assumption-ledger

Claude gets a tool, register_assumption. While it works it uses the tool to write down three kinds of thing:

• assumption: something it took for granted and did not check.
• decision: a fork in the road and the way it went.
• considered-not-done: a better or fuller fix it thought of and chose not to do. This is the one to watch. Models often see the right solution and then skip it, and you never find out.

When the turn ends, the list shows above the prompt, grouped by kind:

Ledger for this turn [ Dismiss ]
Considered, not done (1)
 · Add retry on 429 (out of scope)       [ Do it ]
Assumed (1)
 · Node 20 is the runtime

Do it sends Claude a follow-up asking it to do that item, or to say in one line why it should stay undone. Dismiss clears the band. A new turn clears it too.

Commands

| Command | What it does | | --- | --- | | /assumptions | Prints every entry from this session, grouped by kind, with the turn each came from. | | /assumptions write [path] | Appends the session's entries to DECISIONS-log.md (or path) in the working directory. Nothing is written to disk unless you run this. | | /assumptions off | Stops recording and moves the tool behind ToolSearch, so the model stops seeing it. Remembered across sessions. | | /assumptions on | Turns it back on. |

What it costs

The mod makes no model calls of its own. Its cost is what the model spends using the tool:

| When | Tokens | Notes | | --- | --- | --- | | Every request | about 200 input tokens | The tool's name, description and schema in the tool list. The description never changes, so after the first request this is a prompt-cache read (about a tenth of the price). | | Each entry Claude records | about 40 to 80 output tokens, plus a 2-token result | One tool call. Claude often batches it with other tool calls in the same step; when it doesn't, the step after it re-reads the context from cache. | | A turn where nothing was worth noting | 0 extra | The tool's description tells the model to skip the obvious. | | /assumptions off and on | one cache rebuild each | Changing whether the tool is listed changes the tool list, which is part of the cached prefix. | | Do it | one new turn | You are asking Claude to do more work, so that turn costs what the work costs. |

The band and the session log cost nothing in tokens. /assumptions prints its list as a command-output row in the transcript; whether the model reads that row on the next turn, as it does other command output, was not checked.

Composes with

• Anything drawing above the prompt. The band draws whatever other mods draw there beneath its own list ({await next(e)}), and it steps aside while a survey holds the band or a turn is running.
• A next-steps supervisor or quiz fork. A fork that asks "did this solve the original goal, was it lazy?" can read this mod's state ($.state.get({ plugin: 'assumption-ledger', key: 'pending' }), which holds the current turn's entries until the next turn starts; shown is only filled once this mod's own turn.complete hook has run, so another mod's turn.complete may see it empty) and put the considered-not-done items in its question. That is cheaper and more honest than asking the fork to guess what was skipped.
• A mode selector or model router. A mode switches the ledger off by running /assumptions off through $.command.run: that stops recording, moves the tool behind ToolSearch and remembers the choice. Writing { plugin: 'assumption-ledger', key: 'isOff' } through state.set alone only stops recording: the tool stays in the list (its tool.describe answer is cached until invalidated) and the choice is not remembered.
• A shared project board. /assumptions write gives a file another Claude or a dashboard can read.

Install and load

claude --plugin-dir /path/to/assumption-ledger

Or copy the folder into your mods folder. Check it before loading:

claude plugin validate /path/to/assumption-ledger
claude plugin test /path/to/assumption-ledger

Built and checked against Claude Code 2.1.286.