@jjanczur/tyran 0.1.17 → 0.1.19

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.
@@ -11,7 +11,7 @@
11
11
  {
12
12
  "name": "tyran",
13
13
  "description": "A task conductor for Claude Code: multi-agent orchestration with an enforced evidence contract, repo-specific learning, and update-safe local evolution.",
14
- "version": "0.1.17",
14
+ "version": "0.1.19",
15
15
  "source": "./",
16
16
  "author": {
17
17
  "name": "Jacek Janczura",
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "tyran",
3
- "version": "0.1.17",
3
+ "version": "0.1.19",
4
4
  "description": "A task conductor for Claude Code: multi-agent orchestration with an enforced evidence contract, repo-specific learning, and update-safe local evolution.",
5
5
  "author": {
6
6
  "name": "Jacek Janczura",
package/CHANGELOG.md CHANGED
@@ -1,5 +1,118 @@
1
1
  # Changelog
2
2
 
3
+ ## 0.1.19 — 2026-08-14
4
+
5
+ ### The board answered four questions on one scroll
6
+
7
+ `board.html` is now four tabs. **Overview** — what is waiting on you, agents
8
+ running, progress, tickets that need a human, then the agent strip. **Board** —
9
+ the kanban lanes, where every card is a button that opens a detail panel
10
+ carrying its lane, initiative, agents, note and, once spend has loaded, that
11
+ ticket's tokens and cost. **Waiting on you** — the queue, with the three
12
+ commands that answer it printed above the questions, and its open count in the
13
+ tab label so a pending decision survives being on another tab. **Spend** —
14
+ unchanged in substance, still FETCHED rather than embedded, so `board.json` and
15
+ `board.html` keep their byte-exact `--check` contract.
16
+
17
+ The palette is pulled back. The first version filled whole bars and whole cards
18
+ at full saturation, which reads as an alarm rather than as information. The
19
+ roles are unchanged — one warm accent for the operator's call to action, a cool
20
+ one for the agent strip, red for refusal, green for the ledger's `+` — but
21
+ saturation is now spent on text, edges and 3px rails, and every large fill is a
22
+ muted tone of the same hue. The tokens are `--brass`, `--steel`, `--clay`,
23
+ `--sage`.
24
+
25
+ ### The README explained the problem before it explained the product
26
+
27
+ Rebuilt, 491 lines to 204: what Tyran is, a recorded demo of the board, what it
28
+ gives you, a comparison, install, use, the dashboard. Everything cut was first
29
+ given a home on BOTH doc surfaces — nine sections that lived nowhere else were
30
+ moved rather than deleted, and four inbound links that pointed at README
31
+ anchors now resolve inside `docs/`.
32
+
33
+ The version badge read GitHub Releases, where the newest was 0.1.3 while
34
+ sixteen versions had shipped by tag and by npm. It now reads npm, which the
35
+ release itself maintains and which cannot go stale.
36
+
37
+ **Upgrading: regenerate your board.** `board.html` changed again, so
38
+ `board.mjs --check` reports drift until you run
39
+ `node scripts/board.mjs --dir .tyran` and commit. `board.json` and `BOARD.md`
40
+ are unchanged.
41
+
42
+ ## 0.1.18 — 2026-08-14
43
+
44
+ ### The spend ledger: nobody could say what a run cost
45
+
46
+ `node scripts/cost.mjs` reports what the work cost, in tokens the platform
47
+ itself reported, per model, per agent type and per ticket. It **reads, it never
48
+ instruments**: every number comes from the `message.usage` records Claude Code
49
+ already writes under `~/.claude/projects/`, so there is no hook, no probe, no
50
+ new event type and nothing to keep in sync.
51
+
52
+ It is deliberately **not a projection**, and that is the decision the rest
53
+ follows from. Its inputs are machine-local and differ per clone, so two people
54
+ with the same journal legitimately get different numbers — which is exactly
55
+ what a byte-exact `--check` would report as drift. `cost.json` is therefore
56
+ never committed (it joins the runtime files `.tyran/.gitignore` excludes),
57
+ never byte-compared, and never merged into `board.json`.
58
+
59
+ **Tokens are facts; money is an opinion.** Only tokens are counted. Money comes
60
+ from an operator-written `pricing:` block, and the card's name travels with
61
+ every amount — a figure that cannot say which card produced it is not a
62
+ measurement. Absent pricing gives tokens and no money, which is the honest
63
+ default: Tyran does not know what anyone pays. All four rate keys are required
64
+ per model, because a table carrying three would price the fourth at zero,
65
+ silently, and cache reads alone were measured at roughly three quarters of a
66
+ real session's cost.
67
+
68
+ Gaps are reported rather than zeroed: `unpriced` models are counted in tokens
69
+ and absent from every amount, `unattributed` agents are the ones whose task
70
+ description carried no ticket id, `unreadable` transcripts are named. A model
71
+ that billed zero tokens is not reported as unpriced — that would be a permanent
72
+ false warning about something that cannot cost anything. And partial pricing is
73
+ a partial sum: an amount is blank only when nothing in the row is priced at
74
+ all, because a real rate card always misses something and nulling the grand
75
+ total over one absent row would leave every figure blank forever.
76
+
77
+ The conductor is its own row. Its context is not attributable to any ticket and
78
+ was measured between 44% and 86% of a tree's tokens, so a per-ticket table that
79
+ omitted it would present a fraction of the bill as the cost of the work. That
80
+ share is reported in tokens, never dollars: a partial rate card prices only some
81
+ models, so a dollar share is a ratio over whichever subset happened to be
82
+ priced — the same tree read 86% by tokens and 99% by dollars.
83
+
84
+ Ticket attribution cost nothing to add. The conductor already puts the ticket id
85
+ at the head of a `Task` description; the reader takes it back out of the agent's
86
+ `.meta.json`. No journal change, no event-set change, and an agent whose
87
+ description omits the id lands in a visible `unattributed` bucket rather than
88
+ being guessed at.
89
+
90
+ `board.html` gains a **Spend** section, and it **fetches `cost.json` rather than
91
+ embedding it** — a deliberate amendment to the page's "self-contained, no
92
+ network" description. It is still zero external hosts and zero CDNs; what it
93
+ makes is one same-origin request to loopback. Over `file://` there is no server,
94
+ the request fails, and the section never appears, so `board.json` and
95
+ `board.html` keep their byte-exact `--check` contract and carry no machine-local
96
+ data. The `/cost.json` route is covered by the existing `Host` pin (a foreign
97
+ `Host` gets 403, verified) and returns 503 rather than taking the board down if
98
+ the cost read fails.
99
+
100
+ Transcripts reach tens of megabytes and the page refreshes every 30 seconds, so
101
+ the reader streams in chunks and caches per source file keyed on `(path, mtime,
102
+ size)`. A finished agent's transcript never changes, so only the running
103
+ session's file is re-read. Measured cold: a 226-transcript tree totalling 6.86
104
+ billion tokens scans in 1.7 s.
105
+
106
+ Also reachable as `npx @jjanczur/tyran cost`. Exit 0 ok · 2 usage, I/O, or no
107
+ transcripts found — spend is read, and nothing here estimates it.
108
+
109
+ **Upgrading: regenerate your board.** `board.html` gained the Spend section, so
110
+ its BYTES changed even though its contract did not. A repo that committed a
111
+ board rendered by an earlier version will see `board.mjs --check` report drift —
112
+ and `doctor --state` report `projection-drift` — until you run
113
+ `node scripts/board.mjs --dir .tyran` and commit the result. `board.json` and
114
+ `BOARD.md` are byte-identical to 0.1.17.
115
+
3
116
  ## 0.1.17 — 2026-08-14
4
117
 
5
118
  ### A race test that picked a winner