@honcho-ai/claude-honcho 0.3.0-rc.1

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.
Files changed (67) hide show
  1. package/.claude-plugin/plugin.json +21 -0
  2. package/dist/chunk-084dpgp3.js +556 -0
  3. package/dist/chunk-084dpgp3.js.map +10 -0
  4. package/dist/chunk-49wja30f.js +20 -0
  5. package/dist/chunk-49wja30f.js.map +10 -0
  6. package/dist/chunk-4ryv9310.js +15534 -0
  7. package/dist/chunk-4ryv9310.js.map +87 -0
  8. package/dist/chunk-4zgbdwm4.js +261 -0
  9. package/dist/chunk-4zgbdwm4.js.map +10 -0
  10. package/dist/chunk-6yybt92z.js +106 -0
  11. package/dist/chunk-6yybt92z.js.map +9 -0
  12. package/dist/chunk-9pntsn3x.js +35 -0
  13. package/dist/chunk-9pntsn3x.js.map +10 -0
  14. package/dist/chunk-d2gc13m0.js +161 -0
  15. package/dist/chunk-d2gc13m0.js.map +11 -0
  16. package/dist/chunk-dxbm7kmf.js +331 -0
  17. package/dist/chunk-dxbm7kmf.js.map +10 -0
  18. package/dist/chunk-gnp95nzg.js +438 -0
  19. package/dist/chunk-gnp95nzg.js.map +12 -0
  20. package/dist/chunk-jgtvcc65.js +90 -0
  21. package/dist/chunk-jgtvcc65.js.map +10 -0
  22. package/dist/chunk-jr0fd6n2.js +181 -0
  23. package/dist/chunk-jr0fd6n2.js.map +11 -0
  24. package/dist/chunk-wm6hvkcv.js +245 -0
  25. package/dist/chunk-wm6hvkcv.js.map +10 -0
  26. package/dist/chunk-xdntcesh.js +15450 -0
  27. package/dist/chunk-xdntcesh.js.map +86 -0
  28. package/dist/chunk-xtdvbfxw.js +146 -0
  29. package/dist/chunk-xtdvbfxw.js.map +10 -0
  30. package/dist/chunk-yfrrc1d5.js +838 -0
  31. package/dist/chunk-yfrrc1d5.js.map +12 -0
  32. package/dist/hooks/post-tool-use.js +220 -0
  33. package/dist/hooks/post-tool-use.js.map +11 -0
  34. package/dist/hooks/pre-compact.js +217 -0
  35. package/dist/hooks/pre-compact.js.map +12 -0
  36. package/dist/hooks/pre-tool-honcho.js +38 -0
  37. package/dist/hooks/pre-tool-honcho.js.map +11 -0
  38. package/dist/hooks/save-user-message.js +96 -0
  39. package/dist/hooks/save-user-message.js.map +11 -0
  40. package/dist/hooks/session-end.js +51 -0
  41. package/dist/hooks/session-end.js.map +11 -0
  42. package/dist/hooks/session-start.js +246 -0
  43. package/dist/hooks/session-start.js.map +12 -0
  44. package/dist/hooks/stop.js +158 -0
  45. package/dist/hooks/stop.js.map +11 -0
  46. package/dist/hooks/user-prompt.js +19 -0
  47. package/dist/hooks/user-prompt.js.map +10 -0
  48. package/dist/mcp-server.js +13640 -0
  49. package/dist/mcp-server.js.map +110 -0
  50. package/dist/skills/backfill-runner.js +245 -0
  51. package/dist/skills/backfill-runner.js.map +10 -0
  52. package/dist/skills/setup-runner.js +339 -0
  53. package/dist/skills/setup-runner.js.map +12 -0
  54. package/dist/skills/status-runner.js +86 -0
  55. package/dist/skills/status-runner.js.map +10 -0
  56. package/hooks/hooks.json +108 -0
  57. package/mcp-servers.json +10 -0
  58. package/package.json +19 -0
  59. package/scripts/check-version.sh +50 -0
  60. package/scripts/honcho-statusline.sh +101 -0
  61. package/skills/briefing/SKILL.md +20 -0
  62. package/skills/config/SKILL.md +288 -0
  63. package/skills/import/SKILL.md +73 -0
  64. package/skills/insights/SKILL.md +109 -0
  65. package/skills/interview/SKILL.md +61 -0
  66. package/skills/setup/SKILL.md +60 -0
  67. package/skills/status/SKILL.md +30 -0
@@ -0,0 +1,288 @@
1
+ ---
2
+ description: Configure Honcho memory plugin settings interactively
3
+ allowed-tools: get_config, set_config
4
+ user-invocable: true
5
+ ---
6
+
7
+ # Honcho Configuration
8
+
9
+ Interactive configuration for the Honcho memory plugin. Uses AskUserQuestion for all menus and selections — never dump numbered text lists.
10
+
11
+ ## Step 1: Status Header
12
+
13
+ Call `get_config` to load the current state. The response includes a `card` field — a pre-rendered box-drawing card with perfect alignment.
14
+
15
+ **Output the `card` value exactly as-is inside a code fence.** Do not modify it, re-render it, or add any formatting. Just wrap it in triple backticks:
16
+
17
+ ````
18
+ ```
19
+ {card value here, verbatim}
20
+ ```
21
+ ````
22
+ - Do NOT show cache info, config paths, or raw JSON.
23
+ - Do NOT show warnings unless they indicate something is broken (skip env var shadowing warnings where the values match what's configured).
24
+ - If `configExists` is false, tell the user no config exists and offer to create one.
25
+
26
+ ## Step 2: Menu
27
+
28
+ Present ONE question with these options (the user can select "Other" to reach advanced settings):
29
+
30
+ ```
31
+ AskUserQuestion:
32
+ question: "What would you like to configure?"
33
+ header: "Config"
34
+ options:
35
+ - label: "Peers"
36
+ description: "Your name and AI name (currently: {resolved.peerName} / {resolved.aiPeer})"
37
+ - label: "Session mapping"
38
+ description: "How sessions are named — per directory, git branch, or per chat (currently: {resolved.sessionStrategy})"
39
+ - label: "Workspace"
40
+ description: "Data space and session scope (currently: {resolved.workspace})"
41
+ - label: "Memory injection"
42
+ description: "What Honcho injects at session start and per turn (currently: start [{resolved.injection.sessionStart}], turn [{resolved.injection.perTurn}], shown in UI [{resolved.injection.showContents}])"
43
+ ```
44
+
45
+ For the "Memory injection" description, use the *effective* values: if `injection.sessionStart` is unset it is `["directives", "summary", "peerCard"]`, and if `injection.perTurn` is unset it is `["userContext"]` (user conclusions on). A stored perTurn value of `"context"` is the legacy name for `"userContext"` — treat them as the same. `injection.showContents` is `[]` when unset — render that as `none`.
46
+
47
+ If the user selects "Other", present advanced options:
48
+
49
+ ```
50
+ AskUserQuestion:
51
+ question: "Advanced settings:"
52
+ header: "Advanced"
53
+ options:
54
+ - label: "Host"
55
+ description: "Platform / local / custom URL (currently: {current.host})"
56
+ - label: "Context refresh"
57
+ description: "TTL, message threshold, dialectic settings"
58
+ - label: "Message upload"
59
+ description: "Token limits, summarization settings"
60
+ - label: "Statusline"
61
+ description: "Memory statusLine visibility — on / off (currently: {resolved.statusline})"
62
+ ```
63
+
64
+ Always include current values in the description so the user can see what's set.
65
+
66
+ ## Step 3: Handle Selection
67
+
68
+ ### Peers
69
+
70
+ When selected, use `AskUserQuestion` to ask which peer to change:
71
+
72
+ ```
73
+ AskUserQuestion:
74
+ question: "Which peer to change?"
75
+ header: "Peers"
76
+ options:
77
+ - label: "Your name"
78
+ description: "Currently: {resolved.peerName}"
79
+ - label: "AI name"
80
+ description: "Currently: {resolved.aiPeer}"
81
+ ```
82
+
83
+ Then ask for the new value. Call `set_config` with `peerName` or `aiPeer`.
84
+
85
+ ### Simple fields (Logging, etc.)
86
+
87
+ Use `AskUserQuestion` to ask for the new value if there are known options, otherwise ask the user to type it. Call `set_config` with the appropriate field. Show the result.
88
+
89
+ ### Session mapping
90
+
91
+ ```
92
+ AskUserQuestion:
93
+ question: "Which session mapping strategy?"
94
+ header: "Sessions"
95
+ options:
96
+ - label: "per-directory (Recommended)"
97
+ description: "{peer}-{repo} — one session per project"
98
+ - label: "git-branch"
99
+ description: "{peer}-{repo}-{branch} — session follows branch"
100
+ - label: "chat-instance"
101
+ description: "chat-{id} — fresh each launch"
102
+ ```
103
+
104
+ Do NOT use markdown previews for this menu — descriptions are sufficient and previews truncate in narrow terminals.
105
+
106
+ After strategy selection, ask about peer prefix:
107
+
108
+ ```
109
+ AskUserQuestion:
110
+ question: "Include your name in session names?"
111
+ header: "Prefix"
112
+ options:
113
+ - label: "Yes — {peerName}-{repoName}"
114
+ description: "For teams sharing a workspace"
115
+ - label: "No — {repoName} only"
116
+ description: "Cleaner for solo use"
117
+ ```
118
+
119
+ ### Workspace
120
+
121
+ When selected, present a sub-menu:
122
+
123
+ ```
124
+ AskUserQuestion:
125
+ question: "Workspace settings?"
126
+ header: "Workspace"
127
+ options:
128
+ - label: "Rename workspace"
129
+ description: "Change workspace name (currently: {resolved.workspace})"
130
+ ```
131
+
132
+ #### Workspace > Rename
133
+
134
+ Dangerous field — requires confirmation. First call `set_config` WITHOUT `confirm: true`. The tool will return a description of what will happen. Use `AskUserQuestion` to confirm:
135
+
136
+ ```
137
+ AskUserQuestion:
138
+ question: "Switch workspace to '{value}'?"
139
+ header: "Confirm"
140
+ options:
141
+ - label: "Yes, switch"
142
+ description: "Change to the new workspace"
143
+ - label: "Cancel"
144
+ description: "Keep current workspace"
145
+ ```
146
+
147
+ If confirmed, call `set_config` again WITH `confirm: true`.
148
+
149
+ ### Memory injection
150
+
151
+ Ask all three questions in a SINGLE `AskUserQuestion` call — multi-select each — so the user configures every surface at once. Always include all three, each offering all of its options: the "Show in UI" answers arrive in the same response as "Per turn", so there is no way to narrow them to what the user just enabled. A "Show in UI" pick for a component that is off is harmless — it's stored and takes effect if that component is enabled later. This is the only injection prompt; do not add follow-ups.
152
+
153
+ ```yaml
154
+ AskUserQuestion:
155
+ questions:
156
+ - question: "What should Honcho inject at the start of each session?"
157
+ header: "On start" # ≤12 chars
158
+ multiSelect: true
159
+ options:
160
+ - label: "Memory directives"
161
+ description: "How to use memory — treat as background, search, save insights"
162
+ - label: "Session summary"
163
+ description: "Rolling long summary of prior sessions"
164
+ - label: "Peer card"
165
+ description: "Your identity + attributes list"
166
+ - label: "Representation"
167
+ description: "Honcho's derived prose profile of you"
168
+ - question: "What should Honcho inject on each user turn?"
169
+ header: "Per turn"
170
+ multiSelect: true
171
+ options:
172
+ - label: "User conclusions"
173
+ description: "Fresh, prompt-scoped memory about you pulled every turn"
174
+ - label: "Assistant conclusions"
175
+ description: "Same fetch, but for the AI peer — what Honcho knows about the assistant"
176
+ - label: "Session messages"
177
+ description: "Recent raw messages from the mapped Honcho session — useful when other instances share the session"
178
+ - label: "Dialectic recall"
179
+ description: "A reasoned answer over your history each turn — richer but slower (off by default)"
180
+ - question: "Which of those should print what they injected to the terminal?"
181
+ header: "Show in UI"
182
+ multiSelect: true
183
+ options: # same four labels as "Per turn"
184
+ - label: "User conclusions"
185
+ description: "List each injected conclusion instead of just the count"
186
+ - label: "Assistant conclusions"
187
+ description: "List each injected conclusion instead of just the count"
188
+ - label: "Session messages"
189
+ description: "List each injected message, one truncated line each"
190
+ - label: "Dialectic recall"
191
+ description: "Print the full reasoned answer — prose, can be long"
192
+ ```
193
+
194
+ Map the selections to component names, then call `set_config` once per field:
195
+ - Session start → `injection.sessionStart`, mapping "Memory directives"→`directives`, "Session summary"→`summary`, "Peer card"→`peerCard`, "Representation"→`peerRepresentation`.
196
+ - Per turn → `injection.perTurn`, mapping "User conclusions"→`userContext`, "Assistant conclusions"→`assistantContext`, "Session messages"→`sessionContext`, "Dialectic recall"→`dialectic`.
197
+ - Show in UI → `injection.showContents`, same mapping as per turn.
198
+
199
+ Pass the value as a JSON array (e.g. `["directives","summary","peerCard"]`). An empty selection means `[]`: for `injection.sessionStart` and `injection.perTurn` that surface then injects nothing, and for `injection.showContents` (the default) every enabled per-turn component still injects — it just reports a one-line summary (count, tokens, timing) instead of printing its payload.
200
+
201
+ `showContents` is display-only. It never changes what reaches the model.
202
+
203
+ Retrieval tuning is intentionally NOT asked here. `injection.searchTopK` (default 10), `injection.maxConclusions` (15), `injection.searchMaxDistance` (0.6, cosine — lower is stricter), and `injection.searchQuerySource` ("prompt" | "topics", default "prompt") are all configurable via `set_config`, but keep the defaults; only mention they're tunable if the user brings it up, and never prompt for them.
204
+
205
+ If the user enables "Dialectic recall", note the two knobs that shape it — `injection.dialecticTemplate` (the query, with a `%{user_query}` placeholder) and `injection.dialecticReasoning` (tier, default "low") — both via `set_config`. Flag the trade-off: it fires a `chat()` call every non-trivial turn (~12s at medium), on its own budget under the 30s hook ceiling, so it adds real per-turn latency. Keep it off unless the user wants it.
206
+
207
+ ### Dangerous fields (Host)
208
+
209
+ Host changes require confirmation. First call `set_config` WITHOUT `confirm: true`. The tool will return a description of what will happen. Use `AskUserQuestion` to confirm, then call again WITH `confirm: true`.
210
+
211
+ ### Context refresh
212
+
213
+ Use `AskUserQuestion` to pick which setting to change:
214
+
215
+ ```
216
+ AskUserQuestion:
217
+ question: "Which context refresh setting?"
218
+ header: "Refresh"
219
+ options:
220
+ - label: "TTL"
221
+ description: "Cache lifetime — currently {contextRefresh.ttlSeconds}s (default: 300)"
222
+ - label: "Message threshold"
223
+ description: "Refresh every N messages — currently {contextRefresh.messageThreshold} (default: 30)"
224
+ - label: "Skip dialectic"
225
+ description: "Skip chat() in prompt hook — currently {contextRefresh.skipDialectic} (default: false)"
226
+ ```
227
+
228
+ Then ask for the new value and call `set_config`.
229
+
230
+ ### Statusline
231
+
232
+ ```
233
+ AskUserQuestion:
234
+ question: "Memory statusLine visibility?"
235
+ header: "Statusline"
236
+ options:
237
+ - label: "on (Recommended)"
238
+ description: "Sync status, clickable session link, and live activity"
239
+ - label: "off"
240
+ description: "Hidden"
241
+ ```
242
+
243
+ Call `set_config` with field `statusline` and the chosen value. Takes effect on the next statusLine repaint.
244
+
245
+ ### Message upload
246
+
247
+ Use `AskUserQuestion` to pick which setting to change:
248
+
249
+ ```
250
+ AskUserQuestion:
251
+ question: "Which message upload setting?"
252
+ header: "Upload"
253
+ options:
254
+ - label: "Max user tokens"
255
+ description: "Truncate user messages — currently {messageUpload.maxUserTokens || 'no limit'}"
256
+ - label: "Max assistant tokens"
257
+ description: "Truncate assistant messages — currently {messageUpload.maxAssistantTokens || 'no limit'}"
258
+ - label: "Summarize assistant"
259
+ description: "Use summary instead of full text — currently {messageUpload.summarizeAssistant}"
260
+ ```
261
+
262
+ Then ask for the new value and call `set_config`.
263
+
264
+ ## Step 4: Loop
265
+
266
+ After handling a selection, call `get_config` again to refresh state. Use `AskUserQuestion` to ask if they want to configure more:
267
+
268
+ ```
269
+ AskUserQuestion:
270
+ question: "Configuration updated. What next?"
271
+ header: "Next"
272
+ options:
273
+ - label: "Configure more"
274
+ description: "Return to settings menu"
275
+ - label: "Done"
276
+ description: "Exit configuration"
277
+ ```
278
+
279
+ If "Configure more", go back to Step 2. If "Done", show the final status header and exit.
280
+
281
+ ## Guardrails
282
+
283
+ - ALWAYS use AskUserQuestion for menus and confirmations. Never present numbered text lists.
284
+ - Always show the result of `set_config` including any cache invalidation that occurred.
285
+ - If a warning about env var shadowing is returned, explain that the env var takes precedence at runtime.
286
+ - Never guess values — always ask the user.
287
+ - Include current values in option descriptions so the user sees what's set without expanding anything.
288
+ - If `get_config` returns `configExists: false`, guide the user to set HONCHO_API_KEY first.
@@ -0,0 +1,73 @@
1
+ ---
2
+ description: Import past local Claude Code sessions into Honcho memory (backfill history)
3
+ user-invocable: true
4
+ ---
5
+
6
+ # Honcho Import
7
+
8
+ Backfill the user's existing local Claude Code session transcripts into Honcho so
9
+ past work becomes apart of Honcho's memory. It uploads real conversation content, so it
10
+ always previews first and requires explicit confirmation.
11
+
12
+ Transcripts live at `~/.claude/projects/<cwd-as-dashes>/<session-uuid>.jsonl`. The
13
+ importer reconstructs each conversation, names sessions using the user's configured
14
+ `sessionStrategy`, and uploads with the original message timestamps into the user's
15
+ configured workspace.
16
+
17
+ ## Prerequisite
18
+
19
+ Honcho must already be configured (`HONCHO_API_KEY` set or `~/.honcho/config.json`
20
+ present). If not, tell the user to run `/honcho:setup` first and stop here.
21
+
22
+ ## Optional arguments
23
+
24
+ The user may specify scope in natural language; translate it to flags:
25
+ - window → `--days N` (default 30)
26
+ - a specific workspace → `--workspace NAME` (default: their configured workspace)
27
+
28
+ ## Steps
29
+
30
+ ### 1. Preview (safe — no upload)
31
+
32
+ Run the importer in dry-run mode and show the user the plan:
33
+
34
+ ```bash
35
+ node "${CLAUDE_PLUGIN_ROOT}/dist/skills/backfill-runner.js" --dry-run
36
+ ```
37
+
38
+ If `CLAUDE_PLUGIN_ROOT` isn't set, the plugin lives here (find the version
39
+ directory inside and run the importer from there):
40
+
41
+ ```bash
42
+ echo "$HOME/.claude/plugins/cache/honcho/honcho"
43
+ ```
44
+
45
+ Report the workspace, session count, and message count. Note that any sessions
46
+ already imported are skipped automatically (the importer is idempotent).
47
+
48
+ ### 2. Confirm
49
+
50
+ Ask the user to confirm before uploading. **Do not upload without an explicit yes**
51
+ — this sends real conversation content to Honcho.
52
+
53
+ ### 3. Import
54
+
55
+ On confirmation, run the real import (add `--days`/`--workspace` if the user asked):
56
+
57
+ ```bash
58
+ node "${CLAUDE_PLUGIN_ROOT}/dist/skills/backfill-runner.js" --yes
59
+ ```
60
+
61
+ ### 4. Report
62
+
63
+ Tell the user, honestly and without first-run/restart framing:
64
+
65
+ > Imported <N> message(s) across <M> session(s) into `<workspace>`. Honcho will
66
+ > start reasoning over them (no restart needed).
67
+
68
+ Notes to relay if relevant:
69
+ - Idempotent: already-imported transcripts are recorded in
70
+ `~/.honcho/backfill-state.json` and skipped on re-run, so it's safe to run again
71
+ (e.g. with a wider `--days` window) to pull in more history later.
72
+ - Sessions are named by the configured `sessionStrategy`, so imported history lines
73
+ up with future live sessions.
@@ -0,0 +1,109 @@
1
+ ---
2
+ description: Distill accumulated Honcho memory into concrete Claude Code config updates — CLAUDE.md edits, output-style rules, and skill ideas
3
+ allowed-tools: chat, honcho_remember, get_context, Read, Glob, AskUserQuestion, Write, Edit
4
+ user-invocable: true
5
+ ---
6
+
7
+ # Honcho Insights
8
+
9
+ Run a deep dialectic pass over everything Honcho has accumulated about the user, and turn it
10
+ into changes they can actually make to their Claude Code setup: CLAUDE.md rules, output-style
11
+ adjustments, and skills worth writing.
12
+
13
+ This is a distillation, not a recall. The dialectic already answers "what do you know about
14
+ me". The job here is to convert that into a short, ranked list of edits.
15
+
16
+ ## Arguments
17
+
18
+ `/honcho:insights` takes an optional focus, e.g. `/honcho:insights prose` or
19
+ `/honcho:insights skill ideas for the deploy flow`. When given, aim every query at that focus
20
+ and drop the other axes. With no argument, cover all of them.
21
+
22
+ ## Step 1: Read what already exists
23
+
24
+ Proposals are diffs, not restatements. Before querying, read the current setup so anything
25
+ already written down can be excluded:
26
+
27
+ - `~/.claude/CLAUDE.md` and the project's `CLAUDE.md` (repo root, plus `.claude/CLAUDE.md`)
28
+ - the active output style, if any, under `~/.claude/output-styles/`
29
+ - skill names under `~/.claude/skills/` and `.claude/skills/` (names only, don't read bodies)
30
+
31
+ Missing files are normal. Note what you found in one line, then move on.
32
+
33
+ ## Step 2: Query the dialectic
34
+
35
+ Prefer **one `chat` call at `reasoning_level: "max"`**. The dialectic plans its own retrieval,
36
+ so a single wide question beats several narrow ones, and max reasoning is the point of this
37
+ command. Include the "already covered" list in the prompt so it spends its budget on gaps.
38
+
39
+ With a focus argument, replace the numbered axes below with that focus and keep the rest of
40
+ the framing — the evidence-and-confidence ask and the exclusion list — intact.
41
+
42
+ ```
43
+ chat({
44
+ reasoning_level: "max",
45
+ query: "I am distilling what you know about this user into their Claude Code configuration. \
46
+ Give me behavior-shaping observations, not biography. Cover: (1) how they want an agent to \
47
+ work — autonomy, what needs approval, how they react to being asked vs. told; (2) \
48
+ communication and prose preferences, with concrete examples of phrasing they accepted or \
49
+ rejected; (3) multi-step workflows they have walked through more than once, which could \
50
+ become a reusable skill; (4) corrections they have given — things they asked not to be done \
51
+ again; (5) tooling, environment, and project defaults that should be assumed rather than \
52
+ asked. For each observation, state the evidence you are drawing on and how confident you are. \
53
+ Skip anything on this list, which is already written into their config: <the list from step 1>."
54
+ })
55
+ ```
56
+
57
+ If that call errors or hits the 120s ceiling, fall back to `honcho_remember` with
58
+ `reasoning_level: "high"`. That tool is opt-in and only registered when the `rememberTool`
59
+ config flag is on, so it may not be available — if it is missing, treat this path as thin and
60
+ go straight to `get_context`. With a focus argument, send one query for that focus; with no
61
+ argument, send all five axes (they run in parallel):
62
+
63
+ 1. How does this user want an agent to behave — autonomy, gating, what needs approval first?
64
+ 2. What are their communication and prose preferences, with examples they accepted or rejected?
65
+ 3. What multi-step workflows have they repeated, that could be automated as a skill?
66
+ 4. What corrections or frustrations have they voiced — things to stop doing?
67
+ 5. What tooling, environment, and project defaults should be assumed without asking?
68
+
69
+ Use `get_context` only if both paths come back thin — it is raw material, not analysis.
70
+
71
+ ## Step 3: Filter
72
+
73
+ Keep an observation only if all of these hold:
74
+
75
+ - **Actionable**: it changes what the agent does, not just what it knows. "Works at a startup"
76
+ is out; "wants the cost of the rejected option named" is in.
77
+ - **Recurring**: true across sessions, not a one-off from a single conversation.
78
+ - **New**: not already covered by the files from step 1.
79
+ - **Placeable**: you can say which file it belongs in.
80
+
81
+ Convert relative dates to absolute. Drop anything you cannot ground in something the dialectic
82
+ actually said — an invented rule is worse than a short list. If the dialectic came back thin,
83
+ say so plainly and stop; do not pad.
84
+
85
+ ## Step 4: Present
86
+
87
+ Cap the whole thing at ~10 items, ranked by how much they would change day-to-day behavior.
88
+ Each item is at most three lines: the rule, the evidence, the destination. No preamble.
89
+
90
+ Group under whichever of these have content:
91
+
92
+ - **CLAUDE.md** — proposed lines, written in the voice of the target file, split by global vs.
93
+ project. Show them diff-shaped (`+` for added, `~` for changed).
94
+ - **Output style** — tone and formatting rules. Give one rendered example per rule; a rule
95
+ about prose is unreviewable without a sample of the prose.
96
+ - **Skill ideas** — name, trigger phrase, what it would encode, and the evidence that it
97
+ happens often enough to be worth writing.
98
+
99
+ Then stop and wait. Do not start editing.
100
+
101
+ ## Step 5: Apply, on request only
102
+
103
+ Never write a file before the user picks. Once they have:
104
+
105
+ - Ask with `AskUserQuestion`: apply all, pick a subset, or nothing.
106
+ - Edit in place, matching the target file's existing structure and voice. Append to an existing
107
+ section rather than creating a new one where one fits.
108
+ - Skill ideas get scaffolded only if asked for by name — a skill nobody asked for is clutter.
109
+ - Recap what changed in one line per file.
@@ -0,0 +1,61 @@
1
+ ---
2
+ description: Interview the user to capture stable, cross-project preferences and save them to Honcho
3
+ allowed-tools: chat, create_conclusion
4
+ user-invocable: true
5
+ ---
6
+
7
+ # Honcho Interview
8
+
9
+ Kick off a short interview to learn stable, cross-project aspects of the user and store them in Honcho memory.
10
+
11
+ ## Guardrails
12
+
13
+ - Focus on global traits that are unlikely to change between projects.
14
+ - Avoid project-specific topics, credentials, addresses, or other sensitive information.
15
+ - Ask one question at a time and wait for the answer before proceeding.
16
+ - If an answer is vague, ask one brief clarification before saving a conclusion.
17
+ - If the user declines to answer, skip that topic and move on.
18
+ - Use existing knowledge to avoid repeating questions the memory already covers.
19
+
20
+ ## Pre-Interview Context (Required)
21
+
22
+ Before asking any questions, use the `chat` tool to get a maximally thorough overview of what is already known about the user. Present a concise summary to the user, then tailor the interview to fill gaps or confirm uncertain areas.
23
+
24
+ Example tool call format:
25
+
26
+ ```
27
+ chat({ "query": "Give a maximally thorough overview of what you already know about this user, focusing on stable preferences and cross-project traits. Include any uncertainties or gaps." })
28
+ ```
29
+
30
+ ## Interview Flow (Medium Depth)
31
+
32
+ Ask these questions in order, skipping any that are already answered by the pre-interview context:
33
+
34
+ 1. Communication style: Do you prefer concise answers, detailed explanations, or a mix?
35
+ 2. Tone: Do you prefer a direct, professional tone or a more conversational one?
36
+ 3. Structure: Do you prefer bullet points, step-by-step instructions, or narrative explanations?
37
+ 4. Technical depth: What level of technical detail should I assume (beginner, intermediate, expert)?
38
+ 5. Learning preference: Do you prefer explanations first, examples first, or both together?
39
+ 6. Code quality focus: What matters most by default (clarity, performance, tests, minimal changes)?
40
+ 7. Collaboration style: Should I make changes directly, propose options first, or ask before edits?
41
+ 8. Environment defaults: What OS/shell/tooling should I assume for commands and paths?
42
+
43
+ ## Saving Conclusions
44
+
45
+ After each answer, create exactly one concise conclusion and call `create_conclusion`.
46
+
47
+ Guidelines for conclusions:
48
+
49
+ - Use a single sentence.
50
+ - Make it specific and unambiguous.
51
+ - Avoid hedging if the user gives a clear preference.
52
+
53
+ Example tool call format:
54
+
55
+ ```
56
+ create_conclusion({ "content": "Prefers concise, bullet-pointed responses with a professional tone." })
57
+ ```
58
+
59
+ ## Wrap-up
60
+
61
+ When finished, briefly recap the conclusions you saved and ask if anything should be corrected. Only save a new conclusion if the user explicitly clarifies or corrects a prior answer.
@@ -0,0 +1,60 @@
1
+ ---
2
+ description: First-time Honcho configuration -- set API key, validate connection, create config
3
+ user-invocable: true
4
+ ---
5
+
6
+ # Honcho Setup
7
+
8
+ Walk the user through first-time Honcho configuration so persistent memory works in Claude Code.
9
+
10
+ ## Steps
11
+
12
+ ### 1. Validate the connection
13
+
14
+ Run the setup runner. It's a committed script (`src/skills/setup-runner.ts`) that
15
+ checks for an API key (env var, then `~/.honcho/config.json`), validates the
16
+ connection, writes the config, and installs the memory statusLine:
17
+
18
+ ```bash
19
+ node "${CLAUDE_PLUGIN_ROOT}/dist/skills/setup-runner.js"
20
+ ```
21
+
22
+ If `CLAUDE_PLUGIN_ROOT` isn't set, the plugin lives here (find the version
23
+ directory inside and run the runner from there):
24
+
25
+ ```bash
26
+ echo "$HOME/.claude/plugins/cache/honcho/honcho"
27
+ ```
28
+
29
+ **If the runner reports "No API key found"**, it prints how to set one and exits.
30
+ Relay that to the user:
31
+
32
+ - Get a free key at https://app.honcho.dev
33
+ - **macOS / Linux:** add to `~/.zshrc` (or `~/.bashrc`): `export HONCHO_API_KEY="your-key-here"`
34
+ - **Windows (PowerShell):** `setx HONCHO_API_KEY "your-key-here"`
35
+ - Then restart Claude Code and run `/honcho:setup` again.
36
+
37
+ Do NOT ask the user to paste their key into the chat — it must be set as an
38
+ environment variable outside Claude Code. Stop here and wait for them to come back.
39
+
40
+ **If the runner succeeds**, the key is valid and the config file has been created.
41
+ The statusLine renderer is copied to `~/.honcho/honcho-statusline.sh` and registered
42
+ in `~/.claude/settings.json` (only when no `statusLine` is already configured — an
43
+ existing one is left untouched and the path is printed for manual use). Toggle
44
+ visibility later with the `statusline` key in `~/.honcho/config.json`: `on`
45
+ (default) or `off`.
46
+
47
+ If it fails for another reason:
48
+ - Authentication error: key may be invalid — get a new one at https://app.honcho.dev
49
+ - Network error: check the internet connection
50
+
51
+ ### 2. Confirm setup
52
+
53
+ Tell the user that Honcho is configured and memory will be active on their next session. Suggest they restart Claude Code to see the memory context load.
54
+
55
+ If the setup runner (step 1) reported that past local sessions are available to
56
+ import, mention it as an optional next step — do **not** import here. Point them to
57
+ the dedicated command:
58
+
59
+ > You also have past Claude Code sessions that can be imported into Honcho.
60
+ > Run `/honcho:import` whenever you'd like to backfill them.
@@ -0,0 +1,30 @@
1
+ ---
2
+ description: Show current Honcho memory status and configuration
3
+ user-invocable: true
4
+ ---
5
+
6
+ # Honcho Status
7
+
8
+ Display the current Honcho memory system status: connection health, workspace, peers, observation queue, and conclusion count.
9
+
10
+ ## What It Shows
11
+
12
+ 1. **Connection** - Live connection health and latency
13
+ 2. **Workspace** - Current workspace and endpoint URL
14
+ 3. **Peers** - User peer and AI peer names
15
+ 4. **Observing** - Queue processing status (messages observed, active, sessions)
16
+ 5. **Conclusions** - Total conclusion/memory count for the user peer
17
+
18
+ ## Usage
19
+
20
+ Run `/honcho:status` to see the current state of the Honcho memory system.
21
+
22
+ ## Presentation
23
+
24
+ After running the script, present a concise status card echoing the runner output. Do NOT add prose commentary — the output speaks for itself. Only add a one-line note if something looks wrong (e.g. auth failed, unreachable, 0 conclusions).
25
+
26
+ ## Implementation
27
+
28
+ ```bash
29
+ node ${CLAUDE_PLUGIN_ROOT}/dist/skills/status-runner.js
30
+ ```