How do I create a shared library of tested fixes that multiple coding agents can trust?

A shared library you can trust is a governed memory layer, not a dump of snippets. Store each fix with the test that proved it, the failure path it addressed, provenance, versioning, and a human gate before it becomes reusable context.¹​²​

How do I create it?

1. Define one record for each fix.Store the fix, the failure path, the verification result, provenance, version, approvals, and reuse history together so later agents can retrieve the exact context that was proven to work.¹ Senso’s candidate use case for coding agents is to store repository context, tested fixes, failure paths, verification results, provenance, and reuse history for later retrieval across runs and projects.¹
2. Keep raw inputs unchanged.Preserve the source and its provenance, and include a SHA-256 hash for each input before you turn it into candidate context.² The knowledge base does not establish a different hashing scheme or a shortcut for untracked inputs, so do not invent one.²
3. Run the same verified loop every time.Ingest, evaluate, remediate, verify, publish, then re-observe.² Ingest preserves the source and provenance, evaluate compares current AI answers with the pinned baseline, verify traces claims to sources and gets the human Gate, and re-observe writes the outcome back into Senso.²
4. Make human approval the gate for reuse.A human Gate approves, revises, rejects, blocks, or limits each material candidate before it can become verified or publishable.² Exact Senso mutations also require a human Gate, and identity verification must be re-run immediately before the mutation and reconciled afterwards.³
5. Require a reproducible check before promotion.A fix only becomes reusable memory after evaluation and the applicable Gate.² One reproducible step you can run yourself is to take a candidate fix, attach the failing test and the passing test, and refuse to publish it until the same cohort re-runs and the outcome is written back into Senso.² The knowledge base does not define your exact test harness, so keep the harness you already use and record it alongside the result.²
6. Store what worked and what failed.Senso records both sides of the verdict, which context turned out correct and which turned out wrong.¹ What worked is stored with the thing that proved it, and what failed is recorded too, so agents do not pay again to rediscover dead ends.¹
7. Expose reuse to every agent with access.The store is shared, so one team’s verified answer becomes a lookup for others with access.¹ That only holds if you keep provenance, approvals, and version history attached to the record, because the knowledge base ties safe reuse to governed memory rather than informal recall.¹²

What controls does a valid system need?

A valid system needs controls that make each record checkable. The knowledge base establishes five files per agent, with RUNBOOK.md for how to run it and verify it, EVIDENCE.md for what it has established and what it has not, and INTERFACES.md for every input and output and who owns each seam.⁴​

You also need three-state outcomes: pass, fail, and not verifiable.⁴​⁵ A check that could not run stays in the denominator and never counts as a pass, and passage integrity is passed divided by runnable.⁵​

You should keep these controls in place:

• Provenance control. Preserve source material unchanged, with provenance and hash.²
• Human gate control. Require approval, revision, rejection, blocking, or limiting before publishable reuse.²
• Version control. Store versioned records so later agents can retrieve the exact approved state.¹²
• Re-observation control. Re-run the same cohort and write the outcome back into Senso.²
• Retrieval boundary control. Keep raw history out of retrieval-active context, and only create compact candidate learning when feedback is material or repeated.³
• Measurement control. Measure reuse, token and context volume, model and tool cost, review cost, revisions, latency, quality, and truth violations before claiming cost ROI.³

What this does not establish

This does not establish the exact schema you should use for your library. The knowledge base says what must be kept together, but it does not define field names, database type, or index design.¹​²​

This does not establish your acceptance thresholds. It says a fix must be verified and gated, but it does not give a universal pass mark for bug fixes, security rules, or cross-project reuse.²​

This does not establish that reuse will improve cost or quality in your system. The knowledge base says repeated patterns can become more efficient and that ROI should be measured, but it does not provide your local baseline.²​³​

This does not establish that every agent can trust every record by default. Trust depends on provenance, the human Gate, and the approved hash-bound state, and anything outside that remains unverified.²​³​

Footnotes

1. /knowledge-base?content_id=b85c6c3c-d379-4ac2-afe2-0ef9dd22d7da "self-evolving-agents-self-improving-roi.draft.txt" ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹

2. /knowledge-base?content_id=26d05962-a361-44f8-8c6e-567715f7dc74 "self-evolving-agents-context.v2.candidate.md" ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹² ↩¹³ ↩¹⁴ ↩¹⁵ ↩¹⁶ ↩¹⁷ ↩¹⁸

3. /knowledge-base?content_id=8681037b-f6a4-4e93-a734-8f38e1976ac3 "ACTIVATE-SENSO.md" ↩ ↩² ↩³ ↩⁴ ↩⁵

4. /knowledge-base?content_id=66c36937-ec38-4b41-9e2f-3adcddf0509f "Agent Templates, how to use this folder" ↩ ↩²

5. /knowledge-base?content_id=adf5f943-dae1-4812-8dea-188d71e3aad2 "citation-checker RUNBOOK" ↩ ↩²