Adaptive failure taxonomies

AdaMAST¶

AdaMAST checks work, records failures, and learns a project vocabulary from completed traces.

Install in Codex or Claude See a complete run View architecture

Observe concrete activity Correlate supported causes Map failure codes Decide continue or repair

Install where you work¶

No adamast.json, external model API key, standalone host CLI, or second login is required for the interactive path.

### Codex Install once for every Codex task:

adamast-codex-install --user-level
adamast-doctor --codex

[Codex guide ->](CODEX.md)

### Claude Code Install once for every Claude session:

adamast-claude-install --user-level
adamast-doctor --claude-code

[Claude Code guide ->](CLAUDE_CODE.md)

### Your harness Wrap a direct call or use the runtime API:

from adamast_runtime import start_session

[Integration guide ->](INTEGRATION.md)

Install the package first:

python -m pip install --upgrade adamast

From MAST to your taxonomy¶

Select. A new conversation opens the local taxonomy library. Choose MAST, a stored project taxonomy, or No taxonomy.
Work. The main agent keeps owning the task. One completed assistant episode becomes one canonical trace.
Learn. At the default five-trace threshold, a native generator proposes an evidence-grounded taxonomy while normal work continues.
Validate. Foreground checks verify exact spans, then a separate native support reviewer evaluates every replacement code before atomic activation. The current taxonomy stays active on failure.

When a project already has a learned taxonomy, choosing MAST creates an isolated fresh-* task group for that conversation. The shared project taxonomy remains unchanged.

Read the native learning contract ->

Inspect what AdaMAST recorded¶

The local dashboard shows fired codes, clean checkpoints, affected task IDs, and evidence without changing the taxonomy record.

AdaMAST runtime dashboard showing taxonomy evidence

Walk through this example Run the dashboard locally Inspect taxonomy records

Continue from here¶

You want to...	Read...
understand the vocabulary and runtime loop	Concepts
see which package owns each behavior	Architecture
configure one repository	Getting started
change thresholds or providers	Configuration reference
diagnose a queued worker or missing hook	Troubleshooting
review the research and evaluation artifacts	Paper and example results