove root identity
81
82	Resolve transcripts only for selected candidates. Constrain every path to an
83	approved data root, reject symlinks and unsafe IDs, and compare basenames exactly.
84
85	For tree-shaped stores, accept only root-level project transcripts and exclude
86	every `subagents/` descendant. For rollout-shaped stores, inspect the first
87	projected `session_meta` before reading later records; reject explicit subagents
88	before they can contribute copied parent history.
89
90	### 4. Fingerprint the schema without content
91
92	Project only key names, value types, role/type enums, content-block type names, and
93	counts. Do not print IDs, paths, timestamps, or text values. Collapse unrecognized
94	enum strings to an `unknown-string` marker, then compare the fingerprint to the
95	observed patterns in the reference.
96
97	If require

People who liked this also viewed

6 picks

Type